OpenAI announced its own secret IPO filing. The announcement is the document worth reading.

There are two ways to file confidential IPO paperwork with the Securities and Exchange Commission. The first is the one the mechanism was built for: you submit your draft S-1 in private, you work through the SEC's comments away from the headlines, and the world finds out when you are ready for it to find out. The second is the one OpenAI chose this week. You submit the draft in private, and then you publish a statement telling everyone you have done it — in OpenAI's case with an explanation that deserves to be quoted in full: "We expect it to leak so we're just announcing it."

That sentence is doing more work than anything in the filing. A confidential submission you announce is not confidential. It is a press release with a regulator attached. And a company that announces a secret filing while insisting it has not decided when, or whether, to actually list is not making a disclosure. It is making a move.

The facts, first. OpenAI confirmed on Monday US time that it has submitted a confidential draft S-1 — the registration document that precedes a public listing — to the SEC. Goldman Sachs and Morgan Stanley are reported to be leading the underwriting, with JPMorgan also said to be involved. Reports point to a possible listing window between September and November of this year, with September cited as the earliest plausible target. The company's own statement was careful to commit to none of it: an IPO "may be a while," because "there are things we want to do that are likely easier as a private company."

The sentence doing the work

Hold the two halves of that statement next to each other. We have filed the paperwork to go public. Going public may be a while. Both can be true — a draft S-1 can sit at the SEC for months — but a company does not engage three bulge-bracket banks and start the regulatory clock as a hobby. The filing is an action. The hedge is an announcement. When the action and the announcement disagree, the action is the one that cost something.

The chief financial officer's version is more candid, as CFO statements often are. Sarah Friar has spent the spring saying "I want us to be ready," and observing that public markets are "much bigger than private markets." Translated from the dialect: the company believes it will need more money than private investors can comfortably supply, and it wants the paperwork done before the need becomes visible. "Readiness" is what you call an IPO when you do not want to admit you are in the queue.

Why announce it now, then, if the timing is genuinely unsettled? Because of the calendar everyone at OpenAI can read as well as you can. Anthropic disclosed its own confidential filing on June 1, at a reported valuation of $965 billion. SpaceX priced its offering last week and begins trading on Nasdaq this Friday at roughly $1.75 trillion, in what is set to be the largest raise in IPO history. The window for a generational AI listing is being defined right now, in public, by other companies. "We expect it to leak" is the official reason for the announcement. The unofficial reason is that in a week like this one, silence reads as absence.

A confidential filing you announce is not confidential. It is a press release with a regulator attached.

What the S-1 will have to say out loud

The strategic value of a confidential filing is control over the order of revelation. You get to choose the moment the numbers meet the public, and you get to choose the narrative that arrives with them. It is worth being precise about which numbers OpenAI has been happy to circulate, and which ones have only ever reached the public by leaking.

The circulated numbers are the revenue ramp, and the ramp is genuinely remarkable: roughly $2 billion in 2023, around $6 billion in 2024, about $20 billion in 2025, with the annualized run rate reported to have passed $25 billion early this year. Some 900 million people are reported to use ChatGPT weekly. The last private round, in March, valued the company at $852 billion — a figure that would place it among the fifteen or so largest companies in the S&P 500 the day it lists.

The leaked numbers are the other column. OpenAI's cash burn for 2026 has been reported at roughly $25 billion, with accounting losses reported to run higher still. Reported gross margins have moved the wrong way — to around a third, down from roughly 40% a year earlier — because the cost of actually answering all those queries is growing nearly as fast as the revenue from them. Inference costs, reported at $8.4 billion last year, are projected by outside analysts to rise by two-thirds this year. None of these figures has been confirmed by the company, and that is rather the point. An S-1, once public, is the end of the era in which they do not have to be. A registration statement is audited, GAAP, and signed. It is the one document in this story that will be under oath.

There is a useful precedent for what happens at that moment, and it is trading this week. SpaceX's filing forced into daylight a set of numbers — $18.7 billion in revenue, a $4.9 billion net loss, a profitable Starlink carrying an unprofitable everything else — that years of friendly private-round coverage had left comfortably vague. The company survived the disclosure; the mythology did not, entirely. OpenAI's filing will perform the same conversion on "the fastest-growing product in history," and the company knows precisely what is in the envelope. The confident announcement and the careful hedging are both downstream of that knowledge.

