OpenAI's Jalapeño beats Nvidia Blackwell on benchmarks — so why wait until 2027
openai's jalapeño chip outperforms nvidia's blackwell on the inferencex benchmark, delivering higher tokens-per-user and throughput-per-watt. however, mass deployment isn't expected until 2027, leaving the company to bridge a significant time gap against rapidly advancing competitors.
openai unveiled jalapeño, a custom inference processor co-developed with broadcom, that claims to surpass nvidia's blackwell in both tokens-per-user and throughput-per-kilowatt on the semianalysis inferencex benchmark. designed as a multigenerational, full-stack platform, the chip specifically targets latency in prefill and communication phases by keeping model state like the kv cache local. while this architecture serves more ai work per unit of power, the timeline remains a hurdle, with small volume deployment slated for late 2026 and full rollout expected in 2027. this delay creates a race against the market, as nvidia and competitors will likely release advanced iterations before jalapeño reaches scale.