A data center announcement usually arrives with three numbers: investment, megawatts and jobs.

Each tells us something. None, on its own, tells us who will capture the economic value.

An announced investment does not establish how much capital has been committed, drawn or spent. A capacity figure may describe an ultimate campus design rather than equipment installed and serving customers. An employment forecast describes an expected outcome, not an observed one.

The commercial picture emerges through the relationships underneath those numbers.

Who controls the land? Who supplies the electricity? Who finances construction? Who absorbs a delay? Who has committed to buy capacity? When does that commitment become revenue? Which parties receive payment before the project becomes profitable?

These are the questions behind the Australian AI Data Centre Commercial Tracker, a proof of concept I built with AI assistance.

The initial register contains 62 facilities and campuses, 39 tracking fields and 62 public sources, with records across all eight Australian states and territories. It includes AI deployments, AI-ready facilities and clearly labelled context facilities whose AI use is not established in the reviewed sources.

You can explore it here:

Open the Australian AI Data Centre Commercial Tracker

The purpose is to make the commercial structure easier to inspect, and to show where the public evidence stops.

The full-stack capitalist lens

For Full Stack Capitalist, the relevant unit of analysis is the connected project: its assets, counterparties, obligations and cash flows.

An AI data centre brings several distinct economic interests together.

The landowner has a property interest. The developer coordinates delivery. The operator provides the facility and its services. Equity investors supply risk capital. Lenders provide financing under contractual terms. Contractors and equipment suppliers deliver the physical infrastructure. Energy counterparties provide electricity or connection infrastructure. Customers contract for services or capacity.

Some organisations occupy several of these roles. Others participate in only one.

That distinction matters. A shareholder in an operating platform is not necessarily the registered owner of a particular campus. A property tenant is not necessarily the customer consuming compute. An electricity network provider is not necessarily the energy retailer or power purchase agreement counterparty.

A useful tracker must preserve those differences.

The analytical task is to follow a project through five connected layers:

A public announcement may illuminate one layer while leaving the others largely undisclosed. The tracker makes that uneven visibility explicit.

Start with the identity of the asset

Before analysing economics, we need to know what the record describes.

Is it a single building, a multi-building campus, an expansion phase or a national investment program? Is the capacity figure attached to that specific asset? Does a financing announcement cover the campus, the operating company or a broader portfolio?

Without those boundaries, it is easy to count the same development more than once or attach a corporate commitment to the wrong project.

The register therefore distinguishes facilities and campuses from national and platform announcements. Campus phases sit within their parent record where the reviewed evidence supports that treatment.

This also helps with naming. Similar site codes can appear across different operators. A project name without its operator and location can be an ambiguous identifier.

Identity sounds administrative. It is the foundation of reliable commercial analysis.

Capacity needs a denominator, and a delivery stage

Megawatts are useful only when we know what they measure.

Published figures may describe IT load, utility supply, planned capacity or an operating installation. Other disclosures use maximum demand, MVA, cabinet counts or GPU counts.

Those measures answer different questions.

A planned campus capacity describes an intended scale. Built capacity describes a delivery milestone. Contracted capacity describes a commercial commitment under whatever terms apply. Billing capacity describes a further stage in monetisation.

The relationship between them is central to understanding progress, but they should not be treated as interchangeable.

The tracker retains the source’s capacity basis and separates the relevant disclosures. It does not convert GPUs or cabinets into megawatts. It does not combine battery capacity with data centre capacity. It does not calculate a national total by adding incompatible figures.

Where sources disagree, the discrepancy remains visible.

This makes the dashboard less convenient for producing a single headline number. It makes it more useful for examining an individual project.

Power is a set of arrangements

“Power secured” can conceal several separate questions.

Is there a documented grid connection? What infrastructure must be delivered? Is an energy supplier named? Has a power purchase agreement been disclosed? Are discussions still underway? Is the available figure a design requirement or an established supply arrangement?

Connection infrastructure and energy procurement deserve separate attention.

So do the commercial terms: price, duration, pass-through provisions and responsibility for additional costs. These can matter to the allocation of project risk, but public sources often do not establish them.

The tracker records what is disclosed and preserves the status of the arrangement. A negotiation remains a negotiation. An announced target remains a target.

Where pricing or contractual protections are unavailable, the field remains unknown. That is more informative than treating a named provider as proof that every power-related risk has been resolved.

Financing is not the same as expenditure

A financing headline is another point where scope matters.

An equity investment in a platform does not establish the amount allocated to a particular campus. A debt facility does not, by itself, establish how much has been drawn. An announced development cost does not establish expenditure to date.

To examine the capital structure, we need to distinguish:

• Equity investors and their disclosed ownership interests.

• Lenders, arrangers and the borrower where identified.

• Financing commitments and evidence of financial close.

• Project expenditure and broader investment announcements.

• Debt terms, security and conditions where publicly available.

The tracker keeps platform-level disclosures separate unless the source explicitly assigns them to a named project.

This avoids a particularly consequential mistake: using a large corporate financing announcement to imply that a specific development is fully funded.

The same discipline applies to returns. Investment size is not a substitute for project revenue, cash flow, yield or internal rate of return.

The customer relationship is where many crucial questions remain

A named customer is useful evidence. The nature of the relationship matters just as much as the name.

A memorandum of understanding, a strategic partnership, a property lease and a contracted capacity commitment are different arrangements.

To understand commercial use, we would ideally establish the customer, the relevant facility, the committed capacity, the contract duration, the commencement conditions and the payment obligations.

We would then track how delivery progresses toward utilisation and billings.

Much of that information may be confidential or absent from the reviewed public sources. The dashboard therefore separates customer identity from the underlying commitment and from evidence of revenue.

A disclosed relationship is not automatically proof of rent received or capacity generating billings.

That is an essential boundary for a tracker intended to support commercial scrutiny.

Who is in the stronger negotiating position?

This is one of the most interesting questions, and one of the easiest to answer too confidently.

A company’s position in the infrastructure chain does not, on its own, establish its bargaining strength.

Ownership of a site tells us who holds an asset interest. It does not reveal the alternatives available to a customer. A construction award identifies a contractor. It does not disclose the contractor’s expected margin or its exposure to overruns. A customer announcement identifies a relationship. It may tell us little about termination rights or payment protections.

A factual assessment of negotiating position would require evidence such as exclusivity, pricing, security, take-or-pay obligations, termination provisions and responsibility for delay.

The tracker includes fields for disclosed negotiation rights and commercial commitments. It does not produce a bargaining-power score where those terms are unavailable.

The same reasoning governs “winners” and “losers.”

We can identify a documented contract award. We cannot automatically translate that award into realised profit. We can record a reported valuation change. We cannot treat it as cash proceeds from a completed sale.

The ambition is to trace value capture with evidence, rather than assign it from a company’s visibility in an announcement.

Second- and third-order economics

The economic footprint of a data centre extends beyond the facility itself. Examining that footprint requires the same attention to evidence and timing.

For this tracker, I use three levels.

First-order economics concern the project’s direct arrangements: land, construction, financing, energy and customer commitments.

Second-order economics concern documented effects around delivery and operation, including supplier activity, employment and property outcomes.

Third-order economics concern documented downstream activity, such as research, industry use or wider economic outcomes associated with the infrastructure.

These are organising categories for the register, not a claim that every effect can be cleanly isolated or attributed.

A forecast number of construction jobs is not an observed employment result. An expected wider economic contribution is not realised project revenue. A downstream research partnership is not evidence of a quantified productivity gain.

The register retains those distinctions and labels forecasts accordingly.

The question at every level is the same: what has been disclosed, what has happened, and what remains an expectation?

Value realization is a sequence of milestones

A project can advance physically while important commercial questions remain unanswered.

That is why the tracker follows separate milestones rather than compressing progress into a single score.

Land and tenure establish a site position. Approvals establish particular permissions. Construction records establish delivery activity. Power disclosures establish specific supply or connection arrangements. Customer commitments establish commercial relationships. Utilisation and billings provide evidence of use and monetisation. Cash-flow and return disclosures, where available, help establish financial outcomes.

None of these milestones should stand in for all the others.

This framework creates a practical research agenda. For each project, we can ask which milestones are supported by evidence, how old that evidence is and what disclosure would resolve the next important uncertainty.

An unknown field is therefore useful. It identifies the next question.

What AI accelerated

AI helped me assemble the interface, organise the tracking structure and turn the research into a searchable dashboard.

The result includes project dossiers, filters, side-by-side comparisons and evidence exports. It requires no login or backend database.

That acceleration creates room for more attention to the work that determines whether the tool is trustworthy: reading disclosures, checking scope, identifying conflicting figures and maintaining records as projects develop.

The dashboard is a manually maintained snapshot dated 9 September 2026, not a live feed.

Its 62 records are not an exhaustive census of Australian data centres or undisclosed AI deployments. A source’s access date does not establish that its project status is current. Published claims are attributed to their sources; they have not all been independently audited.

These limitations belong inside the product, where readers can see them.

An invitation to improve the evidence

I want this tracker to become more useful through better disclosures, corrections and clearer project boundaries.

If you work in development, energy, construction, financing or data centre operations, I would welcome primary sources that help establish:

• Project ownership and land tenure.

• Capacity definitions and delivery milestones.

• Named contractors and financing counterparties.

• Executed customer or energy commitments.

• Observed commercial outcomes.

The full-stack capitalist question is how ownership, obligations and payments connect across the project.

Following those connections gives us a more precise way to examine Australia’s AI infrastructure buildout, and a clearer view of what we still do not know.

Explore the Australian AI Data Centre Commercial Tracker

Which missing disclosure would most improve your assessment of an AI data centre project?

We are spending something close to three-quarters of a trillion dollars building the physical infrastructure of AI.

And I think we are asking the wrong question about it.

Does spending is sustainable?

Whether there is an AI bubble?

Whether the hyperscalers are overbuilding?

Whether there will be enough demand for all these GPUs?

Those are reasonable questions.

But there is a much more important one.

Who actually earns the return on all this capital?

Because the company writing the cheque is not necessarily the company capturing the value.

And the company capturing the value today may not be the one holding the power ten years from now.

That distinction matters enormously.

I have spent enough of my career around infrastructure to be suspicious whenever people treat a capital cycle as a technology story.

I worked at Huawei.

Then Cisco.

I have designed infrastructure.

Worked around data centres.

Built racks.

Managed large technology programs.

Sat between vendors, customers, engineering teams and executives trying to make the economics work.

Later, working around enterprise data, analytics, cloud and industrial software, I saw the same pattern from the other side.

Technology changes.

The economics underneath it are strangely repetitive.

There is always a supplier.

There is always an infrastructure owner.

There is always somebody financing the infrastructure.

There is always someone operating it.

There is always somebody using it.

And eventually there is somebody who actually makes money because the infrastructure exists.

Those are rarely the same company.

AI is no different.

Except the numbers are much bigger.

Start with the money

The four largest American hyperscalers are now committing extraordinary amounts of capital to infrastructure.

Alphabet originally guided towards $175–185 billion of 2026 capex.

Meta subsequently lifted its expected range to $125–145 billion.

Microsoft spent $37.5 billion in one quarter alone, with roughly two-thirds going into relatively short-lived assets such as GPUs and CPUs.

Amazon has been operating around the $200 billion capex level.

Not every dollar is AI.

But AI is increasingly determining the architecture of the spending.

And the spending is accelerating.

Nvidia gives us another way of looking at the same phenomenon.

Its latest quarterly data-centre revenue reached $89 billion.

For one quarter.

Gross margin: roughly 75%.

Think about that for a moment.

While everyone debates whether AI applications have found sustainable business models, one layer of the stack is already collecting extraordinary economics.

The shovel manufacturer is doing very well.

But this is where the story gets interesting.

Because infrastructure booms do not distribute returns evenly.

The AI value chain

I think about the AI economy as a stack.

Not a software stack.

An economic stack.

It looks something like this:

Energy

↓

Grid access

↓

Land

↓

Data centres

↓

Power and cooling equipment

↓

Semiconductors

↓

Networking

↓

Cloud infrastructure

↓

Foundation models

↓

Applications

↓

Enterprise workflows

↓

Labour and business process transformation

↓

Final economic output

Money moves downward into the stack.

Value is supposed to move upward.

Those are not the same thing.

This is the mistake I think a lot of investors make.

They follow the expenditure.

But expenditure tells you where money is being spent.

It does not tell you where economic surplus ultimately accumulates.

Layer 1: Nvidia is capturing the first return

The first obvious winner is semiconductors.

More specifically, accelerated compute.

Today, Nvidia occupies probably the most attractive point in the entire AI value chain.

Demand exceeds supply.

Switching costs are high.

Its software ecosystem reinforces the hardware.

Customers are competing against one another for capacity.

And those customers are enormously well capitalised.

That is a beautiful business.

Nvidia doesn’t need to prove that AI increases global productivity.

It needs customers to believe enough in AI to keep buying infrastructure.

There is a huge difference.

The hyperscalers carry the utilisation risk.

Nvidia largely gets paid when the equipment ships.

This is first-order AI economics.

Someone announces another data centre.

Nvidia sells another rack.

A cloud company raises capex.

Semiconductor revenue increases.

Very straightforward.

But first-order economics rarely tell you where an industrial revolution finishes.

Layer 2: the infrastructure companies

When I used to work around physical infrastructure, racks, networks and data centres, one thing became obvious very quickly.

Servers don’t float in the air.

You need switchgear.

Transformers.

UPS systems.

Cooling.

Cables.

Generators.

Networking.

Fire suppression.

Land.

Water.

Construction.

Operations teams.

Maintenance.

And, above everything else, electricity.

This layer isn’t as glamorous.

It is also becoming extremely important.

The IEA expects global data-centre electricity consumption to roughly double by 2030.

More importantly, these loads are geographically concentrated.

Transmission infrastructure can take four to eight years to build.

Transformer and cable lead times have already stretched.

The IEA estimates that roughly 20% of planned data-centre projects could face delays unless grid bottlenecks are addressed.

That changes the economics.

When compute was scarce, Nvidia had power.

When electricity becomes scarce, electricity starts gaining power.

When grid connections become scarce, the grid connection itself becomes an economic asset.

This is already happening.

Texas has been forced to confront hundreds of gigawatts of proposed data-centre load requests, much of it speculative, and regulators have begun tightening access because the grid cannot treat every proposed project as real demand.

That tells you something.

The AI bottleneck is moving.

And economic power normally moves toward the bottleneck.

The first big power shift

For the last three years, the AI hierarchy looked roughly like this:

Models > GPUs > Cloud > Electricity

I don’t think it stays that way.

I think we are moving towards:

Energy + grid access > compute availability > models

Models are becoming more numerous.

Compute architectures will diversify.

Chips will improve.

Inference will get cheaper.

But you cannot prompt your way around a missing 500 megawatts.

Physics eventually gets a vote.

This is something software people occasionally forget.

Then comes the hyperscaler problem

Now we reach Amazon, Microsoft, Google and Meta.

These companies are spending the money.

So naturally people assume they will capture the return.

Maybe.

But the economics are less obvious than they appear.

Suppose Microsoft spends $100 billion building AI infrastructure.

Some of that immediately becomes Nvidia revenue.

Some becomes networking revenue.

Some becomes power equipment revenue.

Some becomes construction revenue.

Some becomes utility revenue.

Some becomes depreciation.

Then Microsoft has to turn what remains into customer revenue.

That requires utilisation.

And utilisation requires applications.

And applications require customers willing to pay enough to cover the infrastructure underneath them.

Suddenly this becomes a much more difficult equation.

The hyperscaler doesn’t merely need AI adoption.

It needs:

AI revenue > depreciation + electricity + cooling + networking + financing + labour + model development + operating costs.

And it needs that equation to remain attractive while the underlying hardware improves so rapidly that today’s state-of-the-art accelerator becomes tomorrow’s legacy asset.

Microsoft disclosed that roughly two-thirds of its $37.5 billion quarterly capex was going into short-lived assets, primarily GPUs and CPUs.

That phrase matters.

Short-lived assets.

We are building some of the most expensive infrastructure in human history using components with technology-like obsolescence curves.

That is unusual.

A railway can operate for generations.

A transmission line can operate for decades.

A GPU may be economically old surprisingly quickly.

I wrote recently that data centres are increasingly behaving like technology stocks.

This is why.

The hyperscalers may still win

There is another side to this.

Amazon, Microsoft and Google aren’t simply renting GPUs.

They control distribution.

That matters enormously.

AWS already sits inside enterprise infrastructure.

Azure sits inside enterprise identity, data, security and productivity.

Google controls Search, Workspace, advertising and enormous consumer distribution.

Meta controls billions of daily human interactions.

That means their return on AI capex does not have to appear as a neat line item called “AI revenue.”

Google can spend $1 billion on AI infrastructure and earn the return because Search gets slightly better.

Meta can make its advertising system slightly more efficient.

Amazon can improve AWS economics.

Microsoft can defend Office.

This is a critical point.

AI capex can generate defensive ROI.

The return may be revenue preserved rather than revenue created.

That makes the economics considerably harder for outsiders to measure.

Then we reach the model companies

OpenAI.

Anthropic.

Google DeepMind.

xAI.

And dozens of others.

This layer gets most of the attention.

I’m less convinced it captures most of the long-term economic surplus.

Foundation models face an uncomfortable structural problem.

The models are extremely expensive to create.

But the marginal differentiation between models can shrink very quickly.

One company spends billions training a breakthrough capability.

Six months later, competitors replicate much of it.

Open-source models close part of the gap.

Inference becomes cheaper.

Customers route workloads between providers.

The intelligence becomes increasingly interchangeable.

That doesn’t mean foundation-model companies disappear.

It means their bargaining power may decline relative to the infrastructure underneath them and the distribution above them.

This is classic commoditisation.

The middle gets squeezed.

The application layer has the opposite problem

Thousands of AI startups are now being built on top of these models.

They have tiny infrastructure requirements compared with Nvidia or Microsoft.

Wonderful.

But many have almost no structural moat.

A model API.

A workflow.

A nice interface.

Some prompts.

Maybe proprietary context.

I’ve written before that many “AI-native” companies may eventually discover that they own little more than a prompt library.

That doesn’t mean they cannot make money.

They absolutely can.

But revenue and economic power are different things.

If your gross margin depends on somebody else’s model…

running on somebody else’s cloud…

using somebody else’s GPUs…

inside somebody else’s distribution channel…

then you don’t own much of the stack.

And businesses that don’t own scarce parts of the stack rarely control the economics forever.

Which brings us to the enterprise

This is where my view changed the most.

At Cisco, I spent years working around data, analytics, migrations and enterprise transformation.

The biggest economic return from AI may not accrue to “AI companies” at all.

It may accrue to ordinary businesses.

A mining company.

A manufacturer.

A bank.

A logistics company.

An insurer.

A retailer.

A marketing agency.

Imagine an industrial company spends $20 million adopting AI.

It eliminates $60 million of recurring operating cost.

Who captured the AI value?

Nvidia captured revenue.

Microsoft captured cloud revenue.

The model provider captured API revenue.

The integrator captured implementation revenue.

But the industrial company captured $40 million of economic surplus.

And it may capture that surplus every year.

That is fundamentally different.

The infrastructure companies get paid for providing intelligence.

The enterprise gets paid for turning intelligence into economics.

That could ultimately be the largest pool of value.

First-order economics

The first-order effects are easy.

More AI demand means:

More GPUs.

More servers.

More data centres.

More electricity.

More networking.

More cooling.

More cloud revenue.

This is where most equity-market excitement currently sits.

And understandably so.

The numbers are enormous.

Second-order economics

This is where things become much more interesting.

AI infrastructure demand creates scarcity somewhere else.

Electricity becomes scarcer.

Grid interconnections become more valuable.

Transformers become harder to acquire.

Gas turbines get longer lead times.

Land beside substations appreciates.

Nuclear projects become economically plausible again.

Utilities gain negotiating power.

Governments start asking whether AI infrastructure should pay for grid upgrades.

Local communities start asking why residential electricity customers should subsidise transmission infrastructure built for trillion-dollar technology companies.

Capital starts moving into energy systems.

The bottleneck leaves Silicon Valley.

It moves into substations.

I’ve said this before:

AI infrastructure is becoming energy policy.

And once that happens, AI stops being purely a technology industry.

It becomes an industrial system.

Third-order economics

This is the part I find most important.

Once countries recognise compute as strategic infrastructure, governments enter the market.

Now we get:

Sovereign compute.

Energy subsidies.

GPU export controls.

AI industrial policy.

National data-centre strategies.

Nuclear investment.

Grid reform.

Strategic semiconductor manufacturing.

Government-backed AI champions.

Compute trade agreements.

Potential AI infrastructure taxes.

Suddenly the $750 billion isn’t merely corporate capex.

It starts reshaping geopolitics.

Countries with abundant reliable electricity gain leverage.

Countries with constrained grids lose competitiveness.

Natural gas becomes an AI input.

Nuclear becomes technology infrastructure.

Copper becomes technology infrastructure.

Transformers become technology infrastructure.

Land becomes technology infrastructure.

Industrial policy and technology policy merge.

That is third-order AI economics.

The consequences spread far beyond software.

So who captures the value?

My current ranking looks something like this.

Today

1. Nvidia and semiconductor infrastructure

Scarcity plus enormous demand plus exceptional pricing power.

2. Power, cooling and electrical equipment

Less visible.

Increasingly important.

3. Hyperscalers

Enormous strategic upside, but carrying enormous capital and utilisation risk.

4. Model companies

Huge strategic influence today.

Questionable long-term margin durability.

5. Applications

Potentially enormous businesses, but highly uneven outcomes.

6. Enterprises using AI

Still early.

Potentially the largest long-term economic beneficiary.

But who captures the power?

This is a different ranking.

And I think investors confuse the two.

Revenue does not equal power.

Profit does not equal power.

Market capitalisation does not equal power.

Control over a scarce dependency creates power.

Right now Nvidia controls an important dependency.

That gives Nvidia power.

The hyperscalers control compute infrastructure and distribution.

That gives them power.

But over the next decade, I suspect power migrates further down the physical stack.

Towards:

Electricity.

Generation.

Transmission.

Grid access.

Semiconductor manufacturing capacity.

Critical minerals.

And ultimately governments.

Because governments control many of the things nobody else can manufacture with software.

Land use.

Energy markets.

Transmission approvals.

Nuclear licensing.

Trade restrictions.

Export controls.

Industrial subsidies.

National security regulation.

The deeper AI penetrates the economy, the more political the infrastructure beneath it becomes.

The strange thing about the next ten years

I think AI will simultaneously become more powerful and less special.

Intelligence will become cheaper.

Models will proliferate.

Agents will become normal.

Inference costs will collapse.

AI will disappear into software.

And precisely because intelligence becomes abundant, the scarce complements around intelligence become more valuable.

This is a basic economic principle.

When one input becomes abundant, value migrates towards whatever remains scarce.

AI makes cognition cheaper.

It does not make electricity cheaper.

It does not create transmission lines overnight.

It doesn’t manufacture transformers instantly.

It doesn’t produce land.

It doesn’t shorten nuclear permitting to six weeks.

It doesn’t magically create semiconductor fabs.

So the economic centre of gravity changes.

My ten-year view of the value chain

If I had to draw the AI value chain today:

Capital → chips → compute → models → applications → enterprises

Ten years from now, I think it looks more like:

Energy → compute → intelligence → automation → economic output

And the biggest winners may be companies that sit at the edges.

At one end:

Those controlling scarce physical infrastructure.

At the other:

Those converting abundant intelligence into real economic productivity.

The middle may be brutal.

