OpenAI’s homemade chip just beat Nvidia on the metric that actually pays the bills

“ChatGPT’s parent just grew its own stomach. It’s a jalapeño, and it eats less power than Nvidia.”

The Story

OpenAI published first Jalapeño numbers: 1.5–1.9× more AI work per watt and 1.7–3.6× lower latency than the then-best public Nvidia Blackwell systems on GPT-OSS, DeepSeek R1, and Kimi K2.5. The ASIC was designed with Broadcom, and OpenAI says its own models helped design the chip. It is inference-only; training still leans on partners.

Why It Matters

The model helped design the body that will run the model. Tokens-per-watt is the new oil, and the chatbot company now has a pepper-shaped organ. Avoid generic chip-race punditry.

Evidence

Hot Chips results 25 Aug; TechCrunch; The Verge; SemiAnalysis InferenceX. HN ~520 pts. Small-volume deploy late 2026, ramp 2027.

Sources

Daily scan: 2026-08-26