Anthropic filed its Form S-1 with the SEC on October 26, 2026, about nineteen days ahead of a rumored initial public offering at a valuation near $2.2 trillion. The number that traveled fastest was a net loss of $80.7 billion for the six months ended June 30, 2026. The number that matters more sits in a footnote: $72.4 billion of that loss is a non-cash accounting charge tied to the rising value of Anthropic's own convertible instruments. Strip it out and the picture is a company with $16.3 billion in first-half revenue, more than $130 billion in liquid reserves, and compute commitments of roughly half a trillion dollars running through 2045. Anyone reading an AI IPO filing needs to hold all three of those facts at once.

The headline loss and what sits beneath it

Anthropic's consolidated statement of operations shows a net loss of $80,748 million for the first half of 2026. Footnote 2 explains that $72,418 million of this comes from the change in fair value of convertible promissory notes, derivative liabilities and SAFEs. A SAFE, or Simple Agreement for Future Equity, is a contract that gives an early investor the right to shares at a later round. When a private company's valuation jumps across successive funding rounds, the obligations owed to those early investors become more valuable, and accounting rules require the company to book that increase as a loss against earnings even though no cash leaves the building.

The S-1 also provides a non-GAAP reconciliation that walks from the statutory loss to an adjusted figure:

Item Amount
Net loss (GAAP), H1 2026 -$80.7 billion
Add back: fair-value change on convertible notes, derivatives, SAFEs +$72.4 billion
Add back: mark-to-market financial instruments +$8.7 billion
Add back: stock-based compensation +$2.2 billion
Add back: non-cash charitable contributions +$0.1 billion
Adjusted net income (non-GAAP) +$2.7 billion

Excluding only the convertible-instrument markups and leaving other costs in place, the core operating loss for the half appears to land somewhere between $7 billion and $9 billion. That is still a large number, but it is of a different order than $80 billion, and the fair-value charge should disappear once the company is public and those instruments convert to ordinary shares.

Two cautions apply. Non-GAAP figures are the company's own presentation, and details of revenue recognition and other expenses deserve further scrutiny as more disclosures arrive. The adjusted profit is a reconciliation, not cash in the bank.

Revenue growth and the compute turnaround

The revenue trajectory in the filing is steep by any standard:

Period Revenue
2023 $41 million
2024 $386 million
2025 $4.6 billion
H1 2026 (six months) $16.3 billion

For the three months ended June 30, 2026, revenue grew 1,368 percent year over year while compute and infrastructure expense grew 300 percent. That gap is the source of the operating leverage the filing emphasizes. The clearest way to see it is revenue generated per dollar of compute spend, which the company reports quarter by quarter:

Quarter Revenue per $1 of compute
Q1 2025 $0.40
Q2 2025 $0.57
Q3 2025 $0.68
Q4 2025 $0.69
Q1 2026 $1.40
Q2 2026 $2.08

Six server racks of rising height with growing coin stacks on top and an upward arrow drawn on a whiteboard behind them

▲ Revenue per dollar of compute, Q1 2025 to Q2 2026

In early 2025 Anthropic lost money on every dollar of compute it bought. Five quarters later it earned $2.08 for each dollar, a 5.2-fold improvement. Research firm SemiAnalysis framed the broader result as the company having raised $141.1 billion in aggregate capital as of June 30, 2026, while burning only $9.7 billion net to date, calling it likely the most efficient scaling in capitalist history. Whether that efficiency holds as compute commitments scale up is the central open question.

Who is paying, and how

The customer base is concentrated at the top. As of September 30, 2026, the filing counts about 250 customers spending $10 million or more on a trailing-twelve-month basis, more than 1,900 spending $1 million or more (up from over 950 a year earlier), and more than 11,900 spending $100,000 or more. Those tiers have doubled or tripled since Anthropic's chief financial officer shared earlier figures a few months before. Concentration cuts both ways: it reflects deep enterprise commitment, but the loss of a handful of mega-accounts would be felt.

Net revenue retention stands at roughly 900 percent, up from about 300 percent three months earlier. In plain terms, a customer spending $1 million on January 1 is on a run-rate of about $9 million by December 31. Revenue per employee reached $6.6 million on a trailing-twelve-month basis, up from $1.4 million a year earlier.

A striking share of this revenue arrives through cloud marketplaces rather than direct contracts. The portion of customer revenue contracted through Amazon Web Services, Google Cloud Platform and Microsoft Azure marketplaces rose from 11 percent in 2023 to 32 percent in 2024, 47 percent in 2025 and 48 percent in the first half of 2026. Enterprises prefer to buy against existing cloud budgets and pre-approved security and procurement agreements. The trade-off is structural: the same hyperscalers hosting Claude are also building competing models and chips, so Anthropic depends on partners who are also rivals.

Cash on hand versus $500 billion in commitments

Earlier rumors suggested Anthropic was running close to zero cash. The S-1 shows otherwise: $112.4 billion in cash and cash equivalents plus $19.3 billion in short-term investments as of June 30, 2026, roughly $132 billion in liquid reserves. The filing states that existing cash, equivalents and credit facilities will fund operations and capital expenditures for at least the next twelve months. One market commentator has projected that an IPO raising about $140 billion, plus $77 billion in committed partner funding and continued revenue growth, could push reserves near $400 billion by the end of 2027. That is a projection, not a filing figure.

