Anthropic goes public: what the prospectus reveals

The annual figures tell the story: $41 million in 2023, $386 million in 2024, $4.6 billion in 2025, then $16.3 billion in the first half of 2026 alone. In the second quarter of 2026, year-on-year growth reached 1,368%. The third quarter is not yet quantified in the filed document, but the prospectus states that it represents 86 times the fourth quarter of 2024: about $16.9 billion, consistent with the $16.8 billion presented to investors according to the Financial Times.

The slope changes character in early 2026: +132% in Q1, +144% in Q2, then about +47% in Q3. Growth remains enormous but is slowing in percentage terms, which the company itself flags: it expects a slowdown, possibly faster than anticipated. The prospectus ties each acceleration to a model release: versions launched in the previous six months accounted for more than 80% of third-quarter revenue.

The "run rate" often quoted in the press (annualized revenue) is computed as the last 28 days of consumption multiplied by 13, plus monthly subscriptions multiplied by 12. It is a snapshot, not booked revenue: consumption can drop from one month to the next. Over the 2025 year-end holidays, customer spending fell by about 20% in two weeks.

The model is consumption first. Customers pay for the tokens they consume, at the price of the model they choose: $13.1 billion in the first half of 2026, or 80% of revenue. Subscriptions (Pro, Max, enterprise plans) account for $3.2 billion, or 20%. This model aligns revenue with actual usage but also makes it more volatile than subscription software.

Large accounts drive growth. The number of customers spending more than $100,000 a year rose from about twenty at the end of 2023 to more than 11,900 at the end of September 2026. More than 1,900 exceed one million and about 250 exceed ten million. More than half of the Fortune 500 are among them. Net retention (what last year's customers spend today, relative to what they spent then) is around 900%: the average customer has multiplied their spending by nine.

Real but declining concentration. The top five customers account for about 25% of revenue for the first nine months of 2026 (none above 10%). In the third quarter, 79% of revenue came from outside those five customers, compared with 64% in the first. The rest of the base, about 40 million customers below $100,000, provides depth.

The clouds are distributors, suppliers and shareholders all at once. 48% of revenue flows through the AWS, Google Cloud and Azure marketplaces (11% in 2023). These same players sell compute to Anthropic, invest in it and offer competing models. The prospectus acknowledges this as a potential source of conflicts of interest. Finally, 72% of revenue is billed in the United States: international business ($4.5 billion for the half, 28%) is presented as a growth driver.

Anthropic presents its expenses by nature, which is unusual and instructive. Compute (chips rented from the clouds, data centers, networking) is by far the largest item: $5.6 billion in the second quarter of 2026, versus $944 million in the first quarter of 2025. But it grew 6-fold while revenue grew 31-fold. As a result, one dollar of compute brought in $0.40 of revenue in early 2025 and $2.08 in the second quarter of 2026.

The reason is mechanical. Compute is bought 12 to 18 months in advance under multi-year contracts: the expense arrives when capacity is delivered, not when customers ask for it. In 2025, capacity arrived ahead of revenue and losses deepened. In 2026, revenue caught up with and then overtook capacity. The company warns that this ratio will fluctuate again, especially because it is buying more expensive short-term compute to meet demand.

The other items tell the story of scaling up. Personnel cost $1.9 billion in Q2 2026, of which $1.3 billion in stock granted to employees: headcount grew by about 170% in a year to reach, by our estimate, some 5,000 people ($6.6 million of revenue per employee). Sales expenses jumped in 2026 ($2.1 billion in Q2): they consist mainly of commissions paid to the clouds that distribute Claude. The $1.5 billion settlement with authors (the Bartz case) explains the Q3 2025 spike.

A note to the financial statements reclassifies the same expenses by function, the way a traditional software company would. It is the most useful lens for understanding the economics of the business.

Cost of revenue (the compute that serves paying customers, plus some personnel and overhead) came to $2.4 billion in the first half of 2026 against $16.3 billion of revenue. The implied gross margin is therefore about 85%, versus 12% a year earlier. At that level, serving Claude economically resembles selling software.

Research and development absorbs $9.7 billion, or 60% of revenue. It includes most of the compute devoted to training, research and internal use of Claude. Measured against total compute of $8.9 billion, serving paying customers accounts for only about 27%: nearly three quarters of compute goes to preparing the next models, running Claude internally or providing free access (classified as marketing).

This is the heart of the debate about AI labs. Taken in isolation, each model is highly profitable once in service. But the next, larger one must constantly be funded to stay at the frontier: 80% of revenue comes from models less than six months old. Lasting profitability therefore depends on the ratio between what it costs to train a generation and what it earns during its short commercial life.

This is the central question for anyone trying to understand the economics of AI. The price of a token is falling very fast. Yet Anthropic reports a soaring gross margin. There is no contradiction: it all depends on which price we are talking about and what the margin measures.

Two very different price declines

At constant capability, prices are collapsing. Epoch AI finds that the price of reaching a given level of performance falls by 9 to 900 times a year depending on the benchmark, for example 40 times a year for GPT-4-level performance on PhD-level science questions. Today's entry-level model beats the top of the range from two years ago for a fraction of the price.

