OpenAI's Jalapeño Inference Chip Beats Nvidia Blackwell on Perf-per-Watt, Says Hardware VP Richard Ho
firstadopter · x · 2026-10-22
Tae Kim's Key Context interviews OpenAI VP of Hardware Richard Ho about the in-house Jalapeño inference chip.
- Jalapeño debuted at Hot Chips in late August to wide acclaim
- SemiAnalysis InferenceX benchmarks show it beating rivals including Nvidia's Blackwell on performance per watt across open models like GPT-OSS, DeepSeek R1, and Kimi K2.5
- Rankings may shift once the newer Vera Rubin systems arrive, but shipping such a strong first chip this quickly is a major achievement
- Ho previously worked on molecular dynamics supercomputers at D. E. Shaw Research, then became the first engineer on Google's TPU from gen one before joining OpenAI
- The author calls Jalapeño the most promising non-GPU AI chip program of the current wave; the Q&A covers chip design and memory technology trends
More from Companies & People
- Ashlee Vance: A New Hard-Tech India Is Rising in Hoskore, Backed by Lachy Groom and Shruti Rajagopalan — kyliebytes · 2026-10-22
- Debate erupts over claims Anthropic contributed 'zero value' to security research vs OpenAI — tszzl · 2026-10-22
- 16-Year Product Strategy Playbook From Epic, Affirm, Apollo Alum, on One Page — aakashgupta · 2026-10-22
- Anthropic researcher tells PhD students degrees are no longer safe bets; skeptics flag conflict of interest — 2C_ornot2C · 2026-10-22
- Every's writers on AI writing: judging by process, plus a classroom logging idea — lmoroney · 2026-10-22
- Bloomberg reporter courts AI researcher on memory, compression and world models — ziv_ravid · 2026-10-22