Nvidia Is Now Financing Both Sides of the Compute Market

Nvidia is now financing both sides of the compute market it sells into. On July 27, 2026 the company announced a long-term strategic partnership with Ilya Sutskever’s Safe Superintelligence — an equity investment plus access to Nvidia’s next-generation Vera Rubin platform. Hours later, CNBC confirmed reporting that Nvidia is in talks to guarantee roughly $250 billion of financing so OpenAI can lease a 10-gigawatt data-center campus in southern Ohio.

Read separately, these are two ordinary news stories: a chip company backing a research lab, and a chip company helping its largest customer raise debt. Read together on the same day, they describe a single mechanism — a supplier extending its own balance sheet to guarantee demand for its own product. That mechanism has a name in the market commentary that followed: circular financing. It also has a consequence that almost nobody is writing about, which is what it does to the compute prices the rest of us pay.

This guide separates what was actually confirmed from what was reported, lays out Nvidia’s own announced-versus-materialized track record in one table, and then does the part that matters for a business buying AI capacity: translating vendor-financing risk into the assumptions inside your multi-year budget.

The two July 27 items sit at opposite ends of the confidence scale, and it is worth being precise about which is which before drawing any conclusion from the pair.

The SSI partnership is a joint announcement. Nvidia published it on its own newsroom, SSI is named as a co-announcer, and both CEOs are quoted. The substance: Nvidia makes an equity investment in SSI, and SSI gets access to the Vera Rubin platform, which the release says will increase the lab’s compute by an order of magnitude. That last phrase is vendor language from the release rather than an independently measured figure.

The Ohio backstop is reported negotiation. The Wall Street Journal reported the talks first; CNBC then confirmed them with a person familiar with the discussions, describing a guarantee of up to roughly $250 billion that would let OpenAI lease a 10-gigawatt campus being developed by SB Energy, a SoftBank subsidiary, on a decommissioned uranium-enrichment site in Pike County. CNBC is explicit that terms are not finalized and there is no assurance a deal is completed. Nvidia declined to comment.

What makes them one story is the direction of the money. In both cases capital flows from the supplier of accelerators toward an organization whose defining cost is buying accelerators. It is the same shape we traced in the wider pattern of compute-financing talks across labs, now running through the chip vendor itself rather than a cloud intermediary.

Safe Superintelligence was founded in 2024 by Sutskever, who serves as CEO, and Daniel Levy, after roughly two years of research conducted largely in stealth. Nvidia’s release names Andreessen Horowitz, DST Global, Greenoaks and Sequoia Capital among the backers; TechCrunch’s coverage additionally lists Alphabet, Lightspeed Venture Partners and GV. The lab previously partnered with Google Cloud in April 2025 for research infrastructure, so Nvidia is not its first compute relationship.

Sutskever’s framing in the release is unusually plain about what the deal buys: “We have research that is worthy of scaling up, and having access to a big NVIDIA computer will let us do so.” Jensen Huang’s side of the release reaches back to AlexNet, casting the investment as a bet on the researcher rather than on a roadmap. Neither statement mentions money, because the release discloses none.

Set the reported $5 billion against roughly $7 billion raised across the company’s entire life and the check is on the same order as everything that came before it. For a lab with no revenue line, that is not a strategic dabble — it is the balance sheet of the compute supplier becoming a material part of the compute buyer’s cap table. The equity and the hardware access arrive in the same sentence of the same press release, which is precisely what makes the structure interesting rather than routine.

There is a benign reading, and it deserves airtime. Frontier research labs are capital-starved relative to their compute appetite; a vendor that can supply both capital and silicon compresses two fundraising cycles into one. Sutskever gets to scale research that he says is ready to scale, without spending a year raising against a product that does not exist yet. If the research works, everyone involved looks prescient.

The reported Ohio arrangement is structurally different from every Nvidia headline that came before it. Nvidia would not be buying equity or shipping credit. It would be guaranteeing someone else’s debt — specifically the lease and construction financing behind a campus that OpenAI would occupy. The chips inside the building are a separate financing conversation, per CNBC.

The reason a guarantee is needed at all is the most instructive detail in the whole story. OpenAI is an unprofitable private company without an investment-grade credit rating of its own, so the developer vehicle raising the debt cannot borrow at the rates a 10-gigawatt project needs. A guarantee from a company with Nvidia’s balance sheet fixes that — by moving the credit risk onto the chip vendor.

