Designed with Broadcom, the chip has a clear lead over Nvidia Blackwell systems in tokens per user and throughput per kilowatt, according to the InferenceX bench.
The first results published on August 25, 2026 give an idea of what Jalapeño is worth, the inference chip developed by OpenAI with Broadcom and announced in October 2025. Measured on InferenceX, the SemiAnalysis test bench, it exceeds the current inference accelerators both on the number of tokens served per user and on the flow rate reported per kilowatt consumed — a difference that is considered to be very significant in relation to Nvidia Blackwell systems.
Gains come less from gross power than from caring for dead times. Jalapeño reduces the time required for pre-filling and flea exchanges, and keeps locally the key-value cache used during the generation, thus avoiding unnecessary data flow. OpenAI presents the whole as an integrated approach where models, silicon and memory are thought together; His own models were also involved in the design work.
The schedule remains cautious: the first copies will arrive at the end of 2026 in very small quantities, and deployment to scale is not expected until 2027.
Read the full article on the original source
Go to the source →