Anthropic files S-1 for IPO at $2 trillion valuation: 85% margin, profitable Q2

Anthropic has publicly filed its S-1 and plans to list on Nasdaq under the ticker ANT. The expected valuation is about $2 trillion, and the offering was forecast for around November 9. Second-quarter revenue came in at $11.6 billion, up 1368% year over year. The quarter's operating profit was $974 million, and gross margin for the first half was 85%. The company has raised $141.1 billion in total, and estimates of its cumulative burn stand at $9.7 billion. • 48% of H1 2026 revenue ran through Bedrock, Vertex and Azure Marketplace • Commitments for NVIDIA compute via SpaceXAI: up to $120.5 billion • Pretraining of a new frontier model has been postponed

AI-processed from Latent Space (swyx); edited by Hamidun News
Anthropic files S-1 for IPO at $2 trillion valuation: 85% margin, profitable Q2
Source: Latent Space (swyx). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceLatent Space (swyx) ↗
Editor-in-Chief
Key metric #148% of H1 2026 revenue ran through Bedrock, Vertex and Azure Marketplace
Key metric #2Commitments for NVIDIA compute via SpaceXAI: up to $120.5 billion
Key metric #3Pretraining of a new frontier model has been postponed

Anthropic published a 350–400-page S-1 on Monday at around 9 a.m. New York time. The company plans to list on Nasdaq under the ticker ANT. Revenue grew from $41 million in 2023 to $16.3 billion in the first half of 2026 alone. In the second quarter it reached $11.6 billion, up 1368% from a year earlier. For the first time in the company's history, a quarter closed with an operating profit: $974 million. Operating cash flow for the half-year is also positive, at +$3.9 billion. The market values the IPO at roughly $2 trillion: after the filing appeared, the expected closing price on Polymarket slipped from $2.1 trillion to $2.0 trillion.

Margin that climbed from negative to 85%

The key figure for anyone calculating the cost of a token is the trajectory of gross margin:

• 2023: 37% • 2024: −95% • H1 2025: 12% • 2025 overall: 40% • H1 2026: 85%

Analyst Aymeric Roucher derived $2.08 of revenue for every dollar of compute from the filing. That implies a compute payback period of under a year.

One caveat: cloud marketplace fees are booked under sales, marketing and partnerships rather than cost of revenue, so they are not included in the 85%. In H1 2026, 48% of revenue, about $7.8 billion, ran through Bedrock, Vertex and Azure Marketplace (in 2023 it was 11%).

Spending on partner channels rose by about $2.7 billion over the year. The hyperscalers, in effect, get paid twice: for compute and for distribution.

The falling "token price index" in the S-1 is also being misread: it mainly reflects the mix of Haiku and Opus 5.5 releases, not a cut in list prices. The company does not disclose margins for individual models.

$141 billion raised, $9.7 billion burned

Anthropic has raised $141.1 billion in total. Its accounts hold $112.4 billion in cash and $19.3 billion in short-term investments. SemiAnalysis puts cumulative burn at $9.7 billion and calls it "probably the most efficient scaling of a business in the history of capitalism." Austen Allred gives $9.4 billion, and cash figures vary between $132 billion and $134 billion across sources. Derived cash-flow estimates are best treated with caution.

The structure of the capital raised, according to Jamin Ball's breakdown:

• about $128.6 billion in preferred stock; • $11.3 billion in convertible notes from Amazon and Google (2023–2025), of which about $5 billion has already converted; • the remaining $6.2 billion of principal is carried as a liability with a fair value of $108.6 billion, because it converts at $2–5 per share while the Series H price is about $59; • about $0.6 billion in SAFEs and common stock, and about $0.7 billion in pre-2023 rounds.

The revaluation of these notes is what produces the headline H1 net loss of $80.7 billion: it includes $81.1 billion in non-cash charges on financial instruments.

Compute commitments and blind spots

The flip side is non-cancellable compute contracts running into the next decade. They include up to $120.5 billion for NVIDIA capacity via SpaceXAI under agreements through 2029. Sarah Guo points to the mismatch: revenue is consumer-driven, while the obligations are multi-year. Skeptics add that a significant share of the money comes from Amazon and Google, which are at once investors, suppliers and competitors. Anthropic also acknowledged that it recently postponed pretraining of a new frontier model. Teortaxes considers this rational: a huge model trained on 100T tokens is wasteful, and Vera Rubin systems will get more expensive.

Some operational details: internal Claude spending exceeded monthly payroll costs (excluding stock) as early as April, and at list prices it grew more than 250x from January to September 2026. Claude wrote more than 90% of contributions to the codebase in September. The S-1 says nothing about in-house chips: the risk section says only that competitors may vertically integrate through their own designs, and that Anthropic depends on third-party design houses for "some custom accelerators." This contradicts Dylan Patel's claim that the company is already making its own chip.

For shareholders: the lockup extends beyond a year, and some holders may be able to sell from day one. 5% of shares go to the Public Benefit Fund, and the board may add up to 2% more with each 50% increase in value. The founders have pledged to give away 80% of their stake.

The strategic pitch is "virtual colleagues" (the document avoids the word "agent") for more than a billion knowledge workers by 2028. They receive access rights, connections to systems and organizational context, and carry out multi-step tasks at a professional level.

If you build a product on Claude, three things from the S-1 are worth taking away. Token prices will likely stabilize: the margin is already 85%, and the pressure is more likely to come as volume discounts. The main purchasing channel is shifting to cloud marketplaces, which means model spend is easier to route through existing committed spend at AWS, Google and Microsoft. And compute remains the main constraint on Anthropic's growth, so factor that into your plans for rate limits and queues.

Story

Key Questions Answered

What are the key details on Margin that climbed from negative to 85%?

The key figure for anyone calculating the cost of a token is the trajectory of gross margin: 2023: 37% 2024: −95% H1 2025: 12% 2025 overall: 40% H1 2026: 85% Analyst Aymeric Roucher derived $2.08 of revenue for every dollar of compute from the filing. That implies a compute payback period of under a year. One caveat: cloud marketplace fees are booked under sales, marketing and partnerships rather than cost of revenue, so they are not included in the 85%. In H1 2026, 48% of revenue, about $7.8 billion, ran through Bedrock, Vertex and Azure Marketplace (in 2023 it was 11%). S

What are the key details on $141 billion raised, $9.7 billion burned?

Anthropic has raised $141.1 billion in total. Its accounts hold $112.4 billion in cash and $19.3 billion in short-term investments. SemiAnalysis puts cumulative burn at $9.7 billion and calls it "probably the most efficient scaling of a business in the history of capitalism." Austen Allred gives $9.4 billion, and cash figures vary between $132 billion and $134 billion across sources. Derived cash-flow estimates are best treated with caution. The structure of the capital raised, according to Jamin Ball's breakdown: about $128.6 billion in preferred stock; $11.3 billion in co

What are the key details on Compute commitments and blind spots?

The flip side is non-cancellable compute contracts running into the next decade. They include up to $120.5 billion for NVIDIA capacity via SpaceXAI under agreements through 2029. Sarah Guo points to the mismatch: revenue is consumer-driven, while the obligations are multi-year. Skeptics add that a significant share of the money comes from Amazon and Google, which are at once investors, suppliers and competitors. Anthropic also acknowledged that it recently postponed pretraining of a new frontier model. Teortaxes considers this rational: a huge model trained on 100T tokens i

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…