Nvidia, which makes the chips, has more than $170 billion at stake in the companies that rent those chips out (called neoclouds), according to a report by The Information. Most of that isn't cash. It's promises: if a data center tenant defaults, Nvidia covers part of the landlord's loss, and for some capacity that goes unsold, Nvidia agrees to buy it. In the same week, Nebius's CEO said some of the financing neoclouds rely on had temporarily dried up. Together, those two facts show who is carrying the risk in the AI buildout.

Update, October 29. A reader asked how the $164.5 billion splits between lease guarantees and capacity purchase promises. It splits unevenly, and the largest piece is neither of those as we first described it. Per The Information's analysis, as relayed by Yahoo Finance: about $105 billion is a guarantee tied to an Ohio data center campus, about $36 billion is a commitment to buy cloud capacity that providers can't sell to others, about $20 billion is data center leases Nvidia expects to pass through to customers, and about $3.5 billion guarantees leases signed by cloud providers.

Two of these are confirmed in Nvidia's own filings. Its August 17 8-K says Nvidia entered into residual value guaranties with SB Energy, the landlord, for leases to an OpenAI affiliate covering about 4.25 gigawatts, with Nvidia's aggregate payment obligation capped at $105 billion. Nvidia pays only if OpenAI's insolvency causes a lease default or OpenAI fails to pay under a lease, and then generally covers the shortfall between a guaranteed minimum value and what the landlord recovers through a new lease or a sale. OpenAI has agreed to reimburse Nvidia for amounts it pays. The guaranties take effect when each lease begins, with ready-for-service conditions expected from 2028, and end at the earliest of the 20th lease anniversary, the lease ending, or OpenAI reaching a satisfactory credit rating. The 10-Q confirms the smallest item: $3,529 million of land, power and shell guarantees, with partners holding $712 million in escrow. We haven't confirmed the $36 billion and $20 billion amounts in the filings. The 8-K also notes Nvidia may provide credit support for about 3.8 more gigawatts at its discretion.

So the lease-default guarantee we described above is a small part of the total. The largest piece is conditional, reimbursable by the tenant and not expected to start until 2028, and the pieces carry different risks.
How we sourced this. We couldn't open The Information's report. The figures below come from Michael Parekh's Oct 23 analysis of it and from Yahoo Finance's coverage of the Nebius CEO's remarks, so treat them as reported secondhand. Nvidia's own filings would be the primary source for its guarantees, and we haven't checked them.

What was reported

Per Parekh's summary of The Information's Oct 20 report, Nvidia had up to $164.5 billion in guarantees and other commitments for new data centers as of its July quarter, plus more than $7 billion invested or agreed to be invested directly in AI cloud providers. That's about $171.5 billion in total. The guarantees include agreements to pay landlords if tenants default and promises to buy capacity. Equity from Nvidia helps neoclouds cover GPU purchases and project costs that lenders won't fully finance.

Examples in that coverage: a $2 billion Nvidia equity investment in CoreWeave in January, after which Jensen Huang said CoreWeave wouldn't need a further backstop. And Firmus, which Nvidia agreed in June to backstop on a purchase of an estimated $15 billion of server racks with at least 170,000 GPUs. Firmus has since pulled an IPO that sought $5 billion at a $30 billion valuation and is trying to raise at least $2 billion privately. Nvidia's CFO said in August that the risk is limited and that Nvidia can redeploy chips to other customers.

The financing side

Nebius CEO Arkady Volozh said at Nvidia's GTC event in Berlin that some capital sources had “temporarily dried up,” including asset-backed financing and debt markets, and acknowledged neoclouds are finding it harder to raise debt. He said Nebius has a healthy balance sheet and enough funding for its planned capacity, and is working with customer advances and other instruments. Nebius announced a $3.5 billion debt facility from Citi the previous Thursday, and its COO said it has $40 billion of long-term customer contracts that could serve as collateral. Per Parekh, CoreWeave, Nebius and IREN shares fell 15 to 20 percent in one week and CoreWeave's bond yields rose above 15 percent, which matches what we've tracked in our neocloud debt piece.

Why it matters

This is SOI's view, not a claim from the sources. Non-hyperscaler customers made up 45 percent of Nvidia's July-quarter data center revenue, about $40 billion, per the same report. If a growing share of those buyers can only finance their GPUs because Nvidia guarantees the lease, then part of Nvidia's demand is supported by Nvidia's own balance sheet. That doesn't mean demand is fake. Neoclouds are signing real contracts with Microsoft, Meta and others. It does mean the risk of an unused cluster doesn't sit only with the neocloud. We described a similar structure in our look at the web of GPU-backed financings.

The cleanest test is what Anthropic's S-1 showed us: revenue per dollar of compute. As long as it stays high, guarantees are rarely called. We covered that in our Anthropic S-1 piece.

A counter-signal

Not all inference capital is tightening. On Oct 28, The AI Insider reported that General Compute, a neocloud built on non-Nvidia inference chips from Cerebras, SambaNova and Positron, raised a $100 million Series A at an $800 million post-money valuation. It reports a customer agreement worth $100 million in annual recurring revenue and 75 MW planned for 2027. That's one report on one company, but it shows money still moving toward inference-specific hardware.

What we're not saying. We aren't saying Nvidia's guarantees will be called, or that any neocloud is in trouble. The CEO of Nebius says it is funded, and Nvidia says its risk is limited. The point is narrower: the size of the exposure, and who holds it, is now public enough to track.

What this means, in plain English

Nvidia isn't only selling the chips. It's helping pay for the buildings they go in. If those data centers fill up, that's fine. If they don't, part of the bill lands on Nvidia. That's why we're watching whether financing loosens or keeps tightening.

What we're watching

Nvidia's own disclosures on guarantees at its next filing. Whether Firmus closes its private raise. Whether neocloud bond yields ease or keep climbing. And whether the inference-focused neoclouds keep raising while GPU-backed debt is hard to get.

Sources: Michael Parekh on The Information's report, Yahoo Finance on Nebius, The AI Insider on General Compute.