$1.1 billion.
That is the headline value of the data centre agreement between Digi Power X and Cerebras.
People will look at a deal like this and focus on the obvious things.
40 MW.
10 years.
AI compute.
Data centre capacity.
Infrastructure.
But that is not what interests me.
What interests me is the power structure underneath the deal.
Who owns what?
Who pays for what?
Who carries the risk?
Who keeps the asset at the end?
And most importantly, who is actually making the better bet?
I went through the agreement from a commercial perspective, and my read is pretty simple:
Digi Power is betting on infrastructure.
Cerebras is betting on AI demand.
And in my opinion, Digi Power has the stronger side of the deal.
Not because Cerebras is making a bad decision.
But because Digi Power is betting on something durable.
Cerebras is betting on something that still has to prove itself over time.
Why this contract caught my attention
I have spent parts of my career around data-centre operations, transformation and project delivery.
When you work close enough to infrastructure, you start to see things differently.
A data centre is not just a building.
It is power.
Cooling.
Redundancy.
Permits.
Connectivity.
Generators.
UPS.
Land.
Grid access.
Construction.
Operations.
Maintenance.
Financing.
And increasingly, it is the physical layer underneath the AI economy.
That is why I am less interested in the contract value and more interested in the ownership structure.
Because when the hype disappears, ownership matters.
Cash flow matters.
Power access matters.
And residual asset value matters.
Start with the basic structure
Digi Power is the operator.
Cerebras is the customer.
Digi Power owns the land.
Digi Power builds the facility.
Digi Power owns the power infrastructure.
Digi Power owns the cooling infrastructure.
Digi Power owns the generators, UPS systems, distribution systems and core building infrastructure.
Cerebras gets access to the facility.
Cerebras owns the servers.
Cerebras owns the networking equipment.
Cerebras owns the software.
Cerebras owns the models.
Cerebras owns the AI workload.
So the first thing that jumps out is that the stack is split in two.
Digi Power owns the physical infrastructure.
Cerebras owns the compute.
That distinction matters a lot more than it looks.
Digi Power owns the durable layer
The contract is explicit.
Digi Power owns the land and the facility.
Cerebras does not get a real-property interest.
It gets a licence to use the facility.
That means Cerebras can operate inside the asset, but Digi Power owns the underlying asset.
This is important because the useful life of the two layers is very different.
A server can become economically obsolete very quickly.
AI hardware can depreciate fast.
One generation of chips can be overtaken by the next.
But land does not disappear.
Grid connection does not disappear.
Power infrastructure does not disappear.
Cooling infrastructure does not disappear.
The building does not disappear.
The asset can be upgraded, repurposed, refinanced, sold or used by another customer.
That is the first reason I think Digi Power has the stronger position.
They are betting on the layer that survives technology cycles.
Cerebras is helping fund an asset it does not own
This is one of the most interesting parts of the agreement.
Cerebras is not simply paying a monthly fee.
It is also contributing upfront capital toward parts of the build through non-recurring charges.
Those charges are tied to design, procurement, installation and construction of the data centre infrastructure.
And under the agreement, Cerebras can potentially pay suppliers directly.
Think about that for a second.
Cerebras is helping fund the build.
But Digi Power keeps the infrastructure.
That is a very attractive structure for the asset owner.
The customer helps finance the development.
The operator keeps ownership.
The customer then continues paying for capacity over the life of the contract.
And the operator still owns the residual asset at the end.
The exact economics are redacted, so I cannot say how generous or aggressive that balance is.
But structurally, it is important.
Cerebras is contributing capital into infrastructure that remains on somebody else’s side of the ownership line.
The product is not floor space
This is another thing the AI data-centre conversation gets wrong.
The product is not really square metres anymore.
The product is megawatts.
Cerebras is contracting for up to 40 MW of IT load.
That is what matters.
Digi Power must make that power available.
The agreement requires firm, non-interruptible power up to the contracted capacity, subject to genuine grid emergencies.
That means Digi Power is not really selling a building.
It is selling reliable access to electricity inside a highly engineered environment.
That is the scarce asset.
And in the AI economy, scarcity creates power.
Land is useful.
A building is useful.
But land with guaranteed power is a completely different asset.
Cerebras takes the utilisation risk
The take-or-pay structure is one of the most important commercial terms in the contract.
Cerebras pays based on contracted capacity.
Not actual usage.
If it contracts for the capacity and does not use all of it, it still pays.
That transfers utilisation risk to Cerebras.
Digi Power does not need to care as much about whether every rack is running flat out every day.
It has contracted revenue.
Cerebras, on the other hand, has to monetise that capacity.
That means Cerebras is betting on AI demand.
It is betting that the workload exists.
It is betting that customers exist.
It is betting that the economics of AI inference, training or whatever workload sits behind this facility remain attractive enough to justify that capacity.
That is a much more volatile bet.
Cerebras is betting on AI demand for a decade
This is the part I keep coming back to.
A ten-year contract is a long time in AI.
Look at how fast the technology has changed already.
Models change.
Hardware changes.
Efficiency changes.
Training architectures change.
Inference costs fall.
New competitors appear.
Workloads shift.
Software improves.
And hardware economics can change brutally fast.
Cerebras is making a long-duration commitment in a market where the underlying technology can change every year.
That does not mean it is a bad decision.
It may be exactly the right decision if access to power becomes the biggest constraint in AI.
But it is still a bet.
And the bet is fundamentally on future demand.
Digi Power does not need to predict which AI model wins.
It does not need to predict which chip architecture wins.
It does not need to predict which application wins.
