Why Inference Matters 12x More Than Training | YC Paper Club
Inference computation is 12x more important than training computation over a model's lifespan
“the amount of inference computation becomes 12 times more important than the amount of training computation”
A YC Paper Club session argues that inference compute dominates training compute by 12x over a 2-year model lifespan, using DeepSeek V3's publicly disclosed figures (2.8M H800 hours to train vs 43K H800 hours/day of inference) as the only available data point. The argument reframes AI infrastructure investment priorities away from training toward serving efficiency. DeepSeek's reported 80% margins are cited as validation that inference optimization unlocks strong unit economics.