OpenAI's Jalapeño chip challenges Nvidia inference dominance

OpenAI's custom Jalapeño ASIC benchmarks at Hot Chips show 1.5-1.9x throughput per kW and 3.6x lower latency vs Nvidia GB300, co-developed with Broadcom. Uses HBM4, competing for supply. Deployment late 2026. Could disrupt Nvidia's inference market share and shift AI hardware economics.