It needs demand for powered infrastructure to remain valuable.
That is a much broader bet.
Digi Power carries the construction risk
This deal is not one-sided.
Digi Power carries real risk.
It has to deliver.
It has to build.
It has to commission the facility.
It has to hit the Ready for Service requirements.
If it is late, it can face delay credits.
If it fails badly enough, Cerebras can terminate.
Digi Power also has to maintain the infrastructure throughout the term.
Power.
Cooling.
Generators.
Fire systems.
Monitoring.
Security.
Permits.
Operations.
So Digi Power is taking project and operational risk.
That matters.
Owning infrastructure is not free money.
The operator has to execute.
But that risk is front-loaded and operational.
Once the infrastructure is built and functioning, the economics become much more attractive if the contract performs as expected.
Phase 2 tells you something else
The deal is split into phases.
Phase 1 is 15 MW.
Phase 2 adds another 25 MW.
But Phase 2 depends on financing.
Digi Power has to secure debt or equity financing for the additional build.
And Cerebras gets visibility over that financing.
This tells me something important.
The real bottleneck is not only technical.
It is financial.
Data centres are becoming capital structures.
You need land.
You need power.
You need permits.
You need equipment.
You need customers.
And then you need enough capital to connect all of those pieces together.
Digi Power is not just acting as an operator.
It is acting as a capital allocator.
Cerebras has more control than a normal customer
This is where the deal gets more interesting.
Cerebras does not own the facility.
But it has significant influence over it.
It has exclusive access.
It has rights over future capacity.
It has protections if Digi Power finances the facility.
It has protections if a lender forecloses.
It has rights around additional occupants.
It has restrictions around who the facility can potentially be transferred to.
So Cerebras has managed to get something very valuable:
control without ownership.
That is strategically smart.
It avoids tying up even more capital in land and real estate.
It allows Cerebras to focus on compute.
But there is a trade-off.
Control rights are not ownership rights.
At the end of the day, Digi Power still owns the physical asset.
The lender may have more power than either company
This is another part people rarely talk about.
Digi Power can finance or refinance the facility.
That means lenders can sit above the asset.
Mortgages.
Security interests.
Debt.
Potential sale-leaseback structures.
If Digi Power gets into trouble, the lender may eventually control the asset.
The contract anticipates this.
It includes recognition and non-disturbance protections designed to preserve Cerebras’ rights even if ownership changes through foreclosure.
That tells you something about the real power structure.
There are at least three important actors.
Digi Power owns the asset.
Cerebras controls the capacity.
The lender controls capital.
And above all of them sits the grid.
The grid is the kingmaker
This may be the most important point.
Everyone can negotiate everything perfectly.
Digi Power can own the land.
Cerebras can bring the chips.
The lenders can bring the capital.
But without power, the entire asset is useless.
That is why I keep coming back to power availability as the real constraint in AI infrastructure.
The contract reflects that.
Digi Power must provide firm power.
It cannot simply divert Cerebras’ capacity to someone else.
The agreement also carefully splits electricity-price risk.
Certain new government charges, transmission charges or regulatory costs can be passed through.
Normal electricity commodity-price volatility generally cannot simply be dumped onto Cerebras.
That means somebody has to manage that exposure.
Somebody has to hedge it.
Somebody has to negotiate with utilities.
Somebody has to carry the risk.
This is where data-centre economics start looking less like technology and more like energy infrastructure.
The hidden third party
One of the most interesting things in the contract is what is missing.
A number of references are redacted.
But those redactions appear in places that matter.
There are references to a third party in relation to end-user rights.
Consent rights.
Transfer rights.
Occupancy rights.
Protection rights.
I am not going to speculate publicly about who that party is.
But commercially, its presence matters.
It suggests that the visible contract between Digi Power and Cerebras may only be one layer of a larger commercial structure.
There may be another customer behind Cerebras.
Another capital provider.
Another beneficiary.
Another party with economic rights.
That would make the structure look something like this:
Digi Power
↓
Cerebras
↓
Underlying end customer or beneficiary
And that raises another question:
Who is Cerebras really building this capacity for?
That question may be more interesting than the contract itself.
So who has the power?
My read is Digi Power.
Not absolute power.
Not all the power.
But the stronger long-term position.
Why?
Because they own the durable layer.
They own land.
They own infrastructure.
They own cooling.
They own power systems.
They own the residual asset.
They have a long-term customer.
They have take-or-pay economics.
And they potentially receive customer contributions toward construction.
Cerebras has substantial contractual control.
But Cerebras is still making the riskier bet.
Cerebras is betting on AI workloads.
AI demand.
Utilisation.
Compute economics.
Hardware economics.
And the ability to monetise that capacity for years.
If AI demand continues exploding, Cerebras may look brilliant.
If compute economics compress, utilisation disappoints or hardware becomes dramatically more efficient, Cerebras still has a long-duration capacity commitment.
Digi Power still owns the infrastructure.
This is why I think infrastructure wins
I am becoming more convinced that one of the most important questions in the AI economy is not:
Who has the best model?
Or:
Who has the most GPUs?
It is:
Who owns the durable bottlenecks?
Power.
Land.
Grid connection.
Cooling.
Transmission.
Physical infrastructure.
Those assets do not care which model wins.
They do not care which AI company is fashionable this year.
They sit underneath all of them.
And that is what makes this $1.1 billion contract so interesting to me.
Cerebras is making an AI bet.
Digi Power is making an infrastructure bet.
One is betting on demand.
The other is betting on scarcity.
Right now, I know which side of that trade I would rather own.