And that brings me back to the $750 billion

Whenever I see these capex numbers, I think back to standing around actual infrastructure.

Racks.

Cables.

Equipment.

Physical systems.

You quickly learn that PowerPoint architecture and physical architecture are very different things.

Every box needs electricity.

Every machine generates heat.

Everything eventually fails.

Everything needs financing.

And somebody somewhere must earn enough money to justify building it again.

AI has spent the last few years feeling almost magical.

But $750 billion of capex has a wonderful ability to remove the magic.

Eventually somebody needs a return.

Nvidia is already getting one.

Electrical-equipment suppliers are getting one.

Utilities increasingly will.

Cloud providers probably will.

Some model companies will.

Many won’t.

Thousands of AI applications won’t.

And perhaps the biggest return of all will appear somewhere almost nobody on Wall Street currently labels an “AI company.”

A manufacturer producing 15% more with the same factory.

A bank processing twice the work with the same headcount.

A logistics company removing kilometres from every delivery route.

A pharmaceutical company compressing years of research.

A mining operation avoiding one day of downtime.

That is where technology becomes productivity.

And productivity is where infrastructure investment finally becomes economic value.

So I don’t think the $750 billion question is:

“Will AI generate enough revenue?”

The better question is:

Who owns the bottleneck between $750 billion of infrastructure and the trillions of dollars of economic value it is supposed to create?

Find that bottleneck.

And you will probably find the next decade’s real winners.

I have designed data centres.

I have built racks.

I have managed data centre operations.

I have worked inside this world at Huawei and Cisco.

So when I hear people describe AI data centres as safe, predictable infrastructure, I get slightly nervous.

Because I know what is inside the building.

From the outside, a data centre looks reassuring.

Concrete.

Steel.

Cooling systems.

Backup generators.

Security gates.

Very serious people carrying access cards.

It looks like infrastructure.

It looks permanent.

But open the door and walk between the racks.

The building may last thirty years.

Almost everything producing its economic value may not.

That distinction matters.

A lot.

I have seen the physical reality

When you build a rack yourself, technology stops being an abstract thing.

Every server has weight.

Every cable needs somewhere to go.

Every device produces heat.

Every system needs power.

Every component has a failure rate.

Every design decision creates another operational dependency.

You learn quickly that there is no cloud.

There are only machines inside buildings.

There are people installing them.

There are people maintaining them.

And there are people panicking when something starts flashing red at 2 a.m.

At Huawei, I saw the infrastructure side.

The physical equipment.

The capacity planning.

The operational discipline required to keep technology running.

At Cisco, I saw the other side.

Networks.

Platforms.

Enterprise systems.

Technology transitions.

Customers trying to modernise without breaking the systems already running their businesses.

Those experiences taught me something simple.

Infrastructure is not valuable because it exists.

It is valuable because someone can use it economically.

A rack with the wrong equipment is just an expensive metal cabinet.

A data centre in the wrong location is a warehouse with a spectacular electricity bill.

And a facility designed around yesterday’s computing architecture can become obsolete long before the concrete begins to crack.

That was already true before AI.

AI has made it much more dangerous.

The old data centre deal

The traditional data centre investment story was fairly simple.

Build the facility.

Connect it to power and fibre.

Lease the capacity to reliable customers.

Sign long contracts.

Collect predictable revenue.

Increase the value of the asset.

It behaved a little like commercial property.

In some cases, it behaved like a bond.

Investors liked that.

Pension funds liked it.

Infrastructure funds loved it.

The returns were not supposed to be spectacular.

They were supposed to be dependable.

AI has changed the deal.

An AI data centre is not just a building with tenants.

It is a giant financial bet on chips, power, cooling, software, customer demand and the future architecture of computing.

All at the same time.

That does not behave like a bond.

It behaves like a technology stock with a concrete shell around it.

The building is slow. The technology is fast.

This is the central problem.

Infrastructure moves slowly.

Technology moves quickly.

A large data centre can take years to plan, finance, approve, connect and build.

During those same years, the chips it was designed to host can change several times.

Power density changes.

Cooling requirements change.

Networking changes.

Rack designs change.

The economics of inference change.

The dominant AI models change.

Even the type of computing customers want may change.

You can spend billions constructing an asset based on assumptions that were sensible when the planning application was submitted.

By the time the doors open, those assumptions may already be old.

I have lived through technology refresh cycles.

The equipment always feels permanent when it arrives.

It is new.

It is powerful.

Everyone wants access to it.

Then a few years pass.

A newer generation arrives.

The old equipment still works.

But working is not the same as being economically competitive.

That is the part financial models often miss.

They model the building over twenty or thirty years.

The technology inside it may have an economically useful life of three to five.

Sometimes less.

GPUs are not tenants

A commercial building can lose a tenant and find another tenant.

An AI facility can lose the economic relevance of its entire hardware configuration.

GPUs are expensive.

They also age differently from traditional infrastructure.

A bridge does not become commercially obsolete because a faster bridge was released eighteen months later.

A transmission line does not lose half its usefulness because Nvidia launched a new architecture.

AI hardware can.

The newer chip may process more work.

It may consume less energy per unit of compute.

It may support a different cooling design.

It may make the previous generation less attractive to customers.

The old GPU does not suddenly stop functioning.

Its economics deteriorate.

That creates a strange asset.

Physically operational.

Technically functional.

Commercially ageing at high speed.

That is technology-stock behaviour.

Not bond behaviour.

Power is now part of the product

During my years around data centres, power was always critical.

No power means no operation.

That has not changed.

What has changed is the scale.

AI does not merely consume electricity.

It concentrates enormous electricity demand in specific locations.

That changes everything.

A data centre site without secured power is not really a data centre site.

Grid access becomes part of the asset.

Transmission capacity becomes part of the asset.

Water availability may become part of the asset.

Political permission becomes part of the asset.

This is where investors can fool themselves.

They may believe they are investing in digital infrastructure.

In reality, they are making a combined bet on:

• Semiconductor demand

• Electricity prices

• Grid expansion

• Planning approvals

• Cooling technology

• AI adoption

• Hyperscaler spending

• Government policy

That is a lot of moving parts for something being sold as a safe infrastructure investment.

The customer concentration problem

Many AI data centres are not supported by thousands of independent customers.

They may depend heavily on one hyperscaler.

Or two.

On paper, this looks safe.

Who would not want Microsoft, Amazon, Google or Meta as a customer?

But customer quality and customer concentration are different things.

A powerful customer is not necessarily a safe customer.

A hyperscaler can negotiate aggressively.

It can delay deployments.

It can change technical specifications.

It can shift workloads to another region.

It can build its own facilities.

It can develop its own chips.

It can decide that it reserved too much capacity during the AI land grab.

The data centre owner carries the fixed asset.

The hyperscaler carries options.

That is not an equal relationship.

When demand is strong, nobody cares.

When demand slows, everyone suddenly discovers who had the negotiating power.

Usually, it was not the infrastructure fund.

The first-order economics

The first-order story is obvious.

AI demand rises.

Companies need more compute.

More compute requires more data centres.

Data centre owners make money.

That is the story in the investor presentation.

It is not wrong.

It is simply incomplete.

The first-order winners are easy to see.

Chipmakers.

Construction companies.

Power equipment manufacturers.

Cooling providers.

Utilities.

Landowners near grid connections.

Data centre operators.

Everyone earns money during the buildout.

But first-order economics tells you what happens when the spending begins.

It does not tell you what happens after the system starts reacting.

That is where things become interesting.

The second-order economics

The second-order effects begin when everyone responds to the same signal.

AI demand is growing.

Capital rushes in.

More facilities are announced.

More land is purchased.

More power is reserved.

More chips are ordered.

Every investor assumes demand will arrive for their facility.

But customers are also becoming more efficient.

Models are getting smaller.

Inference is being optimised.

Companies are building custom chips.

Workloads can shift between regions.

Some AI applications will create huge value.

Many will not survive their pilot stage.

So supply may be built using today’s demand assumptions while demand itself is being transformed by efficiency.

This creates a dangerous possibility.

We could have an electricity shortage and excess data centre capacity at the same time.

That sounds contradictory.

It is not.

The wrong facilities may exist in the wrong places.

They may have the wrong cooling.

The wrong chip density.

The wrong network connections.

Or power that is too expensive to make the workload competitive.

Capacity is not interchangeable.

A megawatt in the wrong location is not the same as a megawatt next to customers, fibre and cheap electricity.

Then financing begins to change.

Lenders become less willing to treat every data centre as a stable infrastructure asset.

They demand higher returns.

Insurance costs rise.

Refinancing becomes harder.

Older facilities receive lower valuations.

Contracts become shorter because customers do not want to commit to yesterday’s architecture.

The cost of capital goes up.

Once that happens, projects that looked profitable at cheap infrastructure financing begin to look much less attractive.

That is the second-order shift.

The risk moves from construction into financing.

The third-order economics

The third-order effects are larger.

They reach beyond data centres.

Governments will realise that AI facilities are competing with households and industry for power.

That turns grid access into a political issue.

A factory creates jobs across a supply chain.

A data centre can consume enormous power with relatively few permanent employees.

Communities will start asking difficult questions.

Who paid for the grid upgrade?

Who receives the economic benefit?

Who carries the environmental cost?

Why are household electricity bills rising while a hyperscaler receives preferential access?

Governments will respond.

They may introduce special grid charges.

They may require data centres to fund transmission.

They may impose local generation requirements.

They may restrict development in power-constrained regions.

They may demand sovereign computing capacity in return for approvals.

They may favour national champions.

At that point, data centre economics becomes industrial policy.

It becomes energy policy.

It becomes national security policy.

And it becomes geopolitical.

Countries with cheap, reliable energy will gain an advantage.

Countries with weak grids will discover that AI ambition cannot be powered by press releases.

Regions may compete for data centres today, then tax or restrict them tomorrow.

That political risk will eventually be priced into the asset.

There is another third-order effect.

Ownership will begin to matter more than capacity.

If a country hosts the building but a foreign company owns the chips, controls the software and decides which workloads receive priority, does that country really own AI infrastructure?

Not necessarily.

It may simply be renting land and electricity to someone else’s intelligence economy.

That is a much bigger question than data centre returns.

It is a question about who captures value from the next industrial system.

Some assets will become stranded

People hear “stranded asset” and think of coal mines or oil infrastructure.

AI may create its own version.

A stranded AI asset may still have power.

It may still have cooling.

It may still contain functioning hardware.

But it may no longer be competitive.

The electricity may be too expensive.

The chips may be too old.

The customer contract may not be renewed.

The network latency may be wrong.

The cooling design may not support newer hardware.

The local government may change the rules.

The facility remains physically present.

The economic value disappears.

I have seen enough technology transitions to know that obsolescence rarely sends a polite calendar invitation.

It arrives while everyone is still depreciating the previous investment.

This does not mean data centres are a bad investment

AI data centres will create enormous wealth.

Some assets will become incredibly valuable.

Facilities with secure power, flexible designs, strong network connectivity and diverse customers may perform extremely well.

But that is exactly the point.

The category is not uniformly safe.

You cannot value every AI data centre as though it were a toll road.

You need to understand what is inside it.

You need to understand its power contract.

You need to understand the hardware cycle.

You need to understand who controls demand.

You need to understand whether the facility can adapt.

You need to understand the customer’s alternatives.

And you need to understand what happens if AI becomes much more efficient.

The winners will not simply own buildings.

They will own adaptable access to power, compute and customers.

The losers will own very expensive boxes built around assumptions that expired before the debt did.

I no longer see a building

When I look at a data centre, I do not just see concrete.

I see racks.

I see cables.

I see cooling.

I see operational dependencies.

I see hardware waiting to become old.

I see power contracts.

I see customers with more bargaining power than landlords.

I see financing models that may be using the wrong definition of risk.

I see all the things that can change while the building stays exactly where it is.

That perspective came from working inside the industry.

From Huawei.

From Cisco.

From designing facilities.

From building racks.

From operating the machinery behind the word “cloud.”

The AI infrastructure boom is real.

The demand is real.

The opportunity is real.

But the safe, bond-like data centre is disappearing.

What is replacing it is more powerful.

More strategic.

More profitable for the winners.

And far more dangerous for anyone who mistakes concrete for certainty.

AI data centres still look like infrastructure.

Economically, they are becoming technology stocks.

The only difference is that when this technology bet goes wrong, you cannot uninstall the building.

AI-native companies will tell you they are building proprietary intelligence.

THEY ARE NOT!

They are renting intelligence from OpenAI, Anthropic or Google. Renting compute from Amazon, Microsoft or Google. Connecting to customer data they do not own. Distributing through channels they do not control. Then placing a workflow, a user interface and a prompt library on top.

That can still create a useful product.

It can even create revenue very quickly.

But usefulness is not ownership. Revenue is not defensibility. And a prompt that produces an impressive result today may become unnecessary after the next model release.

The question for thousands of AI companies is simple:

If the model provider improves the model, the cloud provider controls the infrastructure, the customer owns the data and the incumbent software vendor owns the workflow, what is left for you?

For many companies, the honest answer is: a prompt library, some orchestration code and a monthly API bill.

That is not automatically a bad business.

It is a dangerous business to mistake for a durable one.

The prompt library is the new spreadsheet macro

Prompt libraries are useful. So were spreadsheet macros.

A good prompt encodes judgment. It captures instructions, exceptions, examples and a particular way of turning messy inputs into useful outputs. A collection of prompts can represent hundreds of hours of experimentation.

But prompts have three economic problems.

First, they are easy to copy.

Second, their value depreciates as models improve.

Third, the model provider can absorb the capability into the base product.

Yesterday, getting reliable structured output required elaborate prompting, retries and validation. Today, much of that is a native model or API capability. The better the foundation models become at understanding vague instructions, using tools and handling long context, the less value sits in the exact wording of the prompt.

This creates a strange form of technological depreciation.

In traditional software, accumulated code usually increases the capability of the product. In AI applications, some accumulated logic is deleted by the next model upgrade.

The startup spends six months engineering around a model weakness.

The model lab fixes the weakness.

The startup’s technical achievement becomes redundant.

The customer gets a better product. Society gains. But the company that built the workaround discovers that technological progress can destroy its intellectual property faster than it creates it.

The prompt library is therefore not worthless.

It is closer to perishable inventory than permanent capital.

AI has lowered the cost of creating a company. It has not lowered the cost of building a moat.

Generative AI has compressed the cost of software creation.

A small team can now build a polished application, generate marketing assets, write documentation, support customers and ship integrations at a speed that once required a much larger organisation. Research on new business formation is already pointing in the same direction: generative AI appears to increase the formation of smaller, leaner digital ventures and reduce time to launch.

That is the first-order effect.

More people can build companies.

The second-order effect is harsher:

If everyone can build, building itself stops being scarce.

Product supply expands. Feature differences shrink. Competitors appear faster. Customers face lower switching costs. Every attractive niche attracts ten nearly identical products, often using the same models and producing similar outputs.

The bottleneck moves away from software production and toward distribution, trust, workflow access and customer acquisition.

AI makes the product cheaper to create while making attention more expensive to buy.

That reverses one of the comfortable assumptions of the SaaS era. Software development used to be a meaningful barrier to entry. The engineering effort required to reproduce a product gave incumbents time. AI reduces that protection.

The result is not the death of software.

It is the commoditisation of undifferentiated software production.

The full stack of rented intelligence

To understand the risk, look at the layers of a typical AI-native company.

This does not mean every company needs to train a frontier model or own a data centre. That would be economically absurd for most businesses.

Ownership should not be confused with physical possession.

The issue is control.

Do you control a scarce asset?

Does it improve as customers use the product?

Does it survive a model upgrade?

Can the supplier change its price without destroying your economics?

Can the customer replace you without reorganising the way work gets done?

If the answer to all five is no, the company does not have an AI moat. It has temporary access to someone else’s moat.

The startup problem: speed without accumulation

For startups, rented intelligence is initially a gift.

It removes the need to raise hundreds of millions of dollars for compute, model training and research talent. A founder can reach the market before knowing which model will ultimately win. Multi-model architectures can reduce technical dependence. Falling inference prices can improve gross margins.

This is why application-layer AI companies can grow very quickly. The Financial Times has documented AI application startups reaching substantial revenue at speeds that would have looked abnormal in the old SaaS market. Lower model costs make more use cases viable and allow small teams to serve large markets.

But speed creates a false signal.

Fast revenue growth can look like product-market fit when it is partly curiosity, subsidised usage or a temporary capability gap in the foundation model. Customers may pay because the base platforms have not yet packaged the feature, not because the startup has built something the platforms cannot reproduce.

That creates four startup traps.

1. The demo-to-company trap

AI produces unusually impressive demos.

A demo proves that a task can be performed. It does not prove that the company can own the task.

Investors and founders can confuse model capability with company capability. If 90% of the magic comes from the underlying model, the startup must explain why the remaining 10% deserves the margin, the valuation and the customer relationship.

Sometimes it does. Deep workflow integration, verification, permissions, audit trails and proprietary context can be extraordinarily valuable.

But the burden of proof sits there, not in the demo.

2. The negative learning-curve trap

In a strong software business, every customer should make the product harder to replace.

The company learns. Its data improves. Its workflows deepen. Its distribution expands. Its unit economics strengthen.

Many thin AI wrappers do not accumulate much of anything. They send inputs to a model, receive outputs and discard the interaction. The customer receives the value, while the model provider receives the usage and the startup retains little reusable learning.

The company grows without compounding.

Revenue goes up, but strategic ownership does not.

3. The margin illusion

Traditional SaaS trained investors to expect very high gross margins because the marginal cost of serving another user was small.

AI reintroduces variable production costs.

Every request can consume tokens, retrieval, tool calls, external APIs, vector storage, validation and sometimes human review. Agentic systems multiply this because a single user action can trigger many model calls and interactions between agents.

Falling token prices help. They do not remove the structural issue.

When the model is a material part of the product, the supplier sits inside the cost of goods sold. If the application competes mainly on price, falling inference costs may be passed straight to customers rather than retained as margin.

Cheaper intelligence can improve demand while simultaneously destroying pricing power.

4. The distribution trap

When building becomes cheap, distribution becomes the moat everyone pretends is product.

The startup may rely on Google for discovery, LinkedIn for leads, Microsoft for enterprise access, a cloud marketplace for procurement and OpenAI for the underlying capability.

Each gatekeeper can charge rent.

The startup is not one platform decision away from failure. It is several platform decisions away from failure.

The mature-company problem: paying twice for intelligence it already owns

Large companies face the opposite problem.

They already own many of the scarce assets AI startups want:

• decades of customer relationships

• proprietary operational data

• embedded workflows

• licences and regulatory permissions

• distribution

• domain experts

• historical outcomes

• trusted brands

Yet many incumbents fail to convert those assets into usable intelligence.

Their data is fragmented. Permissions are unclear. Systems do not connect. Institutional knowledge sits in email, presentations and the heads of employees. Procurement buys AI tools department by department. Every vendor creates another interface over the same disconnected organisation.

The mature company then pays an AI vendor to extract value from data and workflows the mature company already owns.

This can be rational. Outside vendors can move faster and bring specialised capability.

But there is a strategic difference between buying a tool and outsourcing the learning loop.

If the vendor observes the prompts, corrections, exceptions, approvals and outcomes, the vendor may learn the workflow faster than the customer does. The enterprise supplies the domain knowledge. The vendor turns it into a repeatable product. The enterprise pays the subscription.

The company has effectively financed the construction of someone else’s intangible asset.

That is the second-order risk for incumbents.

The third-order risk is worse: once a vendor becomes the interface through which employees access corporate knowledge and execute work, the vendor begins to control the organisational operating layer.

At that point, switching is not a software migration. It is institutional surgery.

What actually survives the next model release?

Distribution

A company with trusted access to a customer segment can swap models underneath the product. The model lab cannot instantly reproduce the customer relationship.

Distribution is especially powerful in regulated or fragmented industries where procurement, reputation and local market knowledge matter.

Proprietary context

Raw data is overrated. Context is more valuable.

The useful asset is not a pile of documents. It is the mapping between data, decisions, permissions, exceptions and outcomes. That context tells the system what mattered, what was allowed and what worked.

Workflow control

An AI assistant that comments on work is easy to replace.

An AI system embedded in the transaction, approval or execution path is much harder to remove. The more operational responsibility it carries, the more valuable reliability, integration and governance become.

Outcome data

Generated content is abundant. Verified outcomes are scarce.

Did the contract clause survive litigation?

Did the maintenance recommendation prevent failure?

Did the sales action increase conversion?

Did the medical intervention improve the patient’s condition?

The company that captures the link between recommendation and outcome can build a learning system competitors cannot reproduce with prompts alone.

Trust and permission

In high-stakes markets, the right to act is more valuable than the ability to generate.

Licences, auditability, liability frameworks, security approvals and institutional trust are slow to build. Model capability may commoditise. Permission to use it often does not.

Proprietary economics

Some companies will build an advantage through inference efficiency, routing, caching, specialised smaller models or ownership of critical infrastructure.

The question is not merely whether the product works. It is whether the company can deliver the outcome at a cost structure competitors cannot easily match.

The second-order economics: where the money moves

If thousands of AI companies own little beyond prompts and interfaces, the value does not disappear.

It moves.

More value flows to foundation-model and cloud providers

Every application expands demand for tokens and compute. Even when the application company struggles to defend its margin, its suppliers collect revenue from usage.

This resembles previous platform economies. A crowded field of businesses can compete above the platform while the platform captures the most stable rent below them.

The application layer may generate the experimentation. The infrastructure layer may capture the dependable economics.

Customer acquisition becomes more expensive

When competitors can reproduce features quickly, they compete through advertising, sales teams, partnerships and discounts.

Engineering costs fall. Go-to-market costs rise.

Capital moves from product creation toward distribution warfare.

The irony is brutal: AI lets a five-person company build what once required fifty people, then forces it to spend the savings fighting fifty near-identical competitors for attention.

Services return through the back door

Many AI companies present themselves as scalable software but rely on implementation teams, prompt tuning, data cleaning, integration and ongoing workflow redesign.

That is not necessarily a flaw. The work may be valuable.

But it changes the economics. The company may be a technology-enabled service business carrying a SaaS valuation. Its moat may sit in people and deployment experience rather than code.

Markets will eventually distinguish between recurring software revenue and recurring consulting disguised as software.

Model providers gain bargaining power

Multi-model strategies reduce dependence, but they do not eliminate it. Models are not perfectly interchangeable. Changing providers affects quality, latency, safety behaviour, tool use and customer commitments.

Once the product is optimised around a model, switching becomes costly. The provider can influence the application’s margins, roadmap and service quality without owning a share of the company.

This is vertical power without vertical ownership.

Valuations split between revenue growth and strategic control

Two AI companies may report identical revenue.

One owns distribution, workflow and outcome data.

The other buys leads, calls an external model and returns text through a dashboard.

They should not receive the same multiple.

As the market matures, investors will pay less for revenue that depends on rented capability and more for revenue attached to controlled workflows, proprietary feedback loops and durable customer access.

The third-order economics: what happens after the shakeout

The deeper consequences arrive after the application market becomes crowded.

1. AI markets may consolidate faster than SaaS

SaaS companies could coexist because products accumulated specialised features over many years.

AI compresses feature development and lets platforms move across categories quickly. A general model provider can enter legal research, coding, customer service, analytics or enterprise search without rebuilding intelligence from zero.

That does not mean vertical startups all disappear. It means they must move deeper into the workflow faster than the platform moves outward.

The middle gets squeezed.

At one end sit large platforms with models, compute and distribution.

At the other sit specialised companies with domain data, trust and workflow ownership.

