What Has Actually Been Filed, and When
Anthropic submitted a draft registration statement to the Securities and Exchange Commission on June 1, 2026. That filing was confidential, which is normal: it starts the SEC review clock without putting the company's financials in front of competitors while the terms are still moving.
Reporting on August 20 indicated the company now expects to file publicly as soon as the end of August. Backers are pointing at an October listing, at a valuation of $2 trillion or more. If that holds, it would be the largest initial public offering ever recorded, ahead of SpaceX's record raise.
Those are two different events and it is worth keeping them separate. The confidential draft was a legal option. The public S-1 is the moment the numbers stop arriving through journalists and start arriving as a document somebody signs under penalty of perjury.
The Numbers Behind the Valuation
Preliminary second-quarter revenue came in above $11.5 billion. The same quarter a year earlier was $787 million. Q1 2026 was $4.73 billion, so the company more than doubled again in the space of three months.
The run rate tells the same story from another angle: past $47 billion in May, past $65 billion by the end of July. Backers expect it to finish the year somewhere between $100 billion and $120 billion.
And for the first time, a frontier AI lab reported positive adjusted operating income. That sentence has been repeated everywhere this month, usually without the adjective, and the adjective is the entire story.
Fourteen-Fold Growth Is Not a Normal Input to a Valuation Model
Standard public-market valuation assumes you can extrapolate a trend. A company that grew revenue fourteen times in twelve months does not have a trend, it has a slope that nobody can responsibly extend. Underwriters will price the October listing off a forward multiple, and the forward number is a projection made in a market where enterprise AI budgets are still being set for the first time. This is not an argument that the valuation is wrong. It is an argument that the confidence interval around it is far wider than the precision of the headline implies.
One Word Is Doing an Enormous Amount of Work
The reported figure is positive adjusted operating income. Adjusted means before certain costs are counted, and which costs get excluded is a choice the company makes and the auditors then have to let stand or challenge.
In this industry those excluded lines are not rounding errors. Stock-based compensation at a lab competing for a few thousand researchers is enormous. So is the training compute for models that have not shipped and therefore have not earned a dollar. A presentation that sets both aside can be entirely defensible as a picture of the operating business, and can also sit a very long way from the number a public shareholder ultimately owns a claim on.
The Other Filing
Anthropic is not filing into an empty room. OpenAI submitted its own confidential S-1 on May 22, 2026, targeting a September listing at a valuation reported between $852 billion and $1 trillion, with Goldman Sachs, Morgan Stanley and JPMorgan leading the deal.
The economics read differently. OpenAI is at roughly $2 billion in monthly revenue, against reporting that puts the loss at about $1.22 for every dollar earned, with profitability not expected until around 2030. Analysts have also flagged that the widely quoted $14 billion loss figure for 2026 is a non-GAAP presentation, and that the GAAP number is materially larger.
Two labs, two filings, one quarter apart, in the same window. For the first time since this cycle began, the comparison between them will be made with audited statements rather than anonymous sourcing and selectively released metrics.
What an S-1 Actually Forces Into the Open
For readers who have never had a reason to open one, this is the part worth understanding, because it is where the real information is.
Customer Concentration
The filing has to disclose how much revenue comes from the largest customers. If a meaningful share of that $11.5 billion quarter traces to a small number of enterprise contracts or to one hyperscaler relationship, then the growth story and the risk story are the same story.
Compute Commitments
Multi-year obligations to cloud and chip suppliers appear as contractual commitments. These are the numbers that determine what happens if demand growth slows, because the bill arrives whether or not the revenue does.
The Real Cost of Serving a Token
Cost of revenue and gross margin, stated plainly. Every argument about whether frontier AI is a good business eventually reduces to this line, and until now nobody outside these companies has seen it.
Related-Party Relationships
Investment from major cloud providers that is also spent back on those providers' infrastructure has to be described. Whether that circularity is material is a judgment call, and the filing forces the company to make that judgment in writing.
Risk Factors and Cancellability
Litigation exposure, regulatory dependencies, and, importantly, the terms under which large contracts can be cancelled. Fast-growing enterprise revenue sitting on short, cancellable terms is worth a different multiple than the same revenue under multi-year commitments.
This Listing Is a Stress Test for the Entire AI Trade
Private valuations in this cycle have been set by a small number of investors with strong incentives to mark them upward and no obligation to justify them publicly. A $2 trillion listing hands that pricing decision to a market of buyers who can sell, who read footnotes, and who have no position to protect. Whatever number the book builds at becomes the reference price for every other AI company's next round, up and down the stack, from model labs to infrastructure vendors to the application layer. That is why this filing matters to people who will never buy a single share of it.
What This Means If You Build on These Models
1. Your Vendor Is About to Acquire Quarterly Obligations
A public company reports every ninety days and answers to shareholders about margin. Expect pricing, deprecation schedules and enterprise terms to become more disciplined and less generous over time. That is not cynicism, it is what public-market accountability does to every infrastructure vendor eventually.
2. Read the Risk Factors, Not the Coverage
When the S-1 publishes, the risk factors section will describe the company's own view of what could break, written by lawyers who are liable for omissions. It is the single most honest document any AI lab has ever produced about itself, and it will be free to read.
3. Price Your Own Dependency
If your product's margin depends on one lab's API pricing, the filing gives you the first real basis for estimating how sustainable that pricing is. Gross margin on inference tells you whether current rates are a stable business or a customer-acquisition subsidy that has an expiry date.
4. Separate the Company From the Category
Anthropic reporting positive adjusted operating income does not mean frontier AI is profitable. It means one company, under one presentation, cleared one bar. The category question stays open until several of these filings can be read side by side, which is exactly what the next two months will make possible.
The one thing worth doing when it lands. Open the S-1 and search it for three things: the reconciliation between adjusted and unadjusted operating income, the customer-concentration disclosure, and the contractual compute commitments. Those three items will tell you more about the state of the AI industry than every valuation headline published this year, and they will take about twenty minutes to find.
Frequently Asked Questions
My Take
For three years this industry has been financed by people who never had to show their work. Private rounds, selective disclosure, revenue figures that reached the public through journalists who were handed a number and no context to put around it.
An S-1 ends that, and it ends it in a way no amount of skepticism from the outside ever could. Customer concentration, compute commitments, the real cost of serving a token, the contracts that can be cancelled on thirty days notice: all of it gets written down, and someone signs their name at the bottom.
That is the actual news here, not the $2 trillion.
I have no position on whether the valuation is right, and honestly neither does anyone else, because until this month nobody had the inputs. What I would say is that we are days away from the best data anyone outside these companies has ever had on whether frontier AI is a business or a very expensive research programme with a revenue line attached. If the numbers hold up under audit, the skeptics lose their strongest argument, which has always been the absence of evidence rather than the presence of bad evidence. If they do not hold up, the correction happens in public, with a paper trail, which is infinitely healthier than another two years of private marks nobody can check.
The part I will be reading first is the adjustment reconciliation. Not because I expect something scandalous in it, but because at a company spending this much on compute and on talent, the gap between adjusted and unadjusted is the gap between a profitable business and a growth story, and those get very different multiples once a market rather than a syndicate is doing the pricing.
Either way, the era of taking it on faith ends with a document.
Will you read the S-1, or the headline about it?
Related Articles:
