Anthropic's S-1 Shows Only 25% of Compute Goes to Inference, Sparking Debate
Anthropic's S-1 reveals only about 25% of its 2GW compute (roughly 500MW) serves paid inference, with most going to training. Investor Brad Gerstner cited this to suggest Anthropic may need only 500MW for inference next year, but critics argue the assumption fails as models keep growing larger.
2026-11-18 ~ 2026-11-18 · 2 related posts
- Anthropic's S-1 Shows Only 25% of Compute Goes to Inference — Sparking a Bear Case — toptickcrypto · 2026-11-18
- Debate: will Anthropic really need only 500MW of inference next year? — zephyr_z9 · 2026-11-18