The generic wrapper in between has nowhere safe to stand.

2. Enterprises become the training ground for their future suppliers

Companies adopting external AI tools will generate valuable process data: how decisions are made, where exceptions occur, which recommendations are accepted and which outputs create value.

If contracts do not define ownership and use of this learning, enterprise adoption can transfer institutional knowledge outward.

The strategic procurement question will shift from “Can this vendor access our data?” to “Who owns what the system learns from our organisation?”

That is a much harder question, and most procurement templates are not designed for it.

3. Labour power shifts toward people who control context

If basic software production becomes abundant, value moves toward people who own customer relationships, domain judgment, operational permissions and the authority to change workflows.

The most important employee in an AI transformation may not be the person who writes the prompt.

It may be the person who knows why the process exists, which exceptions matter, who carries the liability and how the outcome is measured.

AI reduces the scarcity of generation.

It increases the value of accountable judgment.

4. National AI sovereignty becomes an application-layer issue

Policymakers often discuss sovereignty in terms of chips, data centres and foundation models.

But a country can host compute domestically and still surrender the interfaces through which its firms make decisions.

If foreign platforms mediate legal work, industrial maintenance, financial analysis, education and public administration, dependency exists above the data-centre layer as well as below it.

The strategic asset is not only domestic compute.

It is domestic control over data, workflows, standards and institutional learning.

5. Regulation may strengthen the largest platforms

Compliance creates fixed costs.

If every AI provider must fund expensive assurance, documentation, security and liability processes, the burden will fall more heavily on smaller companies. Large incumbents can spread those costs across millions of users and bundle compliance into existing enterprise contracts.

Poorly designed regulation can therefore produce the opposite of competitive policy: it can convert trust into another hyperscaler moat.

The answer is not no regulation. It is regulation that distinguishes between model risk, application risk and the actual authority a system has to act.

What investors should ask

Investors should spend less time asking whether a company is “AI-native” and more time asking what compounds.

1. What does the company control that the model provider does not?

2. What becomes more valuable after the thousandth customer?

3. Which data rights are contractual, and which are merely assumed?

4. Does usage create proprietary outcome data or only higher API costs?

5. What happens if the model price doubles?

6. What happens if the model price falls by 90%?

7. What happens if Microsoft, Google, Salesforce or OpenAI bundles the feature?

8. Is the company replacing a system of record, becoming one, or merely reading from one?

9. How much revenue requires human implementation?

10. Does the company own distribution or continuously repurchase it?

The decisive question is not “How good is the product?”

It is “Which part of the value chain can this company prevent others from taking?”

What operators should build

Operators should treat models as replaceable inputs and learning loops as strategic assets.

Build model portability where it is economically sensible. Capture structured feedback. Measure outcomes rather than output volume. Negotiate rights over interaction and improvement data. Embed the product into execution, not merely advice. Own the customer relationship. Make integrations deep enough to create value but not so dependent that a platform can switch off the company.

Most importantly, decide what the company intends to own before adding more AI features.

If the answer is “better prompts,” the strategy is already expiring.

What mature businesses should protect

Large organisations should not respond by building everything internally.

That usually produces slow, mediocre software and years of governance theatre.

They should decide which learning loops are too strategic to outsource.

The enterprise may happily rent the model, the cloud and many applications. But it should preserve control over identity, permissions, core operational context, decision records, verified outcomes and the ability to switch suppliers.

The goal is not technological independence.

It is bargaining power.

What policymakers should understand

The AI economy will not be shaped only by who invents the best model.

It will be shaped by who captures the learning generated when models enter real institutions.

Competition policy should watch bundling and platform self-preferencing. Procurement rules should define rights over derived learning and workflow data. AI policy should avoid imposing identical obligations on a frontier model, a low-risk application and an agent authorised to move money or make consequential decisions.

Governments should also stop treating the number of AI startups as proof of national capability.

A country can produce thousands of AI applications while importing the models, compute, cloud, distribution and capital.

That is startup activity.

It is not necessarily technological power.

The final test

There is nothing shameful about building on someone else’s model.

Every modern company builds on layers it does not own. Airlines do not manufacture every engine. Banks do not build every server. Software companies do not generate their own electricity.

The problem begins when a company rents every scarce layer and mistakes assembly for ownership.

The best AI companies will use commoditised intelligence to capture something non-commoditised: a customer relationship, a regulated permission, a proprietary learning loop, a critical workflow, verified outcomes or a distribution advantage.

The weakest will keep adding prompts and calling the collection a platform.

When the next generation of models arrives, we will discover which companies built an asset and which ones merely documented the limitations of the previous model.

The model is rented.

The cloud is rented.

The intelligence is rented.

If all you own is the prompt, you do not own the company’s future.

Get this: Governments are no longer just regulating AI.

They are starting to fund it, subsidise it, procure it, protect it, finance the infrastructure underneath it, and deliberately shape who wins.

That starts to look less like traditional technology policy and more like a giant form of state-sponsored venture capital.

The interesting part is that the U.S., China and Europe are all doing this differently.

China is the most obvious version. The state has been using guidance funds, state-owned banks, local governments, industrial subsidies and national champions for years. AI is now being plugged into that machinery.

The U.S. tells itself a different story. Silicon Valley supposedly allocates the capital. But once you look underneath the stack, Washington is everywhere: CHIPS subsidies, export controls, defence procurement, tax incentives, energy policy, federal research funding, sovereign-security restrictions, infrastructure permitting and increasingly direct attempts to shape domestic compute capacity.

Europe is trying something else again. It wants sovereign compute, sovereign models and European champions, but it also has less venture capital, fewer hyperscalers and a regulatory structure that can sometimes fight against the industrial ambitions sitting next to it.

I want to ask:

Are we watching the emergence of three different models of state capitalism for AI?

And if so, which one actually works?

Follow the money through the entire AI stack

Don’t just talk about foundation models.

Look at the whole machine:

chips → semiconductor equipment → data centres → electricity → grids → cloud → compute → models → applications → talent → data → defence and government procurement.

At every layer, ask the same questions:

Who is putting up the capital?

Who is taking the risk?

Who owns the infrastructure?

Who gets subsidised?

Who gets protected?

Who gets procurement contracts?

Who captures the upside if the bet works?

And who is left holding the losses if it doesn’t?

That is where the real economics of sovereign AI sits.

Compare the U.S., China and EU as capital allocation systems

I don’t want a generic country comparison.

I treat each region almost like a giant investment fund with a different investment committee.

China:
The government is much more comfortable explicitly picking industries and companies. Look at government guidance funds, provincial AI funds, semiconductor funding, state banks, state-owned enterprises, national compute infrastructure, domestic GPU development and the push for technological self-sufficiency.

But dig into the problems too. Local-government duplication. Bad capital allocation. Zombie companies. Subsidy arbitrage. Overcapacity. Projects created because Beijing wants AI rather than because customers actually want the product.

China may be better at mobilising capital quickly. That doesn’t necessarily mean it is better at allocating it.

United States:
The U.S. still has by far the deepest private technology capital markets, so the state doesn’t need to behave like a Chinese investment fund.

Instead, it changes the economics around private investors.

A semiconductor fab becomes more attractive because of CHIPS Act support.

A domestic GPU supply chain becomes strategically valuable because export controls make advanced compute geopolitical.

A data-centre project becomes partly an energy-policy question.

An AI company can become strategically important because the Pentagon, intelligence agencies or federal government become customers.

So instead of asking whether the U.S. has industrial policy, ask a better question:

How much private AI investment would still happen in exactly the same form if Washington disappeared from the equation?

Probably less than Silicon Valley likes to admit.

European Union:
Europe is the awkward case.

It wants technological sovereignty.

It wants European AI champions.

It wants AI factories and sovereign compute.

It wants local semiconductor capacity.

It wants public supercomputing infrastructure.

But Europe also regulates more aggressively, has shallower VC markets and doesn’t have equivalents of AWS, Microsoft, Google, Nvidia or Meta at comparable scale.

That creates a fascinating contradiction:

Europe increasingly understands that AI is an industrial-capacity problem, but its political system still often treats AI as primarily a regulatory problem.

Explore whether that is changing.

The main mental model

The core framework should be something like:

Sovereign AI is not really about owning a chatbot. It is about controlling enough of the capital stack that your economy cannot be switched off by someone else.

That includes compute.

Energy.

Semiconductors.

Cloud infrastructure.

Models.

Networks.

Capital markets.

Talent.

Government demand.

Once you see it like that, sovereign AI starts looking a lot more like defence industrial policy than software policy.

Another useful mental model:

Governments are moving from referee → customer → investor → market maker

For most of the internet era, governments mostly regulated technology after companies built it.

AI is changing that relationship.

The state can now:

1. fund the science,

2. subsidise the factories,

3. finance the infrastructure,

4. create the demand,

5. restrict foreign competitors,

6. guarantee national-security customers,

7. shape energy access,

8. control exports,

9. and occasionally take direct or indirect investment exposure.

At that point, the government isn’t standing outside the market.

It is helping manufacture the market.

Where I want real disagreement

One camp will say AI is too strategically important to leave entirely to markets.

Markets optimise for return on capital, not national resilience.

If advanced chips, compute or cloud infrastructure become geopolitical chokepoints, governments have legitimate reasons to pay for redundancy that a private investor wouldn’t.

The other camp will say this is exactly how governments waste extraordinary amounts of money.

Politicians are bad venture capitalists.

Once “strategic AI” becomes a funding category, every company suddenly discovers that it is strategically important.

Subsidies attract lobbyists.

Protection creates lazy incumbents.

Local governments chase fashionable industries.

Capital stops following productivity and starts following policy.

That tension should run through the whole essay.

Ask whether governments can actually pick winners

This is one of the most interesting questions.

There are cases where industrial policy clearly mattered.

Semiconductors.

Aerospace.

Defence.

Space.

The internet itself.

Renewable energy.

But AI moves unusually fast.

A government might commit billions to a technology that becomes obsolete before the infrastructure is finished.

Today’s strategically essential accelerator architecture might not be tomorrow’s.

Today’s leading model company might not exist in ten years.

So sovereign AI creates a strange problem:

States are making 20-year infrastructure decisions around technologies whose competitive cycles can be measured in months.

That is an incredible capital-allocation mismatch.

Open source makes sovereignty weirder

If open-weight models continue improving, countries may not need a domestic OpenAI.

They might need:

domestic compute + domestic energy + access to open models + fine-tuning capability + secure inference infrastructure.

That could dramatically reduce the cost of sovereign AI.

And it could shift power away from model developers toward whoever owns the compute and energy underneath them.

This is important.

The most valuable sovereign asset might be the infrastructure capable of running whatever the best model happens to be.

Follow the second- and third-order economics

If governments start underwriting AI infrastructure, what happens next?

Private investors may take more risk because the downside is partially socialised.

Utilities may become AI infrastructure companies without intending to.

Grid connection queues become industrial-policy tools.

Semiconductor fabs become geopolitical assets.

Cloud providers become quasi-national infrastructure.

Defence procurement becomes an AI go-to-market strategy.

VCs start investing around government priorities.

Startups optimise for sovereign contracts.

Countries compete for data centres using land, tax incentives and cheap electricity.

Eventually the competition may move from:

Who has the best AI company?

to:

Which state has the cheapest and deepest capital stack behind AI?

That is a much bigger question.

Bring in actual numbers

Government funding commitments.

Semiconductor subsidies.

AI infrastructure spending.

Public compute programs.

Chinese guidance funds.

EU AI Factory and InvestAI-type initiatives.

CHIPS Act disbursements.

Data-centre capital expenditure.

Electricity investment.

Defence AI procurement.

But don’t just repeat trillion-dollar press-release numbers.

Separate:

money announced

from

money actually committed

from

money actually spent

from

private capital supposedly “mobilised.”

Those are very different things.

Governments love multiplying the last category.

The historical analogy

Compare this with earlier state-backed technology races:

railways,

electricity,

semiconductors,

telecommunications,

aerospace,

nuclear energy,

the internet.

AI is different because the infrastructure is being financed by a strange coalition:

Big Tech balance sheets + private credit + utilities + sovereign governments + venture capital + infrastructure funds + defence budgets.

The AI economy may therefore be producing something new:

a public-private capital stack where it becomes increasingly difficult to separate the market from the state.

The conclusion should land somewhere uncomfortable

For thirty years, the West told itself that governments set the rules and markets picked the winners.

AI is quietly breaking that distinction.

China never really believed it.

Europe is reluctantly abandoning it.

And even the United States, the country most ideologically committed to private capital allocation, is discovering that once compute becomes strategically important, markets alone are apparently not enough.

The AI race is therefore not just a competition between OpenAI, Anthropic, DeepSeek, Google or Meta.

It is increasingly a competition between capital systems.

American private capital backed by strategic state intervention.

Chinese state-directed capital supplemented by private entrepreneurship.

European public industrial policy trying to compensate for weaker private technology capital.

And the winner may not be the country with the smartest model.

It may be the country that figures out how to deploy trillions of dollars of capital into compute, energy, chips and talent without destroying returns in the process.

Because that’s the dangerous part of state-sponsored venture capital.

Governments can fund things markets would never fund.

Sometimes that creates the future.

Sometimes it just creates the world’s most expensive stranded assets.

AI is making intelligence dramatically cheaper. It is also turning intelligence into a metered input. That changes software margins, pricing, power grids, capital markets, and ultimately which countries and companies capture the economics of AI.

There is a strange contradiction sitting at the center of the AI economy.

For most of the software era, the dream was to make the marginal cost of serving one more customer disappear. Build the product once. Put it in the cloud. Add another user. The extra cost existed, obviously, but compared with the subscription revenue it was often tiny.

AI breaks that mental model.

Every time an AI system reads a document, reasons through a problem, generates code, searches a database, calls another tool, checks its own work, or runs another loop inside an agent, somebody has to pay for computation.

That payment might be a fraction of a cent. It might eventually fall another hundredfold. It may be hidden inside a subscription. But it exists.

And as AI moves from answering questions to executing transactions, that variable cost gets embedded deeper into the economy. OpenAI’s latest enterprise data show exactly that migration: enterprise usage is moving from assistance toward execution, and the firms in the top decile of AI usage now generate 8.3 times as many output tokens per active user as typical firms. As of June 2026, Codex alone produced 64% of the combined Codex and ChatGPT output tokens among OpenAI enterprise customers. Usage of enterprise Codex since February grew 108-fold in legal, 41-fold in sales, 41-fold in recruiting, and 26-fold in marketing.

That is the beginning of what I think of as the compute tax.

Not a tax imposed by government. And not literally a GPU charge on every payment made in the economy.

It is a new production cost attached to every economic action that requires machine intelligence.

The distinction matters because AI is simultaneously doing two things that look contradictory.

It is destroying transaction costs by making search, analysis, communication, negotiation and execution cheaper. MIT researchers describe the fundamental economic promise of AI agents in almost exactly those terms: agents can substantially reduce the time and effort involved in searching, communicating and contracting. But the mechanism doing that work consumes compute.

So AI may eliminate a $20 human coordination cost and replace it with 20 cents of machine inference.

That’s an extraordinary economic bargain.

But the 20 cents still has to go somewhere.

And once trillions of decisions, searches, negotiations, customer interactions, software changes, purchases and administrative actions become mediated by AI agents, those pennies start looking less like a rounding error and more like an entirely new layer of the economy.

The tax hiding inside the product

The easiest way to understand the change is to stop thinking about tokens as a technical metric.

A token is not a business outcome.

Nobody cares that an AI consumed 80,000 tokens to resolve an insurance claim. The insurer cares whether the claim was resolved correctly, how long it took, how many humans had to intervene, how much the computation cost, and whether the final cost was lower than the old process.

OpenAI is now explicitly telling CFOs to think this way. Its proposed economic scorecard asks whether the value of work completed grows faster than the cost required to produce it. It argues that cost per token can be misleading because a cheaper model may require retries, longer processing or human review, while a more capable and more expensive model may complete the task correctly in one pass.

BCG has arrived at almost the same conclusion independently. It argues that the useful denominator is cost per successful outcome, because agentic workloads can accumulate cost through long context windows, repeated loops, retrieval, tool use and model selection. A weaker model can therefore cost more in total than a stronger one if it repeatedly fails before reaching an acceptable answer.

This sounds like a minor accounting change.

It isn’t.

It means intelligence is becoming something companies can buy in variable quantities.

Consider a deliberately simple example using OpenAI’s published API pricing in August 2026. GPT-5.6 Terra is currently priced at $2 per million input tokens and $12 per million output tokens for standard processing below the stated context threshold. A workload using 20,000 input tokens and producing 5,000 output tokens therefore costs roughly 10 cents for that individual model invocation: four cents of input and six cents of output. One million identical calls would cost roughly $100,000. A workflow requiring ten such calls would be around $1 of model inference per completed workflow before counting external tools, searches, retries, storage, conventional cloud infrastructure or human review.

That is not an estimate of what an average agent costs. Agent economics vary enormously.

It is simply a way of seeing the scale.

A dollar of machine intelligence can be insanely cheap if it replaces thirty minutes of professional labor.

A dollar can also be catastrophically expensive if the product earns fifty cents from the interaction.

That difference is going to separate a lot of good AI businesses from bad ones.

First-order economics: software gets a meter

The first-order consequence is the easiest one.

AI turns part of software back into a variable-cost business.

BCG makes the accounting point bluntly: tokens used inside products and customer interactions belong in cost of goods sold. Unlike conventional SaaS economics, where incremental software usage could often be served at very low marginal cost relative to revenue, inference rises as customers interact with the model. BCG’s current analysis estimates gross margins of roughly 65% to 80% for AI-enabled software and 50% to 65% for AI-native products, though these are estimates of emerging economics rather than a universal law for every AI company.

That creates a surprisingly old-fashioned problem for supposedly futuristic companies:

What does it cost to produce one unit of what you sell?

The SaaS industry spent twenty years training investors to care about seats, annual recurring revenue, retention and sales efficiency. AI forces management teams to relearn manufacturing economics.

Except the factory is a data center.

For a customer-service company, the unit might be a resolved ticket.

For an AI coding company, it might be a successfully merged code change.

For a legal agent, a reviewed contract.

For a procurement system, a completed sourcing decision.

For an AI-native bank, perhaps a completed underwriting decision.

The important number is no longer merely:

Revenue per user.

It starts becoming:

Revenue per successful outcome minus compute per successful outcome.

And the word successful is doing a lot of work there.

Suppose Model A costs 5 cents and completes the task correctly 60% of the time, while Model B costs 15 cents and succeeds 95% of the time. Once you account for retries, escalation and human review, Model B can easily have better economics despite charging three times more per inference pass. Both BCG and OpenAI now emphasize this distinction between nominal model price and full cost of successful work.

This is why there will not be one compute tax.

There will be millions of them.

Different workflows will have radically different intelligence intensity.

Generating an email subject line and autonomously negotiating a commercial contract may both appear as a single click to the user. Economically, they are completely different products. Agentic systems may retrieve documents, maintain long contexts, reason repeatedly, use external tools and inspect intermediate results before producing a final action. BCG notes that these loops can make apparently similar requests carry very different costs, while OpenAI’s enterprise data show the leading users consuming far more output tokens as they move toward delegated, multistep work.

That creates an immediate strategic consequence.

Model architecture becomes business-model architecture.

Routing a trivial request into an expensive frontier model is no longer just sloppy engineering. It is margin leakage.

Using deterministic software for something an agent unnecessarily reasons about is margin leakage.

Repeatedly feeding gigantic context windows into a model is margin leakage.

Failing to cache reusable information is margin leakage.

Letting an agent loop without an economically rational stopping condition is margin leakage.

The companies that understand this are already moving toward routing, caching, batching and task-specific model selection rather than sending every problem to the biggest available model. BCG explicitly recommends separating deterministic tasks from model-based reasoning and routing simpler work toward lighter models while reserving expensive models for difficult tasks.

In other words, the AI stack is starting to develop something industrial companies have understood forever:

cost engineering.

And that is probably healthy.

Because the real question was never whether AI was expensive.

The question is whether the intelligence purchased for $1 creates $1.10 of value or $100 of value.

Second-order economics: the business model starts mutating

The second-order effect is more interesting because companies cannot simply absorb a new variable cost forever.

Eventually pricing has to adapt.

This is already visible across software. McKinsey argues that inference is introducing recurring compute and infrastructure costs into software economics and that vendors increasingly need pricing structures that scale with usage, outcomes, actions or compute rather than relying purely on traditional seats.

That is why the seat may gradually lose its status as the dominant unit of enterprise software.

A seat makes sense when a human is the worker.

What is a seat worth when one employee launches 500 autonomous tasks overnight?

What happens when a company has 3,000 employees but 40,000 persistent software agents?

Charging $30 per employee starts looking ridiculous if the agent produces $10,000 of work. Equally, an “unlimited AI” subscription starts looking suicidal if one power user can unleash millions of tokens of autonomous work.

So AI pricing is likely to migrate toward something closer to economics itself:

work performed.

Per resolution.

Per transaction.

Per code change.

Per lead qualified.

Per invoice reconciled.

Per claim processed.

Per successful research task.

That does two things at once.

First, it makes AI companies more directly exposed to their customers’ economics. If the AI cannot produce a valuable outcome, it becomes difficult to keep charging merely because someone has a login.

Second, it creates a powerful incentive for AI companies to drive the compute required per successful outcome downward.

That could be a much stronger productivity engine than merely making models cheaper.

BCG points to emerging AI companies already charging around outcomes rather than conventional seats and argues more broadly that AI is reallocating profit pools toward businesses that control valuable customer relationships, proprietary assets and difficult-to-replicate workflows. Its August 2026 work describes AI as a reallocation of capital and margins rather than merely another productivity technology.

This is where the compute tax starts changing competitive strategy.

Imagine two AI legal companies.

Both use roughly the same frontier models.

Company A is basically a wrapper. It buys expensive inference, adds an interface and resells it.

Company B has proprietary legal data, a retrieval layer, specialized evaluation systems, model routing, workflow integration, cheaper models for routine work, frontier models for hard questions and a feedback system that improves as customers use it.

The second company may pay the same headline API price as the first.

Its cost per successful legal outcome can still be dramatically lower.

That means the durable moat in AI software may not be “we have AI.”

Almost everybody will have AI.

The moat becomes the system surrounding the model that converts compute into useful work efficiently.

Data.

Workflow.

Distribution.

Routing.

Evaluation.

Customer context.

Proprietary feedback.

Trust.

And critically, the ability to determine when not to use expensive intelligence.

That last point is underrated.

The cheapest AI call is the one you did not need to make.

This also explains a seeming contradiction in the economics of AI agents.

MIT researchers argue that agents can dramatically reduce traditional transaction costs by lowering the effort involved in searching, communicating, comparing alternatives and contracting. They describe agents as potentially useful even in complex markets such as procurement, investment and real estate because software can inspect far more information than a human economically could.

I think they’re right.

But the transaction cost isn’t disappearing.

It is being converted.

We are swapping human transaction costs for machine transaction costs.

Instead of paying an employee to spend three hours comparing 70 suppliers, you might pay an agent to interrogate 70 databases, read 1,000 documents, request quotations, score the offers and return three candidates.

Human coordination cost collapses.

Compute consumption rises.

The net economic result can be overwhelmingly positive.

This is why calling compute a tax should not be confused with calling it a burden.

A good tax can still be a bargain when the alternative is much more expensive.

The strategic issue is who captures the savings.

Suppose an old process costs $100 in human labor.

AI reduces total production cost to $10, of which $2 is inference.

There is now $90 of economic surplus available.

