Nvidia's $100B OpenAI Commitment, One Year On: What Was Actually Wired

Nvidia announced on 22 September 2025 that it intended to invest up to $100 billion in OpenAI, released one gigawatt at a time. A year later, the Nvidia OpenAI investment that actually closed was $30 billion of equity in an ordinary funding round, with no gigawatt attached. The other $70 billion never became a contract. What replaced it is a $105 billion guarantee that pays nothing unless OpenAI defaults.

The short answer. Nvidia did not invest $100 billion in OpenAI. The September 2025 letter of intent never reached a definitive agreement. Nvidia instead committed $30 billion to OpenAI's March 2026 funding round, reportedly paid in instalments ending around 1 October 2026, and in August 2026 signed lease guarantees for OpenAI capped at $105 billion.

What the $100 billion Nvidia OpenAI letter actually promised

Start with the words, because the rest of the year turns on them. The joint announcement of 22 September 2025 described "a letter of intent for a landmark strategic partnership to deploy at least 10 gigawatts of NVIDIA systems". It added that "NVIDIA intends to invest up to $100 billion in OpenAI progressively as each gigawatt is deployed."

Three qualifiers sit in those two sentences. It was a letter of intent, not a contract. Nvidia "intends", which is not the same verb as "will". And the ceiling was "up to", released only as hardware went live, with the first gigawatt targeted for "the second half of 2026 on the NVIDIA Vera Rubin platform".

The release closed by saying both companies looked forward to "finalizing the details of this new phase of strategic partnership in the coming weeks." The coming weeks became the rest of the year. Markets priced the headline number anyway, and most coverage repeated "$100 billion" without the verb in front of it.

I do not think anyone at either company misled investors here. The document was precise. The failure was in how a conditional framework got read as a cheque, and that misreading is the recurring pattern in how the AI circular deals actually work.

What Nvidia's 10-Q and 10-K said, filing by filing

Press releases describe intent. Securities filings describe obligations, under legal liability for getting them wrong. So the cleanest way to see what the $100 billion became is to read what Nvidia told the SEC each quarter.

Quotations are from the filings' commitments notes, risk factors and MD&A. Our readers of the Q1 and Q2 FY2027 filings could not retrieve every Part II risk factor, so a passing mention there cannot be ruled out. The commitments and investment notes were read in full.

November 2025: a letter of intent, with no number

The first filing after the announcement was the 10-Q for the quarter ended 26 October 2025, filed 19 November. Its commitments note is the most revealing paragraph of the year. Investment commitments were $6.5 billion, including $5 billion in Intel subject to regulatory approval.

OpenAI appears one sentence later, and only as "a letter of intent with an opportunity to invest". The same note records that in November 2025 Nvidia agreed "to invest up to $10 billion in Anthropic", with a figure attached. One AI lab got a number in the commitments note. The other got a description of an opportunity.

Two weeks later CFO Colette Kress told a UBS conference, as reported by Fortune, "We still haven't completed a definitive agreement". That is a CFO choosing her words carefully, and they matched the filing exactly.

February 2026: finalizing, with no assurance

By the time of the annual report filed 25 February 2026, the language had moved. "We are finalizing an investment and partnership agreement with OpenAI," the 10-K says. The next sentence reads "There is no assurance that we will enter into an investment and partnership agreement with OpenAI or that a transaction will be completed."

Note what is missing. There is no $100 billion, no gigawatt schedule and no tranche structure anywhere in the passage. Jensen Huang had already said in Taipei on 1 February, per Fortune, "It was never a commitment," and that Nvidia would "invest one step at a time".

May and August 2026: the name leaves the investment notes

Once the funding round closed, OpenAI stopped appearing where you would look for it. The 10-Q for the quarter ended 26 April 2026 reports "$18.6 billion in private companies and infrastructure funds" invested in the quarter. It notes that some "include AI model makers that may indirectly purchase or use our products in the cloud." It does not name one of them.

