Anthropic's Revenue Grew 14×. Its Compute Bill Grew 3.8×. Did AI Get Cheaper to Serve?
Anthropic's IPO filing shows revenue rising far faster than spending on computing power. In the first half of 2026, each dollar of revenue needed 55 cents of compute, down from $2.01 a year earlier. The improvement is large and real, but it is not the same as a profitable AI business.
Compute and infrastructure covers all of it: training and research, serving customers, free use and internal use. Cost of revenue is the part spent serving paying customers. Source: Anthropic, Form S-1 filed with the SEC (26 October 2026).
Frontier AI companies rarely publish detailed accounts. Anthropic, the developer of the Claude models, changed that on 26 October 2026, when it filed a registration statement (Form S-1) with the US Securities and Exchange Commission for a proposed listing on Nasdaq (SEC filing).
The filing answers a question that has hung over the AI industry: does serving AI get cheaper as it scales? For Anthropic, over the past year, the answer is clearly yes. In the first half of 2026, revenue was $16.3 billion, 14 times the $1.16 billion of a year earlier. Spending on compute and infrastructure grew to $8.95 billion, 3.8 times the year before.
Put another way, every dollar of revenue needed $2.01 of compute in the first half of 2025, and $0.55 in the first half of 2026 Calculated by NodeZero.
The numbers in the filing
The S-1 gives full-year results for 2023 to 2025 and half-year results for 2025 and 2026. These are the main lines, in millions of dollars:
| 2023 | 2024 | 2025 | H1 2025 | H1 2026 | |
|---|---|---|---|---|---|
| Revenue | 41 | 386 | 4,585 | 1,163 | 16,286 |
| Compute and infrastructure | 433 | 2,522 | 7,326 | 2,334 | 8,945 |
| Personnel | 95 | 454 | 1,839 | 538 | 3,152 |
| Sales, marketing and partnerships | 2 | 43 | 568 | 120 | 3,168 |
| Other operating and administrative | 62 | 347 | 1,412 | 306 | 1,714 |
| Legal settlement (one-off) | – | – | 1,500 | – | – |
| Total operating expenses | 592 | 3,366 | 12,645 | 3,298 | 16,979 |
| Operating loss | (551) | (2,980) | (8,060) | (2,135) | (693) |
Source: Anthropic Form S-1, summary financial data Reported. Half-year figures are unaudited. H1 = six months ended 30 June.
The equation: compute × revenue per compute dollar
Revenue can be split into two parts: how much compute the company runs, and how much revenue each dollar of that compute brings in.
Revenue = compute spending × revenue per dollar of compute
In the first half of 2025, each dollar of compute brought in about 50 cents of revenue. A year later it brought in about $1.82 Calculated by NodeZero. That is 3.7 times as much, while compute spending itself grew 3.8 times. Multiplied together, the two make up the 14-fold rise in revenue. Using the same logarithm method as our Meta article, the growth splits almost exactly in half: about 51% came from running more compute and about 49% from earning more per compute dollar Calculated by NodeZero.
The longer view shows how far that ratio has moved:
All compute is included: research and training, serving customers, free use and internal use. 2023 to 2025 are full years; H1 2026 is the six months ended 30 June 2026 (the same ratio for H1 2025 was $2.01). Source: NodeZero calculation from Anthropic Form S-1 (26 October 2026).
In 2023, Anthropic spent more than ten dollars on compute for every dollar it earned. In the first half of 2026, it spent 55 cents.
The cleaner measure: what it costs to serve paying customers
Total compute mixes two very different jobs: training new models and serving customers who pay. A notes table in the filing separates them. Cost of revenue, which the filing says consists substantially of compute for customer-serving workloads, gives the closest view of what serving AI costs.
Cost of revenue is substantially all compute for customer-serving workloads, plus some staff and overhead. Compute for free users is counted in sales and marketing, and compute for training in research and development. Source: NodeZero calculation from Anthropic Form S-1, Note 8 (26 October 2026).
In 2024, serving customers cost almost two dollars for each dollar they paid. In the first half of 2026 it cost 15 cents, which implies a gross margin of about 85%, up from about 12% a year earlier Calculated by NodeZero. Revenue grew 14 times while the cost of serving it grew 2.3 times.
Meanwhile, research and development, which includes training compute, grew 5.3 times, to $9.74 billion. Anthropic now spends about four times as much on research and development as on serving paying customers.
Price and volume: what the filing says
The filing does not report usage in units such as tokens or requests, so price and volume cannot be separated exactly. It does give three clues:
- Volume of customers. Customers paying at least $100,000 a year (trailing twelve months) rose from more than 950 in September 2025 to more than 11,900 in September 2026 Reported. More than 1,900 paid $1 million or more.
- How customers pay. About 80% of first-half 2026 revenue ($13.1 billion) was consumption-based, meaning customers pay for what they use. The rest ($3.2 billion) came from subscriptions Reported.
- Price per task. Anthropic says its Claude Opus 5.5 model cut customers’ cost per task by about 40% compared with Claude Opus 5 Reported. This is the company’s own measure for one model, not an average across all products.
