Opinion thread says trailing frontier models pays under compute limits

What the research found

An opinion thread argues Chinese labs trailing frontier models is rational under compute limits, citing DeepInfer figures of about 1.5 cents blended revenue per million tokens and 80 to 85 percent hardware margin; the figures are the author's own arithmetic from published prices.

Claims and how well the evidence supports them

1. DeepInfer data shows one cluster serving V4.1-Flash handles 155 trillion tokens a day 98%
2. Blended realised price is about 1.5 cents per million tokens 98%
3. Hardware gross margin is 80 to 85 percent 98%
4. Serving cost on a B300 is about a quarter of a cent per million tokens 97%

Evidence

• twitter.com https://twitter.com/jenzhuscott/status/2113108957355229513

Actionability rated low.