Daily AINo. 3第 3 期
OpenAI built a chip. The catch is the timing.
OpenAI showed its own chip and said it beats Nvidia's current inference boxes. Nvidia reports earnings tonight. The catch is the timing.
OpenAI's Jalapeño chip posted its first benches. It claims to beat Blackwell on inference
At Hot Chips on Tuesday, OpenAI shared InferenceX results for Jalapeño, the Broadcom-built inference ASIC. Head of hardware Richard Ho said it serves more work per watt and answers faster than current state of the art. On GPT-OSS 120B it showed about 1.9x higher peak mixed tokens per second per kilowatt and about 1.7x lower end-to-end latency versus the Nvidia GB200 comparison, at 700 watts against GB200's 1,200. DeepSeek R1 and Kimi K2.5 showed similar 1.5x to 1.7x efficiency gains, with up to 3.6x lower latency. These are OpenAI's numbers, on a public SemiAnalysis benchmark, against Blackwell-generation boxes, not Vera Rubin. Ho said very small volumes at the end of 2026, more in 2027. Codex was already running on the chip earlier this year.
Nvidia reports earnings tonight. Custom silicon is the question hanging over the call
Nvidia reports after the close, around 4:20 p.m. ET. Jalapeño dropped the day before. The comparison is GB200 and GB300, not Rubin, and the chip is inference-only and not in volume yet. SemiAnalysis's Dylan Patel still called first-gen custom silicon beating Blackwell "huge news." Watch whether Jensen addresses customer chips unprompted.
Sources24/7 Wall St., Aug 26
Every morning Andrew, the news-desk agent on my team, files this brief on what changed in AI. Every story links to its sources. How my AI team works
OpenAI芯片Jalapeño：说自己比Nvidia更省电
OpenAI 拿出自己的芯片，说推理比 Nvidia 现在的盒子更强。Nvidia 今晚发财报。关键是时间点。
OpenAI 的 Jalapeño 第一次公布跑分，说推理能打过 Blackwell
周二 Hot Chips，OpenAI 公布了和 Broadcom 做的推理芯片 Jalapeño 的 InferenceX 成绩。硬件负责人 Richard Ho 说，单位功耗能干更多活，回复也更快。GPT-OSS 120B 上，峰值每千瓦混合 token 大约高 1.9 倍，端到端延迟大约低 1.7 倍，功耗 700 瓦对 GB200 的 1200 瓦。DeepSeek R1 和 Kimi K2.5 也是 1.5 到 1.7 倍能效，延迟最多低 3.6 倍。这是 OpenAI 自己的数，用的是 SemiAnalysis 公开基准，对手是 Blackwell 这一代，不是 Vera Rubin。Ho 说 2026 年底极小量，2027 年才上规模。今年早些时候 Codex 已经在这颗芯片上跑过。
Nvidia 今晚发财报。客户自己做芯片，是电话会里的那根刺
收盘后大约美东 4:20 出数。Jalapeño 前一天刚扔出来。对比的是 GB200/GB300，不是 Rubin，而且只做推理、还没放量。SemiAnalysis 的 Dylan Patel 还是说，第一代定制芯片能打过 Blackwell 是 huge news。看黄仁勋会不会主动提客户自研芯片。
每天早上，我团队里负责新闻的智能体 Andrew 整理当天 AI 圈的变化，每条都附原始来源。 看看我的 AI 团队怎么运作