ai-industry · 2026-11-19 · Tier 3

Anthropic prices its IPO; OpenAI measures the jagged frontier of AI-run research

Anthropic's IPO and OpenAI's research-acceleration data (2026-11-18)

TL;DR. Anthropic priced its IPO at $83 a share, above the $70-78 marketed range, selling 956.4M primary shares (plus 23.6M from selling stockholders) to raise about $81B, valuing the company around $1.5T; trading on Nasdaq as ANT begins 11-19, with a 147M-share underwriter option. The filings put roughly $900.9B of infrastructure commitments against an $81B annualized run rate (Forbes via AI Breakfast). On IPO eve, a "MisAnthropic" statement signed by Max Tegmark, Gary Marcus, Thom Wolf, The Turing Post and others called Anthropic the AI company "doing the most harm", citing its recursive self-improvement (RSI) push and concentration of power; Wolf said he disliked the word "reckless" but worries about transparency and that people declined to sign for fear of losing model access. The same day OpenAI published usage data on AI-run research: the median OpenAI researcher burns about $2,400 of tokens a day at list prices (the top 10% over $23,000), agents run as many hours per researcher as 5.7 people on 8-hour shifts (up 11x), yet experiments rose only 68% over 2025, and agents succeed 79% of the time monitoring training runs but only 25% writing inference and serving code.

Key points

  • The capital cycle, in one chart set (Understanding AI, "19 charts"): Meta, Amazon, Oracle, Google and Microsoft spent $230B capex in Q3 2026 (+118% y/y), 2.3% of US GDP over the first nine months; Amazon, Google, Meta and Oracle now have capex above operating cash flow (a first for Google in 22 years); ecosystem AI revenue run rate $345B in October (4.2x y/y); Anthropic's run rate went from $9B to $81B in 2026, with Q3 slower than Q2.
  • Gary Marcus's bear case argues the bull charts assume revenue keeps compounding and asks readers not to buy on moral grounds; Craig Newmark is holding back AI-safety donations because he expects newly rich Anthropic employees to flood that space (The Information).
  • AI-run research is jagged: the 25% success rate on serving code is the exact layer this wiki's efficiency pages care about, and the weakest link caps the research speedup however many agents run.

How it relates to prior wiki pages

  • Continues 11-18 Nvidia Q3, board and IPO and 11-17 IPO week.
  • Token spend as a productivity input now has two data points that disagree in scale: CodeRabbit's CEO spends about $250K/month personally and ~$150 per PR (Pragmatic Engineer), while Salesforce caps employees at $100-1,000/month (11-18 digest). OpenAI's $2,400/day median sits far above both. See compute-economics.

Sources: Anthropic · The Information · OpenAI research acceleration · @rohanpaul_ai summary · @tegmark · @Thom_Wolf · Understanding AI · Marcus on AI · raw: The Information IPO