Introducing Benchmark Breakfast with Jack & Ev
Anthropic's S-1 shows frontier AI labs are profitable, code generation is a historic product-market-fit moment pulling whole categories upward, open and closed models coexist, and cyber risk is urgent while bio and robotics lag on feedback loops.
Watch on YouTube ↗Charlie's summary
- ▶ 2:28
Anthropic's S-1 shows frontier labs are money-printing machines, not dollar-for-90-cents businesses
The key chart compares revenue with compute spend. In Q1 2025 Anthropic had about $400 million of revenue against $900 million of compute, a 4x ratio with compute larger. By Q2 2026 it had $11.6 billion of revenue against $5.6 billion of compute, a 2.1x ratio the other way. The speaker says Anthropic did in six quarters what takes software companies a decade, with 8% operating margins. The open question is growth: 145% quarter over quarter slowing to 45%, with 80% consumption-based revenue.
- ▶ 7:01
Buyers pay for the best model because many tasks have unbounded upside
Ev says tasks fall into two types: those that just need to be completed correctly, like filing taxes, and those with unbounded upside, like writing code, sales, drug discovery and robotics. Most tasks people care about are the second kind, so buyers keep choosing the best model even if the last one could do most of the work. Jack adds that token sellers are doing well, but buyers outside code still have to prove they are more productive.
- ▶ 10:06
Code generation is the biggest product-market-fit moment in tech, and it keeps pulling whole categories up
The speaker lays out a chain of derivative trades. In public markets after ChatGPT, Nvidia was followed by servers, cooling, power, memory and CPUs. In venture, Opus 4.5 and Claude Code in November 2025 created a codegen PMF wave. It lifted backend, deployment, inference clouds, data, sandboxes, search, code review and CI/CD companies, many going from 10 to 100 to 300. Unlike Zoom, codegen has no obvious ceiling. New companies build software with agents in a software factory, so growth could double yearly for a while.
- ▶ 18:18
The code pull distorts venture: growth that once excited investors now looks weak
Because capital and attention chase code-adjacent categories, other good businesses find it harder to raise. Datadog went roughly 5 to 18, which would once have excited everyone. Now, the speakers say, investors would ask why it was not 5 to 40. Ev says durability is also hard to call right now, since application software was pronounced dead six months ago and is alive today. So investing means either riding the trend or calling the next one, such as robotics or bio, where returns are highest.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.