Does the customer keep it through lower prices?

Does the AI application capture it?

Does the model provider capture it?

Does the cloud provider?

Does the GPU supplier?

Does the owner of proprietary data?

Does competition push almost all $90 back to the consumer?

That is the real capital-allocation question.

BCG’s latest work argues that AI is shifting profit pools and that AI-native attackers may selectively capture high-margin parts of incumbent value chains without reproducing the incumbent’s entire business.

Which means AI may not simply make existing businesses more efficient.

It may unbundle their margins.

Third-order economics: the token bill reaches the physical economy

Now zoom out another level.

A token looks digital.

Its supply chain isn’t.

Every token ultimately descends through a physical stack:

model software,

servers,

accelerators,

high-bandwidth memory,

networking,

data-center buildings,

cooling,

transformers,

transmission,

electricity generation,

land,

water,

capital.

Once enough intelligence is consumed, the compute tax becomes an infrastructure bill.

The International Energy Agency says global data-center electricity consumption reached roughly 485 TWh in 2025 and projects around 950 TWh by 2030, roughly 3% of global electricity demand. Electricity consumption specifically associated with AI-focused data centers is expected to triple over the same period.

And here is the really important part.

The electricity required per AI task is falling rapidly.

Yet total electricity consumption is rising.

The IEA says per-task efficiency is improving at a rate it describes as unprecedented in energy history, while at the same time more people are using AI and more compute-intensive applications such as agents are spreading. Its current forecast therefore still has data-center electricity consumption doubling by 2030.

That is the entire compute-tax thesis in miniature.

The tax rate falls.

The tax base explodes.

We have seen versions of this before. Make something dramatically cheaper and people don’t necessarily spend less on it. They find more things to do with it.

Cheap bandwidth didn’t cause the world to consume less bandwidth.

Cheap storage didn’t cause us to store less data.

Cheaper computation didn’t cause us to compute less.

Cheap machine intelligence could follow the same pattern.

The biggest economic question is therefore not:

How cheap will one token become?

It is:

How much machine intelligence will the economy demand once intelligence becomes cheap?

That question reaches directly into electricity markets.

The U.S. Department of Energy’s latest Lawrence Berkeley National Laboratory modeling estimates that data centers could reach 11.8% of total U.S. electricity consumption by 2030, with scenarios ranging from 9.5% to 15.3%.

The IEA estimates that more than 40% of the additional electricity required by global data centers through 2030 will still be met by natural gas and coal, even as renewable generation grows rapidly. The mix varies sharply by geography: U.S. data centers currently rely most heavily on natural gas, while Chinese data centers remain much more coal-intensive; Europe is projected to move toward a data-center electricity mix dominated by renewables and nuclear.

Then comes the capital bill.

The IEA says the capital expenditure of five major technology companies exceeded $400 billion in 2025 and is projected to rise a further 75% in 2026. It also warns that data-center expansion has become too capital intensive to rely only on corporate balance sheets, making future construction increasingly sensitive to capital markets, financing conditions and investor expectations about AI returns.

Federal Reserve Governor Lisa Cook pointed to more than $1.5 trillion of announced data-center plans, only a small fraction of which had been realized as of May 2026. She also noted rising prices in chips, high-tech equipment and software and stronger wage pressure in specialized construction trades as the investment boom feeds through the supply chain.

That is third-order economics.

The first-order question was:

How much does this AI answer cost?

The second-order question was:

What happens to the company’s margins and pricing model?

The third-order question is:

What happens when the entire economy tries to buy the chips, transformers, power plants, construction workers and financing required to generate trillions of those answers?

Now the economic consequences escape the technology sector.

A shortage of transformers becomes an AI constraint.

A gas-turbine backlog becomes an AI constraint.

A grid interconnection queue becomes an AI constraint.

A transmission project becomes AI infrastructure.

The IEA says shortages and delays now span transformers, gas turbines, advanced chips and high-bandwidth memory, with planning and grid-connection systems also under pressure.

And once compute begins competing for scarce physical capacity, the tax can spill over onto people who never asked for the AI product in the first place.

That is exactly why U.S. policy has started focusing on tax incidence, even if Washington doesn’t use that phrase.

In March 2026, the White House announced a Ratepayer Protection Pledge asking major hyperscalers and AI companies to “build, bring, or buy” the energy needed for new data centers, pay the cost of required power-delivery upgrades and negotiate structures under which they still pay for capacity created for their facilities even when it is unused. The explicit rationale is to prevent ordinary electricity customers from absorbing infrastructure costs created by data-center demand.

That is worth pausing on.

The government is effectively saying:

The compute tax should stay inside the AI economy rather than leaking onto everybody else’s electricity bill.

That is not really an AI regulation.

It is industrial economics.

And every country is approaching it differently.

Four economies, four ways to pay the bill

The competition between the United States, China, Europe and India is usually described as a race for models.

I think that’s becoming too narrow.

It’s increasingly a contest over who builds the cheapest, deepest and most strategically controlled supply of machine intelligence.

The Federal Reserve’s comparative work identifies compute as one of the clearest measures of national AI capacity. Based on observed high-end AI supercomputer capacity through mid-2025, it estimated roughly 74% in the United States, 14% in China and 4.8% in the EU, while cautioning that the underlying data capture only part of global capacity and that Chinese data are comparatively opaque.

But the countries are not simply building different quantities of the same thing.

They are building different political economies around compute.

The United States is treating compute primarily as a private-capital industry.

The U.S. advantage is extraordinary depth in private technology capital, hyperscalers, model companies, data centers, chips and cloud infrastructure. The policy response has largely tried to accelerate physical buildout while increasingly addressing the externalities that buildout creates. The White House’s 2026 ratepayer framework is a good example: let private companies build aggressively, but push the incremental grid and generation costs back toward the companies creating the demand.

That’s basically a market-led model with industrial-policy support.

Capital allocators decide where to build.

Technology companies decide which architectures win.

Energy markets respond.

Government increasingly tries to remove permitting and infrastructure constraints while preventing some of the cost from being socialized.

The strength of this model is brutally obvious: scale and capital.

The weakness is also obvious: private firms will optimize around private returns, which can collide with grid planning, local politics, electricity affordability and infrastructure lead times. The IEA now identifies U.S. grid queues, power equipment and generation as practical bottlenecks to the pace of data-center construction.

China is thinking about compute more like national infrastructure.

The distinction is striking.

China’s National Development and Reform Commission describes the country’s emerging national compute network as infrastructure that links computing resources across regions. The “East Data, West Computing” program was designed partly to shift compute demand from economically dense eastern regions toward western regions where energy resources are more abundant. Eight national computing hubs and ten data-center clusters formed the initial national structure, and subsequent policy has moved toward integrated scheduling and more market-oriented allocation of compute resources.

China’s current Five-Year Plan framework goes further, calling for an integrated national computing network with resource pooling, monitoring, scheduling and operating standards while coordinating compute-network development with the electricity grid and communications infrastructure.

That is a very different way of conceptualizing the problem.

America largely asks:

Where will private capital build compute?

China increasingly asks:

How should the national compute system be architected?

That doesn’t mean China has eliminated markets. Its policy language explicitly discusses marketization and resource allocation. But the physical architecture is much more deliberately coordinated at the state level.

China also has a potentially enormous energy advantage. Federal Reserve research notes that China’s power-generation infrastructure expanded far faster than that of the United States in recent years, even though the U.S. currently retains a commanding lead in high-end compute capacity.

There is a catch.

China’s current data-center electricity supply remains much more carbon intensive. The IEA estimates that coal supplies close to 70% of the electricity serving Chinese data centers today, with renewables near 20% and nuclear around 10%. Policy is deliberately steering more facilities toward renewables-rich western regions, and both coal and renewable generation are projected to expand materially in response to data-center demand through 2030.

So China is doing something strategically clever but environmentally messy:

It is treating the geography of compute and the geography of energy as the same problem.

That idea may turn out to matter enormously.

Europe is trying to socialize part of the fixed cost of strategic compute.

The European Union’s model sits somewhere else again.

Rather than relying exclusively on a handful of domestic hyperscalers that Europe largely does not have, the EU is building shared AI infrastructure through AI Factories and the newer AI Gigafactory program.

In July 2026, the EU launched a tender for as many as seven AI Gigafactories. The initiative is backed by up to €10 billion in EU and national public funding and is designed to unlock at least €20 billion more from private investors. The planned facilities combine processors, cloud software, connectivity and energy-efficient data centers and are meant to provide compute access to startups, scale-ups, industry, researchers and public authorities.

Europe is therefore making a different bet:

Compute capacity itself has become strategic infrastructure, and some of its fixed cost should be pooled because dependence on foreign compute can become an industrial dependency.

The EU explicitly frames these facilities around technological resilience and strategic autonomy as well as compliance with European rules on safety, data protection and security.

Whether that produces globally competitive AI companies is not guaranteed.

But economically, the logic is clear.

Europe is trying to lower the compute tax faced by domestic innovators by helping finance the fixed infrastructure underneath them.

India is doing something more direct: subsidizing access.

India’s model may be the most explicit acknowledgment that compute access itself is an economic input.

The government’s IndiaAI Mission has a budget of roughly ₹10,372 crore and, by June 2026, had expanded a shared national compute pool to more than 45,000 GPUs. By August, according to the government, 237 projects had used subsidized computing capacity covering 9.318 million GPU hours, specifically to reduce barriers for researchers, startups and innovators.

India cannot currently outspend the American hyperscalers.

So instead it is trying to change the economics faced by the user.

Rather than every startup individually acquiring scarce compute at global market prices, government procurement aggregates capacity and makes it cheaper to access.

My read is that these four models amount to four different answers to the same question:

Who should absorb the fixed cost of making machine intelligence abundant?

America: mostly private capital.

China: coordinated national infrastructure plus markets.

Europe: public-private sovereign infrastructure.

India: subsidized shared access.

That distinction could become as consequential as differences in model capability.

Because in an agentic economy, the country with the cheapest reliable intelligence input may eventually have an advantage similar to the country that once had cheap electricity, cheap labor or cheap capital.

The argument experts are actually having

There are at least three big disagreements here, and they are more interesting than the usual “AI good versus AI bad” debate.

The first is whether the compute tax survives at all.

There is a strong case that inference becomes so cheap that worrying about it is pointless.

The evidence for this view is formidable. Stanford’s AI Index found that the cost of running a model at roughly GPT-3.5-level benchmark performance dropped from about $20 per million tokens in November 2022 to $0.07 by October 2024, a decline of more than 280-fold in roughly 18 months. Across different tasks, Stanford reported inference-price improvements spanning roughly 9-fold to 900-fold annually. McKinsey has highlighted the same underlying cost collapse as one of the defining AI technology trends.

That trend is real.

The mistake is concluding that falling unit costs make aggregate compute economics irrelevant.

The IEA is observing the opposite phenomenon in electricity: energy per task is dropping rapidly while total AI electricity demand accelerates, because adoption and task intensity are growing faster.

Those two things can coexist indefinitely.

Imagine inference becomes ten times cheaper.

Then imagine agents become cheap enough that businesses run one hundred times as many autonomous tasks.

The unit tax falls 90%.

The aggregate compute bill rises tenfold.

That is why I don’t think the right debate is:

Will inference get cheaper?

Almost certainly.

The better debate is:

Will efficiency improve faster than demand for machine intelligence expands?

Right now, nobody can answer that confidently.

The second disagreement is whether all this compute produces an economic return commensurate with the capital going into it.

Here the range of credible expert opinion is enormous.

McKinsey’s latest work on agents and robotics estimates that a midpoint adoption scenario could unlock roughly $2.9 trillion annually in U.S. economic value by 2030, primarily through automation and the redeployment of labor hours. McKinsey is careful to call this potential economic value rather than a forecast of equivalent GDP growth and says realizing it depends heavily on workflow redesign and organizational adaptation.

Daron Acemoglu at MIT is much more skeptical about macroeconomic impact. His modeling places the ten-year increase in total factor productivity from current AI advances at no more than around 0.66%, with an estimate around 0.53% once harder-to-learn tasks are treated more conservatively. His corresponding GDP estimates are roughly 0.93% to 1.16% over ten years under a modest investment response and as high as roughly 1.4% to 1.56% under a larger capital boom.

These estimates are not directly comparable. McKinsey is measuring the potential economic value of work that could be automated or reallocated under an adoption scenario; Acemoglu is estimating economy-wide productivity and GDP effects using a much more restrictive macro framework. But the gap tells you just how unsettled the economics still are.

The Federal Reserve sits somewhere in the middle.

Its July 2026 analysis says task-level experiments can show substantial improvements without those improvements immediately appearing at the company or national level. Making a programmer 10% faster at one task does not automatically make the company 10% more productive if bottlenecks elsewhere remain. The Fed says there was still no large aggregate productivity signal clearly attributable to AI as of 2026, while emphasizing that historically, general-purpose technologies often require years of complementary investment and organizational change before productivity gains become visible.

The OECD reaches a similar conclusion: controlled and workplace studies often report 20% to 40% performance improvements on specific tasks, depending on context, while the economy-wide and long-run consequences remain uncertain.

This disagreement matters enormously for investors.

Because a trillion-dollar compute buildout is very easy to justify if AI ultimately reorganizes several trillion dollars of annual labor and economic activity.

It is much harder to justify if most companies end up buying expensive copilots that make employees moderately faster without fundamentally changing output.

That is why the decisive variable isn’t benchmark intelligence anymore.

It is organizational conversion.

Can companies turn model capability into completed economic work?

OpenAI, McKinsey, BCG and the Federal Reserve, despite coming from very different positions, are converging on that point: deeper workflow redesign and successful outcomes matter more than raw usage.

The third disagreement is subtler.

Does AI raise transaction costs or destroy them?

My answer is:

Both.

AI agents are likely to crush a huge category of Coasean transaction costs: searching for information, finding counterparties, comparing options, negotiating terms, monitoring performance and coordinating activity. MIT researchers explicitly identify this reduction in search, communication and contracting friction as one of the fundamental economic promises of agents.

But they do it by substituting computation for human effort.

That means an old transaction that might have cost:

$50 of employee time

• $20 of administration

• three days of latency

could become:

$0.70 of compute

• $0.10 of database and search access

• thirty seconds

The compute tax rose from zero to 70 cents.

The total economic transaction cost collapsed from $70 to less than a dollar.

There is no contradiction.

And this distinction is critical, because it tells us not to optimize for minimum compute spending.

The objective should be minimum total cost per valuable outcome.

OpenAI calls its version “Useful Intelligence per Dollar.” BCG calls for measuring return against the combined costs of human intelligence and tokens. Both frameworks point in the same direction.

Spend more compute when another dollar of inference replaces twenty dollars of labor.

Spend less compute when a deterministic API can solve the problem for a fraction of the cost.

Use an expensive frontier model when reliability saves five retries.

Use a tiny model when the task is trivial.

Use no model at all when normal software works.

This sounds obvious.

It isn’t how much of the industry has been built so far.

What the capital allocator should take away

I think the AI economy is heading toward a realization.

Intelligence is becoming cheap, but it is not becoming free.

And once intelligence becomes embedded in products, the economics of intelligence start appearing everywhere.

In gross margins.

In pricing models.

In cloud bills.

In electricity markets.

In data-center financing.

In transformer factories.

In grid regulation.

In government industrial strategy.

In national competitiveness.

The winners could be the companies that turn each dollar of computation into more economically valuable work than everybody else.

For founders, that means one of the most important metrics in the company may eventually be something deceptively simple:

Compute cost per successful customer outcome.

Not token consumption.

Not model calls.

Not users chatting with the agent.

Actual finished work.

For CEOs, it means AI expenditure should stop being buried inside a generic technology budget. BCG’s current framework separates AI expenditure among investment in reusable capabilities, operating costs for internal AI work, and COGS when inference sits inside customer-facing products. That distinction becomes increasingly important as agent use expands.

For VCs, I would be increasingly suspicious of AI businesses whose economics are basically:

buy intelligence retail, resell intelligence with a markup.

That is not automatically a moat.

The more interesting companies are the ones that can structurally change the ratio between compute consumed and value produced through proprietary data, workflow integration, model routing, specialized evaluation, distribution or ownership of the customer relationship. BCG’s latest competitive analysis similarly argues that durable advantage is shifting toward assets and relationships that improve with scale and use rather than generic capabilities that AI can easily replicate.

For infrastructure investors, the compute tax moves upstream.

Inference demand becomes server demand.

Server demand becomes electricity demand.

Electricity demand becomes generation, storage and transmission demand.

Those requirements create demand for chips, networking, memory, cooling, transformers, turbines, batteries and capital. The IEA already sees pressure and bottlenecks across many of these categories and expects capital markets to become increasingly important in financing continued data-center expansion.

For policymakers, the problem is even bigger.

The central policy question will increasingly be not merely how much AI a country develops, but what domestic businesses pay to access machine intelligence and who ultimately absorbs the infrastructure costs required to produce it.

America is trying to make private companies internalize more of the power-system costs their data centers create.

China is building a nationally coordinated compute network around the geography of electricity and demand.

Europe is using public capital to build shared sovereign compute capacity and pull in private investment.

India is directly subsidizing compute access for domestic innovators.

Those are not just AI policies.

They are different theories of who should pay for intelligence infrastructure.

And over the next decade, that may turn out to matter far more than most arguments about which chatbot currently wins which benchmark.

Because the important economic transformation isn’t that AI can answer questions.

It is that AI is beginning to do work.

Once an agent searches, reasons, compares, negotiates, codes, buys, sells, monitors and acts, computation stops being something that happens somewhere in the background of the technology industry.

It becomes an input into economic production.

Every AI-mediated action carries some amount of it.

The price of that input will keep falling. Stanford’s data already show how violently inference costs can collapse, while the IEA simultaneously shows how rapidly aggregate demand can rise as efficiency improves.

That is why the central paradox of the next phase of AI is so important:

The compute tax could approach zero per transaction while becoming enormous in aggregate.

That isn’t a bug in the AI economy.

It may be the AI economy.

We spent the software era turning computation into something users barely had to think about.

Now we are turning intelligence itself into a metered commodity.

And when intelligence has a price, the people who control its production, distribution, efficiency and energy supply don’t merely own pieces of the technology stack.

They sit inside the cost structure of everything built on top of it.

Share Full Stack Capitalist

Goldman Sachs put a number on the AI buildout that finally makes the scale legible.

Global AI investment will cross $1 trillion in 2026, with $581 billion of that landing inside the United States.

By 2028, on Goldman’s own extrapolation, AI capex reaches 2.8 percent of US GDP.

Say that number a different way. One dollar in every 36 the American economy produces will be spent building the physical guts of artificial intelligence. Transformers, turbines, copper, land, concrete, switchgear, and the electricians and linemen who install all of it.

Looks like bubble? Whether the valuations hold, whether the revenue shows up, whether Microsoft and Meta and Oracle can actually earn a return on the debt they’re taking on to build it.

That’s the first order question, and it’s the one every analyst on financial television is paid to answer. It’s also, for our purposes, the less interesting one.

What happens to everything else while this is happening. Because 2.8 percent of GDP does not appear out of thin air. It has to come from somewhere.

Capital that goes into a transformer for a data center in northern Virginia is capital that does not go into a transformer for a semiconductor fab in Arizona, a housing development in Texas, or a grid upgrade in Ohio. A construction crew pouring concrete for a hyperscale campus is a construction crew that is not pouring concrete for an apartment building. This is not a hypothetical. It is already showing up in the data, and it is going to reshape which industries get funded and which get starved for the rest of this decade.

The Binding Constraint Has Moved, and It Keeps Moving

The Full Stack Capitalist lens on any technology buildout is simple. Find the thing that is actually scarce, not the thing that gets the headlines, and trace who captures the rent from controlling it. For most of 2023 and 2024 the scarce resource was GPUs.

That constraint is largely gone. TSMC has repeatedly expanded advanced packaging capacity, and chip allocation is no longer the thing standing between a hyperscaler and a bigger cluster.

The constraint moved to power, and specifically to the physical equipment needed to deliver power.

Power transformers were averaging roughly 128 weeks of lead time by 2026, with generator step up units stretching to 144 weeks and the highest capacity units quoted four to five years out.

Before 2020 the same transformer shipped in under two years, often in a matter of weeks. Of the roughly sixteen gigawatts of data center capacity announced for 2026 in the US, only about five gigawatts was actually under construction, with analysts estimating that 30 to 50 percent of the announced pipeline would be delayed or cancelled outright because of transformer and switchgear shortages, not because the money wasn’t there.

That last clause is the whole essay in miniature. Not because the money wasn’t there. This is a capital abundant, physical infrastructure scarce economy. The hyperscalers have functionally infinite access to debt and equity markets. What they cannot do is manufacture grain oriented electrical steel faster, and China controls roughly 60 percent of the world’s transformer manufacturing capacity, which means the bottleneck is also a geopolitical one.

When the binding constraint is a slow moving industrial input rather than a fast moving financial one, the excess capital doesn’t sit idle. It goes looking for the next scarce thing and bids the price up. That’s copper, and it’s construction labor, and increasingly it’s the electricity itself.

Copper Is the Cleanest Proxy for What’s Being Bid Away

A single one gigawatt AI facility requires on the order of 50,000 metric tons of copper.

At current buildout paces of roughly fifteen gigawatts of new capacity a year, data centers alone are adding somewhere around 750,000 metric tons of incremental copper demand annually, and that’s before counting the copper needed for the substations, transmission lines, and grid upgrades required to actually deliver the power, which several analysts now believe will exceed the copper used inside the buildings themselves.

Copper prices have risen more than 60 percent since early 2025 and have repeatedly traded above $13,000 to $14,500 a metric ton, levels that used to be considered crisis pricing.

Copper is inelastic in a way most commodities aren’t.

You can’t substitute your way around a high voltage application the way you can swap aluminum into a low stakes consumer product. So every other industry that needs copper, electric vehicle manufacturing, grid modernization for reasons that have nothing to do with AI, home electrification, general industrial construction, is now bidding against hyperscalers with functionally unlimited balance sheets for a metal whose supply grows at roughly 1.4 percent a year against demand growing far faster.

Mine permitting timelines run 15 to 17 years from discovery to production. Supply cannot respond to price. Demand, backstopped by trillion dollar capex budgets, doesn’t have to care about price. That’s a structural transfer of purchasing power from every copper consuming industry that isn’t AI, straight into the AI buildout, and it happens silently, through a commodities market, without anyone voting on it.

Manufacturers Are Losing to Data Centers for the Same Megawatts

This is the part of the story that gets the least attention and matters the most, because it’s happening at the level of individual grid interconnections, not abstract GDP shares.

In regions across the country, manufacturers are discovering that the power capacity they assumed would be available has already been claimed by a data center campus down the road.

PJM, the regional grid operator spanning thirteen states and Washington DC, received more than 800 project applications totaling roughly 220 gigawatts in its most recent interconnection cycle, a queue that was never designed to process anything close to that volume.

This is not an abstract crowding out story. It is the reshoring story colliding head on with the AI story, and AI is winning. Every speech about bringing semiconductor fabs and battery plants and industrial capacity back to American soil assumed those factories would be able to get power on a normal timeline.

They can’t, because the queue in front of them is now dominated by hyperscale data center requests that arrived with more capital and, often, more political priority attached. Eaton, Vertiv, and Schneider Electric have all announced new manufacturing capacity for the switchgear and transformers everyone needs, but new factories take years to reach output, which means the equipment shortage outlasts the current wave of expansion plans on both sides.

The bottleneck has migrated from the substation to the factory floor of the companies that build substations, and there is no version of that sentence that is good news for a reshoring agenda that depends on getting industrial power online this decade rather than next.

