Two customers, a quarter of the revenue, and a $2 trillion price tag

Anthropic is reportedly heading for public markets at a valuation north of $2 trillion. If it lands there, it would be the largest IPO in history and rank roughly sixth on US exchanges by market capitalisation, behind only Nvidia, Apple, Alphabet, Microsoft and Amazon.

Its draft prospectus is not public. It was confidentially submitted to the SEC and then leaked to Reuters, which published details on 28 September. Several outlets have since reported on the same document.

Buried in that reporting are three things that would get a normal business repriced, delayed or walked away from in diligence:

• Two customers account for roughly a quarter of revenue, and many large customers have no long-term contracts.
• Close to half of revenue flows through two partners who are also major investors and major suppliers.
• The headline profitability figures are struck before costs that look a great deal like cost of sales.

Any one of these would be a finding in a mid-market deal. All three together would normally be a problem.

Anthropic may well get away with all three. That tension is worth understanding precisely, because it explains what buyers actually forgive and why.

First, an honest caveat

The prospectus has not been published. Everything below comes from journalists describing a document they have seen, not from a filing you or I can read.

That matters more than it sounds. As forensic accounting analyst Olga Usvyatsky has pointed out, when financial information reaches investors through leaks rather than a published prospectus, there is no way to assess whether it is accurate, whether it is complete, or who chose to release it and why. Somebody selected these numbers for publication. Nobody has had to stand behind them.

Treat the figures below as well reported, not as audited.

The numbers, as reported

For the 2025 financial year, according to Fortune’s summary of the leaked document: revenue of $4.6bn, operating expenses of $13bn, an operating loss of more than $8bn, a net loss of $42bn, and $20.28bn of cash and equivalents.

The gap between an $8bn operating loss and a $42bn net loss is enormous, and the reporting does not explain what sits in between. That is itself a reason to wait for the real document.

Growth, meanwhile, is extraordinary. Q1 2026 revenue was $4.73bn, more than the whole of 2025. Q2 2026 came in around $11.5bn. By the end of July the company was reportedly at a $65bn annualised run rate. It has also committed to roughly $518bn of spending on cloud services and data centres.

This is the backdrop against which the risks have to be read. Nothing below suggests the business is weak. It suggests the business carries exposures that would be fatal to a company growing at a normal pace.

Risk one: two customers, a quarter of the revenue

Two unnamed customers each generated about 12% of 2025 revenue. Nearly a quarter of the business, sitting in two accounts.

Worse, the filing reportedly warns that many of its largest customers are not tied into long-term contracts and could reduce or stop spending.

Read those two facts together, because that is how a buyer reads them. Concentration on its own is a known quantity. Concentration without contracts is something else entirely. It means a quarter of revenue can leave on no notice, for no reason, with no breach.

In mid-market M&A this combination is one of the most reliable ways to lose value in diligence. Not because buyers refuse to buy concentrated businesses, but because of what concentration does to the structure of the deal. The buyer stops thinking about what you earn and starts thinking about what happens if the phone call comes. The headline price holds, but more of it moves into an earnout. Or the warranty package hardens. Or the buyer asks for an indemnity tied specifically to those accounts.

The honest version of the rule: concentration rarely kills a deal outright. It changes who carries the risk, and the answer is usually you.

Risk two: the channel is also the investor and the supplier

This is the part that has had less attention and deserves more.

Around $2.16bn of 2025 revenue, close to half the total, was reportedly generated through the cloud platforms of Amazon and Google. Both companies are also significant investors in Anthropic. Both are also suppliers of the compute it runs on. The filing reportedly acknowledges that relying on a small number of partners and suppliers could create conflicts of interest.

So the same two counterparties are the route to market, the cost base, and part of the cap table.

That is a genuinely unusual structure, and it is worth being precise about why it is risky rather than simply calling it concerning. Three separate exposures are stacked on the same two names. If they change marketplace terms, revenue is affected. If they change compute pricing or priority, the cost base is affected. If they build competing products, the strategic position is affected.

A single counterparty decision can hit the top line, the gross margin and the competitive position simultaneously. Diversifying customers does not help if the channel stays the same.

Founders should take the general lesson rather than the specific one. Ask not just who your biggest customers are, but where your dependency is actually concentrated. Plenty of businesses with a healthy spread of logos are in fact wholly dependent on one referral partner, one platform, one distributor or one accreditation. That concentration does not show up in a customer list. It shows up in diligence.

Risk three: profitability that excludes the cost of doing business

The headline figures that have circulated are strong. Gross margins above 80%. Adjusted operating income positive for a second consecutive quarter.

Both figures are struck before significant costs.

According to the Financial Times, gross margins are above 80% before accounting for revenue shared with distribution partners, including Amazon, and before the cost of training its models. Adjusted operating income, separately, strips out stock-based compensation.