The 10-Q for the quarter ended 26 July 2026 names OpenAI exactly once in the passages we could retrieve, as the customer under a lease guarantee. The equity stake that dominated a year of headlines is, in Nvidia's own reporting, one undisclosed share of an aggregate balance.

How much of the Nvidia OpenAI investment was actually wired

The equity that did move came through OpenAI's own funding round rather than through the gigawatt framework. OpenAI closed that round on 31 March 2026, raising $122 billion at an $852 billion post-money valuation. Bloomberg reported that Amazon agreed to put in $50 billion, with Nvidia and SoftBank at $30 billion each.

That $30 billion is the honest answer to "how much was wired", with one caveat about timing. It is 30% of the September headline. It bought shares in a priced round, alongside other investors on the same terms, which is a different instrument from a strategic tranche released per gigawatt of Nvidia systems.

The round was a valuation trade, not a deployment trade

Under the letter of intent, Nvidia's money would have followed its own hardware into service. Each gigawatt deployed would have released more capital, which tied the investment to the chip orders it was meant to support. The structure that closed has no such link disclosed anywhere in the filings.

That changes what Nvidia shareholders own. Under the old framework, they held a claim that grew as Nvidia sold more systems. Under the new one, they hold a minority stake whose value moves with OpenAI's next valuation. That valuation question is examined in the analysis of OpenAI's run-rate multiple.

The instalment question, and why it matters

"Committed" and "wired" are different words. Dealroom, summarising a 1 October 2026 report in The Information, said Nvidia and SoftBank "each funded the final $10B of their separate $30B pledges". If accurate, Nvidia's cash went out in parts across roughly six months, ending about one year after the original letter.

The filings are consistent with that, without proving it. Net additions to Nvidia's privately-held stakes were $31.0 billion across the first half of fiscal 2027. The Q2 10-Q also showed $25 billion of equity investment commitments still to fund, $18 billion of them before fiscal year end. A $10 billion payment in early October fits inside that $18 billion. So would several other deals.

I would treat $30 billion as committed and very likely paid in full by the end of September 2026. I would not quote an exact schedule of instalments, because the only source for it is a single secondary report of a paywalled story.

The $105 billion guarantee is not the $100 billion investment

In August 2026 a new number appeared that is easy to mistake for the old one. Reuters reported on 17 August that Nvidia would provide a guarantee of up to $105 billion so OpenAI could lease a campus being built by SoftBank's SB Energy in Pike County, Ohio. The figures are close. The instruments are opposites.

The Q2 10-Q gives the terms. Nvidia "entered into guarantees, capped at a total of $105 billion, to provide credit support" on behalf of an OpenAI affiliate, "related to leases for approximately 4.25 gigawatts of IT load". Each guarantee starts when its lease commences, across nine construction phases, "the first of which is expected in fiscal year 2029."

Payment is "triggered upon certain tenant defaults". Exposure falls across each phase's 20-year lease term. The guarantees end if OpenAI achieves "a satisfactory credit rating". In exchange, the site will "exclusively host NVIDIA AI infrastructure, subject to limited exceptions." Nvidia also holds an option to extend support to about 3.8 more gigawatts.

Read that as a credit enhancement sold for an exclusivity clause. Nvidia lends its balance sheet to OpenAI's landlord, and gets a guaranteed home for its chips. Reuters put the first 800 megawatts of the campus online in 2028. Nvidia separately said it would invest $1.5 billion in SB Energy, which is equity, and which is not part of the cap.

Nvidia equity stakes now rival the size of the original pledge

Step back from OpenAI and look at the whole portfolio, because that is where the money went. Nvidia's privately-held equity stakes stood at $3.4 billion on 26 January 2025. By 26 July 2026 they were $47.9 billion. Add the publicly-held positions and the equity-method stakes, and the total runs close to $99 billion on our arithmetic from the balance sheet.

That last figure is our sum, not one the filing states: $42.8 billion of marketable equity, $5.0 billion of longer-dated public stakes, and $51.2 billion of non-marketable securities. It is close to the $100 billion once promised to one company, spread instead across many.

