Are OpenAI and Anthropic Really Losing Money on Inference?
DeepSeek R1 running on H100s puts input-token costs near$0.003 per millionâwhile output tokens still punch in north of$3. Thatâs a 1,000x spread. So if a job leans heavy on inputâthink code linting or parsing big docsâthose margins stay fat, even with cautious compute. System shift:This lop-sided ..