Could we be right about AI changing the economy and wrong about how much debt its revenues can support?
That question emerged from my discussion with Kevin. It deserves more attention than the familiar argument over whether AI is revolutionary or overhyped.
AI can become more capable, more widely used and more economically valuable while some of the businesses financing it fail. Customers can benefit from falling prices while infrastructure owners struggle to repay their lenders. Useful assets can survive while the original shareholders lose their investment.
My position is that optimism about the technology should come with discipline about its financing.
A full-stack capitalist follows the entire chain: electricity, buildings, chips, cloud services, models, applications and paying customers. At every layer, the questions are concrete. Who supplies the capital? Who controls the asset? Who earns the margin? Who owes the money? And who absorbs the loss when expectations disappoint?
The answers can be very different within the same industry.
Start with the customer’s money and work backwards.
An enterprise pays for an AI application because it expects some combination of lower costs, faster work, better decisions or additional revenue. The application provider pays for models and computing. Those payments help support servers, data centres, electricity and the financing used to build them.
Investment capital can fund this chain before customer revenue arrives. That is a normal and necessary part of building new industries. But the long-term economics ultimately require customers willing to pay, or other durable corporate cash flows willing to support the expenditure.
The distinction matters because revenue at one layer is an expense at another. A cloud provider can report strong sales to a model developer that is still consuming investor capital. Those sales are real, but their durability depends partly on whether the customer develops a sustainable business.
Counting activity across the stack does not establish that the whole stack is profitable.
Nor does a productivity benefit automatically become revenue for an AI supplier. A business might save $100 using a tool but pay only $10 for it because competition keeps prices low. That can be an excellent outcome for the economy while leaving less income available to support the infrastructure than investors expected.
Technological value, customer value and investor returns are related. They are not interchangeable.
Debt makes the timing of success consequential.
Equity investors can tolerate an uncertain path to future profits, although their patience and funding are finite. Debt introduces interest payments, maturity dates and contractual requirements that can become binding before the business reaches its potential.
A project can therefore be directionally right and financially early.
Customers may arrive eventually. Electricity connections may become available eventually. Applications may become profitable eventually. The project still needs enough cash to meet its obligations along the way.
The Reserve Bank of Australia’s October 2026 review distinguishes large technology companies from more vulnerable parts of the AI supply chain, including businesses with weaker balance sheets or concentrated customers. It also notes increasing use of arrangements outside the main corporate balance sheet to finance projects. Those differences make broad claims about “AI debt” inadequate.
RBA Financial Stability Review
A diversified company funding investment partly from established businesses has a different capacity to withstand disappointment from a specialist operator dependent on a few customers and continued borrowing.
The relevant unit of analysis is the company, project and contract, not simply the sector.
The financing structure determines where the risk lands.
A separate project company can own a data centre, equipment or contractual rights. Investors supply equity, lenders supply debt, and customer payments support operating expenses and financing obligations.
This can be sensible. It allows different investors to fund different assets and can limit their exposure to unrelated activities.
But the legal structure does not create additional customer demand.
The building owner may not own the GPUs. The equipment lender may have no claim on the building. The company buying computing services may have payment obligations without guaranteeing the operator’s debt.
A separate company also does not automatically imply securitisation or misconduct. These are established financing techniques. What matters is the actual allocation of rights and obligations.
CoreWeave’s 2023 announcement of a $2.3 billion debt facility provides a straightforward example: the company said the funding would support hardware for contracts already executed with customers, alongside hiring. It illustrates how customer commitments can support investment ahead of the associated cash receipts. It does not, by itself, establish that those commitments eliminate financing risk. CoreWeave announcement
Meta’s Hyperion arrangement illustrates a different allocation. Its October 2025 announcement described a joint venture owned 80% by Blue Owl-managed funds and 20% by Meta, with approximately $27 billion of development costs for buildings and long-lived infrastructure. Meta also described initial four-year leases with extension options and a capped residual-value guarantee applying in specified circumstances during the first 16 years. Meta announcement
Outside capital participated, but Meta retained exposure through ownership, leases and a conditional guarantee. The precise scope of those obligations matters more than a headline about moving infrastructure off a balance sheet.
The financing box allocates legal claims. It does not make the underlying economic risk disappear.
A contract is only as protective as its terms and its counterparty.
A firm minimum-payment commitment can place utilisation risk on the customer: it may have to pay even when it uses little capacity. A cancellable or usage-based agreement leaves more of that risk with the operator.
Neither arrangement makes risk vanish. It changes who bears it.
Even a firm commitment depends on the customer’s ability to pay. A contract with a profitable, well-capitalised business differs economically from one with a customer that needs another fundraising round to meet its obligations.
Construction delays introduce another question. If power arrives late or the facility cannot deliver the agreed service, does the customer still owe the expected payments?
Then comes expiry. A project may enjoy predictable payments for several years but face lower prices, weaker demand or expensive upgrades when the contract ends.
“Contracted revenue” should therefore prompt further investigation: payment conditions, cancellation rights, customer credit quality, delivery obligations and renewal assumptions.
The assets also run on different clocks.
Land, power infrastructure, buildings and computing equipment have different economic lives.
A functioning GPU can lose commercial value without physically failing. New equipment may deliver the same work at a lower cost, making older equipment less attractive at yesterday’s rental price.
That creates a potential mismatch between the time available to earn a return and the time allowed to repay the financing.
Consider a simplified, hypothetical operator earning $100 million in annual revenue, with $55 million of operating cash costs. That leaves $45 million before debt payments. If annual interest and principal payments total $30 million, the apparent cushion is $15 million.
