By the Time Jalapeño Reaches the Racks, the Leaderboard May Have Moved On

SemiAnalysis reports that an internal version of Codex wrote working, performant MLA kernels for OpenAI’s Jalapeño inference chip without the kernel engineering team stepping in. The obvious documentation joke is: ask the model; it wrote the code. No need to pretend the engineers can’t read it.

Now imagine tuning kernels for the model to beat today. Then come the rest of the chip work, manufacturing, shipping and installation. OpenAI says Jalapeño went from initial design to manufacturing tape-out in nine months, with its first systems expected to be deployed by the end of 2026. A startlingly fast hardware schedule; a leisurely age in model years.

A startlingly fast hardware schedule; a leisurely age in model years.

By the time those systems reach the racks, today’s frontier model might be a trivia question about AI in 2026. Not a forecast of three model changes—just the risk of building something durable for a target that moves.

The efficiency could be extraordinary. So could the efficiency of turning a cutting-edge chip into a beautifully optimized monument to yesterday’s leaderboard. All the racks need now is nostalgia for the benchmarks they were built to win.