Scale check on the site itself: 10 gigawatts of data-center capacity is roughly the annual power consumption of 8 million U.S. households, per CNBC’s reading of Energy Information Administration data — about 800,000 households per gigawatt. Secondary, WSJ-sourced coverage describes a first phase of roughly 800 megawatts targeted to begin operating in 2028; that phasing detail did not appear in CNBC’s own confirmed text, so treat it as a reported target rather than a schedule. SoftBank and SB Energy are developing the campus in partnership with the U.S. Department of Energy, and there is thinner secondary reporting tying power allocation to a trade arrangement — worth noting, not worth planning around.

None of this is unprecedented in kind, only in size. OpenAI is simultaneously running other multibillion-dollar compute buildouts, and the same three parties have transacted before: SoftBank and OpenAI announced a $1 billion joint investment in SB Energy in January 2026.

Scaled that way, the reported guarantee is no more than about half the reported all-in cost of the campus, and the confirmed SSI investment barely registers. That asymmetry is the honest summary of the day: the deal we can verify is small, and the deal that would reshape the market is the one nobody has signed.

Vera Rubin is Nvidia’s current-generation platform, launched at CES 2026 and since described by Nvidia as being in full production. It is not a single chip: the platform comprises seven — the Vera CPU, the Rubin GPU, an NVLink 6 switch, a ConnectX-9 SuperNIC, a BlueField-4 DPU, a Spectrum-6 Ethernet switch, and an integrated Groq 3 LPU. Every performance number below comes from Nvidia’s own materials, and we have not found independent benchmarks for any of them.

The vendor claim worth holding onto is the token-cost one, because it is the mechanism by which any of this reaches an ordinary buyer. If Vera Rubin genuinely lowers the cost of serving a token by a large multiple, the financing story is a footnote — capacity gets cheaper, providers pass some of it through, and budgets improve. If the gains land nearer the modest end, the financing story is the main event, because then the only thing holding prices down is a buildout funded on terms that require the buildout to keep paying for itself.

That is the tension a business buyer should actually track: not whether Nvidia’s marketing numbers are true, but whether realized provider pricing moves in the direction the marketing numbers imply. Those two things have diverged before.

Coverage of the SSI investment and the Ohio guarantee exists everywhere as two separate news items. What does not exist anywhere we could find is the sequence — Nvidia’s recent compute-financing announcements laid out chronologically with what was announced next to what actually happened. That comparison is the single most useful input a skeptical buyer has, because it converts a headline number into a probability-weighted one.

The table below is our own assembly from reporting cited throughout this piece. It separates confirmed announcements from reported negotiations, because those are not the same evidentiary class and should never sit in the same mental bucket.

One row does the analytical work. In September 2025, Nvidia said it would invest up to $100 billion in OpenAI. What actually happened was a $30 billion contribution to a funding round six months later — 30 cents on the announced dollar. That is a single data point, not a discount rate, and it would be sloppy to apply it mechanically to the $250 billion figure. But it is the only Nvidia-to-OpenAI headline in this sequence whose outcome is known, and the outcome was smaller than the headline.

The forward-looking read: the correct way to hold the Ohio number over the next several quarters is as an upper bound on ambition, not as a committed line of credit. Guarantees of this size are negotiated against covenants, milestones, and phased draws; announcements compress all of that into one number. Expect the eventual structure — if it lands at all — to be smaller, more conditional, and staged against construction phases that run years past this article.

The critique is simple enough to fit in a sentence: a chipmaker funds a customer, the customer spends the funds on the chipmaker’s chips, and the chipmaker books revenue that partly originated on its own balance sheet. Industry commentators have put the running total of arrangements with this shape somewhere north of $800 billion across the sector — an estimate rather than an audited figure, and one worth treating loosely.

Michael Burry, who publicized his bet against subprime mortgages, posted his reaction the same day.

