A $105B Guarantee Is Not a $105B Investment

Nvidia’s $105 billion commitment to OpenAI’s Ohio data center is a financing guarantee, not an investment — and reading the actual SEC filing instead of the headlines changes the story completely. On August 17, 2026, Nvidia filed an 8-K describing a set of “residual value guaranties” backing leases at a campus in Pike County, Ohio. The $105 billion is a cumulative cap on a conditional obligation, not a sum anyone is spending.

Why this matters beyond pedantry: the difference between committed capital and a contingent guarantee is the difference between money leaving a balance sheet and a promise that may never be called. Coverage that writes “Nvidia invests $105 billion in OpenAI” gets the instrument wrong, and most coverage also misses the more interesting fact — the number has been shrinking through negotiation, not escalating. A reported ~$250 billion backstop in late-July talks became a sub-$120 billion figure a week before filing, then $105 billion in the document itself.

This guide reads the filing so you don’t have to: what the instrument actually is, the three-step shrinking trajectory, the trigger and remedy mechanics buried in the 8-K’s defined terms, who actually pays for the campus, and a reusable three-question rubric for reading any splashy AI-infrastructure dollar figure — because this will not be the last one.

Start with the legal instrument, because everything else follows from it. Nvidia’s Form 8-K, filed August 17, 2026, describes a set of “residual value guaranties” — the filing’s own plural spelling — that Nvidia entered with SB Energy as lessor, relating to leases for approximately 4.25 gigawatts of IT load at the Ohio site. Not a loan. Not an equity stake in OpenAI. Not a credit line that gets drawn down as construction proceeds. A guaranty of the residual value of leases that OpenAI, the tenant, is expected to pay for out of its own operations.

The filing is unusually quotable on the two points that matter most: the cap, and the conditionality.

Two phrases carry the whole story. “Cumulatively capped” means $105 billion is the most Nvidia could ever be obligated for under the initial commitment — a ceiling, not a budget. “Subject to specified conditions” means even that ceiling only becomes a live obligation as the site clears ready-for-service milestones, which the filing expects to begin in 2028. Beyond the initial 4.25 GW, the filing gives Nvidia the ability to provide credit support for roughly 3.75 gigawatts more — “exercisable in its sole discretion,” so an option, not a commitment. (The filing’s legal body text rounds the option to approximately 3.8 GW while its own press-release exhibit says 3.75 GW; we use 3.75, the figure in the exhibit, in OpenAI’s announcement, and in press reporting.)

The distinction is not academic. When Nvidia contributed $30 billion to OpenAI’s March 2026 funding round, cash moved. Here, in the base case — OpenAI pays its rent for twenty years — Nvidia’s guaranty expires without a dollar changing hands under it. The instrument’s economic function is to make SB Energy’s construction debt financeable at scale, by putting an investment-grade balance sheet behind the tenant’s lease obligations. That is genuinely valuable. It is also genuinely not an investment of $105 billion.

Here is the part almost nobody foregrounds, and some coverage gets backwards. The figure did not balloon through negotiation the way AI-infrastructure numbers are assumed to. It contracted — sharply — across roughly three weeks, and the campus it covers shrank with it.

On July 27, CNBC reported — crediting the Wall Street Journal as first to the story — that Nvidia and OpenAI were in discussions about a backstop of up to $250 billion for a planned 10-gigawatt campus. Around the week of August 10–14, per CNBC’s August 17 account of the Journal’s follow-up reporting, Nvidia was set to cut that guarantee to less than $120 billion. And the 8-K, filed August 17, landed at $105 billion for an 8-gigawatt campus: 4.25 GW committed, roughly 3.75 GW optional.

Run the arithmetic and the filed ceiling sits roughly 58% below the opening reported figure — a majority haircut in about three weeks. Two caveats keep this honest. First, the $250 billion and the sub-$120 billion figures were reported from talks, sourced to people familiar with the discussions, not to filings — only the $105 billion is a legal document. Second, the scope changed alongside the number: a 10 GW plan became an 8 GW plan, so part of the reduction is a smaller campus, not just a harder bargain. The same July 27 CNBC report carried an estimate that the full 10 GW campus could cost more than $500 billion all-in — a reminder that guarantee size and total project cost were never the same number at any point in this saga.

The direction still matters. A figure that shrinks 58% between leak and filing tells you the leak was a negotiating position, not a plan. That is the single most useful calibration this story offers for reading the next one — and there will be a next one.

The coverage we checked — CNBC’s reporting and Nvidia’s own press release — quotes the $105 billion figure without reproducing the filing’s defined terms at this granularity. The filing’s actual mechanics — the trigger, the remedies, the exits — are where the “not an investment” thesis stops being an assertion and becomes arithmetic. Here they are, translated from Item 1.01.

Read as a whole, the anatomy describes a credit enhancement with layered escape hatches, not a capital deployment. The honest skeptical read cuts the other way, though: the reimbursement clause is only as good as OpenAI’s solvency in exactly the scenario where it gets invoked, and a guaranty that lets a not-yet-rated tenant anchor a multi-hundred-billion-dollar buildout is precisely the kind of interlocking exposure that has drawn “circular financing” concern — CNBC’s own coverage flags the pattern of chipmakers, labs, and financiers increasingly transacting with and guaranteeing each other. Both readings are true at once. Neither makes $105 billion a sum being spent.