There’s a second casualty inside the power story that shows up on household bills rather than corporate balance sheets.

PJM’s electricity suppliers paid $14.7 billion at a recent capacity auction to guarantee they’d have enough power for their customers, up from $2.2 billion the year before, a nearly sevenfold jump driven substantially by the slow moving interconnection queue that data center demand has helped clog. Ratepayers, most of whom have no stake in the AI trade whatsoever, are the ones covering that spread. That is a direct, traceable transfer from households to the AI buildout, mediated by a grid queue nobody outside the industry is watching.

Credit Markets Are the Newest Front, and It’s Bigger Than People Realize

The capex crowding story usually gets told through physical inputs. The credit market version is just as real and less discussed.

Hyperscaler capex in 2026 is consuming close to 100 percent of operating cash flow, compared with a ten year average closer to 40 percent, which means the era of AI being funded out of Big Tech’s spare cash is over.

The five largest hyperscalers issued around $121 billion in US corporate bonds in 2025, more than four times their 2020 to 2024 annual average, and issuance has accelerated further this year, with Goldman projecting roughly $250 billion in hyperscaler bond sales for 2026 rising toward $400 billion in 2027. Technology has climbed to roughly 10 percent of the Bloomberg US Corporate Index and, in several major benchmarks, has overtaken banking in weight for the first time in the index’s history.

That reshuffling matters beyond the tech sector because investment grade credit markets are not infinitely deep in any given window.

Overall US corporate bond issuance is expected to hit $2.46 trillion in 2026, and passive bond funds that track investment grade indexes will mechanically absorb more hyperscaler debt as the weighting shifts, which means every other investment grade borrower, industrials, utilities, healthcare systems, is now competing for spread and investor attention against the highest quality, most aggressively marketed issuers in the market.

AI debt is also colliding with a federal government running toward a $2 trillion annual deficit and no longer backstopped by a Federal Reserve willing to buy Treasuries at scale, which means every dollar of capital markets appetite is being split three ways between sovereign debt, AI infrastructure debt, and everyone else. Early signs of investor fatigue are already visible.

A recent Amazon bond sale needed to offer 18 to 21 basis points of extra yield to clear, with orders covering the deal only 2.5 times versus 3.2 times earlier in the year. When the most creditworthy borrowers on earth have to sweeten the deal to move paper, that’s the credit market’s way of saying the well has a bottom.

Who Actually Gets Defunded

Put the physical and financial constraints together and a pattern emerges. The industries that lose in this environment are the ones that share AI’s inputs but not its capital access.

Reshored manufacturing loses because it needs the same transformers, the same grid interconnections, and the same skilled electrical labor, but arrives at the queue with a fraction of the balance sheet.

Housing and general construction lose because they compete for the same copper, the same construction crews, and increasingly the same land near power infrastructure, without the pricing power to bid AI campuses off a site.

Mid tier industrial and healthcare borrowers lose in the credit markets because investment grade capacity that used to flow to them now flows through indexes increasingly weighted toward hyperscaler paper. Utility ratepayers lose directly, through capacity auction costs that get passed straight to household bills.

And clean energy and grid modernization projects that have nothing to do with AI, projects that in a fairer queue would be interconnecting today, sit for five years or more behind data center requests that jumped the line with bigger checkbooks.

None of this required anyone to make a decision to defund those sectors.

That’s what makes it a genuine economic phenomenon rather than a policy failure with an obvious villain. It’s what economists mean by crowding out, capital and physical inputs flowing toward the highest expected return sector in a supply constrained system, with everyone else absorbing the opportunity cost whether they had a vote in the matter or not.

The externality here isn’t carbon or noise. It’s every other capital intensive industry in the country discovering that its cost of capital, its equipment lead times, and its construction timelines just got worse because of a buildout it has no stake in.

What This Means If You’re the One Bidding, or the One Losing the Bid

If you’re a founder building anything that touches physical infrastructure, energy, industrial equipment, construction, manufacturing, treat 2026 through 2028 as a period where your input costs are structurally elevated regardless of what your own sector is doing. Transformer and switchgear lead times aren’t a data center problem you can ignore. They’re now your problem too, and the companies solving for it, whether through behind the meter generation or vertically integrated equipment supply, are building a genuine moat.

If you’re an operator running a company that depends on grid interconnection, get in the queue earlier than you think you need to, and underwrite your project assuming the equipment timeline, not the permitting timeline, is your critical path. The site with land control and zoning is worthless if the transformer arrives in 2029.

If you’re an investor, the interesting trade isn’t only the hyperscalers or the chipmakers everyone already owns. It’s the picks and shovels one layer further down, the transformer and switchgear manufacturers, the copper producers who can actually bring new supply online inside a normal timeframe, and the credit story itself, where investment grade spread widening on hyperscaler paper is a signal worth watching closely rather than dismissing as noise from AA rated borrowers.

If you’re in government, the reshoring agenda and the AI agenda are currently fighting each other for the same substations, and pretending otherwise doesn’t change the interconnection queue. Any serious industrial policy in 2026 has to reckon with the fact that a semiconductor fab and a hyperscale data center are now direct competitors for the same transformer order book, and a queue that runs first come, first served with no strategic prioritization will keep handing the win to whichever project has the biggest balance sheet, not the one that matters most to national competitiveness.

The trillion dollar number was never really about AI. It’s about what an economy looks like when one sector can outbid every other sector for the physical and financial inputs all of them need at once.

Something has to give. The Goldman Sachs report just told you the size of the bill. It didn’t tell you who’s paying it. That part you have to work out yourself, and increasingly, the answer is everyone who isn’t in the room.

Share Full Stack Capitalist

On August 10, Jensen Huang sat down with five of the most powerful capital allocators on earth.

NVIDIA signed memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize over $500 billion in third-party capital for AI infrastructure. Huang’s framing on CNBC was: this is the first time technology chips have become an investable asset class.

Read the announcement again.

NVIDIA isn’t lending its own balance sheet to customers, the way it awkwardly tried to do weeks earlier when reports surfaced that it might backstop $250 billion for a single OpenAI buildout and the stock dropped hard on the news.

This is the opposite structure. NVIDIA gets six independent, deep-pocketed asset managers to underwrite the exposure instead, so nobody can accuse NVIDIA of manufacturing its own demand.

Apollo brings roughly a trillion dollars of assets under management. Blackstone brings north of $1.3 trillion. Brookfield brings another trillion. BlackRock’s Larry Fink is out there calling the AI buildout “unprecedented investment” that only long-duration capital can fund.

Long-duration capital means pension funds, insurers, and increasingly retirement products. NVIDIA is trying to convert a chip replacement cycle into a bond market.

The binding constraint just moved

Until this year, AI capex was constrained by what Microsoft, Meta, Amazon, Oracle, CoreWeave and OpenAI could fund off their own cash flow and equity. That ceiling was already straining.

Moody’s has been flagging that unprecedented capital expenditures are squeezing hyperscaler free cash flow and pushing them toward heavier debt loads.

The BIS put a number on how much of that debt isn’t even showing up where you’d look for it: in its March 2026 Quarterly Review, the central bank of central banks documented that hyperscalers are increasingly funding data centers through special purpose vehicles that raise private debt while the hyperscaler merely signs a long-term operating lease.

Economically that lease is a financial commitment. On the income statement it’s rent. Moody’s estimated hyperscalers are sitting on roughly $662 billion of these signed-but-not-yet-started lease commitments, off balance sheet, a figure bigger than those same companies’ combined on-balance-sheet debt.

The BIS calls this “shadow borrowing.” I’d call it the AI industry discovering what every capital-intensive industry eventually discovers: you can’t fund $400 billion a year in infrastructure out of $60 billion in revenue forever, so you either slow down or you find someone else’s balance sheet.

NVIDIA’s answer is to institutionalize the someone-else’s-balance-sheet part. Instead of ad hoc SPVs negotiated deal by deal, you get standing platforms with Wall Street’s biggest infrastructure investors, designed to keep pumping capital into the system indefinitely.

That’s the binding constraint shift: AI buildout stops being limited by tech company cash flow and starts being limited by how much of the world’s retirement and insurance capital is willing to underwrite GPU clusters.

Once that gate opens, the more interesting question isn’t whether the money shows up. It’s who eats it if the bet is wrong.

Scenario one: NVIDIA is right and this becomes infrastructure

If AI demand genuinely compounds for another decade, if enterprise adoption becomes ubiquitous and agent workloads keep inference demand climbing, this financing innovation looks like one of the more important structural shifts in the industry’s history.

Capital costs fall because a trillion-dollar pool of patient capital is cheaper than equity. Smaller countries can finance sovereign compute without waiting on hyperscaler goodwill. Enterprises rent capacity instead of building it.

AI infrastructure becomes a standard institutional allocation next to real estate, utilities, and airports, and NVIDIA quietly becomes something stranger than a chip company: an architect of a global compute capital market, collecting hardware revenue on volumes that its own balance sheet never had to risk.

That’s the bull case, and it’s coherent. It’s also the case everyone currently pricing NVDA is implicitly assuming.

Scenario two: demand disappoints, but only a little

This is the realistic base case if you’re being honest about how these cycles usually run.

AI demand is real, but investors have overestimated utilization, pricing power, GPU useful life, or customer credit quality. Nothing collapses. Returns just come in worse than modeled.

GPU rental prices soften. Residual values on aging clusters get marked down.

Operators refinance at worse spreads. Private credit spreads widen. Marginal data center projects that only worked at 80% utilization assumptions get canceled or need more equity to pencil. Neocloud consolidation accelerates as weaker operators get absorbed.

Nobody blows up. Ownership just transfers from the optimistic first-round financiers to distressed-capital specialists who buy the assets for less than they cost to build.

Apollo and Blackstone, notably, are the ones best positioned to be on both sides of that trade: underwriters of the original financing and buyers of the eventual distress.

Scenario three: the credit channel actually breaks

Now push harder. Model efficiency improves faster than compute demand. Custom silicon eats real share. Inference prices fall further and faster than underwriting assumed.

Corporate customers won’t sign the long-duration contracts the debt was structured against. A facility underwritten at 80% utilization runs at 40%.

Cash flow falls. Debt service doesn’t. Here’s the chain, and it’s worth sitting with because this is where a hardware demand problem becomes a macro problem:

An AI company can’t honor its compute contract. The AI cloud operator loses revenue. The SPV can’t service its debt. The GPU collateral gets marked down. The private credit fund holding that debt takes losses.

The pension or insurance investor that owns a slice of that fund marks down its exposure. New lending to the sector freezes. Data center construction stops. GPU orders fall. NVIDIA’s revenue falls. Suppliers cut capex. Power projects tied to those data centers get canceled. Construction employment falls.

The BIS has already mapped these transmission channels explicitly: refinancing stress, shifts in private credit risk appetite, guarantees getting triggered, and connections looping back into banks through warehouse lending lines.

What started as a GPU utilization problem becomes a credit problem, then a capex problem, then an employment problem. That’s what financialization does in every cycle. It doesn’t just fund the boom. It manufactures the amplitude of the bust.

Scenario four: it goes systemic

This requires several things stacking at once: heavy leverage, correlated exposure across the sector, underwriting that assumed too much for too long, meaningful retirement and insurance capital concentrated in the exposure, banks financing the private credit vehicles that financed the SPVs, and a real, not cosmetic, disappointment in AI revenue.

Put those together and you get a liquidity crisis around assets everyone modeled as long-duration infrastructure but that behave, collateral-wise, like depreciating technology equipment.

Pension funding ratios deteriorate. Insurers take write-downs. Private funds gate redemptions. Banks pull warehouse financing. Credit spreads jump across the sector, not just at the margin. Infrastructure projects with nothing to do with AI become harder to finance because the capital that would have funded them is repricing risk everywhere.

At that point the exposure has migrated into institutions that are too politically important to let fail, which sets up the fourth-order problem: what does a government do when public pension plans own the funds, insurers hold the debt, banks financed the vehicles, utilities built power infrastructure around the facilities, and thousands of construction jobs depend on the projects continuing?

History says the answer is rarely “let the capital structure clear.” It’s restructuring, guarantees, tax relief, cheap financing, and possibly government-backed compute demand. Not because GPUs deserve rescuing, but because the exposure moved into places democracies can’t easily let fail.

Why the history keeps rhyming

None of this is new, which is exactly why it’s worth taking seriously instead of dismissing as bubble-talk.

Britain’s canal mania in the 1790s saw the successful canals, like the Bridgewater, cut coal transport costs into Manchester so dramatically that investors piled into every canal scheme that followed. Many of those follow-on canals never produced the promised returns. The investors lost money. Britain kept the canal network.

Railways in the 1840s ran the same play at greater scale.

The technology worked, demand was real, and the Railway Mania still wiped out enormous amounts of private capital when valuations collapsed. Britain kept the railways.

American railway expansion in the 1860s and 1870s was financed almost entirely on debt, with railway bonds reportedly reaching close to a third of GDP by 1890. The Panic of 1873 hit, defaults spread, iron demand collapsed, factories closed, and unemployment rose.

Nobody looked back in 1880 and concluded railways were a mistake. The infrastructure was transformational. The capital structure and the timing were the problem.

Electricity is the closest thing to a clean analogue for what “compute as infrastructure” is supposed to look like, and it’s also the comparison that exposes NVIDIA’s real vulnerability.

Generators, transmission, and distribution became close to ideal pension assets because they have very long useful lives, predictable demand, regulated pricing, and low technological obsolescence.

A transformer built today doesn’t become worthless because a better transformer ships next year. A GPU cluster is a different animal entirely. The data center shell might last thirty years.

The power connection might last fifty. The frontier compute inside it has a competitive life measured in a few years, with some of NVIDIA’s own performance metrics historically doubling roughly every eighteen months to two years.

That’s the structural mismatch sitting underneath the entire “compute is an investable asset class” pitch: the building and the chip inside it depreciate on completely different curves, and the financing is being priced as if they don’t.

Telecom in the 1990s finished the pattern. Fiber got built on the correct bet that internet traffic would explode, traffic did explode, and investors still got wiped out because too much capacity got built too fast at prices that assumed permanent scarcity. Capacity became abundant, prices collapsed, companies went bankrupt, and the next generation of internet companies inherited extraordinarily cheap bandwidth built by someone else’s losses.

The BIS’s own data on the current cycle already rhymes with this pattern more than the industry wants to admit.

Private credit loans to AI-related companies have gone from near zero to reportedly over $200 billion in a few years, with projections of hundreds of billions more over the next two years.

Hyperscaler bond issuance has passed $100 billion in a single year while credit default swap spreads on that debt have been climbing at the same time, which means bond investors are already quietly pricing in more risk than the private credit spreads on AI loans currently reflect. That gap between what bond markets and private credit markets think the same risk is worth is exactly the kind of mispricing that shows up right before someone discovers it the hard way.

The sentence that matters more than the pension warning

There’s a temptation to write this up as “your retirement account is buying GPUs,” and that framing is too clean.

Direct 401(k) exposure to private credit is still underdeveloped, though the Department of Labor moved in March 2026 to make it easier for 401(k) plans to hold alternatives including private credit and private equity, which tells you which direction this is heading even if it hasn’t fully arrived.

The sharper claim is this one: the biggest mistake investors can make is assuming that because AI will transform the economy, AI infrastructure must therefore be a good investment.

History says close to the opposite.

Transformative infrastructure attracts too much capital precisely because everyone correctly recognizes it will transform the economy, and the resulting competition transfers most of the economic surplus from the people who financed the buildout to the people who use it afterward. Railway passengers won.

Manufacturers won on cheap freight. Internet companies won on cheap fiber. Future AI companies may well win on GPU capacity nobody expected to get this cheap this fast. The losers are usually whoever financed the first generation.

That’s what NVIDIA’s August 10 move is actually about. Not “is AI real.” It obviously is. The question is who owns the depreciation risk while the revolution gets built, and NVIDIA just spent $500 billion of other people’s underwriting to make sure the answer isn’t NVIDIA.

What this means if you’re building, running, investing, or regulating

Founders and operators building on rented compute: the era of GPU scarcity dictating your unit economics is not permanent. If utilization disappoints anywhere in this financing stack, capacity gets dumped by distressed holders faster than you can plan around, and your input costs could fall faster than your business model assumes. Build cost structures that benefit from cheaper compute rather than ones that depend on scarcity pricing holding.

Investors: the equity story and the credit story on AI infrastructure are being priced by two different markets right now, and they disagree. Private credit spreads on AI loans are running close to spreads on ordinary private credit, which prices AI risk as roughly average. Equity valuations on the same companies price in outsized returns. One of those markets is wrong, and it’s worth having a view on which one before you’re on the wrong side of the repricing.

Governments and policymakers: the too-important-to-fail characteristic isn’t building around NVIDIA. It’s building around the capital stack: the pension exposure, the insurance holdings, the bank financing lines behind the private credit funds. That’s the thing worth stress-testing now, well before anyone needs a bailout conversation, because by the time the exposure is visible it’s already systemic.

Everyone with a retirement account: you may end up simultaneously exposed to AI as a threat to your labor income and as a source of your retirement returns. That’s a genuinely new distributional position, and it’s worth understanding what your pension or 401(k) actually holds rather than assuming “diversified” means “insulated” from this specific cycle.

Share Full Stack Capitalist

NVIDIA is partnering with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to mobilize more than 500 billion dollars in third party capital for AI infrastructure.

The framing is clean and confident. Compute is becoming an investable asset class, like toll roads or commercial real estate. NVIDIA provides the platform, the financial institutions provide the capital and the underwriting discipline, and everybody wins.

Read it again, slower. What actually happened is that NVIDIA found a new balance sheet to stand on. And the balance sheet it found belongs, eventually, to your 401k.

That is not a metaphor. It is the literal mechanical path the money takes. To understand why this deal matters, you have to walk through three layers of economics, because each layer solves the exact problem created by the layer before it, and each solution pushes the same underlying risk one step further from anyone who chose to take it.

First order: the circular loop everyone already knows about

Start with the thing people have been complaining about for over a year.

NVIDIA invests in its own customers, those customers use the money to buy NVIDIA chips, and NVIDIA books the sale as revenue. Between 2020 and 2025, NVIDIA made roughly 170 investments worth about 53 billion dollars across the AI ecosystem, including stakes in OpenAI, CoreWeave, and Nebius, companies that turn around and spend heavily on NVIDIA hardware.

In 2025 alone the pace accelerated to 59 deals worth 23.7 billion dollars. NVIDIA holds a 91 percent stake in CoreWeave. It committed up to 100 billion dollars to OpenAI. It just put up to 10 billion into Anthropic alongside a 5 billion dollar Microsoft investment, with Anthropic agreeing to buy 30 billion dollars of Azure compute in return.

Jensen calls the circularity accusation ridiculous, and he has a point in the narrow sense that his equity checks are small relative to what these companies raise elsewhere.

But the deeper issue was never the size of any single check. It was that the whole ecosystem’s reported demand was being partially financed by the same company selling the product, which makes it structurally hard to tell how much of the AI buildout reflects end user willingness to pay versus vendor financed order backfilling.

That is first order circularity, and it’s the version of this story that’s been in the financial press for a year.

Second order: the banks said no, so the debt went private

First order circularity created a second, less visible problem: concentration risk on bank balance sheets.

Oracle’s roughly 300 billion dollar buildout commitment tied to OpenAI pushed major banks toward their single counterparty concentration limits.

Data center construction loans are unusually large relative to typical infrastructure lending, so a handful of delayed or canceled projects can blow out a bank’s exposure to the entire sector. Once banks hit those limits, they stopped being willing or able to originate new data center debt at the pace the buildout needed.

So the financing moved somewhere banks don’t have to answer to regulators about single name concentration: private credit.

Loans from private credit funds to AI related companies went from near zero to over 200 billion dollars in a few years, and Morgan Stanley projects another 800 billion in private data center financing on top of that.

CoreWeave’s 7.5 billion dollar debt facility, arranged through Blackstone’s tactical opportunities group, is secured by the company’s GPUs and customer contracts, carries a variable rate around 11 percent, and started requiring repayment just as the value of that GPU collateral was already softening. CoreWeave’s interest payments now eat about 26 percent of revenue and 46 percent of adjusted EBITDA.

This is the second order move. Risk that used to sit on regulated bank balance sheets, subject to stress tests and capital requirements, got repackaged as private credit and moved to firms that don’t face the same oversight.

Third order: private credit needs duration capital, and duration capital means your retirement

Private credit funds don’t hold this debt forever on their own money.

They raise it from limited partners who want long duration, stable yield assets to match long duration, stable liabilities.

Who has liabilities like that? Pension funds and insurance companies.

That’s not incidental, it’s the entire point of the structure. Larry Fink said this part back in May, describing the roughly 10 trillion dollars of infrastructure investment the US needs over the next decade as money that has to come from savings accounts, pension accounts, and insurance companies, because the private sector is where the scale actually lives.

The channel is already well established.

New York and Pennsylvania state pension plans have money in Blue Owl’s 7 billion dollar digital infrastructure fund, the same fund behind Meta’s data center financing vehicles and multiple Oracle deals.

Major life insurers now hold nearly a trillion dollars in private credit overall. When hyperscaler backed data center bonds have come to market, insurance companies and pension funds have been the primary buyers, and demand has run multiples over what was offered, in one case more than three times oversubscribed for what were otherwise speculative grade bonds.

The Bank for International Settlements flagged this pattern in July, noting that when project failure risk lands on institutional investors rather than banks, there’s no equivalent resolution mechanism to absorb the shock in an orderly way.

Now look at what NVIDIA just announced. Apollo, BlackRock, Blackstone, Brookfield, Goldman, and KKR are not just asset managers.

They are the exact firms that specialize in taking illiquid, long duration credit exposure and repackaging it for pension plans and insurers who need yield to meet actuarial obligations decades out.

A 500 billion dollar financing platform built with these six firms isn’t NVIDIA discovering a new customer base. It’s NVIDIA institutionalizing the pipeline from GPU demand directly into retirement capital, at a scale and with a repeatability that ad hoc private credit deals never had.

The binding constraint was never capital. It’s who eats the loss.

Every version of this story from Jensen’s team makes the same argument.

NVIDIA compute is fungible, software upgradeable through CUDA, and backed by a deep pool of potential buyers if any single customer fails, so it behaves like durable infrastructure rather than a depreciating asset.

The A100 example gets rolled out constantly: introduced in 2020, still in active commercial use six years later. That’s true, and it’s also not the relevant question. The relevant question is what a GPU is worth relative to the newest generation, not whether it still runs.

One year H100 rental pricing went from about 1.70 dollars per GPU hour in October 2025 to 2.35 dollars in March 2026, while Blackwell capacity already commands 5.30 to 7.05 dollars per hour. Older silicon doesn’t stop working, it just stops earning what it used to, and every financing model built on residual value assumes someone will keep paying a premium for compute that’s no longer the frontier.

NVIDIA’s own announcement quietly acknowledges this. It’s offering a residual value support mechanism covering up to 25 percent of a given financing opportunity. That’s NVIDIA underwriting a slice of the downside on the assets it’s selling into these platforms.

Fair enough as far as it goes, but it means 75 percent of the residual value risk, plus all of the demand and utilization risk, sits with capital that has no equity upside if the trade works and full exposure if it doesn’t.

That’s the actual second and third order economics here.

First order, the ecosystem financed its own demand. Second order, that financing moved from regulated banks to unregulated private credit because banks hit real limits.

Third order, private credit funded itself with pension and insurance money that needs the yield but has no real way to independently verify utilization, demand durability, or GPU obsolescence risk fifteen layers removed from the actual data center.

