OpenAI says Jalapeño chip delivers up to 1.9× more tokens per watt than Nvidia
OpenAI unveiled its first in‑house inference chip, Jalapeño, built with Broadcom and assembled by Celestica. The chip targets data‑movement efficiency, claiming 1.5 to 1.9 times more tokens per watt than Nvidia’s Blackwell systems on a 700‑watt package (measured draw at or below 550 W). Specs shared at Hot Chips include 13.4 PFLOPS MXFP4 compute, 216 GiB HBM4 memory at 15.4 TB/s, 64 core slices…
Key points
- OpenAI says Jalapeño yields 1.5‑1.9× more tokens per watt than Nvidia Blackwell on a 700 W package
- Chip specs: 13.4 PFLOPS MXFP4, 216 GiB HBM4, 64 core slices, up to 2,048 chips per pod
- Design completed in nine months using AI‑assisted hardware language XLS
OpenAI tested Jalapeño on the InferenceX benchmark, reporting about 1.9× more tokens per kilowatt and 1.7× lower latency versus Nvidia GB200 for a 120 B GPT‑OSS model, and similar gains against GB300 for DeepSeek R1 and Kimi K2.5. The tests were run internally, and no independent verification is cited. The chip will first appear in very small volumes at the end of 2026, with broader rollout in 2027, and will power OpenAI services such as ChatGPT, Codex and its API.
The design was completed in nine months, from RTL code in February 2025 to tape‑out in November 2025, aided by AI‑driven design tools that wrote more than half the core in the XLS language. A team of roughly 100 engineers handled design, verification, packaging and software, but OpenAI still does not trust the models completely, keeping human validation for all changes.
The story so far
5 episodes →- OpenAI says Jalapeño chip delivers up to 1.9× more tokens per watt than Nvidiathis story
OpenAI’s Jalapeño chip bets that cheaper AI is about moving less data
i-scoop.eu · 26 October 2026
Loading the full article…
This text was published by i-scoop.eu and written by Tom Cuylaerts. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Hardware & Compute
All →- Cloudflare fixes GLM-5.3 bug exposing 155 customer responses · 2 src
- Vertiv unveils SmartIT platform for high-density AI infrastructure · 1 src
- NVIDIA reportedly halts RTX 5090 orders for board partners · 1 src
- Google and Amazon launch PyTorch tools for their AI chips to compete with Nvidia · 1 src
- IBM makes Spyre a first-class PyTorch backend · 1 src
Comments
via GitHub Discussions