Anthropic’s AI business model turns compute into paid intelligence: it trains Claude models on capacity it commits to years ahead, then sells their output by the token through the Claude Platform, apps such as Claude Code and Claude Cowork, and cloud marketplaces. Its S-1, filed on 26 October 2026, shows the economics turning: compute and infrastructure expense fell from 251% of revenue in the first quarter of 2025 to 48% in the second quarter of 2026, a quarter with operating income of $1.0 billion.
Anthropic signs multi-year compute and lease agreements across four chip families and orchestrates workloads for price and performance. The S-1 notes that these commitments generally run for years, while much of its revenue is not under long-term contracts.
Mythos (including Fable), Opus, Sonnet and Haiku each sit at a different point on what the S-1 calls the capability-cost-speed curve. The model chosen and the volume used are the main drivers of customer spend.
The Claude Platform gives customers the same building blocks Anthropic uses for its own apps, Claude Code and Claude Cowork. Usage is billed by the token; some apps are sold on subscription.
Compute and infrastructure is the largest expense: $8,945 million in the six months to 30 June 2026, against revenue of $16,286 million. Revenue growth outpaced the committed compute cost base, so its share of revenue fell.
Operating loss was $935 million in the first quarter of 2025 and $3,089 million in the third, a quarter that included the Bartz settlement. The second quarter of 2026 showed operating income of $974 million, which the S-1 describes as $1.0 billion.
Net loss for the six months to 30 June 2026 was $80,748 million, driven by a $72,388 million change in the fair value of convertible notes and similar instruments. Anthropic’s company-defined non-GAAP Adjusted Net Income was $2,671 million for the same period; it excludes stock-based compensation, the Bartz settlement and those fair-value changes, among other items.
Compute deals at industrial scale
As filed: at least $110.0 billion with Amazon from May 2026 to April 2036, at least about $114.5 billion with Google from April 2026 to July 2033 and at least $44.9 billion with Microsoft from November 2026 to May 2033, all minimums; up to $120.5 billion with SpaceXAI into 2029; Broadcom equipment-lease rent of about $161.2 billion excluding interest. The S-1 leaves the total blank, and the figures are of different kinds.
Agents as products
The S-1 calls Claude Code the product that has most concretely defined autonomous AI work. Claude Cowork, in research preview from January 2026, extends the same agentic design to knowledge workers, and its adoption outpaced Claude Code’s at the equivalent point.
Investors: track compute and infrastructure as a share of revenue each quarter; the S-1 presents expenses by nature and has no gross profit line, so this is the clearest unit-economics measure it gives.
Analysts: reconcile company-defined Adjusted Net Income to the GAAP net loss before you compare Anthropic with other companies.
Builders: price your own AI product on tokens consumed, because your largest input cost moves with model choice and volume.
Compute suppliers and lenders: read each deal by type. Amazon, Google and Microsoft are minimums, SpaceXAI and Nscale are caps, Lambda fees accrue as capacity is delivered, Broadcom is lease rent and AMD has no fixed amount.
Operators: watch concentration; about 79% of revenue in the third quarter of 2026 came from customers outside the top five.
The Business Engineer lens
Anthropic’s AI business model is a bet that revenue per unit of compute rises faster than a compute bill committed years ahead. The Business Engineer’s view is that the S-1 is the first public evidence of that crossover: compute fell below revenue in the first quarter of 2026, and operating income followed in the second.
It commits to compute years ahead, trains Claude models on it, and sells their output by the token through its platform, apps and cloud marketplaces, with some apps on subscription.
How much does Anthropic spend on compute?
Compute and infrastructure expense was $7,326 million in 2025 and $8,945 million in the six months to 30 June 2026; in the second quarter of 2026 it equalled 48% of revenue.
Is Anthropic profitable?
It reported operating income of $1.0 billion for the second quarter of 2026, but a GAAP net loss of $80,748 million for the first half, mostly from fair-value changes on convertible notes.
Who supplies Anthropic's compute?
The S-1 describes agreements with Amazon, Google, Microsoft, SpaceXAI, Nscale, Lambda, Broadcom and AMD, across AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs.
The complete explainer, cleaned up and reformatted for mobile.
BUSINESS MODEL
Anthropic AI Business Model
Anthropic sells use of its Claude models (Mythos, Opus, Sonnet and Haiku) by the token, through its own platform, its apps and cloud marketplaces, and sells some apps on subscription. Its S-1 shows revenue of $16,286 million in the six months to 30 June 2026, 80% of it consumption-based.
Anthropic’s S-1 shows a compute-to-revenue business: compute and infrastructure expense fell from 251% of revenue in the first quarter of 2025 to 48% in the second quarter of 2026, a quarter with operating income of $1.0 billion.
Exec Package + Claude OS Master Skill | Business Engineer Founding Plan
FourWeekMBA x Business Engineer | Updated 2026
Anthropic sells use of its Claude models (Mythos, Opus, Sonnet and Haiku) by the token, through its own platform, its apps and cloud marketplaces, and sells some apps on subscription. Its S-1 shows revenue of $16,286 million in the six months to 30 June 2026, 80% of it consumption-based. Partnerships with organizations like Slack, Quora, and Notion broaden applications.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