At the frontier, prices fall much less. Anthropic's Opus model cost $75 per million output tokens from 2024 to mid-2025. It dropped to $25 with Opus 4.5, then to $20 with Opus 5.5: −73% in two and a half years, while capability leapt ahead. Sonnet stayed at $15 for two years before dropping to $10. And a new tier has appeared above: Mythos and Fable, at $50. Anthropic mostly sells frontier models: that is where the revenue is made.

Why revenue rises anyway

Revenue is the number of tokens times their price. The price per token is falling, but the number of tokens per task is exploding: a coding or research agent consumes hundreds of times more tokens than a chat question, because it reads, reasons, tries and checks. What matters to the customer is the price per completed task, not per token. This is the Jevons effect applied to AI: the cheaper intelligence becomes, the more of it we consume.

Why the margin rises anyway

Gross margin depends on the ratio between what a token costs to produce and what it sells for. Cost is falling faster than price, for three reasons: new chip generations (Trainium, TPUs, recent GPUs), algorithmic inference gains that the prospectus calls "compute multipliers" (distillation, caching, more efficient decoding) and above all an exceptional utilization rate. Anthropic was short of compute throughout 2025–2026: every chip it bought ran flat out. Revenue per dollar of compute went from $0.40 to $2.08 in five quarters despite the price cuts.

The prospectus makes an important point: when selling through the clouds, Anthropic keeps control of its pricing and retains the benefit of its efficiency gains. It is not the cloud that sets the price of a token.

What the 85% leaves out

The figure is accurate but its scope is narrow. Cost of revenue only includes the compute that serves paying customers. Three items remain outside:

• Cloud commissions. About $2.7 billion in the first half (17% of revenue) is classified as marketing. For a distributor, many analysts would count it in cost of sales: the margin then falls to around 69%.
• Free access. The compute given away to free users is classified as marketing. It is not quantified.
• Training. A model that generates 80% of revenue for six months looks more like perishable inventory than software amortized over ten years. Charging half of R&D to cost of sales brings the margin down to around 55%, according to the calculation by the OnlyCFO newsletter.

For comparison, OpenAI reportedly shows an adjusted gross margin of about 33% according to investor documents reported by The Information (unconfirmed figure, different scope). The gap is partly explained by the weight of ChatGPT's free users and by Anthropic's strongly enterprise customer base.

What could push the margin down

• The end of the shortage. The 85% was achieved during a period when demand exceeded supply. When the massive capacity contracted for 2027 comes online, utilization will drop if demand does not keep up. The company also expects greater reliance on more expensive short-term compute.
• Agents. The prospectus acknowledges that agentic uses are "structurally more compute-intensive" and that, if pricing does not reflect this extra cost, compute efficiency could deteriorate. Flat-rate subscriptions (20% of revenue) are exposed: a developer running agents all day can consume more than they pay.
• Competition. Open models, Chinese labs, Google and OpenAI: the prospectus explicitly cites the risk of price pressure and of commoditization in which the advantage shifts from the model to compute.
• A fragile slice of revenue. Anthropic estimates that 5 to 7% of its revenue for the first nine months of 2026 comes from customers who violate its terms of use and will be cut off.

The headline figure, a net loss of $80.7 billion over six months, is explained almost entirely outside operations. Anthropic issued convertible notes (notably to Google and Amazon) that will turn into shares. For accounting purposes, their fair value tracks the company's valuation: when it rose to $965 billion in the Series H, the liability to those holders rose accordingly. Hence a $72.4 billion charge, plus $8.8 billion from remeasuring note purchase agreements. None of these charges involve cash going out.

Operations, by contrast, were close to breakeven for the half (a $693 million loss) and profitable in Q2 ($974 million). Excluding employee stock-based compensation and non-cash donations, adjusted operating income reached $1.6 billion for the half and $2.2 billion in Q2, or 19% of revenue. The most robust indicator remains cash: $3.9 billion of positive operating cash flow in the first half of 2026, versus $1.5 billion burned a year earlier.

The balance sheet provides the means for what comes next: $131.7 billion of cash and investments at the end of June 2026, after $141 billion raised since inception. The IPO, described in the press as the largest in history, could add up to $100 billion according to Reuters.

The prospectus details seven quantified contracts totaling about $637 billion. The four most binding (Broadcom, Google, Amazon, Microsoft) account for $431 billion on their own: they are owed even if Anthropic does not use the capacity. The SpaceXAI, Nscale and Lambda contracts are more flexible, though not free. On top of that come $145–170 billion of TPU and AMD system purchases whose amount will depend on deliveries. Aggregate amounts are not yet filled in in the filed version: the press puts totals between $417 billion and more than $600 billion depending on the scope.

The strategy is shifting. After renting everything from the clouds, Anthropic now leases chips directly, installed in data centers operated on its behalf ("Anthropic Direct"). This requires payments before go-live and will transform its cash profile: the company expects strongly negative investing cash flows "for the foreseeable future".

An order of magnitude

If each contract is spread evenly over its term (a simplifying assumption: ramp-ups are in fact gradual), the annual compute bill would reach about $115 billion once all capacity is delivered, five times the Q2 2026 pace ($22 billion a year). The SpaceXAI contract alone represents $37 billion a year from December 2026. To keep today's 48% ratio between compute and revenue, annual revenue of about $240 billion would therefore be needed. Even with 50% more headroom on that ratio, the requirement remains around $150–200 billion. The Q3 2026 pace is about $68 billion.