Anthropic's $2 Trillion Question

For Anthropic's IPO price to make sense, its revenue needs to grow roughly sevenfold over the next five years, and then nearly triple again, while it keeps software-like profit margins. The bigger threat to that path is not the headline loss. It is the possibility that AI models become a cheap commodity.

Anthropic is aiming for a valuation of more than $2tn, more than double its $965bn valuation in May (TNW, citing Reuters). The listing is expected after the US midterm elections in November (TNW).

What the prospectus shows

The headline $42bn loss overstates the operating problem. About $34bn of it was a non-cash accounting charge from the rising value of financing that could convert into shares (TNW). The operating loss, $8.06bn, is the better guide to the business.

Growth has continued sharply in 2026. Preliminary second-quarter revenue exceeded $11.5bn, up from $4.73bn in the first quarter (Inc.). Annualised, the second quarter implies a run rate of about $46bn (author's calculation: $11.5bn × 4).

The customer base is concentrated. The two largest direct customers each generated about 12% of 2025 revenue (Daily Caller). The filing also warns that many of its largest customers have no long-term contracts and could cut or stop spending (TNW).

The growth the price requires

An investor buying at $2tn and wanting 12% a year needs Anthropic to be worth about $6.2tn in ten years. At a 30% net margin, that implies roughly $830bn of annual revenue by about 2036.

The figures in this section are the author's calculations, not reported numbers. They rest on these assumptions:

• Required return: 10–15% a year, typical for a high-risk growth stock.
• Mature valuation: 25 times earnings in year ten.
• Net margin: 30% for a software-like business, 20% for a compute-heavy one.
• Dilution: ignored. New shares and stock compensation would raise every target.

The five-year checkpoint is just as demanding. At 12% a year, Anthropic must be worth about $3.5tn by 2031. At an assumed 10 times revenue for a still-fast-growing company, that means about $350bn of revenue. Getting there from the $46bn run rate takes about 50% growth a year for five years, then about 19% a year to 2036.

Against the US and world economies

Measured against the whole economy, the required revenue is small. The required profit pool and valuation are not.

US nominal GDP was $30.76tn in 2025 (BEA data via Wikipedia). The IMF projects $32.38tn for 2026 and $39.03tn for 2031 (Worldometer, IMF April 2026 WEO). It projects world nominal GDP of $126.3tn in 2026 and $158.4tn in 2031 (StatisticsTimes, IMF April 2026 WEO). The IMF's projections stop at 2031, so the 2036 figures below extend each economy at its projected 2026–31 growth rate (author's calculation).

Profit is the more telling comparison. US corporate profits totalled $4.0tn at the end of 2024 (St. Louis Fed). The ~$250bn of annual earnings needed in the 12% case equals about 6% of that total (author's calculation). The measures differ: the Fed figure is pre-tax and national, while Anthropic's would be global net income. So treat this as a sense of scale, not a precise share.

The revenue would also be earned worldwide, not just in the US. Against world output, even the 2036 target is under half a percent. The size of the ask is less about the economy's capacity to pay. It is about whether one company can capture that much of it at high margins.

The $518bn compute bill sets a floor

Whatever the share price, Anthropic has already committed to a very large compute bill. It expects to spend at least $518bn over about a decade with six partners, and about 80% of that cannot be cancelled or must be paid regardless of use (TNW, citing Reuters).

Spread evenly, $518bn over ten years is about $52bn a year (author's calculation). That is roughly seven times the $7.33bn spent on compute in 2025, and more than the entire $46bn current revenue run rate. Operating leverage cuts both ways here. If revenue keeps compounding, the fixed bill shrinks as a share of sales. If growth or pricing disappoints, most of the bill still has to be paid.

The bigger risk: models become a commodity

The valuation assumes Anthropic keeps software-like margins. That fails if AI models become interchangeable and cheap, while the real value moves to the layer around them: the harnesses, agent frameworks, classifiers and filters that turn a model into a business tool.

When one layer of a technology stack commoditises, profits usually move to its neighbours. Here they would move up to the application and harness layer, which owns the customer. They would also move down to whoever owns the scarce inputs: chips, power and data centres. A model maker in the middle would earn something closer to a utility margin.

Why Anthropic is exposed. An estimated 75–85% of its annualised revenue comes from usage-based API sales, with consumer subscriptions at only about 5% (SemiAnalysis, via TradingKey). Selling raw model access is the part of the business most exposed to price competition. Its biggest customers, who account for about a quarter of revenue and often lack long-term contracts, could switch models or split spending across several. The fixed compute bill described above would then fall on a shrinking margin. One consultant quoted by Inc. describes AI labs as "usage businesses with a handful of very large purchase orders and a utility-size bill" (Inc.).

Why it may hold up. Anthropic is partly a harness company already, through Claude Code, its apps and its agent tools. It can train models to work inside its own harness, which third parties on commodity models cannot easily copy. The margin trend so far also points the other way. SemiAnalysis estimates blended gross margin rose from negative 94% in 2024 to the mid-60s, with API margins above 80%, and an EBIT margin of 36% in the second quarter of 2026. It argues Anthropic could become the first company worth $6tn (SemiAnalysis, via TradingKey). These are outside estimates, not audited figures.

The figures in the table are the author's illustrations, not forecasts. Cheaper tokens would probably raise demand, so revenue could keep growing even in the weaker cases. The damage comes through margins and fixed costs, not lack of demand.

What the prospectus says, and what to watch

The reported risk factors touch the ingredients of the commoditisation scenario but, in coverage so far, not the scenario itself. About 80 of the filing's 261 pages are risk factors, and they go well beyond the usual warnings on competition, regulation and customer demand (eWeek). Anthropic says customer usage, and therefore revenue, is driven by new models, so a continuous cadence of releases is needed to stay at the frontier (CTech). CTech notes that it competes directly with OpenAI, Google, Meta and xAI (CTech). No outlet has yet quoted language on open-weight models, commoditisation or falling token prices.

The public S-1 should settle the open questions. Four numbers matter most:

1. Gross margin trend. Rising with scale, or competed away?
2. Revenue mix. First-party products (Claude Code, subscriptions, enterprise) versus raw API sales.
3. Price per token over time. Whether volume growth is outrunning price cuts.
4. Customer behaviour. Net revenue retention, and whether the largest customers are spreading spend across several model providers.

If first-party products are a growing share and margins hold, Anthropic is on the integrated-winner path. If API sales dominate and margins fall, the market is moving towards commoditisation.

Conclusion

The $42bn headline loss is mostly an accounting item and not the real obstacle. The $2tn price needs three things at once. Revenue must compound at about 50% a year to reach roughly $350bn by 2031. Margins must settle near software levels rather than utility levels. And both must happen while Anthropic pays down a compute bill of about $52bn a year.

If models stay differentiated and Anthropic owns the harness its customers use, the price could look reasonable in hindsight. If models become commodities and the value moves to the layer around them, the same revenue growth could leave shareholders with a much smaller business than the one they paid for.

Attribution. This article is published by Human Freedom Foundation. AI models were used in the authoring of this article.