None of this makes the AI buildout fake.

Enterprises are shipping real products with this compute, and usage is genuinely growing.

But real usage growth and a well underwritten financing structure are two different claims, and the entire architecture of this 500 billion dollar deal exists because the first three years of AI infrastructure financing already ran into limits that regulated capital wouldn’t cross.

Pension funds and insurers are not being brought in because they have superior judgment about GPU depreciation curves. They’re being brought in because they’re one of the last remaining pools large enough to absorb the scale this buildout requires, and because their liabilities are long dated enough that any reckoning is somebody else’s problem for a decade.

If the revenue shows up on schedule, this looks brilliant in hindsight, a natural infrastructure asset class finding its natural capital base.

If it doesn’t, and there are already senators writing formal letters about opaque debt markets and Moody’s publishing notes questioning whether AI data center leases will hold up as advertised, the loss doesn’t land on NVIDIA’s equity holders, who’ve had a decade long run to build a cushion.

It lands on whoever’s retirement plan bought the bond, years after the fact, with no idea the exposure was ever there.

What this means if you’re building, operating, investing, or governing

If you’re a founder building on top of AI infrastructure, the message is that compute pricing has a floor set by financing economics, not just supply and demand, which means the “AI is getting cheaper” trend line is less guaranteed than it looks once older GPU generations need to keep earning their financing costs.

If you’re an operator running AI transformation inside a company, pay attention to which vendors are structurally exposed to this financing chain, because a private credit stress event in the neocloud layer would hit availability and pricing for everyone downstream, not just the company that defaulted.

If you’re an investor, the tell to watch is not NVIDIA’s revenue growth, it’s the spread and coverage ratios on hyperscaler and neocloud bonds. Forbes already flagged softening demand and declining coverage ratios on that paper earlier this summer. That’s the canary, well before anything shows up in NVIDIA’s own numbers.

If you’re in government or policy, the four senators who wrote that letter in January were asking the right question before it was fashionable.

The exposure here isn’t concentrated in a few AI companies anymore. It’s distributed across pension funds and insurers whose beneficiaries never consented to a GPU depreciation bet, and there is currently no equivalent of a bank resolution regime for what happens when that bet goes wrong at scale.

NVIDIA found a new place to put the risk that the first two rounds of financing couldn’t hold anymore. That’s worth understanding clearly, because retirement money doesn’t get to renegotiate the way a hyperscaler balance sheet can.

Share Full Stack Capitalist

The lazy version of the AWS artificial-intelligence thesis is that Amazon arrived late, Microsoft owns OpenAI, Google owns DeepMind, Nvidia owns the chips, and AWS is desperately trying to catch up.

Read more

You may think Nvidia is winning AI and nobody can reach them.

That’s partly true. But I want you to know AWS is doing something weired here.

They are trying to control the ground that every chatbot has to stand on.

That’s the binding constraint here, and once you see it, the whole AWS strategy stops looking scattered and starts looking like one of the cleanest infrastructure plays in tech history.

Here’s the simple version.

Amazon doesn’t need Claude to beat GPT, or GPT to beat Claude, or its own model, Nova, to beat either one. It just needs to be the toll road that all of them drive on.

Power contracts, chips, the software layer that connects a company’s data to a model, and increasingly the models themselves. If you want to build or run serious AI, there’s a good chance you’re paying Amazon somewhere along the way, regardless of who wins the model war.

Think of it in four layers, stacked on top of each other like a toll road with checkpoints at every mile marker.

At the bottom is power and land.

AI runs on electricity before it runs on anything clever, and Amazon has been locking up power deals, nuclear capacity, and data center real estate across the US, Europe, and beyond, faster than almost anyone else. Whoever controls the electrons controls the ceiling on how much AI can actually get built.

Above that sits chips.

This is where the strategy gets interesting. Amazon still rents out Nvidia’s best hardware because customers need it and won’t switch easily.

But it’s also building its own chips, called Trainium, specifically to claw back the margin Nvidia has been keeping for itself. Anthropic is the proof of concept here, using more than a million Trainium chips and committing over a hundred billion dollars to Amazon’s infrastructure.

Amazon financing a partner who then hands the money right back to Amazon.

The third layer is called Bedrock, and it’s basically the checkout counter for AI.

Instead of picking one model and betting the house on it, Amazon lets you rent access to more than a hundred different models, including its competitors’ models, all running through Amazon’s billing, security, and data pipes.

Once your company’s tools are wired into that plumbing, switching away from Amazon gets a lot harder than switching which model you’re using.

At the top are the models themselves, Claude, OpenAI’s systems, Amazon’s own Nova family.

It’s also, honestly, the layer Amazon cares about least, because it wins either way. If your favorite model wins, it probably runs on Amazon’s chips and gets sold through Amazon’s storefront.

That’s the toll road. Amazon isn’t betting on one horse. It’s charging admission at the racetrack.

Now, this strategy isn’t bulletproof, and that’s exactly where it gets interesting from an investor and operator standpoint.

Amazon is spending an almost unbelievable amount of money to build this out, its free cash flow has actually gone negative because of it, and the bet only pays off if AI demand keeps compounding for years.

There are real cracks worth understanding: how much of Amazon’s growth is organic demand versus money it’s essentially lending to its own partners, why its custom chips face a real adoption ceiling, and why regulators in the UK have already flagged how hard Amazon makes it to leave once you’re locked in.

That’s the part I dig into in the full piece for paid subscribers: the bear case, the pricing mechanics Amazon uses to extract more from bigger customers, the anchor tenant risk with Anthropic and OpenAI, and the specific numbers on where this could break.

If you’re making capital allocation decisions, running a company that depends on cloud infrastructure, or just want to understand where the leverage in AI sits, that’s the piece to read closely.

For everyone else, the headline is this: watch who owns the road underneath AI.

AI is not software, it’s industrial revolution.

Are you still thinking Apple missed AI?

And on the surface, sure, that story writes itself. No frontier model. Siri has been a punchline for two years. OpenAI, Google, and Anthropic get all the headlines while Apple shows up late to its own keynote.

But I think that story is looking at the wrong scoreboard.

Here’s the simpler way to think about it. In any system, the thing that actually determines who wins is the binding constraint, the one resource everyone else needs and can’t easily get around.

Right now the industry treats “best model” as that constraint. Whoever has the smartest AI wins. That’s the assumption baked into almost every AI headline you read.

Apple’s bet is that this assumption is wrong. Apple thinks the binding constraint isn’t going to be intelligence. It’s going to be access. Who gets to sit between you and the AI. Who owns your phone, your identity, your payments, your camera roll, your location, and the little permission prompt that decides whether an app gets to touch any of it.

If that’s the real constraint, Apple already owns it. Two billion devices worth of it.

Let’s walk through this.

Apple stopped trying to build the one best model

At its developer conference this June, Apple rebuilt its whole AI framework around a simple idea: don’t force one model to do everything. Apple’s Foundation Models system now lets any outside model plug in, including “Anthropic’s Claude and Google’s Gemini, through a shared protocol any model provider can build against”. A developer can start with Apple’s own on-device model and swap in a different provider with barely any code changes.

That is a genuinely strange thing for Apple to do. This is a company that has spent forty years trying to control every layer of its stack, chips, software, glass, aluminum, all of it. And here it is, essentially saying, we don’t care whose model answers the question, as long as it answers it inside our house.

That’s the tell. Apple isn’t trying to win the model race anymore. It’s trying to make the model itself replaceable.

The Google deal makes the point even harder to miss

In January this year, Apple did something that would have been unthinkable a few years back. It signed a multi year deal to run the new Siri on “a custom 1.2 trillion parameter Gemini model, roughly eight times larger than what Apple’s own cloud models were running”, reportedly paying “something in the neighborhood of a billion dollars a year for the access”.

Read that again. Apple is paying its biggest search rival to power its own voice assistant. That is not the move of a company trying to out build Google on model quality. That’s a company saying, fine, license the smartest available brain, and let us keep the thing that actually matters, which is the wrapper the brain sits inside.

And here’s the part that really shows the strategy. The Gemini model doesn’t even show up as Gemini. Apple runs it through its own private cloud, under its own branding, invisible to the user. You’ll just think it’s Siri. The supplier gets swapped out like a part in an engine, and the driver never notices.

Meanwhile the old partner is getting quietly demoted

Back in 2024 Apple leaned on OpenAI to give Siri a boost for complicated questions.

That relationship is clearly cooling. Apple’s new Siri, launching this fall, runs on Google’s models instead, and one analyst summed up the shift bluntly, saying “the Google deal pushes OpenAI into a supporting role, with ChatGPT sticking around for complex, opt in requests rather than being the default brain behind the phone”.

Apple hasn’t cut OpenAI off completely. But it’s telling that the company that used to be the star player is now riding the bench.

And then it gets messy. Really messy.

OpenAI knows exactly what Apple is doing, which is why it’s trying to become Apple

Here’s the other half of the story, and it’s the half that turns this into a genuinely great essay instead of just a tech news recap.

OpenAI’s president, Greg Brockman, confirmed that the company is building what he called a family of devices, physical hardware designed to be the way people talk to ChatGPT, not just an app you open on someone else’s phone. He said the products are coming soon, without giving a date, and he’s been clear that “the goal is a much more natural way to talk to a machine, with voice as the focus rather than a screen”.

This didn’t come out of nowhere. Last year OpenAI bought Jony Ive’s design studio, io, in a deal worth somewhere around $6.5 billion dollars, paid entirely in OpenAI equity.

Ive is the guy who designed the iPhone. Bringing him in wasn’t a talent hire. It was OpenAI trying to buy the one thing it doesn’t have, which is a seat at the hardware table.

Why would a software company that just wants to sell you a chat subscription spend billions building a physical gadget? Because Brockman and Altman clearly understand the same thing Apple understands. If ChatGPT only ever lives inside Apple’s phone, as a feature Apple can promote or bury whenever it wants, then Apple owns the relationship with the customer, not OpenAI. The only way out of that trap is to stop being a guest in someone else’s house and build your own house instead.

So you’ve got this beautiful mirror image happening. Apple is trying to turn frontier models into an interchangeable commodity that plugs into its ecosystem. OpenAI is trying to build its own ecosystem so it never has to be anyone’s commodity again.

Apple wants OpenAI to become a supplier. OpenAI is racing to become the next Apple before that happens.

And now there’s a lawsuit that makes the whole thing feel personal

In July, Apple actually sued OpenAI. Not over anything to do with AI models. Over hardware. Apple’s complaint accuses OpenAI of running a systematic campaign to recruit more than four hundred former Apple employees, and of walking off with confidential product designs and manufacturing details in the process.

The filing goes after OpenAI’s chief hardware officer by name, and it alleges that some of the stolen material touched on Apple’s proprietary metal finishing techniques, the kind of manufacturing know-how Apple guards as closely as anything it makes.

Interestingly, the lawsuit is careful to leave Jony Ive’s name out of it entirely, even though his studio, io, is named as a defendant. Nobody quite knows why. Maybe there isn’t enough evidence tying him directly to the recruiting, maybe Apple doesn’t want the optics of suing its own former design chief by name. Either way, it’s a strange, almost respectful omission in an otherwise aggressive filing.

Apple didn’t have to file this lawsuit. Filing it while OpenAI is heading toward what’s expected to be a massive IPO is a very deliberate kind of pressure.

Apple defending the one thing it has that OpenAI desperately wants and doesn’t have yet, which is decades of knowing how to design and manufacture a physical object people are willing to put in their pocket.

So what’s the actual fight here

It’s tempting to frame this as Apple versus OpenAI on model quality. That’s not it.

Apple has basically conceded that fight already, which is exactly why it licensed Google’s model instead of grinding away trying to out build it internally.

The real fight is over who owns the interface. Who decides which AI gets to see your messages, your photos, your location, your calendar, your face when you unlock your phone. Whoever holds that position doesn’t need to have the smartest model in the room. They just need to control the door the smartest model has to walk through to reach you.

Apple’s bet is that it already owns that door, and that the smartest move now is to keep the door and stop worrying about which brain sits on the other side of it. OpenAI’s bet is that owning a brain isn’t enough, and that it needs to build its own door before Apple’s door becomes the only one that matters.

Neither company can fully win this without controlling the layer the other one owns. That tension, more than any benchmark score, is the actual AI story happening right now.

If you’re building anything in this space, the lesson generalizes. Don’t ask whether your product is the smartest one out there. Ask who owns the interface your product has to pass through to reach a real customer. That’s usually where the leverage actually sits, and it’s usually not where anyone’s paying attention.

Here’s the question people keep asking: whose model is going to win?

Wrong question.

It’s the question a normal tech reporter asks, and it’s the question that keeps you staring at leaderboards instead of at the thing that actually matters.

The model layer is not where this gets decided. Models converge. Every few months the gap between the top labs on any given benchmark shrinks to a rounding error, and the moment one lab ships something clever, the other three copy it within a quarter.

That’s a commodity racing to its natural price, which is close to zero.

The AI race is being decided one layer down, in the physical and financial infrastructure that has to exist before any model can be trained or served to a single user.

And that infrastructure runs through five chokepoints: power, chips, data centers, capital, and distribution.

Whoever sits at each chokepoint gets to charge on everyone downstream. That’s the whole game. Not who’s smartest. Who’s scarce.

Let me walk you through each one the way I’d explain it over a beer (or coffee!), because once you see the pattern you can’t unsee it, and it changes how you read every AI headline from here on out.

Chokepoint one: power

Two years ago the constraint on AI was GPUs. You couldn’t get them.

Now you can, or close to it, and the constraint has moved to something much more boring and much harder to fix: electricity.

A single AI data center site is now asking regional grids for 100 to 750 megawatts of power, and those grids were never built for loads like that. Gartner is projecting that 40 percent of AI data centers will be power-constrained within the next year. Global data center electricity use is on pace to roughly double by 2030, and some projections have data centers alone consuming as much power in 2026 as the entire country of Japan.

This is why Jensen Huang stood on stage at GTC 2026 and reframed Nvidia’s whole business with a formula: revenue equals tokens per watt times available gigawatts.

Read that twice. He’s telling you, in public, that the ceiling on AI revenue isn’t chip supply anymore.

It’s megawatts.

That’s a stunning admission from the guy who runs the chip company. And it’s why you’re now watching Microsoft commit $15 billion to a UAE buildout, Meta drop $10 billion on a Louisiana campus, and every hyperscaler sign direct power purchase agreements instead of waiting in line for the grid like everybody else.

Whoever can conjure electricity out of the ground faster than their competitor gets to build faster, train faster, and serve more customers. Everyone else waits in an interconnection queue that can run three years long.

Chokepoint two: chips

You already knew this one, but it’s shifted shape. Nvidia still owns somewhere between 75 and 90 percent of the AI accelerator market depending on whose number you trust, and it’s still the default answer to “who wins AI hardware.”

But the real chokepoint underneath Nvidia isn’t Nvidia.

It’s TSMC, because TSMC is the only company on earth that can manufacture the advanced packaging, called CoWoS, that every serious AI chip needs.

Nvidia alone is expected to consume something like 595,000 CoWoS wafers in 2026, more than the entire industry used in 2024.

That’s a company that has locked up the scarce input the entire market depends on.

So the hierarchy looks like this: TSMC sits underneath everyone, Nvidia sits on top of TSMC’s packaging capacity, and everyone else, AMD, the custom silicon teams at Google and Amazon and Meta, is fighting over what’s left.

This is also where geopolitics stops being background noise. TSMC is in Taiwan.

Export controls already cut a $4.6 billion quarterly China revenue line out of Nvidia almost overnight. Whoever controls Taiwan’s fabs, or builds a credible alternative to them, controls the entire AI hardware stack.

Chokepoint three: data centers

This is where a new class of company has muscled into a game the hyperscalers used to have to themselves.

They’re called neoclouds, CoreWeave, Nebius, Lambda, Crusoe, and their entire pitch is: we can get you GPU capacity faster than Amazon, Microsoft, or Google can build it themselves.

It’s worked. Microsoft has struck roughly $60 billion in commitments with CoreWeave, Nebius, and Nscale combined. Meta signed $35 billion with CoreWeave and up to $27 billion with Nebius.

The funny part is the hyperscalers are simultaneously the neoclouds’ biggest customers and their biggest long-term threat, because those contracts let hyperscalers book AI spend as an operating expense instead of piling more capex onto their own balance sheets.

But watch what just happened. When Meta signaled it might resell its own excess compute through something called Meta Compute, Nebius and CoreWeave stock dropped 15 percent in a single morning.

That’s the whole fragility of this layer in one data point. Neoclouds don’t own a chokepoint of their own. They’re renting scarce power and scarce chips and repackaging them faster than the giants can. That’s a real business, but it’s a business built on somebody else’s scarcity, which means it’s the most exposed layer of the five.

Chokepoint four: capital

Here’s the number that should stop you: the five biggest hyperscalers are on pace to spend over $600 billion on infrastructure in 2026 alone, roughly three quarters of it aimed straight at AI.

Multiply that out and Goldman Sachs is now talking about $5 trillion in hyperscaler AI and data center spending by 2030.

No company generates that kind of cash from operations. So they’re borrowing it.

Hyperscalers issued a record $428 billion in bonds in 2025, and estimates for total AI-related debt issuance over the next few years run as high as $1.5 trillion.

A lot of that debt isn’t even sitting on the hyperscaler’s own balance sheet. It’s structured through special purpose vehicles and leases, serviced by private credit funds and insurers, in arrangements the Bank for International Settlements is now openly calling “shadow borrowing.”

Translation: the AI buildout is being financed less like a tech company expanding and more like a utility building a power plant, except the debt is scattered across a web of private credit vehicles instead of sitting in one regulated place where anyone can see it clearly.

Whoever controls that capital, the infrastructure funds, the private credit shops, the sovereign wealth funds writing nine figure checks, has leverage over the entire buildout regardless of who’s got the best model.

Money is the chokepoint that makes the other four chokepoints possible.

Chokepoint five: distribution

This is the one people get most wrong, because they assume the best model wins the user.

It doesn’t.

Distribution wins the user. ChatGPT still leads on raw traffic, but its share of AI chatbot web visits has fallen from around 87 percent to roughly 53 percent in about eighteen months, not because the product got worse, but because Google started placing Gemini directly inside Search, Android, Chrome, and Workspace, tools billions of people already had open.

Gemini’s referral traffic grew 388 percent year over year. Google didn’t win users by being smarter. It won them by being everywhere already.

That’s the whole distribution chokepoint in one example.

Whoever owns the pipe the user is already standing in, the operating system, the search bar, the messaging app, gets AI usage for free. Everyone else has to buy it, one download at a time. It’s the same reason Grok is climbing off the back of X Premium bundling instead of model quality.

So who actually holds the leverage right now

Walking chokepoint by chokepoint, here’s how I’d rank who’s sitting pretty, who’s exposed, and who could flip the board.

On power, the utilities and grid operators in power-rich regions, think Texas, parts of the Gulf South, the UAE, hold real pricing power for the first time in decades, because they’re now the thing everyone is begging for access to.

Their vulnerability is regulatory and political, since ratepayer backlash over AI-driven price hikes is already showing up in state utility fights. The challengers here are the hyperscalers building their own dedicated generation, nuclear partnerships and on-site gas turbines, specifically to route around the utilities altogether.

On chips, TSMC has the strongest structural position of any company in this entire essay, full stop, because it is the sole supplier of the packaging every serious AI chip needs.

Its vulnerability is almost entirely geopolitical, sitting ninety miles from a country that claims it.

Nvidia sits just below TSMC, extraordinary pricing power, 80-plus percent margins on its top chips, but its vulnerability is customer concentration risk running the other direction, since Amazon, Google, Meta, and Microsoft are all racing to build custom silicon specifically to reduce their Nvidia dependence. AMD and Broadcom are the clearest challengers, AMD by building a real merchant alternative, Broadcom by co-designing the custom ASICs the hyperscalers are building to escape Nvidia in the first place.

On data centers, the hyperscalers, Amazon, Microsoft, Google, Meta, own the strongest position because they control both the balance sheet and the customer relationship. Oracle has carved out a genuinely strong niche as the value option for bare-metal AI clusters. The neoclouds, CoreWeave and Nebius foremost, have real revenue and real backlog, but the Meta Compute scare showed exactly how thin their moat is the moment a hyperscaler customer decides to become a hyperscaler competitor.

On capital, the infrastructure funds and private credit shops, the ones that have quietly grown into a $1.7 trillion asset class, hold more leverage than most people realize, because they’re now writing the checks that let hyperscalers keep spending without blowing up their own credit ratings.

Their vulnerability is that they’re underwriting demand assumptions nobody has stress tested against a slowdown. Sovereign capital, Gulf state funds especially, are the emerging challengers, trading capital for guaranteed access to compute their own economies don’t yet produce.

On distribution, Google holds the strongest hand of anyone in this whole list, not because Gemini is necessarily the best model, but because it’s about to be the default AI experience on both Android and, through the Apple deal, iPhone, covering north of 90 percent of the global smartphone market.

OpenAI’s vulnerability is the mirror image of Google’s strength, extraordinary product but no owned distribution layer underneath it, which is exactly why it keeps cutting platform deals instead of just growing organically. Anthropic’s position is the interesting anomaly, weak on consumer distribution but dominant on enterprise API spend, which tells you distribution isn’t one chokepoint, it’s actually two, consumer and enterprise, and right now different companies own each half.

The binding constraint, stated plainly

None of these five chokepoints is fixed. They rotate. Two years ago it was chips. Right now it’s power. Next year it might be capital, if debt markets start pricing AI infrastructure risk the way they’re starting to price everything else with concentrated counterparty exposure.

The skill that matters, whether you’re a founder, an operator, or an investor, isn’t picking the best model. It’s identifying which chokepoint is binding right now and positioning yourself as close to it as you can get.

For founders, that means asking a different question than “what can I build with AI.” Ask instead: which of these five layers am I renting from, and what happens to my margins if that layer gets more expensive next year.

For operators, it means treating your AI vendor contracts as infrastructure exposure, not software procurement, because you’re now several layers deep in someone else’s power and capital risk whether you priced it in or not.

For investors, it means the boring picks-and-shovels companies, the ones nobody’s writing hype pieces about, the transformer manufacturers, the grid interconnection specialists, the advanced packaging suppliers, are frequently the ones with the actual pricing power, because they get paid regardless of which model or which chatbot wins the popularity contest upstream.

And for anyone thinking about this at the policy level, the lesson is that national AI competitiveness is now an energy and manufacturing policy question wearing a technology costume. The countries that win aren’t the ones with the best AI labs. They’re the ones that can permit a gigawatt of new power generation in eighteen months instead of five years.

Full Stack Capitalist is not another newsletter about which AI model won this week.

I write about the infrastructure beneath AI: chips, energy, data centres, capital and geopolitical power.

Free subscribers receive the major essays.

Paid subscribers receive the deeper research: company rankings, infrastructure maps, investment implications and the systems determining who wins the AI economy.

The next paid report will rank the companies controlling the five critical chokepoints of AI and examine where their power is strongest, and where it could break.

Become a paid subscriber to receive the full report.

Share Full Stack Capitalist

Let me tell you what actually happened this week, because the headline undersells it.

Simply Wall St put out one of those “here are three stocks” screener pieces on July 26, the kind that usually blends into the noise.

Bloom Energy, Siemens Energy, Vertiv.

Power grid technology picks for the AI buildout. Normally I’d scroll past it. But sitting underneath that boring little roundup is basically the whole story of where AI infrastructure money is actually going right now, and almost nobody is explaining it in a way that makes the incentives clear.

