Anthropic 提交 IPO 招股書，披露營收急升與龐大算力承諾

Anthropic filed its IPO prospectus, and the market seems to like it.

The AI lab is the first major company of its kind to publicly file an S-1 with the SEC. Anthropic plans to list on the Nasdaq under the ticker "ANT," with Goldman Sachs and Morgan Stanley leading the offering. Wall Street expects a valuation of roughly $2 trillion, according to Yahoo Finance, up from the company's most recent private valuation of about $965 billion, which already put it ahead of rival OpenAI at roughly $852 billion.

The first-half numbers show huge growth. Revenue jumped from $376 million in Q1 2025 to $11.6 billion in Q2 2026, a 1,368 percent year-over-year increase. For the first half of 2026, Anthropic brought in $16.3 billion total, more than triple its full-year 2025 revenue of $4.6 billion. Government contracts account for only about one percent of that total, while roughly 5 to 7 percent comes from customers who violate usage policies, primarily through model distillation, which Anthropic has already been cracking down on.

In Q2 2026, Anthropic posted its first positive GAAP operating profit at $974 million. Compute costs rose 300 percent over the same period, but that's slower than revenue growth. Revenue per employee sits at $6.6 million, higher than any Magnificent 7 company except Nvidia, with about 5,300 people on staff. Chip stocks like Nvidia, Broadcom, and AMD rallied after the filing went public.

Existing customers spend nine times more year over year

Anthropic's net revenue retention rate sits at roughly 900 percent, meaning existing customers are spending nine times what they did a year ago. That figure speaks to AI's scaling potential, especially considering most deployments are likely still shallow, depend on human oversight, and don't run AI at full speed yet.

More than 11,900 customers pay over $100,000 annually, and about 250 spend more than $10 million. No single customer accounts for more than ten percent of revenue, but five customers together make up 25 percent. If those include companies like Meta, Google, or Microsoft that could switch to their own or cheaper models, that concentration could become a risk.

Claude Enterprise users who also use Claude Code are the biggest spenders, paying about 20 times more than other customers, according to the S-1. Eighty percent of revenue is billed on actual usage, per token processed as the smallest billing unit for AI models, while the remaining 20 percent comes from fixed subscriptions. That split might signal how heavily subscriptions are still being cross-subsidized, but it also suggests Anthropic has room to raise prices. The company's forecast for next year will be primarily usage-based, according to the FT.

Compute commitments top $780 billion

Anthropic has taken on huge financial obligations to fuel its growth. Compute capacity contracts with Broadcom, Google, Amazon, Microsoft, and other providers add up to roughly $780 to $810 billion, and those costs are more or less fixed and locked in long-term, while revenue remains variable and usage-dependent.

That means the company needs to keep growing fast to cover its future bills. If it can't, the AI-fueled capital market could start to wobble. To gain more flexibility and lower costs over time, Anthropic is building its own data centers under the name "Anthropic Direct".

Anthropic also left Q3 numbers out of the S-1, even though the company already has them. One possible reason is that growth slowed significantly in Q3. Momentum may have already picked back up, though. OpenRouter data shows demand for Anthropic models climbed again after the October release of the 5.5 series, especially Opus 5.5. OpenRouter only captures a tiny slice of total token demand and tends to skew toward open-weight models, so if Anthropic is dominating there, it's likely doing even better everywhere else.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.