Anthropic Pre-IPO Full Disclosure: 85% Gross Margin, Only 30% of Its Compute Is Generating Revenue

Anthropic has submitted its S-1 filing to the SEC, paving the way for a Nasdaq listing. With a valuation range between $1.5 trillion and $2 trillion, it will immediately rank as one of the most valuable public companies on the planet once it goes public.

But the hundreds of pages of documents are not focused on the valuation itself. For the first time ever, they lay bare the full financial inner workings of a cutting-edge AI lab: revenue structure, gross margins, compute costs, customer concentration, and how a company claiming to build AGI frames its growth story for public markets.

Joey Brookhart, analyst at SemiAnalysis, breaks down the filing in The Verge's Decoder podcast. His team previously modeled Anthropic's finances from external data, and now that they have access to audited official figures, several key findings stand out.

85% gross margin, down from -94% just two years prior

Anthropic's annualized revenue now exceeds $700 billion. Its inference gross margin hits 85%, a dramatic turnaround from the -94% figure recorded two years ago. The shift to profitability came far faster than most industry observers predicted.

Revenue per megawatt of inference capacity currently sits at roughly $65 million, up from around $15 million in the same period last year. SemiAnalysis' internal model shows that running Opus 5.5 on Nvidia Blackwell chips can push per-megawatt revenue as high as $100 million to $120 million, while average compute costs per megawatt sit just above $10 million.

The jump in inference efficiency comes from combined optimizations across software and hardware. Token throughput on Trainium chips has risen 5 to 10 times year-over-year. Anthropic runs a very diverse chip stack, utilizing Trainium, TPUs, and GPUs all at once.

The $80 billion loss costs the company zero actual cash

The most eye-catching figure in the S-1 is the $80 billion loss posted for the first half of 2026. This is a purely paper loss unrelated to core operations: it stems from the meteoric rise in the company's valuation, which requires marking up the fair value of previously issued warrants and convertible bonds on the income statement. It is 100% non-cash, and no actual funds were spent to generate this loss.

Anthropic has only burned $100 billion in actual cash to date, and still holds more than $1.2 trillion in cash and marketable securities on its balance sheet. It expects to raise an additional $100 billion in proceeds after the IPO.

Only 30% of total compute capacity drives revenue

Even more revealing is Anthropic's compute allocation. Only roughly 30% of its total compute pool is dedicated to revenue-generating inference services. The remaining 70% is directed to training, post-training fine-tuning, and R&D experiments. This is the classic operational structure for frontier AI labs: monetize current generation models to pour the vast majority of resources into building next-generation systems.

The core problem here is that this business's revenue is deeply tied to the novelty of its models. The S-1 discloses that 80% of its revenue in the most recent three months came from models released in the prior six months. Older models lose their value extremely quickly. The recurring subscription logic that works for traditional software businesses does not apply here: this is a usage-driven consumer-style business, where users vote directly with their consumption choices.

Its customer base is also highly skewed. Half of the Fortune 500 are Anthropic customers, but the average spend from these clients only accounts for 0.1% of their total operational expenditure. The top 5% of Claude Code users spend an average of $18,000 per year, while non-Claude Code users spend less than $75 per month. Its top 1% of enterprise clients spend $50,000 to $100,000 per employee per year. Traditional enterprise IT procurement budgets have not fully opened up to AI spending yet.

Competition is heating up, and model switching costs are extremely low

Anthropic captured 90% of the industry's net new ARR in the first half of 2026. That share dropped sharply in Q3, as OpenAI took back most of the new incremental revenue. Open source model inference revenue jumped from $1 billion in Q1 to over $10 billion in Q3.

SemiAnalysis' own internal usage data serves as a perfect microcosm of this trend: at the end of Q2, over 90% of its workload ran on Anthropic models. After OpenAI launched ChatGPT 5.6 and Astra, that share dropped to 70%, before bouncing back above 90% after Anthropic released Opus 5.5. Switching costs are negligible in code generation use cases.

Anthropic's net dollar retention rate stands at 900%, which looks staggering on paper. But this reflects its pure usage-based business model: over a two-week holiday period, total token spending dropped 20% year-over-year. There are no multi-year enterprise contracts, no enforced 95% renewal rates, customers only pay for exactly what they consume.

Cloud providers are both landlords and direct competitors

Anthropic does not own any of its compute infrastructure, it leases all of it. 48% of its total business runs through third-party cloud marketplaces: AWS Bedrock, Google Vertex, and Azure Foundry, where cloud platforms take a 16% cut of revenue. Google is both its TPU supplier and the developer of competing Gemini models, while AWS acts as its infrastructure landlord and also pushes its own in-house foundation models. The S-1 explicitly flags that this conflicting relationship could turn hostile at any time.

Anthropic's hedge against this risk is to buy chips directly. It is purchasing 3.5 to over 4 gigawatts of TPU capacity via a dedicated SPV, then leasing the hardware back for its own use, moving steadily toward a rent-to-own compute model.

The S-1 also discloses $500 to 600 billion in total compute commitments stretching across 7 to 10 years. Brookhart's analysis finds that at an expected input cost of $11 million per megawatt, the total scale of these committed investments is not unreasonably large relative to Anthropic's projected future revenue — assuming its top line keeps growing at its current pace.

This is also its biggest bet. Dario Amodei has previously stated he sees a clear path to $1 trillion in annual revenue. The S-1 does not lay out a timeline for that target. The path to $1 trillion runs through full automation of white-collar work, expanding from coding use cases into legal services, finance, cybersecurity, and biotech. All of these vertical use cases are still in beta testing stages today.

OpenAI's S-1 will tell a very different story

OpenAI only shifted its business mix from consumer-focused to enterprise-led in August 2026. Its Q3 enterprise revenue grew 100% quarter-over-quarter, with total overall business up 77% in the same period. The Wall Street Journal reported OpenAI runs a -100% EBITA margin, compared to Anthropic's positive 11% EBITA margin.

Brookhart's assessment is that OpenAI's financials will not look nearly as strong as Anthropic's, but it has a far larger and more mature enterprise sales team that will deliver solid Q3 results. OpenAI has pushed its S-1 filing back to early 2027, almost certainly to give its financial metrics more time to improve.

Key points to watch during the IPO roadshow

Anthropic will have to explain why its Q3 performance underperformed expectations. Meta was once a major customer accounting for over 10% of its total revenue, but Meta cut its Claude Code seat count from 60,000 to 30,000, shifting almost all of its token workload to its own in-house models. Anti-distillation protection measures also reduced its revenue by 5% to 7%.

Net new monthly ARR hit $10 billion in Q2, before dropping to $4 billion in Q3. Public markets want clarity on whether it can bounce back to $8 billion to $10 billion in new monthly ARR in Q4, as well as its full 2027 growth roadmap. There is also an unavoidable core contradiction: the same company that publishes papers warning Claude could gain consciousness and pose existential risks to humanity also has to deliver predictable quarterly earnings guidance to public investors.

One critical question the S-1 does not answer: is the $2 trillion valuation priced on the cash flows Anthropic would generate if it successfully builds AGI, or the far more mundane 85% gross margin inference business it operates today? The former is a bet on faith, the latter is simple arithmetic. Public markets will price both, just to different degrees.