First benchmark results show the chip delivers more tokens per user and higher throughput per kilowatt than Blackwell systems.

OpenAI revealed the first Jalapeño benchmark results at the Hot Chips conference on Tuesday, showing more tokens per user and higher throughput per kilowatt than Nvidia Blackwell, the current state-of-the-art inference processor.

Jalapeño was first announced in October and developed with Broadcom, using OpenAI’s own models to assist in the design process. OpenAI plans to make the chip a multigenerational platform, with models, chips, and memory developed together.

For operators, the design means more AI work per unit of power and lower latency. By keeping model state and the KV cache local, Jalapeño minimizes data movement and communication delays, which often bottleneck inference processing.

Jalapeño will enter production in very small volumes at the end of 2026, with broader deployment following in 2027. Nvidia will have time to push Blackwell replacements, so the competitive picture may shift before the chip scales.

What matters

• OpenAI presented the first Jalapeño benchmark results at the Hot Chips conference.
• Jalapeño reduces prefill and communication delays, the main bottlenecks in inference processing.
• Initial deployment starts in small volumes at the end of 2026 and scales through 2027.

Why it matters

Initial deployment starts in small volumes at the end of 2026 and scales through 2027.

This GenAI News article was prepared in original wording using reporting and materials published by TechCrunch AI. Source reference: https://techcrunch.com/2026/08/25/openais-jalapeno-chip-is-built-for-fast-inference-at-scale-benchmarks-show/.

Drafted by the GenAI News review pipeline.