OpenAI Files Its Draft S-1, and the Interesting Number Is Not the One Everybody Is Talking About

OpenAI submitted a confidential draft registration statement to the SEC this week, with Goldman Sachs and Morgan Stanley advising, targeting a September listing at a valuation above $1 trillion.¹ This is, by any measure, a large IPO. It is also, on the disclosed numbers, a slightly strange one, and I want to walk through why.

The headline figure everyone is reaching for is the trillion. That is the wrong number to start with. The number to start with is minus one hundred and twenty-two.

What was actually filed (as far as we can tell). Almost nothing, publicly. The JOBS Act confidential filing mechanism allows OpenAI to begin the SEC review process without disclosing financials for 60 to 90 days; the public S-1 must land at least fifteen days before the roadshow.¹ So the earliest the market sees audited numbers is late July. Until then, the valuation conversation is being conducted on leaked figures, prior-round comparables, and prediction-market probabilities. The reporting we have says OpenAI's annualised revenue run rate sits somewhere between $12.7 and $14 billion, the Q1 non-GAAP operating margin is approximately -122%, and ChatGPT user growth has stalled.² The last priced private round, October 2024, set the company at $157 billion.³

The implied move from that round to the IPO target is roughly 6.4x in under two years, with no public revenue confirmation and a deeply negative operating margin. This is not, on its own, disqualifying, the comparables for large frontier-AI rounds have been generous, but it is the structural fact the prospectus will eventually have to argue against.

The inference economics frame, applied. The -122% margin is the diagnostic figure. It says OpenAI is spending roughly $2.22 to generate $1 of revenue at the operating line, before financing costs. Some portion of that is R&D and training, which can plausibly be characterised as investment. The rest is inference: the cost of actually running the models for paying users. Inference is the part that scales with usage, and inference is the part where ChatGPT's stalled growth becomes structurally interesting rather than merely disappointing.

If usage growth had continued compounding, the inference cost curve could be defended as a leading indicator of distribution: spend now to capture a market that will later monetise. If usage growth has stalled while inference costs continue to compound on the existing base (a base whose queries are getting longer, more multimodal, and more agentic), then the operating margin is not a J-curve waiting to invert. It is the steady state, and the question becomes whether compute unit costs can fall faster than per-user query complexity rises.

What this is a case of. Confidential filings ahead of large, loss-making technology IPOs are not new — WeWork, Uber, Snowflake all used the mechanism. The pattern is consistent: the confidential window is used to negotiate disclosure scope with the SEC, refine the comparable set, and shape the roadshow narrative before the financials become public. It does not, however, change the financials. The -122% operating margin will appear in the public S-1 in audited form, against PCAOB-reviewed cost of revenue line items. Whatever narrative is being built now has to survive contact with that document.

The Anthropic problem. Here is where the map gets odd. Anthropic's most recently reported private valuation is $900 billion and its claimed ARR is $45 billion.⁴ If both figures hold up to audit, and the ARR figure is, to be clear, a claim rather than a disclosure, Anthropic is currently a higher-revenue business than OpenAI at a slightly lower valuation, and is itself expected to pursue a public listing later in 2026.

This inverts the usual narrative. The category-leader-deserves-a-premium argument that typically supports a first-mover IPO valuation works best when the first mover is, in fact, leading on the metric public investors care about. If Anthropic's ARR figure is real, OpenAI is not listing as the largest frontier-AI revenue business; it is listing as the most recognisable one. Those are different valuation arguments, and they support different multiples.

Whoever lists first gets to set the multiple. Whoever lists second gets to say whether that multiple was right.

Coinbase's prediction market currently puts OpenAI's probability of listing first at 85% and Anthropic's at 28%.² These are sequencing odds, not pricing odds, and the coverage that has conflated them has done so badly. The interesting question is not who files first. The interesting question is what happens to Anthropic's eventual prospectus if OpenAI lists at $1T into a public market that decides, two quarters later, that the margin profile does not support the multiple.

Altman's seven percent. The PBC restructuring completed earlier in 2026 included an equity grant to Sam Altman of approximately 7% of the company, worth roughly $70 billion at the target valuation.⁵ This is the first time Altman has held material equity in OpenAI, and it was granted as part of the conversion that made the IPO structurally possible.

I want to be careful here, because the grant is not in itself improper, founder equity is conventional, and Altman had famously held none, but it does create a specific incentive structure heading into the listing. The founder's economic interest is now tightly aligned with maximum valuation at IPO, which is not always the same as maximum valuation supported by fundamentals. The S-1 will need to disclose the grant, the vesting terms, the lockup, and the board composition; what those documents say will tell us whether the conversion was structured to serve the listing or the business. These are not the same thing, and the document will be read for the difference.

The Microsoft overhang. Microsoft's reported pre-restructuring economic interest of approximately 49% is the largest unresolved structural question in the offering.³ No public document confirms how that stake converted through the PBC restructuring, what registration rights Microsoft retains, or what lockup applies. The float on day one, and therefore the supply-demand dynamic that sets the opening multiple, depends entirely on these terms. A 49% holder with short lockup and broad registration rights produces a very different first-day book than a 49% holder locked for two years.

The prospectus has to resolve this. Until it does, every valuation conversation is being held with one of the two largest variables unspecified.

What to watch.

The public S-1, when it lands in late July or early August. Specifically: the audited operating margin trajectory across 2024-2025-Q1 2026, the segment breakdown between consumer ChatGPT and API/enterprise, the disclosed Microsoft conversion terms, and the lockup table.

Anthropic's response. If Anthropic accelerates its own filing timeline, it confirms the sequencing-matters reading. If it slows down, it suggests Anthropic would rather be priced against a settled OpenAI tape than a fresh one.

The compute cost line in the cost-of-revenue note. This is the inference economics frame's load-bearing data point. If unit compute costs are falling faster than per-query complexity is rising, the -122% margin is a J-curve. If not, it isn't.

The trillion will get the headlines. The minus one hundred and twenty-two is the number that will set the price.

Footnotes

¶