How OpenAI's own models helped build its Jalapeño inference chip
OpenAI is handing chip design to its own models, Google is pausing an open-source security program under a flood of AI-written reports, and a new benchmark puts numbers on how Chinese models handle political taboos.
OpenAI's hardware team used the company's own models to help design Jalapeño, the custom inference chip it co-designed with Broadcom — and the models cut real work, not just slideware. In a Q&A published Sunday, VP of Hardware Richard Ho told Ian Cutress that in one example the internal models saved over 13% of die area, and that one attention implementation went from under one percent of the memory/compute roofline to nearly ninety percent in roughly forty hours. The silicon is an inference accelerator carrying 216 GiB of HBM4, rated at 700 W peak, and Ho says the project is now in volume ramp toward production, with the Gen 1/Gen 2/Gen 3 roadmap already sketched at Hot Chips. The telling detail is restraint: Ho says OpenAI's internal model outperformed the EDA vendors' own model efforts, yet the team still signs off through standard tooling — "If the model is 99.99% correct, you can't tape that out." Why it matters: the chip program is how OpenAI loosens Nvidia's grip on its inference bills, but the design process is the more transferable story — model-assisted engineering with a human signature is the pattern other hardware teams will copy next.
Google has paused product-vulnerability submissions to its open-source bug bounty program after a wave of low-quality, AI-generated reports overwhelmed engineers and maintainers. The OSS Vulnerability Reward Program stopped accepting product-flaw reports on October 1 and says an updated process will arrive by the first quarter of 2027, while supply-chain reports keep flowing. It's a small program carrying a large message: model-written reports now arrive faster than humans can triage them, and the cheap-guessing incentive has finally landed on a channel where wasted maintainer time is itself a security risk.
A new benchmark from European vendor Aleph Alpha says most Chinese frontier models toe the state line on politically sensitive questions. Across 967 hand-picked taboo topics — Tiananmen, Taiwan, Xinjiang — the company's own AI scoring system rated only 17% to 41% of responses from Alibaba's Qwen, DeepSeek and Moonshot's Kimi as balanced; the rest repeated official doctrine, deflected, or refused, with DeepSeek's V4 Pro declining two-thirds of questions while comparison models Claude Sonnet 5 and Mistral Small scored 70% and 92%. The finding that travels beyond China: Nvidia's Nemotron Cascade 2 showed party-line patterns in 17% of responses, which Aleph Alpha traces to about 3,500 of its 9.3 million training examples generated with DeepSeek and Qwen — distillation quietly carries one government's editorial line into a competitor's model. We covered the company's model release on Saturday — Aleph Alpha open-sources Kolibri, a sovereign German MoE model — and the vendor interest here is the same: Aleph Alpha sells "sovereign AI" and competes directly with these labs, though its numbers track China's own rules requiring socialist core values in public-facing models.
What to watch: Jalapeño's production ramp and the next generations of the chip, plus Google's bug-bounty redesign when it lands.
OpenAI still tapes out with a human signing off at 99.99% model confidence — is that the right threshold for AI-designed silicon? Tell us in the comments.
