SemiAnalysis tested the OpenAI-Broadcom Jalapeño inference processor in InferenceX. It reported more tokens per user and higher throughput per kilowatt than Nvidia Blackwell systems in the same test. OpenAI says the chip targets the prefill stage. Very small shipments are due at the end of 2026, with wider deployment planned for 2027.
SemiAnalysis ran Jalapeño, OpenAI's first custom inference processor designed with Broadcom, on its InferenceX benchmark. It reported more tokens per user and higher throughput per kilowatt than Nvidia Blackwell systems in the same test.[1]
OpenAI says the design targets the prefill stage, where data movement and communication carry much of the load. That explanation identifies the workload aspect emphasised by the benchmark; the report does not claim the same result for every inference workload.[1]
Very small volumes are due at the end of 2026, with wider deployment planned for 2027. The timetable makes clear that Jalapeño is not yet a generally available product. The report does not detail shipment volumes or the scope of the wider deployment. The InferenceX result therefore shows processor performance in a particular comparison while the product timetable sets a separate limit on availability. The Blackwell comparison and shipment plan appear in the same report, but they do not describe the same stage of maturity. SemiAnalysis tested the OpenAI-Broadcom Jalapeño inference processor in InferenceX. The announced small-volume shipment timetable shows that, despite the benchmark result, the chip is not yet broadly available. This briefing keeps together the actor, reported scope and stated uncertainties of the event described by the source. It adds no separate development, definite outcome or comparison absent from that source; the headline isolates only this event.[1]
