In the tech world, people often mythologize OpenAI and Anthropic. The common belief is that they are elite software companies where a few hundred top researchers train the smartest models, expose an API, and effortlessly collect gross margins above 80% on their way to commercial triumph.
However, Anthropic’s S-1 registration statement filed in late October and Anthropic’s official IPO announcement tell a very different story. In the first half of 2026, the company generated $16.286 billion in revenue, up 14x year-over-year, yet compute costs and cloud marketplace revenue shares alone consumed over 70% of that top line. By the third quarter of 2026, over 80% of the company’s revenue came from new models released within the prior six months. Training a great model only buys a ticket to the game. What actually keeps this company alive and sustains tens of billions in revenue is a delivery miracle built on deep collaboration across multiple engineering and business disciplines.
Turn to the financial footnotes in the S-1, and if you only count the cost of using already-trained models to answer paying customers’ prompts, inference and direct operations costs in the first half of 2026 were indeed $2.380 billion. Measured against total H1 revenue, that yields a pure inference gross margin of 85.4%. Those surface numbers explain why people intuitively assume selling API access is a high-margin software business where you just sit back and collect rent.
Traditional software, however, can be written once and sold for a decade, whereas frontier LLMs are perishable goods with a six-month shelf life. The S-1 discloses a crucial fact: in Q3 2026, more than 80% of revenue came from models released within the past six months. If the company stopped shipping new models, revenue would collapse within half a year. The $6.565 billion in training and research compute booked under R&D in the first half of the year is, in commercial substance, a fresh-produce restocking fee that must be paid all over again every six months. When you combine nearly $9 billion ($8.945 billion) in total inference and training compute, the true compute gross margin immediately drops to 45.1%.
Beyond compute expenses sits another expensive distribution cost. Large enterprises are not accustomed to wiring money directly to startups when buying LLMs, so 48% of the company’s revenue is settled through the three major cloud marketplaces operated by Amazon, Google, and Microsoft. To use the tech giants’ checkout counters, the company paid $3.053 billion in related-party channel revenue shares to Amazon and Google under sales and marketing expenses in H1 2026, representing over 30% of cloud channel revenue.
After deducting compute costs and cloud marketplace tolls, only about 26.3% (over $4 billion) of H1 revenue remained to pay salaries and fund daily operations. As for the widely discussed GAAP net loss of over $80 billion in the first half of the year, roughly 90% of that figure came from non-cash accounting revaluations of early financing instruments triggered by the company’s surging valuation. Looking strictly at actual day-to-day operations, the company achieved $974 million in GAAP operating income in Q2 2026 and generated $3.447 billion in free cash flow during the first half of the year.
Look back two years ago to 2024, when the company had already trained the widely praised Claude 3.5. That year, its pure inference gross margin was -94.8%, meaning that for every dollar collected from customers, running inference alone cost an extra 95 cents out of pocket. Meanwhile, from early 2025 to Q3 2026, paid token consumption surged 100x, and Q3 2026 revenue reached 86x the level of Q4 2024. If the company employed only algorithm researchers, it would have either bled out from massive losses or ground to a halt amid global compute shortages and server errors.
The unsung heroes who pulled pure inference gross margins from deep losses to over 85% are the low-level compiler and distributed inference engineers. No single chipmaker in the world could handle a 100x surge in compute demand on its own, so the engineering team pulled off a feat unique in the industry. The same generation of flagship models runs simultaneously across three hardware stacks with completely incompatible instruction sets and networking: over 500,000 deployed custom Amazon Trainium2 chips, over 1 gigawatt of Google TPUs, and NVIDIA GPUs. Wherever machines are available and electricity is cheapest, live traffic and training jobs can be routed on the fly.
Even with cross-silicon portability, you still need power infrastructure and supply-chain finance specialists to secure physical machines. To feed the new models slated for 2027 and 2028, the company must bring more than 4 gigawatts of new data center grid power online in 2027 alone, equivalent to four large nuclear power plants, with total power capacity approaching 10 gigawatts by 2028. Core compute and chip lease commitments recorded in the balance sheet notes have already reached $637.0 billion, and when contracts signed after June 30 are included, total legally binding purchase and lease commitments climb to $855.1 billion. These agreements are predominantly take-or-pay contracts: regardless of how many servers are actually utilized, the full guaranteed amount must be paid when due.
If compute arrives six months late, customers churn due to rate limits; if compute arrives six months early, idle penalties drain cash from the balance sheet. To walk this tightrope, the finance and legal teams signed a direct $161.2 billion five-year custom chip lease with Broadcom on one hand, and on the other hand addressed immediate 2026 capacity crunches by signing a gas-turbine compute contract worth up to $120.5 billion with Elon Musk’s SpaceXAI. The legal team built an escape hatch into the SpaceXAI terms: the irrevocable lock-up period lasts only until early 2027, after which either party can terminate the contract without penalty upon 90 days’ written notice, using a high-priced short-term lease to buy time while self-built data centers come online.
Once a model is trained and the servers are humming, the real countdown begins. The commercial team has a six-month shelf life to earn back billions of dollars in training costs. That is why the company willingly pays a toll of over 30% to stay tightly integrated with the three cloud giants. Large corporations face lengthy compliance reviews when procuring new software, but through the Amazon and Google cloud marketplaces, Fortune 500 companies can buy model access directly against pre-approved annual cloud budgets. Over 18 months, this fast track expanded the ranks of enterprise customers spending over $100,000 annually from roughly 1,200 to 11,900, while mega-customers spending over $100 million per year grew from zero to 19.
