OpenAI's Jalapeño Chip Takes on NVIDIA

🔥 What happened OpenAI built its own inference chip, Jalapeño, with Broadcom. At HotChips 2026, hardware VP Richard Ho showed first silicon data: 1.5–1.9x better performance per watt than NVIDIA's GB200/GB300. 💡 Why it matters The chip packs 216 GiB HBM4 at 15.4 TB/s and scales to 27 EFLOP/s at 4-bit precision. OpenAI's own models helped design it – one attention kernel jumped from under 1% to nearly 90% of the roofline in 40 hours. ⚡ Our take OpenAI is becoming a vertical chipmaker. NVIDIA should worry – Jalapeño isn't a prototype, it's a measured blow to the status quo.