Ev Randle says Anthropic's S-1 showed immense growth and operating leverage, with Q2 2026 revenue of $11.6bn against $5.6bn of compute and infrastructure costs.
Anthropic reported an 8% GAAP operating margin in Q2 2026 and grew 145% quarter over quarter, though growth slows to 45% in Q3.
Jack Altman argues the frontier labs have proven their business models work, since revenue now covers training costs, while open-source models may still capture 5% to 10% of AI revenue.
Anthropic's S-1 shows AI labs have working economics
Anthropic's S-1 gives hard evidence that frontier labs are real businesses, with revenue well above compute costs and a positive operating margin. Demand for the most intelligent models keeps compounding, and coding is the clearest case. The open versus closed question, security risk and the slow physical world decide where value lands next.
Takeaways
Anthropic's S-1 shows lab economics now work. Q1 2025 brought about $400m of revenue against about $900m of compute and infrastructure costs. Q2 2026 brought $11.6bn against $5.6bn, a 2.1 times ratio, with an 8% GAAP operating margin. Revenue now covers training.
Buyers keep paying for the smartest model. Anthropic grew 145% quarter over quarter, slowing to 45% in Q3. Durability is hard to judge, but Jack Altman says people pick the highest-intelligence model despite price sensitivity, and Ev Randle puts an individual's value from AI at thousands or tens of thousands of dollars.
Code generation is an iPhone-scale product market fit moment. Claude Opus 4.5 in November 2025 supercharged Claude Code. Ev argues coding has a high ceiling, citing software factories where agents build products. Supabase, ClickHouse, Mercor and Exa are growing fast, and venture pricing in these areas has become obscene.
The AI bottleneck trade runs through public markets. Super Micro, Arista, Vistra, GE Vernova, memory and storage stocks have all risen sharply. Bloom Energy rose four times in 2025 and Intel is up 120% in 2026. Venture investors are running the same bottleneck trade in infrastructure startups.
Frontier models make cybersecurity urgent and fast-moving. CEOs who pointed frontier models at their own systems found company-ending vulnerabilities. A French supermarket attack spoiled food. Jack argues slowing frontier releases would let defenses harden, and he expects security buying to move within twelve months instead of thirty-six.
Robotics and biology will work but move slower. Ev says AI improves fastest where feedback is cheap and verifiable. Robots are costly to deploy, and biology needs lab experiments. Getting AI into physical domains will take Manhattan Project scale effort, including training data infrastructure like Mercor's.
Open versus closed source models
Ev Randle makes the closed case through cost. SpaceX launches a kilogram most cheaply and AWS used scale to cut costs, so the labs with the most demand may be the lowest-cost producers of intelligence.
Jack Altman argues open models carry no amortized training cost and enterprises may want to own their models. His warning case is a hypothetical Eli Lilly keeping its data inside Anthropic.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Ev says founders report that small open models, fine-tuned on their own trace data, reach Kimi K3 or Sonnet level performance at very low cost.
Jack estimates inference companies are mid single digits of AI revenue today. Both project open source at 5% to 10% of revenue, about $100bn if the labs combined reach a trillion dollars.