OpenAI's $665B Sits in a Footnote

Introduction

In April, OpenAI's chief financial officer told other company leaders she wasn't sure the company could pay for the computing contracts it had signed. Sarah Friar's concern, first reported by the Wall Street Journal and relayed by Reuters, was that OpenAI might not be able to pay for future computing contracts if revenue doesn't grow fast enough. Two months later, on the day OpenAI announced it had confidentially submitted IPO paperwork to the SEC, Sam Altman and Jakub Pachocki published a blog post promising "everyone on Earth a personal AGI." It runs 8,021 characters and never uses the words capex, financing, funding, or data center.

The commitments were already documented, in a wider scope than the cloud-server figure Friar cited internally. OpenAI's leaked confidential financials, reported by The Information, put roughly $665 billion of chip, energy, and data-center purchase commitments in the footnotes, next to $46 million of capital expenditures for the quarter. More than $3 billion of the $122 billion round that closed in March came from ordinary people, raised from individual investors by JPMorgan, Morgan Stanley, and Goldman Sachs at three times the $1 billion target. Two of those banks are now working with OpenAI on a draft IPO prospectus. If you bought in, or you're waiting to buy when OpenAI lists, OpenAI's own written risk factors, whatever they say about funding $665 billion in commitments, don't reach you until 15 days before a roadshow.

None of that breaks a rule. OpenAI is private, owes the SEC nothing under the Exchange Act, and long-term purchase commitments belong in footnotes under standard accounting. The gap is the story: the CFO's doubt and the company's confidence shared a calendar, and the document that would settle which one was right sits with the SEC until OpenAI schedules a roadshow. It has pushed the listing itself to 2027.

From a Sidelined CFO to "Phase Three"

Friar's April position was specific. She didn't think OpenAI would be ready for a 2026 listing, per reporting from The Information summarized by Implicator.ai, citing more than $600 billion committed over five years to cloud servers against monthly revenue of about $2 billion. Altman then left her out of a high-level meeting with a major investor about server procurement, an absence The Information called notable and awkward given an earlier conversation on the same subject had included her. Since August 2025, per the same Information reporting (relayed by Fortune), OpenAI's CFO hasn't reported to the CEO but to Fidji Simo, who ran the applications business.

The Journal's story landed April 27 and the market read it as information. Oracle fell 3.4%, Arm 6.3%, and CoreWeave 2.8% in the same session, per Reuters. Altman and Friar answered jointly: "This is ridiculous. We are totally aligned on buying as much compute as we can." Spokesperson Steve Sharpe called the report prime clickbait. At an August 19 all-hands, Friar told employees the company will be public in 2027, sooner if the business "continues to inflect."

$46 Million of Capex, $665 Billion of Commitments

The Information reviewed OpenAI's confidential draft and reported the underlying figures on June 22 (walkthrough via TheNextWeb). As of March 31, the balance sheet carries virtually no debt and lease liabilities under $750 million. Capital expenditures for the quarter: $46 million. The purchase commitments for chips, energy, and data-center capacity, covering deals with Microsoft, Oracle, Amazon, and Stargate, come to roughly $665 billion, sitting in the notes rather than on the balance sheet. That's GAAP working correctly, not concealment, and all of it is unaudited.

One number stopped me here, and it's the one most likely to be misread. That $46 million looks small because the company rents nearly all of its compute rather than owning it, so the spending that matters runs through operating expense and contracted cloud services, where capex accounting never sees it. TIME reports OpenAI expects to spend $50 billion on compute in 2026 alone.

The rest of the quarter doesn't describe distress: $5.7 billion of revenue, $3.7 billion of cash burn, a $9.3 billion operating loss, gross margin up from 33% to 39%, and more than $73 billion in cash and securities. OpenAI lost $2.37 per revenue dollar in 2024 and $1.60 in 2025, per Fortune's summary of the leaked full-year figures, then $1.63 in the first quarter of 2026 on the same operating basis, flattening right where Friar was asking her question.

The Blog Post That Never Mentions Money

Read the June 8 post start to finish and the omission looks structural. Altman and Pachocki describe a "third phase" after research and product deployment, built on three goals: an automated AI researcher, accelerating the economy, and a personal AGI for everyone on Earth. It carries one dated, falsifiable commitment, and even that one is hedged: "by March of 2028 we may have a significant fraction of our research being done by AI systems in tandem with our own researchers." The other two goals arrive with no metric and no date.

The charge here is narrow: what the June 8 post leaves out, not what OpenAI has said elsewhere about money. Altman told TIME on August 26 that he thinks OpenAI is "going to be able to use all of the compute very profitably that we are planning to build." Sachin Katti, who runs OpenAI's compute, told the same reporters the company is "still short of compute." Altman announced the $1.4 trillion, 30-gigawatt buildout himself in November 2025. (TIME discloses a licensing and technology agreement with OpenAI; Salesforce, where TIME's owner is CEO, invests in Anthropic.) The document OpenAI published on the day it told the world it had filed still says nothing about what any of it costs.

