Who carries the compute risk

2026-11-05

The AI build-out has reached the stage where the equipment vendor is also, in part, the financier. The market leader in accelerators has said in public that next year's growth, which it has framed at roughly 70%, is a figure it is comfortable with, and that without supply limits the business could more than double. Demand, in its account, is not the constraint. Wafers and memory are.

To make sure the demand keeps arriving in a form its customers can pay for, the vendor has put three financing arrangements in the market. Each one shifts risk somewhere. It pays to know where.

The three arrangements

**Revenue-share agreements with specialist clouds.** For selected customers that lease out compute, the vendor backstops a base price for the capacity and takes a share of the upside if the capacity is leased above it. For the cloud, this converts an uncertain rental stream into something closer to a floor. For the vendor, it creates a recurring line on top of the hardware sale and a direct exposure to utilisation.
**A dedicated data-centre campus programme.** The vendor is sponsoring large sites directly rather than waiting for customers to build them, which brings forward capacity and brings the vendor's name into the project finance.
**A 500 billion dollar private-capital platform.** Announced with a group of the largest alternative managers, the platform is intended to fund AI infrastructure for the vendor's customers at a scale the bank market would not reach on its own. The managers raise the money; the vendor supplies the pipeline.

Add to that the customer mix the vendor now describes. Specialist AI clouds are more than half of the non-hyperscale business. The two largest frontier-model developers account for a fifth or more of end consumption, though they buy through the clouds, not directly. The vendor sells to the intermediary; the intermediary leases to the model builder; the model builder pays with revenue it is still growing into.

Why the circularity debate misses the point

The criticism is easy to state. A vendor that finances its customers to buy its products is booking revenue it helped pay for. The vendor's answer is that the arrangements are measured, capped and backed by the credit of the ultimate buyers of compute, and that model builders' margins are improving as each generation cuts the cost per token.

Both can be true. The relevant question for a lender is not whether the structure is circular but which link in the chain breaks first if utilisation disappoints, and who has recourse to what.

The specialist cloud owns the equipment and the lease. If rental prices fall, the revenue-share floor protects the cloud on the covered capacity, and the vendor absorbs the difference. On uncovered capacity the cloud is exposed, and so is anyone who financed that cloud's equipment.
The model builder has a contract for capacity and a business that must generate the cash to pay for it. Its credit is the final backstop for everything above it, and it is the newest and least tested credit in the stack.
The alternative managers on the platform are raising long-dated money against assets with two-to-three-year useful lives and one dominant supplier. Their return depends on the equipment being re-leased, upgraded or sold on, and the vendor controls the cadence of all three.

What the supply picture adds

Because the vendor is supply-limited, allocation is a lever. Capacity goes first to the customers and the financing structures the vendor prefers. A specialist cloud that sits inside a revenue-share agreement and a platform-funded campus is likely to receive hardware before one that does not. That is good for the insiders and makes the outsiders, often the borrowers approaching private credit on their own, a different and weaker credit.

Our read

The build-out is now a layered financing chain with the equipment vendor at the top of it, supplying hardware, floors and deal flow. The chain works while compute stays scarce and rental prices hold. It is the uncovered middle, the clouds and builders outside the vendor's arrangements, where the leverage is highest and the backstop thinnest.

We spend our time on access to the private companies in this chain and on how the structures around them are built. The test we apply to any data-centre paper is simple: find the floor, find who stands behind it, and ask what happens on the day the vendor is no longer supply-constrained.