If Nvidia isn’t spending $105 billion, who pays for the buildout? The filing and the parties’ own announcements lay out three structurally different instruments — plus the tenant’s rent, which is where the real money comes from. Keeping them separate is the difference between describing this deal and misstating it.

One more commitment rounds out the structure: OpenAI agreed to host Nvidia AI compute exclusively at the site, subject to limited exceptions. That exclusivity is a contractual commitment given in exchange for the guaranteed capacity — not a cash outlay by anyone, but very much part of what Nvidia is buying with its balance sheet. The deal sits under the broader Stargate buildout, and the advisor roster signals its real nature: Goldman Sachs and JPMorgan advised SB Energy, Morgan Stanley advised Nvidia — a project-finance cast, not a product-purchase one. For a framework on modeling what commitments like these actually cost the party making them, see our guide to how CFOs should model AI infrastructure payback.

The physical project is the PORTS-Pike Technology Campus in Pike County, Ohio, built on the decommissioned Portsmouth Gaseous Diffusion Plant — a former uranium-enrichment site spanning private land and federal land controlled by the U.S. Department of Energy. SB Energy will build, own, and operate the data center under a 20-year lease to OpenAI, with phasing expected to begin in 2028. Per OpenAI’s own announcement, the first 800 megawatts are expected to come online in 2028, largely using existing AEP infrastructure — an OpenAI-stated figure that Nvidia’s filing does not break out, and a usefully modest near-term number against the eventual 8 GW ceiling.

The local-impact figures deserve the same labeling discipline. OpenAI says the project is expected to create 35,000 construction jobs over a six-year buildout through 2032 and 2,500 long-term operating jobs — vendor-stated projections, repeated by press coverage but not independently audited, and worth treating exactly that way. On community benefits, OpenAI is contributing $40 million to a community grant fund on top of SB Energy’s earlier, separately announced $40 million fund, plus up to $84 million in Codex and ChatGPT credits for roughly 844,000 eligible Ohio college and technical-school students in the 2026–2027 academic year — a package OpenAI frames as more than $160 million in benefits for Pike County and Ohio residents.

The Ohio guaranty is the second Nvidia–OpenAI headline number in under a year to end up materially smaller than its opening framing. In September 2025, Nvidia announced it would invest up to $100 billion in OpenAI as part of a strategic partnership tied to 10 GW of Nvidia systems. CNBC’s July 27 report states plainly that the investment never materialized — Nvidia instead contributed $30 billion to OpenAI’s March 2026 funding round, and Jensen Huang said in March 2026 that it might be the last time Nvidia invests in OpenAI before its IPO. That is a separate, earlier transaction — its math should never be combined with the Ohio guaranty — but as pattern evidence it is direct: with this pair of companies, the opening number has twice been the high-water mark, not the floor.

The structure is part of a broader shift, too. CNBC’s August 17 coverage notes that just the prior week, Nvidia teamed up with six large asset managers to build financing platforms deploying $500 billion in third-party capital for data center projects. The AI buildout is increasingly financed through guarantees, special vehicles, and other people’s money rather than direct corporate cash — which is exactly why the reading skill in the next section matters. Even the early $250 billion talks figure explicitly excluded the cost of the Nvidia chips inside the campus, which the companies were discussing separately: none of the headline figures in this saga were ever meant to represent total project spend.

Brockman’s framing explains why deals like this exist: if compute is the scarce input, then locking twenty years of guaranteed capacity — with someone else’s balance sheet absorbing the financing risk — is worth real contractual concessions, like hosting Nvidia compute exclusively. And it explains why the skeptical read exists too: when the scarce input’s dominant vendor guarantees the leases of its own largest customer, the market concern CNBC’s coverage surfaces — circular financing, in the phrase Jim Cramer has publicly used with a dot-com comparison — stops being abstract. Our take: the concern is legitimate as a portfolio-level question about the AI buildout, and this specific filing is nonetheless a more conservative instrument than its headlines suggest. Both things are worth holding at once.

The transferable skill from this filing is a three-question rubric that works on any splashy AI-infrastructure dollar figure. It takes about two minutes against a primary source and reliably separates what a headline says from what a document binds.

Apply the rubric habitually and the genre gets easier to read. The same skepticism applies to AI valuation leaks, where the number that reaches the press is also a negotiating artifact, and to soft commitments dressed up as hard numbers in vendor price schedules. On the buying side, the same three questions belong in the questions to ask before signing any AI infrastructure deal, and in the broader discipline of vendor financing structures worth understanding before you sign. None of this requires a law degree — the Ohio 8-K’s operative language runs a few paragraphs, and every fact in this post about the instrument itself came from reading them.

Looking forward, expect more of these, not fewer. A buildout financed through guarantees and third-party capital generates a steady stream of headline figures that are ceilings, options, and backstops — each one reported as if it were spend. The organizations that read the instrument correctly will price vendor risk, negotiate compute contracts, and brief their boards better than the ones that read the headline. If your team is making AI infrastructure or platform commitments and wants that discipline applied to your own stack, our AI transformation engagements start with exactly this kind of primary-source diligence.