Against that balance stand compute commitments that run to 2045:

Provider Commitment
Broadcom (TPU leases) $161.2 billion
SpaceXAI (NVIDIA GPUs, through 2029) up to $120.5 billion
Google (TPUs) at least $114.5 billion
Amazon at least $110.0 billion
Microsoft at least $44.9 billion
Nscale up to about $44.6 billion
Lambda about $41.3 billion
AMD no fixed amount; more than $20 billion if the full first tranche is taken

Roughly $245 billion of this comes from new and amended deals signed after June 30, 2026.

Half-open vault of currency bundles and a conveyor belt carrying crates to eight warehouses linked to a central office

▲ Cash reserves and compute commitments across eight suppliers

The SpaceXAI terms deserve a close read. SpaceXAI LLC, formerly xAI LLC, has become a large neocloud supplier by securing bulk NVIDIA allocations, and is reported to be bringing 200,000 to 300,000 GPUs online each month, about $13.1 billion worth of compute monthly. Anthropic's agreement works out to about $3.1 billion per month once capacity is fully online in December 2026. The contracts are non-cancelable until January and March 2027, after which they can be ended with 90 days' written notice. With GPUs scarce, the supplier appears to hold real pricing leverage during the locked period.

Google's position is unusual. It holds an estimated 17 to 21 percent stake in Anthropic and is committing at least $114.5 billion in TPU capacity to a company that competes directly with Gemini. Amazon, with an estimated 12 to 14 percent stake, backs its investment with a similar compute commitment. The diversification across NVIDIA GPUs, Google TPUs and custom ASICs mirrors the approach OpenAI has taken and reduces dependence on any single chip supplier, though it does not reduce the total bill.

What the filing says about products and the roadmap

Beyond the numbers, the S-1 reveals that Claude Cowork, an agentic workspace that lets non-technical workers run multi-step file and desktop tasks without code, is growing adoption faster than Claude Code, the developer tool that has driven most of Anthropic's token consumption. Cowork launched as a research preview in January 2026 and was folded into Claude.ai in September 2026. When a user submits a work-related prompt, the system routes it to the Cowork agentic backend automatically. This appears to be both a user-experience simplification and a deliberate way to drive backend usage; its effect on earnings will only become visible in public quarterly reports.

The filing also forecasts "Powerful AI" within a few years, defined as systems matching or surpassing leading human experts across nearly all scientific and professional fields, with direct computer control and multi-week autonomous task execution. A section titled "Closing the Loop" describes current Claude models running experiments, generating training data, conducting safety audits and evaluating outputs for their successors. Anthropic co-founder Jack Clark has said the industry is entering the takeoff phase of this recursive paradigm, and Google DeepMind's Demis Hassabis has projected meaningful recursive self-improvement within six to nine months. Translating such capability into drug design or disease intervention is likely to be slower and harder than benchmark progress suggests.

The growth strategy section contains a line worth remembering: "the gap between what Claude can do and how broadly it is used represents one of our largest growth opportunities." The argument is that existing models already hold enough untapped utility to generate billions in additional annual value, even before new releases.

The risks the filing itself flags

The S-1 states plainly: "We do not expect our revenue will continue to grow at comparable rates." A 1,350 percent year-over-year pace cannot persist. Investors including Chamath Palihapitiya and Apollo Global Management have noted that a meaningful share of enterprise generative AI revenue currently comes from venture-funded startups rather than self-sustaining businesses, which raises the question of how durable that spending is if venture funding slows. Physical constraints such as grid capacity, data center power, cooling and geopolitical friction could delay delivery of the compute Anthropic has contracted. And because every dollar of Anthropic revenue depends on generative AI, it is a far more concentrated bet than Alphabet or Meta, which can absorb AI losses with other businesses. A $2.2 trillion valuation may be stretched relative to current fundamentals even if the underlying business is stronger than skeptics expected.

How to read an AI IPO filing

The Anthropic S-1 offers a template for judging any AI company heading to market:

  1. Start with the footnotes. Separate statutory net loss from non-cash charges such as convertible-note and SAFE revaluations, then estimate operating loss on its own.
  2. Track revenue per dollar of compute. A business that earns less than it spends on hardware is subsidizing usage; one that earns more has operating leverage.
  3. Check cash and liquid reserves against the schedule of fixed commitments, and note which contracts are non-cancelable and for how long.
  4. Look at who the customers are. Ask how much revenue comes from large concentrated accounts, how much flows through partner marketplaces, and how much depends on venture-funded buyers.
  5. Read the company's own growth warnings and physical-constraint disclosures as seriously as its growth charts.

Anthropic's filing shows a company that is losing far less cash than its headline implies, growing faster than its compute bill, and holding reserves comparable to established tech giants, while carrying obligations larger than any venture-backed company has taken on before. Whether the IPO price makes sense is a separate judgment that each reader must make, ideally after reading the footnotes rather than the headlines.