Picture this.
Two sponsors are building two very similar AI campuses. Both will own the compute inside and sell capacity to an AI lab under a five-year take-or-pay contract. The first is buying NVIDIA racks. The second is buying a hyperscaler's custom accelerator.
The term sheets come back looking very different.
One lender gives some credit to the hardware. The other barely mentions it and spends four pages on the offtake and who stands behind it.
Same megawatts. Same market. Very different credit story.
The difference is in the racks.
Even on the NVIDIA side, the chips aren’t the main thing being underwritten. CoreWeave described its $8.5 billion facility in March as the first investment-grade financing secured by HPC infrastructure and an associated customer contract.
The second half of that sentence is important.
What the GPU adds is a recovery floor. NVIDIA chips have a secondary market, though one constrained by export rules, software licensing, and warranty terms.
Custom silicon thins that floor.
The more proprietary the accelerator and the system around it, the smaller the pool of buyers able to redeploy it economically.
That pool is growing. Custom silicon isn’t new to training, and recent hardware and software improvements are broadening the workloads for which it can compete with merchant GPUs. One major cloud provider is taking its proprietary AI accelerators beyond its own estate, while another is moving its in-house training chips towards a much more familiar native-PyTorch environment.
But portability and training capability are different things.
A chip can be very good at training a model and still have a much narrower pool of potential buyers if the original user walks.
That is the financing issue… and so the credit moves.
Into the offtake, where term, minimum commitments, assignment, step-in and termination rights do more of the work.
Into counterparty support, because whoever stands behind the user’s commitment, or behind the hardware’s value, matters more when the hardware is less portable.
And back into the real assets: land, shell, power, substation, interconnection, and water. The parts of the campus that hold their value whatever silicon is installed inside.
This isn’t merely theoretical. Recent custom-silicon financings have already used SPV ownership, long-dated customer commitments, and third-party credit support to allocate the risk between the hardware, the user and the parties standing behind them.
The structures are beginning to show where the credit goes when the chip itself cannot do all the work.
Then there’s the refresh question.
What happens to a five-year-old accelerator designed around one hyperscaler's stack? Who carries the remaining value? Who funds its replacement? And what must a ten-year capacity agreement say about a hardware cycle that may be much shorter?
Those are no longer just engineering questions. They are financing questions.
The move from merchant GPUs towards proprietary compute reallocates residual risk rather than removing it.
And somebody must be paid to hold it.
That, to us, is the interesting part of the custom-silicon story. Not whether the technology is better or worse, but who ultimately carries the residual risk when increasingly specialised compute sits inside increasingly expensive campuses.
The two term sheets aren’t inconsistent.
They may be financing two different asset classes wearing the same hard hat.
The campus hasn’t changed.
The collateral has.
That’s why land, power, offtake, tenancy, technology, and the capital stack increasingly need to be structured as one problem.
Best to know that before the term sheet arrives.
Authors: Jane Rueger, Partner, Co-Lead, Data Center Infrastructure; Michael Herrera, Partner, Co-Lead, Data Center Infrastructure; Jamie Ng, Global Clients & Markets Partner.