Paper gains are now a material share of income

The stakes also pay, at least on paper. Gains from equity securities reached $23.7 billion in the first half of fiscal 2027, against $2.1 billion a year earlier. Much of that is unrealised, and $36.9 billion of the public positions sat under short-term lock-ups at quarter end.

This is where I part company with the bullish reading. A chip supplier whose other income depends on its customers' private valuations has taken on a second exposure to the same demand cycle. The precedent for that pattern is laid out in the comparison with Lucent and Nortel's vendor financing.

Announced versus disclosed: the Nvidia OpenAI deal gap stated plainly

Put the year in one table. The left column is what was announced. The right column is what the filings, or the most credible reporting where filings are silent, can support.

The gap is $70 billion of equity that never became an obligation. My reading is that this is not a broken promise, because there was no promise in the legal sense. It is a correction of how the market read the first announcement, and the correction took a full year to become visible in filings.

The last row matters as much as the first five. On gigawatts, the ambition has not shrunk as much as the cash has. Nvidia secured exclusivity over roughly 8 gigawatts of potential capacity at one site, and it did so without spending $100 billion to get it.

Where this reading is weakest

A post about the gap between announcements and filings owes you its own gaps. There are three.

Filings are a lagging and aggregated record

Nvidia reports private stakes in total, not by name. That means the $30 billion figure comes from Bloomberg's reporting of the round, not from Nvidia's own disclosure. If a filing ever breaks the stake out, it may show a different carrying value, and the gains attributed to it may already be large. We could not retrieve every risk factor page in the two most recent 10-Qs either.

The instalment timeline rests on one secondary report

The final $10 billion paid around 1 October is sourced to Dealroom summarising The Information, which cites a person with knowledge plus a SoftBank statement. It is plausible and consistent with the filings. It is still one chain of reporting, and six restatements of it would not make it six sources.

The smaller deal may be the better deal

The strongest case against framing this as a shortfall is about structure. On that view, $30 billion of equity plus a contingent guarantee may serve Nvidia better than $100 billion tied to its own sales. Gigawatt-linked investment was the textbook circular structure that drew the sharpest criticism. A priced stake on the same terms as Amazon and SoftBank is cleaner. I think the restructured version is the better one for Nvidia shareholders, which is an awkward conclusion for a post about what went missing.

What the Nvidia OpenAI deal status tells you about AI vendor commitments

Most of you will never sign a nine-figure letter of intent. You will read about dozens of them, and some will be cited in sales decks as proof that a vendor's ecosystem is safe. The same reading discipline applies at any scale.

First, find the verb. "Intends", "up to" and "opportunity" are not "will", "agreed" and "committed". Second, find the filing. A public company's commitments note is where a deal becomes an obligation, and if the figure is not there, it is not one yet. Third, find the instrument. Equity, guarantees and purchase commitments carry different risks, and adding them together produces a number that describes nothing.

The same three questions apply to any vendor partnership a supplier puts in front of you. A related pattern, where leases sit off the balance sheet until they commence, is covered in the breakdown of hyperscaler lease commitments still unrecognised. The Ohio guarantees are exactly that kind of obligation, waiting for a lease start date.

I would also watch who carries the residual risk. In Ohio it is Nvidia, not OpenAI's equity investors, if the tenant defaults. That is the question to ask of CoreWeave's debt stack and of any infrastructure partner your own business depends on.

Frequently asked questions

Where to start

Pick the one AI partnership your business leans on most, whether it is a model provider, a cloud credit programme or a reseller agreement. Find the public filing or the signed contract behind it, and write down the verb it uses. If the strongest word you can find is "intends", treat the partnership as a plan, not a fact, in your own risk register.

Then set one reminder for Nvidia's next 10-Q. Read Note 10, the commitments note, and nothing else. Whether OpenAI's lease guarantees grow toward the 3.8 gigawatt option will tell you more about the deal than any statement from either company.

Related reading