OpenAI Is Filing to Go Public. The Math Has to Hold in Public Now.

On June 8, OpenAI published a three-paragraph statement announcing it had submitted a confidential S-1 to the SEC. Buried in the second paragraph: "We have not decided on timing yet; it may be a while because there are things we want to do that are likely easier as a private company."

That hedge is more interesting than the filing itself.

OpenAI is the most valued private technology company in the world. Its last private raise — $122 billion in March 2026 — put the post-money valuation at $852 billion. Goldman Sachs and Morgan Stanley are the joint underwriters. The September 2026 target would make OpenAI's IPO the second-largest public offering in US history by market cap, behind SpaceX's $1.75 trillion listing in June.

And yet Sam Altman's official announcement of the filing is a hedge. "Things we want to do that are easier as a private company." He didn't name them. He didn't say when the timing would resolve. He announced a filing and simultaneously reserved the right to delay it.

I don't know what those things are. Neither does anyone covering this story. They could be a major acquisition requiring public shareholder approval. They could be a government partnership that would invite scrutiny in a public company context. They could be a model release or licensing arrangement that doesn't survive daylight. The hedge is a tell that September is not a commitment — it's a window, contingent on conditions that Altman is not naming.

What the S-1 will force into public view

OpenAI has never disclosed its actual unit economics. It has confirmed $25 billion in annualized revenue as of early 2026 and $2 billion per month. It has not disclosed what it costs to generate that revenue at the model-by-model level.

The S-1 will require disclosure of all of the following, for the first time:

Gross margins by segment. OpenAI sells API access, ChatGPT consumer subscriptions, and enterprise plans. These have completely different cost structures. API calls against frontier models cost more to serve than subscription access that is rate-limited. If frontier API gross margins are 20% and consumer subscription gross margins are 60%, the blended figure tells investors nothing about the actual business. The S-1 must break this out.

Compute cost structure. OpenAI runs inference primarily on Microsoft Azure, which is a related-party transaction. The April 27 partnership renegotiation ended exclusivity and eliminated OpenAI's revenue share obligation to Microsoft — but did not eliminate OpenAI's dependency on Azure compute. The S-1 must disclose what OpenAI pays for compute, to whom, and on what terms. That number has never been public.

Sam Altman's personal equity stake. The PBC conversion from a capped-profit structure to a Public Benefit Corporation was required to make OpenAI IPO-eligible. In the process, Altman received an equity stake — the first time he has held equity in OpenAI. The original nonprofit structure was designed specifically to prevent any individual from extracting personal wealth from the mission. The S-1's equity table will disclose Altman's stake for the first time. It will also disclose how OpenAI's nonprofit foundation retains governance control over a company in which Altman now holds a financial interest.

Customer concentration. What percentage of $25B ARR comes from OpenAI's top 10 customers? If Microsoft, one customer or enterprise relationship, or a small number of government contracts represent more than 20% of revenue, that is a concentration risk. This has never been disclosed.

Microsoft post-renegotiation revenue. The April 27 renegotiation ended the exclusive arrangement and the revenue share. But OpenAI still has commercial relationships with Microsoft through Azure, GitHub Copilot, and other products. The S-1 will need to characterize what revenue, if any, OpenAI derives from the company that holds 26.79% of its equity and is now building competing models (SIGNAL-035).

The $115 billion math

OpenAI's own internal projections, as reported by multiple sources, show $14 billion in projected losses in 2026 and approximately $115 billion in cumulative losses through 2029 before reaching cash-flow positive in 2030. At $25B ARR and a $1T IPO valuation, the implied revenue multiple is 40x.

For comparison: Google IPO priced at approximately 7x ARR. Microsoft currently trades at 15x. Nvidia at peak was 30x. A 40x revenue multiple on a company losing more than a dollar for every dollar it earns is not a value investment. It is a bet that $25B ARR grows to $100B+ by 2029 — a 4x increase in 3 years — while burn rate stabilizes and the compute cost structure improves.

That bet might be correct. The $100B ARR trajectory extrapolates cleanly from OpenAI's documented growth rate. If AI API consumption grows as the model capability frontier advances and enterprise adoption deepens, the revenue trajectory is defensible.

What is not defensible is presenting this as a conventional tech IPO. The $14B annual loss, the 4x revenue growth required, the 2030 profitability target — these are venture capital underwriting criteria, not public equity underwriting criteria. Public market investors, who are the marginal buyer at the IPO price, have historically demanded different risk profiles than the institutional investors who have been funding OpenAI's private rounds.

The dual-IPO supply problem

One week before OpenAI's June 8 filing, Anthropic filed its S-1 on June 1, targeting an October 2026 Nasdaq listing at approximately $965 billion. Goldman Sachs is on both underwriting teams. Morgan Stanley is on OpenAI's.

Two AI companies at approximately $1 trillion each, filing within a week of each other, targeting listings within six weeks of each other. Combined market cap at issuance: approximately $2 trillion. The entire S&P 500 technology sector is roughly $15 trillion — two new $1 trillion AI companies in six weeks represents a 13% increase in institutional AI equity supply.

Institutional AI equity demand is finite. The largest sovereign wealth funds, pension funds, and mutual funds have AI allocation targets. Those allocations were sized for a market where OpenAI and Anthropic were private companies accessible only through secondary markets. When both go public simultaneously, every institutional investor must decide: do I buy both at full valuation, one and not the other, or wait for the secondary market to settle?

The banks handling both deals need both to succeed. If one prices well and the other struggles, the weaker deal's underwriters absorb reputational cost. The structured incentive is for Goldman and Morgan Stanley to coordinate the timing and valuation so that neither deal cannibalizes the other. Whether that coordination is achievable in a six-week window — when both companies have export-control risk factors, burn rates that require disclosure, and governance structures that will face public scrutiny for the first time — is an open question.

The risk factor neither S-1 can safely omit

The Bureau of Industry and Security suspended Anthropic's Claude Fable 5 and Mythos 5 globally on June 12 — the first use of export control authority against a consumer AI model API. The model has been offline for 13 days. The directive was served under ECRA § 4817(b)(1) and EAR § 744.22(b), statutes written for physical goods.

OpenAI's GPT-5.5-Cyber scored 85.6% on CyberGym — above the ~84.3% that Mythos 5 scored before its suspension. No BIS directive has arrived for OpenAI as of this writing (SIGNAL-033). But the legal authority that suspended Anthropic's models applies to OpenAI's models by the same statutory logic.

Any OpenAI S-1 that omits "US government may use export control authority to take our primary products offline without advance notice" from its risk factors is legally incomplete. Any investor who reads the S-1 without that risk factor is mispriced.

Fortune put it plainly after the Anthropic suspension: "The government can shut it down." That sentence applies to OpenAI with equal force. The September IPO window assumes no repeat. The BIS has given no indication it is done.

The "things we want to do that are likely easier as a private company" — I wonder if one of them is: respond to a BIS directive without filing an 8-K.