Rohan Paul

@rohanpaul_ai

Anthropic’s S-1 shows that research and development (R&D), not serving customers, uses most of the company’s compute: at least two-thirds of its chip spending goes to model training, research and internal use of Claude. $8.95B on compute in the first half of 2026, but only $2.38B of its costs went to serving paying customers Serving paying customers takes barely a quarter of that compute and already brings in far more revenue than it costs. Those customers cost $2.38B to serve against $16.29B in revenue, leaving 85 cents per dollar, up from 12 cents a year earlier. Because Anthropic buys most capacity 12 to 18 months ahead, surging demand lifted revenue per compute dollar from $0.40 in Q1 2025 to $2.08 in Q2 2026.
打开原帖#511482
  1. Research

    Fei-Fei Li: @jpatel41 Well said Jeetu! Experiments like this take courage and willingness to trial and error, but it’s so important at this moment!
  2. Research

    Jeff Dean: @irenetrampoline @LaudeInstitute Congrats, from someone on the committee that selected the finalists! 🎉 All of the final proposals were excellent, and we're excited to…
  3. Industry

    Mistral AI: Three weeks of training camp for Le Chonk