Benchmark Breakfast: Reading Anthropic's S-1, Codegen's PMF Moment and the AI Derivative Trade
Key takeaways from Uncapped with Jack Altman — the inaugural Benchmark Breakfast between Jack AltmanDealroom has a profile for this one. Try Dealroom → and Everett RandleDealroom has a profile for this one. Try Dealroom →, general partners at Benchmark, recreating the firm's no-agenda Monday partner meetings (October 2026):
Anthropic's S-1 answered the profit question. The pair walked through Anthropic's filing: revenue rose from about $400 million in Q1 2025 to $11.6 billion in Q2 2026 while compute and infrastructure expenses went from ~$900 million to $5.6 billion — a 4x revenue-to-compute deficit flipped to a 2.1x surplus, with ~8% GAAP operating margins and what the partners call 'money printing machines'.
Why public buyers may still hesitate. Growth is volatile — 145% quarter-over-quarter in Q2, decelerating to 45% in Q3 — and roughly 80% of revenue is consumption-based rather than locked in multi-year contracts, unlike Snowflake or Databricks revenue.
The best-model-wins dynamic. Even when last-generation models are 'pretty good', the leading edge commands outsized momentum — people always buy the higher-intelligence option in unbounded-upside tasks (coding, sales, biology, robotics), a 'Keith Rabois' framing where two-task-type thinking doesn't saturate.
Codegen is the PMF moment. Opus 4.5 supercharged Claude Code in November 2025, creating a mega product-market fit moment for code generation — Benchmark's portfolio (Vercel, Railway, Cloudflare, Fireworks, Base 10, Modal, E2B, Exa, and an early CI/CD GitHub Action bet) are all growing 10-100-300x as code derivatives multiply.
The AI derivative trade. Following the public-market 'egg through a snake': NVIDIA, then servers and racks (Super Micro 10x, Dell), cooling (Vertiv), networks (Arista), power (Vistra 3.5x, Bloom 4x, GE Vernova 2x), memory and storage (SanDisk, Micron, Seagate 3x+), and now CPUs again (Intel +120% in 2026).
Token sellers win, buyers still hope. The partners note the sellers of tokens are clearly doing well, but it's not yet proven the buyers are that much more productive in all cases — 'there's still a lot of hope going on' as diffusion to the rest of the economy plays out.
Software factories. Post-Opus 4.5 startups build with a handful of engineers running many agents at once — a 'software factory' organisational structure that is not even human-driven anymore, which Altman and Randle expect to diffuse across the economy.
Read more: Uncapped with Jack Altman