Now assume revenue falls 20%, while costs decline by only $5 million because many expenses are fixed or committed. The operator has $30 million left before debt payments, exactly enough to cover them, with nothing remaining for unexpected costs or equipment replacement.
This is an illustration, not a forecast. It shows how a moderate revenue disappointment can consume financial flexibility.
Accounting profit alone will not answer the question. Investors need to understand the cash required to maintain competitive capacity, as well as the cash required to service debt.
Some apparent protections depend on the same underlying forecast.
A lender might take comfort from customer contracts, equipment collateral and a strong industry outlook.
Those protections can weaken together.
If customers reduce spending, operating cash flow falls. If many operators then try to sell equipment, resale values may decline. If lenders become cautious at the same time, refinancing becomes harder.
The project experiences weaker income, weaker collateral and tighter credit simultaneously.
This is why concentration matters. Several projects with different owners may still depend on the same small group of customers. Several funds may hold different instruments exposed to the same underlying spending decisions.
Diversification across company names does not necessarily provide diversification across economic risks.
BIS researchers have highlighted the growing role of debt and private credit in financing AI investment, while stressing that sustainability depends on firms meeting demanding earnings expectations. Their January 2026 assessment described financial stability risks as moderate at that point, which supports scrutiny rather than certainty about an impending collapse. BIS analysis
Supplier investment deserves careful interpretation.
A supplier can invest in a customer that buys its products. A cloud provider can invest in a model developer that purchases computing from it. These arrangements can finance expansion, align incentives and accelerate useful development.
They can also connect risks that look separate on paper.
Nvidia’s January 2026 announcement of a $2 billion equity investment in CoreWeave is a concrete example of a supplier participating in its customer’s financing. The investment alone establishes neither wrongdoing nor unsustainable demand. Nvidia announcement
The economic question is how much spending ultimately rests on independent customers receiving enough value to keep paying.
If investment supports computing purchases, and those purchases justify further infrastructure financing, the system can expand rapidly. Its resilience depends on what happens when fresh financing becomes more expensive or less available.
A useful stress test is simple: if one major customer cannot raise its next round, how many suppliers, projects and investors are affected?
Agents and cheaper models could improve AI while disrupting its financing assumptions.
An agent that makes more model calls generates more computing activity. Whether it creates more economic value depends on what it successfully accomplishes.
For a business, the useful measure is cost per completed, acceptable task. That includes computing, integration, human review, retries and correction.
An agent can look inexpensive per call and expensive per successful outcome. Conversely, an agent that reliably completes valuable work can support substantial spending because the customer’s return justifies it.
This distinction connects application quality directly to infrastructure demand. Reliable outcomes help turn experimentation into recurring budgets.
Chinese open-weight models, and other cheaper or more efficient alternatives, introduce a further possibility. They could lower the cost of building applications and increase competition among model providers. That would benefit many businesses using AI while potentially pressuring prices assumed elsewhere in the stack.
Lower costs could also stimulate much greater usage.
Neither “efficiency destroys computing demand” nor “usage growth guarantees infrastructure returns” follows automatically. The result depends on how much demand expands, which equipment serves it, and how prices and margins change.
For a full-stack capitalist, this is where the analysis becomes interesting: progress can move profits between layers. A breakthrough that benefits application developers may weaken the economics of an infrastructure investment made under different assumptions.
Retirement capital can participate, but exposure needs to be traced.
Retirement savings can reach AI through listed shares, bonds, infrastructure funds and private credit. The extent depends on actual holdings and financing arrangements.
It is inaccurate to assume that an asset manager’s participation means every pension investor is exposed, or that technology companies have transferred all their risk to retirees.
The type of retirement arrangement also matters. In a defined-contribution plan, investment losses generally affect the participant’s account value. In a traditional defined-benefit pension, losses generally weaken funding first; promised benefits do not automatically decline dollar for dollar. US Department of Labor
The right questions concern transparency, concentration, liquidity and compensation for risk. A long investment horizon can help an investor tolerate volatility. It cannot make an overpriced asset or an unpayable loan economically sound.
Useful infrastructure can outlive its original financing.
Sydney’s Cross City Tunnel offers a local reminder. The original concession holder entered receivership in 2006–07, and the asset was subsequently sold. The physical tunnel survived the failure of its original financial arrangement. NSW transport annual report
The comparison has limits: a road tunnel and a fleet of GPUs have very different replacement requirements and economic lives. But the distinction between an asset and the claims financing it remains useful.
An AI facility could continue operating after restructuring, with creditors accepting losses and new owners paying a lower acquisition price. At that lower cost, the business might become viable.
That would not vindicate the original investment. It would show that useful capacity can survive after financial expectations are reset.
Nor would restructuring be painless. Suppliers could lose payments, employees could lose jobs, customers could face disruption and future investment could slow.
So, is financing AI’s biggest risk?
For a heavily indebted operator, financing may be the immediate existential risk. For a business deploying agents, reliability and measurable returns may be more pressing. For the financial system, the issue is whether losses spread through borrowing, concentrated exposures and interconnected institutions.
The Bank of England’s July 2026 assessment makes the distinction clearly: relatively modest debt stocks had helped contain financial stability risks, while rapidly growing external financing could make future shocks more consequential. Bank of England Financial Stability Report
My full-stack capitalist position is to support productive investment while demanding clarity about the promises attached to it.
I would want to see independent paying demand, credible customer obligations, repayment schedules suited to the assets, realistic replacement costs and enough financial flexibility to withstand lower prices or delays. I would also want to know who takes the first loss and whether the ultimate investors understand the exposure.
AI does not have to disappoint technologically for an investment to disappoint financially.
The technology can succeed while the original capital structure fails. What matters is whether the revenues support the promises, and whether the people bearing the risk understand those promises.