Anthropic IPO. The Real Test of the AI Trade.

On June 1, 2026, Anthropic filed a confidential draft S-1 with the SEC. Reporting puts a potential listing near two trillion dollars, which would rival the largest offerings ever, against the $1.77 trillion valuation SpaceX carried into its June IPO. That number is not the story. The document behind it is.

For the first time, the leading closed-model frontier lab has to open its books. Audited revenue. Gross margin. Customer concentration. The full scale of its forward compute obligations. None of that has ever been public for a pure-play frontier model company, from anyone, at any size. The reporting now points to a prospectus public in late September, an investor day in mid-September, marketing beginning in mid-October, and a listing timed just ahead of November's midterm elections. When it lands, we stop guessing. We get audited numbers, not investor-deck figures.

That is why this one filing matters more than any single AI headline this year. It is the first audited stress test of the entire AI trade.

Treat the valuation as a reading, not a fact. Share count and price range are not set. The figure moving through the press, reported in a range up to around two trillion dollars, is anchored to a recent primary round and to secondary-market activity. What it tells you is not what Anthropic is worth. It tells you what the public market is willing to assume about AI in aggregate.

If Anthropic prices anywhere near where the reporting suggests, that price embeds a set of assumptions about revenue durability, margin trajectory, and the useful life of a frontier model. Those assumptions do not stay inside Anthropic. They set the water level for every private AI company below it. A mark that high says the public market has decided the AI trade is real, durable, and worth underwriting at scale. A soft debut says the opposite, loudly, to everyone holding AI paper at private-round prices.

This is one of the most important listings of our lifetime, and not because of the number. Anthropic is the tip of the spear. Behind it stands every AI company, every private mark, every data-center buildout underwritten by the belief that this trade holds. When the leading frontier lab prices in public, the entire complex gets repriced against the reading. One company files. The whole asset class finds out what it is worth.

Here is the part most coverage skips because it is less fun than a trillion-dollar headline.

Anthropic has committed an extraordinary amount of future spending on compute, and the pace is accelerating. In a single week in late August it committed forty-five billion dollars to Nscale and thirty-five billion to Lambda, roughly eighty billion of contracted compute arriving on dated schedules. That sits on top of more than one hundred billion dollars committed to Amazon Web Services over roughly a decade, tens of billions of Azure capacity, and a gigawatt-scale commitment alongside Nvidia.

These are contracts. Multi-year obligations. They are set against revenue that is growing fast, but has not yet reached the scale those contracts assume. That gap is the whole question. Not whether Anthropic can build great models. It obviously can. The question is whether the revenue curve catches the cost curve before the obligations come due.

The financing around the deal tells you how seriously the market takes that gap. Anthropic is finalizing a revolving credit facility reported at fifteen billion dollars, six times its size a year ago, with a syndicate seventeen banks deep. At the same time, bankers for Anthropic and OpenAI are lobbying the rating agencies for investment-grade ratings once they are public, which would cut borrowing costs and open the door to the pension funds and insurers that cannot hold speculative-grade debt. An investment-grade rating soon after listing would be a remarkable outcome for two companies still in the red, and the agencies are reportedly waiting to see the IPOs before deciding.

That waiting is the tell. Much of the AI buildout's debt financing is structured around these listings. Nvidia's credit support for a large OpenAI data center terminates once the startup earns a satisfactory rating. Google and Broadcom have extended tens of billions in credit backing Anthropic's use of their chips, on the bet it needs less help once public. Broadcom's chief executive put it plainly last week, describing the frontier labs as brilliant but not yet ready to borrow on their own, and the IPO as the thing that changes their credit. The S-1 is where that bet gets tested against audited numbers rather than assumed.

No pure-play model company has ever demonstrated in public markets that it can scale revenue faster than the cost of training and inference. The private market has taken it on faith. The S-1 is where faith meets an auditor.

Anthropic is not going public in a vacuum. OpenAI filed its own confidential S-1 about a week later, but has reportedly signaled it may wait until next year. That sequencing is not a footnote. It is the second reason this filing matters.

The first pure-play frontier lab to print audited financials sets the comp. And the disclosure fight is already happening in private. Reporting says large investors are pressing Anthropic for far more than securities rules require: AI tokenomics, revenue per gigawatt of compute, revenue per millions of tokens processed net of discounts, the cost of serving tokens, and net revenue retention. Those metrics roll up to one number that matters more than the valuation, the inference margin, what it costs to run the models rather than train them. What Anthropic chooses to disclose becomes the template. OpenAI's own IPO advisers are reportedly already debating how much they will have to match.

That is real power, and it cuts both ways. A clean S-1 with defensible margins lifts the entire cohort and makes the private marks look conservative. A messy one, heavy losses against those compute obligations, drags the comp down and forces a repricing that reaches every AI position held at last-round value.

When the public S-1 hits EDGAR, ignore the valuation headline for the first hour. Read in this order.

Inference margin, and the trend, not the snapshot. What it costs to run the models, as opposed to train them, tells you whether this is a software business or a capital-intensive one wearing software multiples. It is the single number the sophisticated buyers are asking for, and the one the growth-rate crowd will skip.

The compute obligations against the revenue ramp. Total contracted spend, the timing, and whether the revenue curve is on pace to meet the dated schedules. The gap between those two curves is the entire risk, and the fifteen-billion-dollar revolver is a hint at how the company plans to bridge it.

Net revenue retention, read skeptically. It is a flattering metric and an easy one to dress up. It is also genuinely hard for Anthropic to produce cleanly, since much of its revenue is API usage that swings month to month as developers move between models, not recurring subscription revenue. Take any tidy retention figure with suspicion.

Customer concentration. A frontier lab leaning on a handful of enterprise contracts and its own cloud partners is a different risk than one with broad, durable demand.

The risk factors. Reporting suggests the filing will name AI backlash and active litigation directly. How a company describes its own tail risks tells you how honest the rest of the document is.

None of this is about whether Anthropic is a good company. It is. The models are excellent and the enterprise traction is real. One large investor summed up the bull case as: do not overthink it, just look at the revenue growth rate. This entire filing is the test of whether that is true. The growth rate is what got Anthropic here. The margin, the obligations, and the disclosures are what tell you whether it stays.

We are about to find out. For the first time, on the record, with numbers that count. That is why this filing, not the valuation, is the most important event in AI this year.

Bashar Aboudaoud
Managing Member, UpRound

Live Webinar: Thursday, September 17: You leave with three things. Where private valuations actually sit right now. Which sectors are repricing fastest. How I am reading prices into 2027. 20 minutes, then live Q&A. 20 slots only for UpRound readers.

Share with 10 people to receive a quarterly call with Bashar.