So let’s do that.

Here’s the thing you need to hold in your head as we go: everyone talks about the AI power crunch like it’s a single problem. It’s not. It’s three stacked problems, and each one is being solved by a different set of companies who are all fighting over who gets to own the toll booth.

Figuring out which layer is genuinely scarce, and who’s positioned to extract rent from that scarcity, is the whole game.

Why Bloom Energy is up 1,100% and also down 39% from its peak, at the same time

Bloom Energy makes solid oxide fuel cells. You drop them next to a data center, feed them natural gas, and they make electricity on site without needing to touch the grid at all. Just power, fast.

That speed is the entire product. Bloom recently delivered a hyperscale AI factory order in 55 days against a 90-day commitment, and that kind of turnaround is why Oracle signed on for up to 2.8 gigawatts of Bloom’s systems and why Brookfield Asset Management quintupled its AI infrastructure financing deal with Bloom to $25 billion.

The stock has done something like 194% year to date at various points this year, north of 1,000% over twelve months.

It’s also round-tripped hard, down roughly 39% from its June high after a short seller started poking at customer concentration and growth assumptions, and current net margins are sitting around 0.2%, funded heavily by external borrowing.

Wall Street’s consensus is a “Hold” with a price target below where the stock trades. So you’ve got a company solving a real, urgent problem, priced like it can never miss a step, in an industry where missing a step is basically the base case.

That tension, demand plus fragile execution plus a valuation with zero room for error, is the story of almost every AI infrastructure stock right now. Bloom’s just the cleanest example because its whole pitch is speed, and speed is exactly what’s scarce.

The bottleneck isn’t gas. It’s the turbine slot.

Here’s where it gets interesting: which stakeholders are positioned to gain power as we lean harder on these energy solutions?

Not the gas producers. Not even the utilities, really.

It’s the handful of companies that physically manufacture the equipment that turns fuel into grid-scale electricity.

There are exactly three companies on earth that make heavy-duty gas turbines at scale: GE Vernova, Siemens Energy, and Mitsubishi Power.

That’s it.

That’s the whole club.

And that club has a combined backlog stretching toward 100 gigawatts, against manufacturing capacity of roughly 10 gigawatts a year from the largest of the three.

Do the math. Lead times for the efficient combined-cycle turbines have stretched to five, sometimes seven or eight years. GE Vernova is already taking reservation conversations for 2030 delivery slots.

This is the binding constraint, and it’s a genuinely different kind of scarcity than the ones we’re used to talking about in AI.

It’s not chips. It’s not capital, there’s plenty of capital chasing this. It’s forged steel, specialized manufacturing lines, and the handful of engineers who know how to build these machines. You cannot solve that with a bigger check. You solve it by waiting in line, or by finding a workaround.

And the workaround is exactly what’s driving the second-order effect. Because AI load needs power in 2026 through 2028, not 2031, developers who can’t get an efficient combined-cycle turbine are defaulting to simple-cycle peaker turbines instead, which can be built in eighteen to twenty-four months but burn fifty to sixty percent more gas per megawatt generated.

So the turbine shortage isn’t just delaying buildout, it’s making the buildout that does happen dirtier and more gas-intensive than it would otherwise be.

That’s a structural, multi-year lift to natural gas demand that has nothing to do with gas supply and everything to do with who can pour steel fast enough.

This is also exactly why Bloom Energy’s fuel cell pitch works: it sidesteps the turbine queue entirely by using a completely different generation technology that Bloom itself controls the manufacturing of.

Who actually pays for all this, and who decided that

Now, the part that turns this from an interesting supply chain story into an actual economics and incentives story: somebody has to pay for the transmission lines, substations, and grid upgrades that all this new load requires.

And for a while, the honest answer was “everyone on the grid, whether they use AI or not.”

PJM Interconnection, the grid operator covering thirteen mid-Atlantic states, saw its capacity prices jump 833% between the 2024-25 and 2025-26 delivery years.

That’s not a typo, and it’s not abstract, it shows up on residential electric bills.

Regulators noticed, ratepayers noticed, and it became politically toxic fast enough that in March 2026, seven of the biggest AI companies, Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI, signed a White House-facilitated Ratepayer Protection Pledge committing to directly fund the grid infrastructure improvements their own load requires, rather than socializing the cost across everyone’s bill.

FERC separately ordered PJM in December 2025 to build new transmission service categories specifically for data centers, effectively treating them as a distinct class of grid citizen with their own rules.

That’s your answer on government’s role here: not funding the buildout, but drawing the line on who’s forced to pay for it. And it’s also where the fights over equitable access are actually happening, not in some abstract fairness debate, but in the specific mechanics of interconnection cost allocation.

Worth watching whether the Ratepayer Protection Pledge holds up once the actual invoices start arriving, because voluntary pledges and binding cost obligations are very different things when the number gets big enough.

There’s a second conflict brewing between AI companies and grid operators over the interconnection queue itself.

American Electric Power’s raw interconnection queue includes 190 gigawatts of requested new demand, but its actual firm commitments sit around 24 gigawatts.

That gap exists because developers routinely file speculative interconnection requests at multiple sites, sometimes five to ten times more capacity than they’ll ever build, just to hold a place in line while they decide where they’re actually going to build. Utilities are now having to build planning models around commercial letters of agreement rather than raw queue numbers, because the queue itself has become a strategic decoy. That’s a resource allocation fight hiding in plain sight, and it’s a big part of why grid planning has gotten so much harder even as the dollars committed to it have gone up.

Where the experts actually disagree, and why both sides have a point

This is where I think most coverage of the AI power story gets lazy. It treats “the grid is the bottleneck” as a settled fact and moves on.

It’s not settled. Here’s where the disagreement lives, and honestly, I think both sides are onto something true.

On the pace of the buildout. Goldman Sachs projects data center power demand will rise 165% by 2030 and estimates roughly $720 billion in grid investment is needed to keep up, and to be fair, capital is showing up at basically that scale. But that projection assumes permitting and construction happen on schedule, and regional transmission line permitting alone typically takes seven to eleven years.

Seven of thirteen major U.S. grid regions are projected to fall below their critical safety margins by 2030. So you can be right that the money is there and still be wrong that the power arrives on time. I lean toward the skeptics on timing, not because the capital commitment isn’t real, but because physical construction timelines don’t bend to press releases.

On whether the current regulatory patchwork can actually handle this. FERC’s PJM directive is the first real attempt at federal-level rules built specifically for data center load, and depending who you ask that’s either the beginning of a coherent national framework or a one-off fix for the one region screaming loudest.

FERC’s own chair has pointed out that each of PJM’s thirteen member states has fundamentally different regulatory structures and politics, which means a rule built for PJM doesn’t automatically translate to Texas’s ERCOT or the Southwest Power Pool, both of which are inventing their own large-load interconnection rules independently right now.

Texas alone is facing a 438 gigawatt queue. We’re not building one national playbook. We’re watching a dozen regional experiments run in parallel, and some of them are going to produce genuinely different rules for who gets power first.

On whether the Ratepayer Protection Pledge is actually good policy or just privatized infrastructure. The optimistic read is that it keeps grandma’s electric bill from subsidizing a hyperscaler’s training run, which is fair.

The skeptical read is that when seven trillion-dollar companies are directly financing the specific grid infrastructure that serves them, you’ve built a two-tier grid, premium infrastructure for AI load, legacy infrastructure for everyone else, without anyone voting on that outcome.

I don’t think this is resolved yet, and I think it’s the single most underappreciated governance question in the whole AI energy conversation.

On whether we’re building a dangerous concentration of power in three turbine companies. GE Vernova, Siemens Energy, and Mitsubishi Power effectively control the entire market for the equipment that generates most new grid-scale power.

That’s the textbook definition of a chokepoint, and their pricing power shows it, quotes are running ten to twenty points above where they were a year ago. The counter-argument is that this concentration is exactly what’s pushing capital and innovation toward alternatives that bypass turbines entirely: fuel cells like Bloom’s, battery storage, and eventually small modular nuclear.

A genuine monopoly doesn’t usually spawn three or four well-funded categories of competitors trying to route around it. I think the turbine oligopoly is real and dangerous in the near term, and also the best argument for why Bloom Energy’s technology bet has a real structural tailwind independent of any single contract.

On what this does to labor markets. The bull case is straightforward, hundreds of billions in construction and equipment orders means real jobs, real fast, in engineering, manufacturing, and skilled trades.

The problem is execution capacity, not appetite. Roughly a third of data center operators report losing staff to competitors amid a broader specialized-technician shortage.

So even where the capital and the equipment eventually show up, there’s a real question of whether there are enough qualified people to build and run it. This is the disagreement I think gets least attention and probably matters most for how fast any of this actually gets built.

What this means depending on where you sit

If you’re a founder building something that touches AI infrastructure, the lesson isn’t “invest in energy.” It’s that speed to power has become a genuine competitive moat, the same way speed to compute was in 2023. If your business model depends on deploying compute somewhere, understand your power sourcing strategy as seriously as your GPU strategy, because the multi-year queue is not hypothetical anymore.

If you’re an operator inside a company running AI workloads at any real scale, get ahead of your interconnection timeline now, not when you need the power. The gap between what utilities can promise and what they can deliver on schedule is widening, and behind-the-meter generation, fuel cells, on-site gas, batteries, is shifting from a backup plan to a primary strategy for anyone who can’t tolerate a multi-year wait.

If you’re an investor looking at this trio of stocks, or others like them, separate the technology bet from the valuation bet. Bloom Energy solves a real problem with real contracts, but it’s priced for flawless execution in an industry defined by permitting delays and community pushback. Siemens Energy and Vertiv both sit closer to diversified infrastructure plays with less single-customer concentration risk, which is worth something when the story this exciting eventually meets a quarter that disappoints.

If you’re in government or policy, the fight worth watching isn’t renewable versus gas. It’s who bears the cost of the grid upgrade, and whether the current wave of voluntary pledges from AI companies holds up once the real bills start landing, or whether it takes actual binding rules, like FERC’s PJM directive, to make cost allocation stick. The turbine oligopoly is also a national competitiveness issue hiding in plain sight, whichever country solves domestic power generation manufacturing capacity first has a real structural advantage in how fast it can stand up AI infrastructure, independent of who has the best models.

The chips got all the attention for two years. The chokepoint moved. It’s steel, slots, and who gets to decide who’s first in line.

NVIDIA just shipped something called Vera Rubin. Ten times better performance per watt. One tenth the cost per token. 350 factory sites in 30 countries building the thing. CoreWeave ran it against DeepSeek-R1 and got ten times the tokens per megawatt compared to the old Blackwell racks.

On paper, that’s 800,000 tokens a second out of a rack that used to give you 80,000, same power draw.

Cool numbers.

But you know what’s important now.. the story is who owns the thing that was scarce before the chip ever mattered: the power to run it. And once you see that, the whole Vera Rubin launch reads differently.

Who gets to convert electricity into intelligence at the lowest cost, and what happens to everyone else.

Why “Performance Per Watt” Is the New Scoreboard

Here’s the thing that changed in the last two years.

For a long time, the bottleneck in AI was just getting your hands on chips. You wanted H100s, there was a line, NVIDIA controlled the line. That’s over.

TSMC can ramp chip production in twelve to eighteen months when demand shows up. What TSMC and NVIDIA cannot do is build you a new electrical substation or get you a grid connection, because that’s not a manufacturing problem, it’s a permitting and physics problem, and it runs on a completely different clock.

Interconnection queues in the biggest US markets, Northern Virginia, Phoenix, Dallas, are running four to seven years right now. There’s over two thousand gigawatts of generation and storage sitting in queues waiting to connect to grids that were never built for this.

So when NVIDIA says Vera Rubin gives you ten times the tokens per megawatt, they’re not selling you a faster chip. They’re selling you a way to get more intelligence out of the same electrical connection you already fought years to secure.

That’s the whole game now.

If you’re stuck with, say, 150 megawatts of power and no realistic path to more before 2030, the only lever left is squeezing more tokens out of every watt you’ve got.

Revenue basically becomes tokens-per-watt multiplied by however many gigawatts you actually control. Not GPUs you own. Watts you can plug into.

That reframes the question: it’s not “how much faster is this chip,” it’s “how does this change who wins in a world where the grid, not the wafer, is the ceiling.”

A cloud provider sitting on a fixed power allocation just found a way to serve ten times more customers or ten times more tokens without asking the utility for another megawatt.

That’s a direct, structural advantage over anyone still running Blackwell-era hardware on the same power budget. And because government-run energy governance, permitting boards, utility commissions, is what decides who gets new watts at all, the companies that already have power contracts locked in are about to become dramatically more valuable, because efficiency gains compound on top of an allocation nobody else can get.

That’s the answer to how energy governance evolves here: it doesn’t need new laws to matter more, it already controls who gets to play, and Vera Rubin just raised the stakes on every megawatt already spoken for.

The Empire Gets Bigger, Not More Crowded

Now follow the money on who actually benefits.

CoreWeave, Google Cloud, Microsoft Azure, and Oracle are the first four names getting Vera Rubin racks.

Not fifty companies. Four.

And NVIDIA’s own numbers say a rack that used to deliver one tenth the token throughput now costs one tenth as much per million tokens to run. If you’re one of those four, your cost structure for serving AI inference just fell off a cliff, and your competitors haven’t caught up yet, because Rubin allocation itself is scarce and follows existing NVIDIA relationships.

Smaller neo-clouds and marketplace providers are looking at 2027 before they see meaningful Rubin capacity.

That’s the mechanism by which this technology could concentrate market share and influence rather than spread it around.

The efficient frontier doesn’t get more crowded, it gets owned by the four or five players who already had the capital, the power contracts, and the NVIDIA relationship to get first access.

A smaller AI startup trying to compete on inference costs isn’t just fighting for market share, it’s fighting a cost curve that just moved ten times in the other direction for its biggest competitors, and it can’t buy its way onto that curve for another year or two even if it has the cash, because the hardware simply isn’t allocated to it yet.

This is where the smaller players actually have a move, though, and it’s not “wait for Rubin.” It’s picking your battles on the workload. The efficiency gains are concentrated in large mixture-of-experts models doing high-concurrency inference, that’s where the ten-times number comes from.

If you’re running smaller models, under maybe seventy billion parameters, on existing Blackwell or even Hopper hardware, you’re not compute-bound in the first place, so the Rubin gap barely touches you. The leverage for a startup is choosing not to compete where the giants just built a moat, and instead building on model sizes and use cases where raw rack efficiency isn’t the deciding factor yet. That’s a genuinely different competitive game than trying to out-infrastructure Google Cloud.

Five Fights Nobody’s Settled Yet

This is the part of the story that gets skipped in every breathless press write-up, because the people writing them have a stake in the “everything is fine” version. It isn’t fine, it’s contested, and here’s where the actual disagreement lives.

Fight one: does cheaper, more efficient infrastructure create monopolies or competition?

One camp says this consolidates power hard: whoever gets first access to Rubin racks locks in a cost advantage that compounds, and the four names getting first shipments become nearly impossible to displace on price.

The other camp points out that cheaper tokens usually mean more applications get built, because things that were too expensive to run suddenly pencil out, and historically cheaper infrastructure has expanded who can compete, not shrunk it, the same way cheaper cloud compute in the 2010s created a wave of new companies rather than just entrenching AWS.

Both things can be true at once, actually, concentration at the infrastructure layer and explosion at the application layer, which is exactly what happened with cloud computing.

Fight two: can the grid actually handle this, or is gigascale AI running into a wall?

Some people look at the water-saving cooling design and the ten-times efficiency and conclude the industry is solving its own energy problem just fast enough.

Others point to forty percent of AI data centers projected to be power-constrained by 2027 and multi-year interconnection queues in every major market and say efficiency gains at the chip level are getting eaten alive by the sheer scale of buildout, so the wall is coming regardless of how good the racks get.

This one isn’t really an opinion disagreement, it’s a math disagreement, whether efficiency gains grow faster than demand, and right now demand looks like it’s winning.

Fight three: does the “cheapest token wins” model eventually lock out smaller players?

One side says a race to the cheapest token per unit of compute becomes a scale game, and if you can’t afford the rack-scale hardware, you simply can’t compete on unit economics no matter how good your model is.

The other side argues token costs eventually become commoditized and cheap for everyone, the way storage and bandwidth did, and today’s frontier advantage is tomorrow’s baseline, so the barrier is temporary, not permanent.

History leans toward the second view eventually happening, but “eventually” has been known to take years that kill a startup’s runway first.

Fight four: is localized, distributed manufacturing across 350 sites a strength or a new fragility?

The optimistic read is that spreading production across thirty countries makes the supply chain more resilient to any single country’s disruption, unlike the old model where a shortage in one region stalled everything.

The skeptical read is that this many interdependent sites and seven co-designed chips built as a single system creates more points of failure, not fewer, because now you need every piece of a complex, tightly coupled chain to work in sync, and a disruption anywhere in thirty countries has more surface area to hit than a disruption in three.

Whether this reshapes trade dependencies in a good or bad direction really depends on whether any single node in that network becomes a chokepoint, which nobody knows yet because the network is brand new.

Fight five: does AI efficiency like this kill jobs or create new ones?

The standard worry is that if a rack can now do ten times the inference work per watt, you need proportionally fewer people managing that infrastructure, and automation absorbs roles that used to require humans watching dashboards and provisioning capacity.

The counter-argument is that cheaper, more available inference means way more AI gets deployed everywhere, and someone has to build, monitor, secure, and govern all of that new deployment, which is a bigger job market than the narrower one it replaces.

Both sides are usually right about different timeframes, contraction in the specific old roles happens fast, expansion in new AI-operations roles happens slower and unevenly, and the people who lose the first job aren’t always the ones who get the second one.

The Geography of Who Controls This

The 350-site, 30-country manufacturing footprint deserves its own beat.

Spreading fabrication and assembly across that many countries changes which governments have leverage over the AI buildout, and it changes what “supply chain risk” even means.

A single-country chokepoint used to be the nightmare scenario, think Taiwan and advanced chip fabrication. A thirty-country web is harder to disrupt on purpose, but it’s also harder to fully audit or secure, and it means more national governments now have a stake, and potential leverage, in how this technology gets built and deployed.

Watch which countries start attaching conditions, export requirements, local employment mandates, data sovereignty rules, to hosting a piece of that manufacturing chain. That’s where the second-order geopolitics of Vera Rubin will show up, not in NVIDIA’s press release, but in trade ministries over the next eighteen months.

There’s also a quieter effect worth naming: what does a ten-times jump in AI factory efficiency do to industries that aren’t AI at all.

Cheaper, denser compute per megawatt makes AI-driven automation viable in sectors that couldn’t previously afford the compute bill, logistics optimization, industrial quality control, scientific simulation.

That’s a slow-burn effect on traditional manufacturing sectors, not because robots show up on the floor tomorrow, but because the compute cost of running sophisticated optimization and simulation just dropped enough that projects which didn’t pencil out a year ago suddenly do.

What This Actually Means for You

If you’re a founder: don’t try to out-infrastructure the four companies who just got first access to this. Build on the model sizes and workloads where the efficiency gap doesn’t apply yet, and treat every year of “we’re not compute-bound” as a genuine head start, not a consolation prize.

If you’re an operator: the tokens-per-watt framework is now the real unit economics of your AI stack, not tokens-per-dollar. If your power allocation is fixed, that’s your actual ceiling on growth, and squeezing efficiency out of your existing infrastructure matters more than it did eighteen months ago.

If you’re an investor: the interesting bet isn’t NVIDIA, that’s priced in. It’s whoever controls power contracts and grid interconnection rights in constrained markets, because that scarcity doesn’t get solved by a better chip, and it’s the actual gate everyone else has to get through.

If you’re in government: the leverage you have isn’t the chip, it’s the grid connection and the manufacturing footprint sitting inside your borders. Whoever writes the rules for interconnection queues, energy allocation to data centers, and manufacturing conditions is deciding, right now, who gets to build the next generation of this stuff and who waits in line. That’s a bigger strategic lever than any AI policy paper you’re about to commission.

The chip got ten times better. The bottleneck just moved to the thing nobody can manufacture their way out of.

Share Full Stack Capitalist

Let me tell you what actually happened in Indianapolis this week.

On July 15, the Indiana Utility Regulatory Commission stood up and said, in effect, we’re going to open the hood on how much profit your electric company is allowed to make, and we’re going to look hard at these financial mechanisms called trackers that let utilities collect money from you outside the normal rate case process.

Two new investigations. One on return on equity, the profit rate regulators guarantee investor-owned utilities on their capital spending. One on trackers, the surcharge lines on your bill that let a utility like NIPSCO or AES Indiana recover certain costs almost automatically, without going through the full public scrutiny of a rate case.

If you only read that as a consumer protection story, you’re missing the real one. This is a story about who gets to build the physical infrastructure that AI runs on, and at what price, and it’s happening in a state that just landed a 3.8 billion dollar SK Hynix memory fab that will be feeding chips straight into Nvidia’s supply chain. Indiana is not a bystander in the AI buildout. It’s one of the places where the buildout is actually happening. Which is exactly why the state is now the place where the fight over who pays for it is happening too.

Here’s the mechanism, explained the way I’d explain it to you over a beer.

A regulated utility doesn’t make money like a normal business. It doesn’t sell electricity for a markup and pocket the difference. Instead, regulators let it earn a guaranteed rate of return on whatever capital it spends building poles, wires, substations, and plants.

That guaranteed rate is the ROE. AEP’s Indiana Michigan Power was pulling a 12.6 percent return on equity over the twelve months ending in March, the highest of any AEP subsidiary anywhere. Think about that for a second. In a year when Hoosier households are furious about their bills, one utility subsidiary is running a return that would make a private equity partner jealous, and it’s fully sanctioned by the state.

That’s the balancing act at the center of this whole story. Every rate case is a negotiation between two things that are structurally in tension: the utility’s shareholders want the highest guaranteed return the commission will bear, because a higher ROE means every dollar of capital spending is more profitable, which means the utility wants to spend more capital, not less.

Meanwhile the ratepayer wants the lowest bill possible. Those interests do not converge on their own. They only converge if a regulator forces them to. So question one that everybody should be asking, the one IURC is now actually trying to answer, is what specific levers get pulled to make sure the commission is representing the family in Fort Wayne and not just rubber-stamping whatever the utility’s finance team modeled.

Earlier this year the commission already trimmed AES Indiana’s rate case down from what the utility asked for, approving 70 million against a larger request, and Commissioner Bob Deig dissented on the AES order because he didn’t think it went far enough. So this isn’t performative. There’s already a track record of the commission actually cutting utility asks.

Now here’s where it gets interesting for anyone thinking about AI infrastructure specifically, which is the whole reason this story is worth 2,500 words instead of a tweet.

Trackers exist because building for a predictable, slow-growing grid used to be a predictable, slow business. You’d forecast demand, file a rate case every few years, true things up. But data center load doesn’t move at the pace of a traditional rate case.

NIPSCO just signed a supply settlement with Amazon. AES Indiana’s resource plans are explicitly modeling accelerating large-load growth, and its own environmental critics have flagged that the utility’s updated capacity plans lean harder into natural gas specifically because data center demand keeps pushing the load forecast up.

Trackers are the tool utilities reach for when they need to recover costs faster than the traditional rate case cycle allows, and there’s no faster-moving cost driver on an Indiana utility’s balance sheet right now than the capital it’s spending to serve hyperscaler and chip-fab load. So when regulators say they’re investigating trackers, they’re not investigating some obscure accounting footnote. They’re investigating the exact financial plumbing that lets a utility pass the cost of building for AI onto the general ratepayer base faster than anyone can publicly object.