If the average price customers pay for a given amount of work fell at all, as the last point suggests, then the volume of work Anthropic served grew by more than revenue’s 14 times Estimated. That fits the filing’s own description: lower costs per task drive greater usage.
The filing reports both values as 'more than' the figure shown. Source: Anthropic, Form S-1 filed with the SEC (26 October 2026).
Why the ratio could rise again
The filing is explicit that compute costs do not move smoothly with revenue:
- Compute is bought in advance. Most of it is contracted under multi-year commitments, typically 12 to 18 months before it becomes available, and costs are recognised when the capacity arrives, not when it is ordered.
- It is shared. Anthropic runs one pool of compute for training, serving and internal use, and plans it as a whole.
- So the ratio depends on timing. In the company’s words, compute as a percentage of revenue depends primarily on how revenue scales relative to committed compute costs. The filing says margins will fluctuate, particularly when large amounts of new compute arrive.
In other words, part of the improvement comes from revenue growing into capacity that was already paid for. If the next large block of compute arrives before revenue catches up, the 55-cent figure would rise. The filing also warns that revenue growth is expected to slow, possibly sharply.
The loss: why spending still matches revenue
Lower compute per dollar does not make the company profitable yet. In the first half of 2026, total operating expenses were $16.98 billion, or about $1.04 for each dollar of revenue Calculated by NodeZero. A year earlier it was $2.84. Two lines grew faster than revenue:
| Expense, H1 2026 vs H1 2025 | H1 2025 | H1 2026 | Growth |
|---|---|---|---|
| Sales, marketing and partnerships | $120M | $3,168M | 26.4× |
| Revenue | $1,163M | $16,286M | 14.0× |
| Personnel | $538M | $3,152M | 5.9× |
| Other operating and administrative | $306M | $1,714M | 5.6× |
| Compute and infrastructure | $2,334M | $8,945M | 3.8× |
NodeZero calculation from the S-1 Calculated by NodeZero.
The filing reports a much larger net loss: $80.7 billion for the first half of 2026. Of that, $72.4 billion was a non-cash change in the fair value of convertible notes and similar instruments, driven by the rise in the company’s valuation. It reflects how the company’s financing is accounted for, not the cost of running AI, which is why this article uses operating figures.
Anthropic also reports an “Adjusted Operating Income” of $1.6 billion for the period. That is a non-GAAP measure that excludes some costs; the filing reconciles it to the GAAP operating loss of $693 million.
What the numbers do not tell us
- Usage in units. Without tokens, requests or tasks, price and volume cannot be separated exactly.
- Whether it lasts. Six months is a short period, half-year figures are unaudited, and the company expects growth to slow.
- Other AI companies. This is one company’s disclosure. Others may have very different costs, prices and contracts.
- The final prospectus. The S-1 has been filed but is not yet effective, and figures can change in amended filings.
- Self-reported claims. Statements such as the 40% lower cost per task are the company’s own and are not independently verified.
- This is not investment advice. The article explains how the reported revenue and costs fit together; it does not judge the proposed share offering or the company’s value.
What to watch next
- Compute per dollar of revenue after new capacity arrives. The filing says more committed compute will become available under multi-year contracts; later results will show whether revenue keeps pace.
- Cost of revenue. The cleanest signal of serving efficiency. A rise from 15 cents would mean serving got relatively more expensive, perhaps because of price cuts or new, heavier products.
- Amended filings. Amended S-1 filings may add a later quarter, which would show whether the trend held through September 2026.
Key terms
- Compute and infrastructure: the cost of the computers, data centres, networking and power used to train and run AI models.
- Cost of revenue: costs directly tied to delivering what customers pay for; for Anthropic, mostly compute for serving paying customers.
- Gross margin: revenue minus cost of revenue, as a share of revenue.
- Consumption-based revenue: revenue from customers who pay according to how much they use, rather than a fixed subscription.
- Trailing twelve months (TTM): the twelve months up to a given date.
- Non-GAAP: a measure that does not follow standard US accounting rules, usually excluding some costs; companies must reconcile it to the standard figure.
Notes on the data
- All revenue, expense, customer and loss figures come from Anthropic’s Form S-1, filed with the SEC on 26 October 2026. Full-year figures are audited; half-year figures are unaudited.
- Ratios (compute per dollar of revenue, cost of revenue per dollar, gross margin, expense growth multiples, revenue per compute dollar) are NodeZero calculations from those figures. The growth split uses logarithms, so the two parts add up to 100%.
- The S-1 also states that revenue grew 1,368% and compute expenses 300% in the three months ended 30 June 2026 compared with a year earlier. This article uses six-month figures, which come with a full expense breakdown.
- Cost of revenue figures come from the functional breakdown of operating expenses in Note 8 of the financial statements.
- The 2025 legal settlement of $1.5 billion is shown separately in the filing and is included in 2025 total operating expenses.
- Figures marked “more than” in the filing are shown at their stated minimum.
Sources
Primary: Anthropic, PBC – Form S-1 registration statement, SEC EDGAR (filed 26 Oct 2026) · Filing index
Every figure in this article links to its source and was checked against that source before publishing. How we check figures