Guest Post by Bernie Margulies
Three kinds of buyers are absorbing the world’s AI GPUs: hyperscalers, neoclouds, and enterprises building their own clusters.
Much of the money behind these purchases is debt, and these GPU chips are increasingly financed as a separate asset class from the buildings that house them.
Analysts expect the four biggest hyperscalers to spend around $700 billion on capital projects in 2026, up from roughly $400 billion in 2025. A large share of that spending goes toward GPUs, the single biggest line in an AI data center’s budget.
How any given chip gets paid for depends on who ends up holding it. That is also what decides who takes the loss when a faster chip ships.
Hyperscalers pay for the chips themselves
Microsoft, Google, Amazon, and Meta pay for their GPUs the simplest way available, out of their own operating cash flow and corporate bonds.
What they push off their own balance sheet is the building. Meta’s $27 billion joint venture with Blue Owl, signed in October 2025, finances the Hyperion data center campus in Louisiana. The debt matures in 2049 and amortizes like real estate, because that is what it is. But the GPUs that go inside the data center are Meta’s own capital expenditure, bought separately.
Hyperscalers are also the only buyers who can swap NVIDIA out for their own silicon. Google runs its TPU line, now on the Ironwood generation. Amazon has deployed more than a million Trainium chips. Microsoft has Maia and Meta has MTIA.
Neoclouds borrow against the chips
Neoclouds are the GPU-focused cloud providers, CoreWeave, Lambda, Crusoe, Nscale, Fluidstack, and they do not have hyperscaler balance sheets. So they borrow against the chips themselves.
The structure is consistent. The GPUs are assigned to an SPV, a standalone company that owns nothing but the hardware and the contracts. A lender funds that vehicle and gets repaid from two sources: take-or-pay contracts, where the customer pays for reserved compute capacity whether it uses it or not, and the resale value of the chips if the borrower stops paying.
CoreWeave is the market’s template. It raised a $7.5 billion facility led by Blackstone and Magnetar in May 2024. In March 2026 it closed an $8.5 billion facility that Moody’s rated A3, the first GPU-backed loan to reach investment grade. That one is secured by CoreWeave’s clusters and a $19.2 billion contract backlog from Meta, and it runs to 2032.
The contracts are what make the chips bankable. CoreWeave signed OpenAI to $11.9 billion over five years in March 2025 and Meta to roughly $14 billion that September. Those same contracts carry the concentration risk, with Microsoft accounting for about two-thirds of CoreWeave’s 2025 revenue.
The rest of the field runs the same playbook at a smaller scale. Lambda raised a $500 million GPU-backed facility from Macquarie. Nscale closed a $1.4 billion term loan in early 2026. Fluidstack lined up a facility backed by roughly $6.7 billion of Google-contracted revenue. NVIDIA quietly de-risks its biggest customers too, holding about 11% of CoreWeave and agreeing to backstop up to $6.3 billion of its unsold capacity through 2032.
Here the residual, or the resale value of the GPUs, is the whole game. If a neocloud defaults, the lender’s recovery is whatever the used GPUs can fetch that quarter. Used H100s have been listing around 60% to 70% of their original price in 2025. It’s unknown whether these prices are sustainable, with many industry insiders expecting the resale market to crash in the next few years.
Enterprises buy on their own books
The third buyer is an organization that wants a cluster for itself: a bank, a drugmaker, a carmaker, a telecom. It buys the hardware for its own AI workloads.
Eli Lilly built an internal supercomputer it calls LillyPod, 1,016 Blackwell GPUs that went live in early 2026. Hyundai is standing up an AI factory of 50,000 Blackwell GPUs. These are corporate buyers funding the hardware from their own budgets.
When they do finance, the tools are ordinary corporate ones: cash, equipment loans, and operating leases that hand the residual back to the lessor. The server vendors run finance arms for exactly this, among them Dell Financial Services and HPE Financial Services.
A new tier now sits beside them, governments. Saudi Arabia’s HUMAIN has committed to buy up to 600,000 NVIDIA GPUs over three years, and India is funding a national GPU pool through its IndiaAI mission. NVIDIA’s sovereign customers spent close to $30 billion in its last fiscal year, paid out of state budgets and sovereign wealth.
For all of them the residual is a secondary concern. An operating lease moves it to the lessor, and an owner like Lilly treats the depreciation as a cost of doing research. They do not need the chip to hold its value, only to do the work.
Who is left holding the residual
The chips are the same in every one of these deals. What changes is who is exposed when NVIDIA’s next generation ships and makes the previous generation obsolete.
For a hyperscaler, their balance sheet can eat the loss. For an enterprise, it is likely the equipment financing firms they are working with. For a neocloud, it is the private credit lenders behind them.
Everybody is rushing to buy compute hardware and adopt AI. But a looming question remains unanswered: have lenders and equipment financiers priced for the day residuals fall, or hoped it would not come.
Author
Bernie Margulies is the founder and CEO of American Compute (amcompute.com), which helps GPU buyers get access to better GPU financing, and structures residual value insurance for lenders/lessors.