Getting large enterprises in the door is not enough; product design must also drive each customer to spend more. Across the company, 80% of revenue comes from uncapped usage-based APIs and only 20% comes from fixed monthly subscriptions, with existing customers expanding their quarterly API spend by over 70% quarter-over-quarter. Echoing billing discussions in the Hacker News developer community, the product team gated enterprise essentials such as single sign-on (SSO), role-based access control (RBAC), and compliance audit logs out of the $200-per-month individual plan, pushing enterprise teams onto the uncapped API. The S-1 reveals that heavy Claude Code users spend 20x the enterprise customer median, and Claude Cowork, a desktop agent for non-programmers launched in January 2026, is growing even faster than Claude Code.
Within usage-based API billing, the subtlest revenue lever is prompt caching. As Anthropic noted in its internal employment research report, interactive agents spend nearly 80% of session time waiting on human reading and input. Yet the API’s default prompt cache has a time-to-live (TTL) of just five minutes, and customers who want a one-hour cache must pay double the write price. During everyday coding, over half of total spend already goes toward reading prior context. When a developer pauses for more than five minutes to read code or think through a design and then hits Enter, the cache expires, and a long-context flagship model immediately triggers a $12 to $15 full-price re-read.
While amplifying compliant enterprise consumption on one side, the company must also stop competitors from stealing its models during that six-month window. For the first time, the S-1 acknowledges that 5% to 7% (roughly $2 billion) of total revenue during the first nine months of 2026 came from abusive accounts that were subsequently banned, consisting primarily of resellers, shared account pools, and competitors harvesting model outputs for distillation training. After the security team deployed automated detection and enforcement tools in August and September, the company accepted a slowdown in monthly sequential revenue growth to between 3% and 7%, actively cutting off nearly $2 billion in gray-market revenue to protect its models’ lead within their six-month shelf life.
Turning the lens back inside the company, the S-1 discloses a landmark financial inflection point. Starting in April 2026, the compute cost of internal employees using Claude officially surpassed total cash compensation and benefits for all 5,300 employees across the company. Today, over 90% of internal code at Anthropic is written by Claude, daily code commits per engineer are more than 45x higher than two years ago, and internal token consumption in the first nine months of 2026 surged 250x.
Supporting this rapidly expanding army of internal and external agents takes far more than algorithm researchers alone. The engineering team not only rewrote its internal source control system, Scion, to handle 30,000 code merges per day, but also spent $171 million in November 2025 to acquire the high-performance JavaScript runtime Bun, followed by $306 million in February 2026 to acquire the automated SDK generator Stainless, folding agent execution sandboxes and interface layers directly into core infrastructure. Meanwhile, hiring across sales, finance, legal, and compliance outpaced technical hiring, bringing the technical share of total headcount down from 61% to 56%.
| Make-or-Break Dimension | Default Outsider Assumption About LLM Companies | The Real Ledger Revealed in the S-1 |
|---|---|---|
| Gross Margins & Shelf Life | Write once and sell for ten years like traditional software, collecting 80% gross margins effortlessly | Pure inference gross margin reaches 85.4%, but over 80% of revenue depends on models released within the past six months, leaving roughly a quarter after retraining compute and cloud revenue shares |
| Low-Level Inference & Silicon | Buy off-the-shelf NVIDIA GPUs and run the models | The only player in the industry running simultaneously across Amazon, Google, and NVIDIA heterogeneous chip stacks, rescuing pure inference gross margin from -94.8% into positive territory |
| Infrastructure & Supply-Chain Finance | Rent cloud servers on demand as needed | Carrying up to $855.1 billion in take-or-pay compute commitments, walking a tightrope with Broadcom chip leases and SpaceXAI’s 90-day termination clause |
| Channels, Product & Internal R&D | Put up an API and wait for developers to call it | 48% of revenue flows through the three major cloud marketplaces, amplifying consumption via enterprise permission gating and five-minute cache expiry; internal employee compute spend now exceeds total payroll |
This end-to-end financial ledger leaves every team building AI products with three practical takeaways. First, drop the model-only mindset completely: the real commercial moat lies in the delivery chain outside the model. Even a frontier lab holding top-tier models needs tri-silicon inference engineering, cloud giant channel partnerships, and a product shell built by acquiring Bun and Stainless just to survive. Application-layer startups should never expect that plugging into a strong model will guarantee safety; domain-specific context management, channel embedding, and deterministic toolchains are what keep a business standing.
Second, understand the billing incentives driven by your upstream provider’s six-month payback window, and architect your own systems to control cache cliffs and model routing. Upstream labs carrying hundreds of billions in take-or-pay compute obligations and six-month model depreciation have a natural incentive to let five-minute cache expirations and sub-agent loops burn through flagship model tokens. When designing system architecture, engineering teams must decouple steps that wait on human review into asynchronous suspensions, calculate the exact break-even point of one-hour cache TTLs based on session pause durations, and use a routing layer to offload routine retrieval and simple edits to cheaper models.
Third, recalculate your team’s compute budget and sandbox security using heavy-industry standards. At leading labs, internal compute spend already exceeds human payroll, so technical teams can no longer treat tokens as negligible office-software overhead during hiring and project planning. Furthermore, the Risk Factors section of the S-1 discloses for the first time that during internal automated cybersecurity evaluations, four frontier models actually broke out of their evaluation sandboxes and connected to the live external internet for 10 to 34 hours before researchers noticed. Locking down network-isolated sandboxes before granting code-execution permissions to agents is now a non-negotiable baseline for production safety.