Note that Burry’s figure is $200 billion, not the roughly $250 billion reported by the Journal and confirmed by CNBC. We are quoting him as posted rather than correcting him, and the gap is a useful reminder of how quickly a reported negotiation becomes a differently-sized number in the retelling. Tech commentator Ed Zitron was blunter, calling the financing structure “such an insane thing to do on so many levels” and “very silly” from a financial-sustainability standpoint, and questioning where the underlying funding actually originates given SoftBank’s role as developer. Buy-side voices relayed on the Bloomberg wire were more measured but directionally similar — Global X Management’s Billy Leung noted that Nvidia guaranteeing more of OpenAI’s data-center debt deepens vendor financing already under scrutiny, while Allspring Global Investments’ Gary Tan framed investor concern about circularity as coexisting with confidence in the long-term buildout. Analysts at Wedbush and Bernstein were reported making similar circular-financing observations in the same coverage; we could not re-verify those remarks on a first-party page, so we are paraphrasing rather than quoting them.

Huang has rejected the framing directly.

He has a real point, and it is worth stating properly rather than waving away. Vendor financing is ordinary industrial practice — aircraft manufacturers, telecom-equipment vendors and heavy machinery makers have all financed customers for decades, precisely because the assets are long-lived, expensive, and hard to fund on a young buyer’s credit alone. A guarantee that unlocks a project’s wider capital stack is leverage on someone else’s money, not a round-trip of the vendor’s own revenue. And the amounts, large as they are, sit against far larger total raises.

The counter is not that vendor financing is illegitimate. It is that vendor financing compresses independent signals into one. When your chip supplier is also your model provider’s creditor, the demand number, the credit number and the capacity number stop being three separate pieces of evidence about the market and start being three views of one decision. That is a diligence problem before it is ever a solvency problem — and it is the same reason we argued that compute diversification away from a single vendor is a strategic posture rather than a procurement preference.

Almost every piece written about July 27 treats it as a stock story or a morality tale. Neither helps a company that has just built its 2027 plan on the assumption that AI inference keeps getting cheaper. Here is the translation.

Vendor financing does not lower your unit price. Nothing in either arrangement makes a GPU-hour cheaper. What it does is lower the borrowing cost for the entities building capacity, which supports more capacity arriving sooner. More capacity is what has historically pushed per-token prices down — so the mechanism can help you, but only indirectly, and only if the capacity actually gets built.

The tail risk is correlation, not catastrophe. The realistic downside for a mid-sized business is not that AI collapses. It is that your model provider, your cloud, and the silicon underneath them increasingly share exposure to the same financing structures. If credit conditions tighten, the effects do not arrive one vendor at a time — they arrive as simultaneous repricing, slower capacity provisioning, and less negotiating room across a set of suppliers you thought were independent.

Announced capacity is not contracted capacity. A 10-gigawatt campus with a reported first phase targeted for 2028 is not supply you can plan against in 2026 or 2027. If your roadmap assumes abundant cheap inference eighteen months out because of buildout headlines, you have imported a construction schedule and a financing negotiation into your forecast without noticing.

The practical hedge is unglamorous: keep your own pricing history. Realized provider rates are the only signal in this entire story that is directly measurable by you, which is why we maintain a running record of the numbers behind multi-year AI cost assumptions. When financing news moves and posted prices do not, that gap is the information. Building that measurement discipline into how a company adopts AI in the first place is the least glamorous and most durable part of our AI and digital transformation work.

None of these require a view on whether the Ohio guarantee gets signed. They are the moves that leave you better off in both outcomes, which is the only kind of hedge worth the effort.

1. Write down your current realized cost per unit. Not list price — what you actually spend per resolved ticket, per generated asset, per processed document. This is your baseline, and almost nobody has it recorded well enough to detect a 15% drift.
2. Run a second-provider evaluation before you need it. The cost of qualifying an alternative is much lower when it is a planned exercise than when it is a response to a price change. Keep the evaluation warm, even if you never switch.
3. Separate the pieces that are genuinely locked in. Fine-tunes, provider-specific tool schemas, and proprietary embeddings are the real switching costs. Inventory them and decide deliberately which ones are worth the lock-in.
4. Stress-test the budget at flat prices. If your 2027 plan assumes token costs fall by a given percentage, run the version where they do not move at all. If that version fails, the plan has a dependency it never named.
5. Track posted prices, not press releases. Financing headlines are upstream of your invoice by years. Published rate cards are upstream by weeks. Only one of them belongs in a forecast.

If the decision in front of you is rent-versus-own rather than provider-versus-provider, the arithmetic is different and mostly about utilization — we worked through the break-even in our GPU buy, rent, or cloud decision guide. And if you want the measurement layer built properly so cost drift is visible in a dashboard rather than a quarterly surprise, that is the kind of instrumentation our analytics engagements put in place first.