The renegotiation that tells you the direction of travel

One reported detail deserves more attention than it has received, because it is the kind of unglamorous term that says what the glossy language will not. OpenAI's revenue-sharing arrangement with Microsoft — once on a trajectory worth well over $100 billion to Redmond through the end of the decade — is reported to have been renegotiated and capped at around $38 billion through 2030.

Read that as a document. A company swimming in margin does not spend negotiating capital to claw back a revenue share; it pays the toll and grows. A company that renegotiates the toll is telling you the toll had become material to its survival math. The cap is good news for OpenAI's income statement and precisely because it is good news, it is a concession about how much room the income statement had. The strategic partnership language survives. The economics underneath it were rewritten by someone who had done the arithmetic on the burn.

The same arithmetic explains the venue change. OpenAI has raised more private capital than any company in history — the March round was backed by Amazon, Nvidia and SoftBank, among others — and the reported forward need is still staggering. One bank's analysts have estimated the company could require more than $200 billion in additional capital by 2030 even under optimistic assumptions. There is no private pool of that depth left to draw on. Friar's "much bigger than private markets" is not a preference. It is a census.

It is also worth translating the company's stated reason for delay, because it is the most specific sentence in the announcement. "There are things we want to do that are likely easier as a private company." The things a company can do more easily in private are not mysteries; they are a short list. It can sign decade-long compute commitments that would look reckless beside quarterly earnings. It can settle litigation without the settlement becoming a line item analysts annualize. It can reprice its products experimentally, finish an unflattering reorganization, or absorb a bad quarter around a model launch without a sell-off attached. Every item on that list describes something OpenAI is reported to be doing or facing right now. "Easier as a private company" is not a reason to doubt the IPO is coming. It is a description of the housekeeping you do before the guests arrive.

The structure the bankers have to sell

What will actually be offered to public investors is one of the more novel corporate objects ever brought to market: a public benefit corporation still controlled by a nonprofit foundation, born of a restructuring that was litigated, regulated and editorialized about for the better part of two years. Elon Musk's suit against the conversion was dismissed in mid-May, clearing the legal runway; the New York Times copyright case and a cluster of authors' suits remain live, with settlement estimates that span an unhelpfully wide range. Sam Altman's own equity position has never been formally disclosed, and reporting on the March cap table suggests it remained unsettled then.

Every one of those items will become a risk factor with page numbers. So will the competitive picture, which is the quietest concession in the whole affair: OpenAI's share of the market it created is reported to have slipped from roughly 60% to 51% over the past year, while the rival that filed its S-1 a week earlier is the one whose coding products have become the developer default. One prominent investor has already described the implied valuation as "priced for a monopoly outcome that does not yet exist." That sentence will not appear in the prospectus. Some version of its underlying data will have to.

And consider who, mechanically, ends up holding the stock. At anything like the contemplated valuation, OpenAI enters the major indices within quarters of listing, at which point every index fund — which is to say, every retirement account — becomes a buyer regardless of its view on the burn rate. That is the practical meaning of Friar's observation about the size of public markets: not that public investors are more enthusiastic than private ones, but that a meaningful share of them are obligated. The deepest pool of capital on earth is deep precisely because much of it does not get to say no. A company that needs hundreds of billions of dollars, and would rather not renegotiate its story with sophisticated counterparties every round, has exactly one venue with buyers like that.

None of this means the offering fails. SpaceX's numbers were worse than its mythology and the book was covered anyway; scarcity is its own thesis, and there will be exactly one OpenAI. The point is narrower. The order of events — banks engaged, paperwork filed, announcement issued, commitment withheld — is not the order of a company ambivalent about going public. It is the order of a company that wants the option fully loaded and the narrative fully managed before the numbers leave its hands.

What the announcement actually announced

Strip the framing and this week's statement says three things. The first is that OpenAI's capital requirements have outgrown every private structure available to it, including the most generous ones ever constructed. The second is that the company would rather control the leak than suffer it — which tells you how many parties now hold information about its finances, and how confident it is in their discretion. The third is that the race to list is now part of the race itself: when your chief rival files on June 1 and the decade's biggest IPO prices the same week, "we may wait a while" is a sentence you issue from inside the starting blocks, not from the stands.

The company describes a filing made for flexibility, at a moment of strength, with no particular urgency. The actions describe a burn rate that requires the deepest capital market on earth, a partner whose economics had to be renegotiated, and a filing timed to the week the window opened. The announcement is confident. The sequence is not. One of them is telling the truth.