Consider what has been removed. Revenue share with distribution partners is the cost of the channel that delivers close to half of sales. Model training is the cost of producing the thing being sold. Stock compensation is how a frontier AI lab actually pays the researchers it competes for.

These are not unusual one-off items. They are the recurring economics of the business.

There is a specific rule about this. The SEC’s guidance on non-GAAP measures, Question 100.01, says that presenting a non-GAAP performance measure which excludes “normal, recurring, cash operating expenses necessary to operate a registrant’s business” is an example of a measure that could be misleading. A misleading non-GAAP measure may breach Rule 100(b) of Regulation G, in which case it cannot be presented at all.

Two precedents show how this plays out in practice.

Groupon, 2011. Its IPO prospectus presented “Adjusted Consolidated Segment Operating Income”, which excluded the online marketing costs of acquiring subscribers. The SEC objected on the basis that those costs appeared to be normal, recurring operating expenditure and that excluding them created a potentially misleading measure. Groupon changed the presentation.

WeWork, 2019. Its pre-IPO filings presented a “Contribution Margin” measure, one version of which excluded straight-line lease costs. The SEC said this could be misleading because it excluded a material component of the company’s primary cost of sales. WeWork initially defended the measure, then dropped it. The IPO was withdrawn.

Neither company is a perfect analogue, and it is worth saying clearly that no regulator has objected to anything here, because there is no public filing to object to yet. The point is narrower and more useful. When a company going public strips recurring costs out of a profit measure, the SEC has a documented history of asking it to stop. The number that survives that conversation is the number investors actually get.

So why is it being priced at a premium?

At a $2 trillion valuation against a $65bn run rate, Anthropic would be priced at roughly 31 times sales. That is against a $965bn valuation at its May 2026 private round, meaning the IPO ask is roughly double the last private mark. Forbes notes SpaceX went public at 49 times forward sales, so there is precedent at the top of the market, and that on a projected $200bn of 2028 revenue the multiple falls to about 10 times.

That last figure is the whole argument. The premium is not a judgement about this year’s numbers. It is a bet that growth makes this year’s numbers irrelevant within two years.

The real lesson: who gets forgiven, and why

It would be easy, and wrong, to read this filing and conclude that concentration and adjusted metrics do not matter any more.

They matter enormously. They are simply being outweighed by two things most companies do not have.

Overwhelming growth. Revenue went from $4.6bn for the full year to roughly $11.5bn in a single quarter. At that rate a buyer is pricing a business that will be unrecognisable by the time the deal closes. Current-year risk gets discounted because the current year is almost irrelevant.

Genuine scarcity.There are very few frontier AI labs, and an investor who wants direct exposure to this specific thing has almost nowhere else to go. Scarcity suppresses the buyer’s willingness to walk.

Now apply that honestly to a business doing £2m of EBITDA and growing 15% a year.

Growth is good but not transformative, so this year’s numbers are exactly what the buyer is buying. And there are likely a dozen comparable businesses the acquirer could look at instead, so walking away costs them very little. Both of the things buying Anthropic forgiveness are absent.

In that situation the same three findings do not get waved through. They get priced.

Two customers at 24% with no contracts becomes an earnout, an indemnity, or a lower multiple. An undisclosed channel dependency becomes a renegotiation when it surfaces late in diligence. Adjusted EBITDA that strips recurring costs becomes a credibility problem, and once a buyer doubts one number they re-examine all of them.

That last point is the one founders underestimate most. The damage from an aggressive adjustment is rarely confined to the adjustment itself. It changes how carefully everything else gets read.

What to do about it

If you expect to sell in the next few years, the practical response is not complicated.

Know your real concentration numbers before a buyer does. Revenue by customer, by channel, by referral source and by contract type. Run it for the last three years, not just the current one, so you can see the direction of travel.

Separate contracted revenue from habitual revenue. Being someone’s biggest supplier for eight years running is not the same as having a contract. In diligence, only one of them counts.

Fix the dependency or document the defence. If a customer is 20% of revenue, the two credible answers are to grow around it or to show evidence of why it is sticky, such as switching costs, integration depth or contractual lock-in. “They love us” is not evidence.

Make sure your adjusted numbers survive scrutiny. The test is simple. Would this cost reappear next year if you did nothing? If yes, it belongs in the number. Adjustments that remove recurring costs do not just get rejected, they cost you credibility on everything else.

The gap between a business a buyer trusts and one they discount is rarely dramatic. It is usually a handful of exposures the founder knew about and had not got round to fixing.

Anthropic can carry those exposures into a $2 trillion listing because growth and scarcity are doing the heavy lifting. Most businesses are not in that position, and the sale process is a poor time to discover it.