August 20, 2026 · Issue 4 · 5 min read
OpenAI and Anthropic are both sitting on confidential S-1s, and one of them could be public by September
At an all hands on Wednesday, OpenAI CFO Sarah Friar told employees the company will be a public company in 2027(opens in a new tab), or sooner if the business keeps inflecting. She also told them not to worry if Anthropic gets there first. Both companies filed prospectuses confidentially with the SEC in June. Friar said Anthropic could pull the cover off its filing in the coming weeks and be public in September.
The numbers under that are large enough to explain the urgency. Anthropic told investors its annualized revenue run rate reached $65 billion at the end of July(opens in a new tab), a sevenfold increase in a year, with $11.5 billion in preliminary second quarter revenue against a $965 billion valuation. OpenAI reported $6.7 billion for the second quarter, a run rate above $40 billion, and an enterprise run rate up 50 percent quarter to date.
Here is why that matters to a budget owner rather than to a market watcher. An S-1 is a procurement document. It contains, under penalty of securities law, most of what a vendor questionnaire has been unable to extract: customer concentration, gross margin on inference, the size and duration of compute purchase commitments, and a risk factors section where the things nobody would put in writing during a negotiation finally get put in writing.
One of those risk factors is already a matter of public record. In June the Commerce Department barred Anthropic from providing Fable 5 and Mythos 5 to any foreign national, including its own non-citizen employees, and the company disabled both models for everyone(opens in a new tab) because it could not separate those users in real time. Access came back after roughly two weeks. Every enterprise standardized on those models lost them by government order, not by vendor choice, and no service credit covers that.
The practical change is in how these vendors will behave afterward. A private company optimizing for growth can absorb a bad-margin contract to win a logo. A public one guides to a quarter. Discounting discipline tightens, list prices firm, and enterprise revenue becomes the line analysts punish when it slows. That makes renewal timing a financial decision rather than a calendar one. A three-year agreement signed this quarter and the same agreement signed eighteen months from now are not the same instrument.
Researched and drafted by an automated workflow, then reviewed and edited by a human editor before publication. Every source is linked. See how we use AI here.
There is a narrow window here that closes on its own. Until a prospectus goes public, both vendors are still private companies competing on growth, which is the last stretch in which an unusually favorable multi-year rate is easy for a salesperson to justify internally.
So the question worth asking finance this week is which AI contracts come up for renewal in the next four quarters, and whether any of them can be pulled forward. Not because a price increase has been announced. Because the incentive that produced the current price is about to be replaced by a different one.
The second question is for legal, and it is about the export order rather than the IPO. Ask what the contracts say happens when a model becomes unavailable for reasons the vendor did not choose and cannot appeal. Most will say nothing, which is itself the answer to write down.
Also worth knowing
- OpenAI's CFO tells employees the company will be public in 2027(opens in a new tab)
CNBC
OpenAI filed confidentially in June, and Friar told staff Anthropic may unveil first and debut in September. Multi-year contracts signed before an IPO price differently than the ones signed after it.
- Anthropic's annualized revenue run rate hit $65 billion at the end of July(opens in a new tab)
CNBC
Sevenfold growth in a year, $11.5 billion in preliminary Q2 revenue, a $965 billion valuation. Enterprise seats are what has to justify that multiple, which is exactly where renewal pressure will land.
- Anthropic disabled Fable 5 and Mythos 5 after a Commerce Department export order(opens in a new tab)
Fortune
The order covered any foreign national, including Anthropic's own staff, so both models went dark globally for about two weeks. Model availability is a government variable now, and it belongs in continuity planning.
- Marvell grants Google an option to buy up to $12.2 billion of its shares(opens in a new tab)
CNBC
Google can buy almost 59 million shares at $206.58, tied to chip purchase targets running through fiscal 2033. When compute is paid for partly in equity, published inference prices stop describing real cost.
- FTC proposes that undisclosed AI output steering is deception under Section 5(opens in a new tab)
Covington & Burling
Issued under Executive Order 14365, the proposal treats tuning model outputs to satisfy state law as potentially deceptive absent prominent disclosure. Comments closed July 31 and general counsel should track where it lands.