Fifteen Days Before a Roadshow Nobody Has Scheduled

Confidential draft review is open to any issuer under a 2017 Division of Corporation Finance policy, and the company agrees in writing to the condition the SEC spells out in its own FAQ: it "must publicly file its registration statement, the initial nonpublic draft registration statement and all draft amendments thereto at least 15 days before it conducts its road show." Item 105 of Regulation S-K then requires disclosure of material risk factors, so if the funding question is material, it has to appear in the eventual public filing.

The clock doesn't start until the company sets a roadshow date, and OpenAI hasn't. A direct EDGAR query on August 27 returns no public S-1 and no CIK for OpenAI Group PBC. Friar told employees on August 19 that the IPO "is not a finish line, it is a milestone, another fundraise," and that the March round "gives us flexibility." Flexibility is doing a lot of work there, and it's exactly what confidential review provides. The delay itself has a documented explanation: the New York Times reported on June 25 that OpenAI's own bankers advised pushing to 2027 over tech-stock volatility, while Altman pushed advisers toward a $1 trillion valuation.

Meanwhile, the clearest picture of OpenAI's exposure this month came onto EDGAR from somebody else. Nvidia's August 17 Form 8-K (accession 0001045810-26-000069) discloses a $105 billion capped, conditional residual-value guaranty on OpenAI's Ohio data-center leases, with OpenAI as tenant and indemnitor, terminating once OpenAI reaches a satisfactory credit rating. That exposure reached the public record because Nvidia has a filing obligation and OpenAI doesn't.

Who Benefits

OpenAI and Altman get narrative control, which is cover rather than cash. Publishing "phase three" on filing-announcement day put mission language into the news cycle on a day when the numbers were with the SEC and nobody else. For as long as confidential review runs, OpenAI's balance sheet is a matter of leaks rather than a document the company signed, and that window's length is OpenAI's to set. OpenAI's own statement on the filing cuts the other way, saying confidential status "gives us the option to go public sooner."

For the banks, the benefit is a retail pipeline assembled before the disclosure exists. JPMorgan, Morgan Stanley, and Goldman Sachs ran the placement that pulled more than $3 billion out of individual investors, three times the $1 billion target, in what Friar described to CNBC as the largest private placement those banks had ever done. Goldman and Morgan Stanley are now on the draft prospectus, and Friar has publicly committed to reserving IPO shares for retail buyers. All of it was arranged before any of those Item 105 risk factors are public.

The fees don't explain it; the Times reporting has these bankers arguing for delay, not speed. SpaceX's June IPO paid its banks a 0.67% gross spread, $500 million split across the syndicate. The better-documented incentive is simpler: of every class of OpenAI investor, the individuals those three banks recruited are the ones who cannot read the company's written account of its own risk, and the ones least able to absorb a loss.

Why the Rebuttals Keep Arriving in Pairs

Altman and Friar issued joint alignment statements twice in three weeks in April: "completely aligned on compute strategy" after the April report, then "totally aligned on buying as much compute as we can" after the Journal's. Either they genuinely agree and keep saying so because reporters keep asking, or, as Implicator.ai argued, joint statements arrive precisely when alignment has already fractured. Simo stepped down in July, and COO Brad Lightcap ended an eight-year run in August, two days before Chief Revenue Officer Denise Dresser left after eight months on the job.

The accounting isn't unique to OpenAI. Anthropic's confidential filing reportedly carries a $45 billion Fluidstack commitment and a Broadcom-guaranteed $350 billion chip lease from Google, per the same Information reporting, and it has since passed OpenAI on annualized revenue and private valuation. The industry's capex risk keeps becoming public through a counterparty's 8-K or a leak of a confidential draft, almost never through the issuer's own words, because until a roadshow is scheduled, nothing has to be published.

The Bottom Line

The document that answers April's question already exists. OpenAI announced it in early June, Item 105 requires it to carry whatever funding risk OpenAI itself deems material, and the only people who've read it sit inside the SEC, inside OpenAI, or inside the banks drafting it. What the rest of us know about that balance sheet came through a leak, or through a CFO's private doubts reaching a reporter.

Anthropic submitted its own confidential draft on June 1, a week before OpenAI announced it had submitted one, and Friar told employees there's a chance Anthropic pulls the cover off in the coming weeks. If it does, the first full public accounting of an AI company's compute commitments inside a signed registration statement will belong to OpenAI's closest competitor. Which sets up a comparison OpenAI didn't choose: Anthropic's commitments in a signed filing, next to OpenAI's in a leak.

This story is developing. Details may change.