That’s the second core tension people miss: financial trackers exist to stabilize utility revenue against demand swings, but the swing driving Indiana’s grid right now isn’t residential demand fluctuating with the weather. It’s a small number of enormous, well-capitalized customers, chip fabs and data centers, whose load additions dwarf anything a normal household does. A tracker built to smooth out ordinary volatility becomes something very different when it’s actually smoothing out the cost of building gigawatts of new capacity for Amazon and Nvidia’s supply chain. The mechanism hasn’t changed. What it’s being used for has.

Which gets us to the part of this story that should actually keep you up at night if you’re a founder or an investor betting on U.S. compute buildout: the political tolerance for this arrangement is visibly running out.

Governor Mike Braun didn’t just appoint a new IURC chairman, Anthony Swinger. Braun publicly said Hoosiers can’t take it anymore and called for a rehearing of a prior AES rate hike approval. Swinger, who came out of the state’s ratepayer advocate office, is now the one running these investigations, and he’s already recused himself from twenty pending dockets because of his prior work fighting utilities on behalf of consumers.

This is a governor who ran on affordability installing a former ratepayer advocate to go after utility profits, in a state where NIPSCO’s residential bills jumped nearly 27 percent in a single year and CenterPoint’s jumped almost 25. Whether that’s good policy or regulatory overcorrection is a real debate, but don’t mistake it for background noise. It’s the leading edge of a fight that’s going to show up in every state with heavy data center load, because the affordability math and the AI buildout math are now the same math, and politicians have figured that out faster than most tech investors have.

So who actually wins and loses if this goes all the way through?

Utility shareholders lose in the direct sense, a lower authorized ROE mechanically compresses the return on every dollar of capital spent, and it reduces the sector’s attractiveness relative to other capital-intensive plays competing for the same investor dollars.

But here’s the twist that a lot of affordability advocates don’t want to say: a utility with a squeezed ROE and tighter tracker rules doesn’t necessarily build less.

It just gets pickier about what it builds and who it builds it for. The households and small businesses. Which is why Amazon signed a supply settlement with NIPSCO instead of just complaining about it. Which is why AES Indiana’s rate case increasingly gets negotiated with an eye toward large-load customers who can offer contractual certainty in exchange for capacity.

If general ratepayers are going to push back harder on absorbing the cost of the AI buildout, and Indiana regulators are now explicitly signaling they will, then the marginal dollar of new grid capacity increasingly gets financed by the hyperscaler and the fab directly, through negotiated large-load tariffs and dedicated service agreements, not socialized across the residential rate base through a tracker. That’s not a side effect of this story. That is the story, and it’s the same pattern I’ve written about with nuclear pipelines and co-location deals elsewhere: whoever can pay for their own dedicated power gets to build. Whoever can’t gets stuck in the queue behind a rate case.

That reshapes the competitive landscape among AI infrastructure players in a way that has nothing to do with model quality.

If you’re a hyperscaler or a chip company that can walk into a state utility commission with a balance sheet strong enough to sign a direct supply agreement, bypass the socialized cost fight entirely, and effectively buy your way to the front of the interconnection queue, you have a durable advantage over a smaller AI infrastructure player who’s stuck depending on the general grid and therefore stuck depending on the political mood of whichever state government is running the affordability fight that year.

This is the binding constraint again, the one I keep coming back to in this publication: it was never really about chip supply, and it’s increasingly not even just about megawatts. It’s about who can secure a bespoke financial relationship with a regulated utility that insulates them from the affordability politics everyone else has to navigate.

For the innovation question, whether tighter ROE rules choke off grid innovation, I’d push back on the framing most consumer advocates use and most utility lobbyists use in the opposite direction. A lower guaranteed return doesn’t stop a utility from investing in genuinely differentiated infrastructure, flexible interconnection tech, storage, faster permitting workflows.

What it stops is low-differentiation capital spending that only ever made sense because the guaranteed return made spending itself profitable regardless of whether the underlying asset was the smartest way to serve the load. If Indiana tightens ROE and tracker rules, you should expect utilities to get more selective and more creative about how they finance data center-driven buildout, not less active. The lazy capital gets squeezed out first.

And that’s ultimately the second-order price effect everyone should be watching. If regulators successfully compress ROEs and tighten tracker recovery, the headline story is lower bills for residential customers, and that’s probably true in the near term.

But the capital still has to get raised somewhere to build the gigawatts Indiana’s data center and chip-fab pipeline actually needs. If it can’t come from socialized rate base recovery anymore, it comes from direct large-load contracts, and those contracts get priced to reflect the exact risk and financing cost that used to be spread across everyone.

Consumer advocates should be careful what they wish for here. Squeezing the tracker doesn’t make the capital need disappear. It just changes who negotiates the price of that capital and who has the leverage to negotiate well. Right now, that’s Amazon and SK Hynix, not the median Hoosier household, and definitely not the smaller AI infrastructure players who don’t have a balance sheet big enough to cut their own deal.

If you’re building or investing in this space, here’s the takeaway that matters by audience.

If you’re a founder in AI infrastructure, this is your reminder that your unit economics increasingly depend on your ability to negotiate directly with a regulated utility rather than ride the general grid, and that capability is becoming a moat in its own right, separate from your model or your product.

If you’re an investor, watch state affordability politics the way you’d watch a supply chain risk, because a governor with an affordability mandate can move faster than a utility’s five-year capital plan, and that volatility now sits directly upstream of your portfolio’s compute costs.

If you’re an operator inside a utility or a hyperscaler’s infrastructure team, the lesson from Indiana is that the socialized-cost path is closing faster than most people in the industry expect, and the teams that lock in direct large-load agreements now are buying themselves years of certainty that everyone else will be fighting state legislatures for later.

If you’re in government, Indiana just handed you a playbook, not a warning. You can run an affordability investigation and still land the SK Hynix fab. The two aren’t in tension. What’s in tension is who bears the cost of the grid that fab and its neighbors need, and that’s a choice regulators are now making explicitly instead of letting it happen by default through a tracker line item nobody reads.

I used to be technical.

I was an engineer. I wrote code for a living, the kind of code where if you got it wrong, something broke and you found out immediately.

Then my career did what most careers do. It moved up and away from the keyboard.

Programs. Transformation. Strategy. Leadership. All the things you get promoted into once you’re good enough at the thing you stop doing.

I haven’t written a meaningful piece of code in about fifteen years.

Right now I’m building a full agentic AI control plane. Frontend. Backend. Database. Agent orchestration. Cost agents, risk agents, value agents, challenger agents, human approvals, audit trails. The whole stack, end to end, by myself.

Am I confident I can pull it off? Not completely. I have real doubts, the kind that show up at 11pm when something isn’t working and I don’t have the muscle memory to know why.

But I believe something more strongly than I doubt myself: AI can make you radically better at almost anything, provided you have enough agency to keep moving when you don’t fully understand the path in front of you.

That’s the part nobody wants to say out loud, because it sounds like it’s letting people off the hook. It isn’t. It’s actually a harder standard than the old one.

For thirty years, the operating assumption was simple. Technical work required technical people. You either had the years in, the syntax memorized, the mental model of the compiler, or you didn’t build the thing. Competence was gated by accumulated, hard-won, narrow expertise. That gate is what made engineers valuable and made everyone else defer to them.

AI doesn’t remove that gate so much as it moves it.

The binding constraint isn’t syntax anymore. Syntax is now the cheapest part of the entire process, something the model produces on demand, correctly, faster than I could type it even in my sharpest engineering years. What’s expensive now, what actually determines whether the control plane gets built or dies in a half-finished repo, is something else entirely.

It’s knowing what to build. It’s being able to break an ambiguous goal into a sequence of concrete, checkable steps. It’s asking the model the right question instead of a vague one, and knowing enough to smell it when the answer it gives you is confidently wrong. It’s testing, re-testing, and having the stomach to keep moving when the first three approaches don’t work and you can’t fully explain why.

None of that is coding. All of it used to require coding to develop.

That’s the uncomfortable part.

AI hasn’t just closed my fifteen-year gap. It’s exposed the fact that the gap was never purely technical to begin with. A huge share of what made “technical people” valuable was never the syntax. It was the judgment that came from years of hitting walls and learning to diagnose them.

AI can now hand you the wall-hitting experience on demand, compressed, without the fifteen years. What it cannot hand you is the willingness to keep hitting walls until something works. That part is still yours. That part was always yours.

So I’m not pretending AI has made me a software engineer overnight.

It hasn’t, and anyone who tells you a tool alone rebuilt fifteen years of missing reps is selling something. What I’m actually testing is narrower and, I think, more honest: whether judgment, curiosity, systems thinking, and high agency can bridge a technical gap that used to be uncrossable without putting in the years.

If the answer is yes, and I think it’s trending yes, then the real scarce resource in the AI era was never technical skill. It was always agency. Technical skill was just the expensive proxy we used to measure it, because for decades it was the only proxy available.

That proxy is gone now. What’s left is the real thing, standing there with nowhere to hide.

Let’s see.

Let me tell you what actually happened last week, because the headline undersells it.

GridMarket, a company most people outside the energy industry have never heard of, signed a deal with a microreactor startup called Deployable Energy.

The press release says $22.5 billion. Read the actual filing and the real number is $145 billion in lifetime contract value over 40 years, built on more than 3 gigawatts of nuclear capacity that GridMarket’s data center and hyperscaler customers are expected to buy through 2035.

Deployable’s reactor, called the Unity Nuclear Battery, hit criticality the week before the announcement, which is the nuclear industry’s version of a startup shipping its first product to a paying customer. The plan is 500 megawatts of new capacity every year starting in 2030, running through 2035.

That is not a partnership between two energy companies. That is a bet that nuclear power is about to become a commodity product that gets sold the way compute gets sold, and that whoever controls the sales funnel for that product owns one of the most valuable positions in the entire AI economy.

I want to walk you through why this deal matters more than it looks like it does, because underneath the press release language is the clearest example yet of the binding constraint thesis playing out in real time. The AI race stopped being about who has the best model a while ago. It is now about who can get electrons to a building fast enough, reliably enough, and cheaply enough to keep GPUs running.

GridMarket and Deployable just showed you what the next five years of that fight looks like.

Why nuclear, and why now

Here is the mental model you need first. Every hyperscaler procurement strategy used to run on a simple assumption: buy power from the grid, sign a power purchase agreement with a solar or wind developer if you want to look green, and let the utility handle the hard part.

That assumption is dead. Data center power demand is growing faster than transmission lines can be built, faster than gas turbines can be manufactured, and faster than utilities can get interconnection studies approved.

Microsoft, Google, Amazon and Meta have all responded by going straight to the source and signing nuclear deals that would have looked insane five years ago. Microsoft locked in an 835 megawatt, $16B dollar, twenty year deal to restart Three Mile Island Unit 1.

Google committed to 500 megawatts from Kairos Power’s molten salt reactors. Amazon put $700M dollars into X-energy for up to twelve reactors. Meta has stacked commitments across TerraPower, Oklo, Vistra and Constellation that could add up to 6.5 gigawatts.

That is the first mental model: energy procurement for hyperscalers has stopped being a cost center decision and become a strategic infrastructure decision, on par with choosing which cloud region to build in or which chip vendor to lock into a multi-year supply agreement.

Nobody signs a 40 year, $145 billion dollar contract to save money in the next fiscal quarter. They sign it because they have done the math on AI compute demand through the 2030s and concluded that firm, dispatchable, carbon-free power is going to be scarcer than the chips themselves.

This directly answers the first deep question worth asking about this deal: what happens economically when data centers stop being fossil fuel customers and become nuclear customers. The answer is that the cost structure of AI compute starts to look less like a technology cost curve and more like a utility cost curve, locked in decades in advance, insulated from natural gas price swings, and backed by contracts that behave more like infrastructure bonds than software subscriptions.

The matchmaker gets paid more than the reactor builder

Now here is the part that should make you sit up if you actually care about where value gets captured in this stack.

GridMarket does not build reactors. GridMarket does not own uranium. GridMarket is, functionally, a lead generation and deal facilitation platform. It has spent a decade building relationships with Fortune 500 to Fortune 2000 industrial customers, data centers, and hyperscalers who need energy solutions, and it uses that funnel to route demand toward supply.

In this deal, GridMarket is the party that gets Deployable Energy in front of buyers who are already qualified, already have deployment-ready sites, and are already desperate for power. Deployable gets distribution. GridMarket gets a cut of a one hundred and forty five billion dollar pipeline without ever pouring concrete for a reactor.

This is the second mental model, and it is the one that should sound familiar if you have read anything I have written about AI infrastructure before: the party that controls demand aggregation captures disproportionate value relative to the party that owns the physical asset.

We have seen this exact pattern with NVIDIA’s software stack sitting on top of commodity silicon, with cloud providers sitting on top of commodity chips, and now with GridMarket sitting on top of commodity nuclear megawatts. Owning a chokepoint in the deal flow is worth more than owning the thing being sold through that chokepoint, as long as the thing being sold is scarce enough that buyers will pay for faster, more certain access to it.

That answers the second deep question, the one about competitive dynamics among energy providers. GridMarket is not competing with Kairos Power or Oklo or TerraPower. It is positioning itself as the layer above all of them, a customer acquisition and deployment facilitation platform that could, in theory, plug in a different reactor vendor tomorrow if Deployable stumbles. That is a much better competitive position than being a single-technology nuclear developer racing against five other single-technology nuclear developers for the same handful of hyperscaler contracts.

A crowded reactor race with very different bets

Speaking of that race, it is worth understanding just how crowded this field has gotten, because it changes how you should read any single deal, including this one.

Google went with Kairos Power’s fluoride salt design. Amazon backed X-energy’s high temperature gas reactor. Meta split its bets across TerraPower’s liquid sodium Natrium reactor and Oklo’s liquid metal fast microreactor, the Aurora, which Meta committed one point two gigawatts to in January. Equinix, the colocation giant, has spread even further, signing agreements with Stellaria, Radiant, and Oklo simultaneously.

As of a few months ago, industry trackers counted thirteen separate nuclear deals across the major hyperscalers totaling nearly ten gigawatts of committed capacity.

Deployable Energy’s Unity Nuclear Battery is a microreactor, smaller and theoretically faster to deploy than the larger SMR designs from Kairos or X-energy, and it is explicitly targeting the segment of the market that needs power at individual sites rather than utility-scale campuses.

That is a real point of differentiation, but it also means Deployable is competing directly with Oklo’s Aurora and with newer entrants like Aalo Atomics, which is targeting an experimental reactor criticality of its own this July. The microreactor sub-market is arguably more crowded than the large SMR market right now, precisely because the barrier to entry looks lower on paper, even though the regulatory and manufacturing barriers remain enormous in practice.

Where the experts genuinely disagree, and why you should care

I am not going to pretend this is a settled question, because it is not, and the disagreements matter more than the marketing copy suggests.

The first disagreement is about deployment speed. The entire pitch for microreactors and SMRs is that they can be built in three to five years instead of the decade or more required for a traditional large reactor. But several industry analysts think even 2030 is optimistic for first-of-a-kind projects, and that mid-2030s is more realistic for anything resembling broad commercial availability. Deployable is targeting 500 megawatts a year starting in 2030. That is an aggressive schedule for a technology that achieved criticality for the first time only weeks ago.

The second disagreement is about whether nuclear can actually outrun solar, wind, and battery storage in a straight race to meet data center demand this decade. Nuclear proponents argue that intermittent renewables cannot provide the twenty four hour, seven day a week firm power that a GPU cluster needs without absurd amounts of battery overbuild.

Renewable advocates counter that solar and storage costs keep falling every year while nuclear costs keep rising with every regulatory delay, and that betting forty year contracts on unproven reactor designs is a much riskier capital allocation than stacking gigawatts of solar and batteries that are already commercially proven today.

The third disagreement is about public acceptance and safety, and this one is not going away just because the industry wants it to. Building reactors, even small ones, near communities that have no history hosting nuclear infrastructure is going to run into local opposition, permitting fights, and the kind of political friction that has slowed nuclear projects in the United States for fifty years. A microreactor that fits on a truck is still a nuclear reactor, and “factory-made” does not mean “friction-free.”

The fourth disagreement is about whether government incentives or pure market demand are actually driving this wave. Deployable’s criticality milestone happened specifically in fulfillment of an accelerated nuclear deployment executive order, and the company went through the Department of Energy’s Reactor Pilot Program. Some analysts think hyperscaler demand alone would have gotten nuclear projects this far. Others think the regulatory tailwind from Washington is doing more of the work than anyone in Silicon Valley wants to admit, which matters a lot if the political winds shift.

The fifth disagreement is about energy independence versus concentration risk. Is it better for a handful of hyperscalers to build a decentralized fleet of on-site microreactors that reduce reliance on the fragile public grid, or does that just recreate the same chokepoint problem in a new form, where a small number of reactor vendors and deal facilitators like GridMarket end up controlling the power supply for the entire AI economy? Decentralization at the site level can still mean concentration at the ownership level.

This directly answers the third and sixth deep questions worth asking here. Regulatory change is not a hypothetical future need, it is the mechanism that already got Deployable to criticality in about one hundred and fifty days from project start, which is unheard of in nuclear timelines. And public perception is going to be shaped less by press releases and more by whether the first handful of pilot deployments happen without incident, on schedule, near communities that were told this technology was different from the reactors their grandparents feared.

Who wins, who gets squeezed, and what it does to prices

Let’s talk about the pricing question directly, because it is the one every operator reading this actually cares about.

When a hyperscaler locks in a forty year nuclear contract, it is buying price certainty in a world where natural gas prices, transmission costs, and grid congestion charges are all becoming less predictable.

That certainty has value even if the per-megawatt-hour price of nuclear power ends up higher than the current grid average, because it removes a massive variable from a company’s long-term compute cost model.

Over time, as more of these contracts get signed, expect the regions where hyperscalers site their reactors to see local industrial power prices diverge sharply from regions without that kind of capital showing up. Communities near these projects could see cheaper, more reliable power as a byproduct of hyperscaler investment. Communities without it get left further behind on the grid modernization curve, because the capital that used to go toward general grid reliability increasingly gets earmarked for point-to-point deals between one buyer and one reactor.

That is the fourth deep question, on regional pricing structure, and it has a clear answer: this is not going to be an even benefit. It is going to concentrate cheap, reliable power in specific geographies chosen by GridMarket’s existing customer funnel and Deployable’s siting decisions, the same way data center buildout itself has concentrated in Virginia, Texas, and a handful of other states with favorable land, water, and grid access.

On the labor question, the fifth deep question, the honest answer is that nuclear construction and operations require a completely different labor pool than the electricians and HVAC technicians currently staffing data center buildouts. Nuclear licensing, reactor operations, and specialized manufacturing for factory-built microreactors demand a workforce the United States has been under-investing in since the last nuclear construction wave decades ago.

If this pipeline actually executes at scale, it creates real upward pressure on wages for nuclear engineers and licensed operators, and it creates a new category of infrastructure jobs tied directly to AI demand, which is a strange and underappreciated second-order effect of the compute race.

The power question, the seventh deep question, is where this gets genuinely interesting from a Full-Stack Capitalist lens. The companies that stand to gain the most structural power here are not the hyperscalers, who are simply buying an input they need, and it is not even the reactor developers, who are taking on enormous execution risk for a promised payout years away. It is platforms like GridMarket, sitting in the middle with existing customer relationships on one side and a portfolio of interchangeable reactor vendors on the other. That middle position is the least risky and most durable place to sit in this entire value chain, and it is exactly the kind of chokepoint this publication has been arguing is the real prize in the AI infrastructure buildout, more valuable long term than owning any single layer of the physical stack.

The people who get squeezed are the traditional utilities, who now have to compete for hyperscaler dollars against direct-to-developer deals that bypass the regulated utility model entirely, and smaller data center operators who lack the balance sheet to sign forty year nuclear contracts and will be stuck buying grid power at whatever price is left over once the hyperscalers have locked in their firm capacity.

Data centers as political actors, not just customers

The eighth deep question asks what role data centers might play in advocating for nuclear policy, and the honest answer is that they already are political actors in this space, whether they say so publicly or not. When a company signs a sixteen billion dollar deal to restart Three Mile Island, or commits to eight reactor plants with TerraPower, it has every incentive to lobby for faster NRC licensing timelines, favorable siting rules, and continued executive branch support for accelerated nuclear deployment. The interests of Big Tech and the interests of the nuclear industry, an industry that has spent fifty years trying and failing to get meaningful new construction off the ground in the United States, are now more aligned than at any point in American history. That alignment is going to show up in lobbying spend, in state-level permitting reform pushes, and in federal policy long before it shows up as electricity flowing to a data center.

Resilience, and what happens if this actually works

The ninth deep question, about infrastructure resilience, cuts both ways. On one hand, on-site nuclear generation reduces a data center’s exposure to grid outages, transmission bottlenecks, and regional demand spikes, which is a genuine resilience win. On the other hand, it introduces new single points of failure. A microreactor going offline for maintenance, a licensing dispute freezing a pilot deployment, or a safety incident anywhere in the industry could trigger regulatory pauses that ripple across every hyperscaler’s nuclear pipeline simultaneously, the same way a single chip fabrication problem can ripple across the entire AI hardware supply chain. Concentrating power generation and power demand in the same handful of technologies creates efficiency, but it also creates correlated risk.

And that brings us to the tenth question, the one about what happens globally if this pipeline actually delivers. If GridMarket, Deployable, Oklo, Kairos, X-energy, and TerraPower collectively prove that factory-built nuclear can go from criticality to commercial deployment inside a five to seven year window, the implications go well beyond American data centers. Every country racing to build sovereign AI compute capacity, and there are a lot of them right now, will look at this playbook and try to replicate it, because energy access, not chip access, will be the actual constraint on national AI ambitions by the early 2030s. Countries that can permit and build nuclear fast will have an AI infrastructure advantage that has nothing to do with how many GPUs they can import. Countries that cannot will find themselves renting compute from someone else’s grid, which is its own form of dependency.

What this means if you actually build or invest in this stuff

If you are a founder or operator building anything that depends on compute access, the takeaway is not that you need to go sign a nuclear deal tomorrow. It is that power availability has quietly become a due diligence question you should be asking any cloud or colocation partner you depend on, the same way you would ask about their chip supply relationships. Whoever you rent compute from is making decisions right now about their power stack that will determine your costs and your capacity access five years from now.

If you are an investor, the interesting opportunity is not necessarily in the reactor developers themselves, who carry enormous execution and regulatory risk on unproven first-of-a-kind technology. It is in the demand aggregation layer, the GridMarkets of the world, and in the specialized labor, manufacturing, and services businesses that will need to scale alongside a nuclear construction wave the United States has not seen in half a century.

If you are in government, the lesson from this deal is blunt. The executive order that helped get Deployable to criticality in one hundred and fifty days did more to move this industry forward than decades of subsidy programs managed to do through normal channels. Whichever country, state, or municipality figures out how to permit nuclear fast without cutting corners on safety is going to become the default site for the next wave of AI infrastructure investment, and that is a competitiveness lever most policymakers have not fully internalized yet.

The binding constraint keeps moving, but it never disappears

Six months ago the binding constraint conversation in AI was about chip supply. A year before that it was about model quality. Right now, in the middle of 2026, the binding constraint is standing in a field in Idaho, waiting for a truck-sized reactor to reach commercial licensing. GridMarket and Deployable Energy did not just announce a nuclear deal. They announced who they think is going to own the next chokepoint in the AI economy, and it is not the company that builds the reactor. It is the company that decides who gets access to it first.