Nvidia has formalized a program in which it acts as a financial backstop for the cloud companies buying its AI chips, guaranteeing them a minimum revenue floor in exchange for a cut of their sales above that floor. The arrangement lets smaller 'neocloud' firms secure the bank financing they otherwise couldn't - and analysts at SemiAnalysis say it helps set the stage for an AI debt market exceeding $7 trillion by 2029, second only to US mortgages.
Key facts
- The backstop guarantees a revenue floor (Nvidia rents unsold GPUs back at a fixed rate) plus a revenue share above the floor, over a roughly six-year contract.
- SemiAnalysis's TCO model projects over $7 trillion of AI debt outstanding by 2029 - the second-largest asset-backed debt market after US mortgages (~$13 trillion).
- Announced July 1, 2026 by Nvidia CFO Colette Kress in an official blog post; analyzed in depth by SemiAnalysis and DatacenterDynamics.
- First named adopters: Sharon AI (up to 40,000 GB300 GPUs, a reported $4.88B deal) and Firmus (a 360MW site in Indonesia, up to 170,000 GPUs).
To understand why this is a big deal, start with the problem it solves. Building an AI data center requires assembling three things at once, in what SemiAnalysis calls the 'AI Project Trinity': capital, offtake (a customer contract for the compute), and the data center itself. Each depends on the others - lenders won't provide debt without a guaranteed customer, you can't win a customer without capital to place deposits, and you can't raise that capital without lenders lined up. For a small cloud startup with no track record, it is a chicken-and-egg trap.
Nvidia's backstop breaks the trap using its own balance sheet. In Nvidia's own words from the announcement, the model lets AI clouds buy Nvidia infrastructure 'through economic alignment with a revenue-sharing and credit-support model,' with Nvidia 'earning both standard product revenue and a share of the cloud revenue on the supported capacity.' Concretely: Nvidia guarantees the neocloud a floor of revenue - if the company can't rent out its GPUs, Nvidia commits to renting them back at a pre-agreed rate. That guarantee, backed by Nvidia's AA credit rating, is enough for banks to lend against. In exchange, Nvidia takes a percentage - modeled at around 40% - of revenue earned above the floor, working out to roughly an 18% average take over six years.
Think of it like a landlord who both sells you the building and promises to rent it back from you if your tenants don't show up - then takes a cut of your rent when they do. The developer can now get a mortgage, because the risky part is covered.
The scale is what makes it a headline. SemiAnalysis's proprietary cost model projects that AI debt financing 'will become a multi-trillion-dollar credit market, with over $7T of debt outstanding by 2029,' making it the second-largest asset-backed debt market after US mortgage financing. (That $7 trillion figure is SemiAnalysis's own estimate, not Nvidia's - an important attribution.) The firm projects annual AI capital spending will top $2 trillion in 2028 and cumulative spending near $11 trillion across 2024 to 2029, mostly funded by credit. The first named adopters are Sharon AI, deploying up to 40,000 of Nvidia's Grace Blackwell GB300 chips, and Firmus, building a 360-megawatt campus in Batam, Indonesia.
Why Nvidia would take on this risk comes down to customer diversification. A handful of giants - Amazon, Microsoft, Google, Meta, Oracle - buy most of Nvidia's chips, and nearly all of them are designing in-house silicon to eventually replace those purchases. Every dollar of demand Nvidia can shift toward independent neoclouds is a dollar less exposed to a customer who is also a future competitor. SemiAnalysis calls Nvidia the emerging 'Central Bank of AI,' supplying liquidity the banking system won't.
The honest caveat is the circularity, and it is the sharpest critique. As The State of AI put it, 'a chipmaker is writing put options' - Nvidia sells the chips, guarantees the revenue those chips produce, sometimes takes equity, and now collects a share of the buyer's sales. In isolation each deal is manageable, but in aggregate Nvidia is accumulating something that looks like a lender's loss reserves without a lender's regulatory capital requirements. If AI compute demand cools, Nvidia is on the hook for a growing web of guarantees on capacity nobody wants. It is the supply-side mirror of the same tension in the GLM-5.2 margin-collapse debate: is the AI economy's growth organic, or is it partly Nvidia's balance sheet reflected back at itself?
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)