Pre-training compute share collapses from 67% to 7% as inference and RL take over
robleclerc · x · 2026-11-15
Rob Leclerc comments on a chart of combined Anthropic+OpenAI compute allocation (1Q24–4Q26E): pre-training's share collapses from 67% to single digits, while post-training/RL and inference dominate (4Q26E: 55% post-training, 38% inference).
- Falling pre-training costs explain why Chinese labs keep shipping strong models, shortcutting post-training via distillation
- The real bottlenecks are now post-training and inference — both bandwidth-bound workloads
This also answers the quoted question of why Anthropic still commits so much compute: demand shifted from pre-training to inference and RL.
More from Infra
- Starship Flight 42 booster caught by the same launch tower it launched from — ChrisGPT · 2026-11-15
- InferSSD streams MoE experts from SSD so you can run models bigger than your VRAM — AIFlow_ML · 2026-11-15
- Nvidia's Huang told executives his Megaspeed interactions were standard business dealings — ShakeelHashim · 2026-11-15
- Three years of chip curbs: Huawei claims its AI chips now outsell Nvidia in China — Scobleizer · 2026-11-15
- JP Morgan sees AI accelerator shipments doubling to 31.2M units by 2028, custom silicon leading — Beth_Kindig · 2026-11-15
- Home Lab: 8x RTX 3090 Rig Runs 10 Parallel Agents at 370 tok/s — dazzou5ouh · 2026-11-15