“OpenAI published measured performance results showing its custom AI inference chip Jalapeño outperforms Nvidia's GB300 on key inference metrics, delivering 1.5-1.9x more AI work per watt at peak throughput, 1.7-3.6x lower end-to-end latency, and 2.1-4.1x higher performance for interactive workloads like AI agents, per the InferenceX…”

The claim's numbers are reproduced accurately from OpenAI's own published post. OpenAI states that across GPT-OSS 120B, DeepSeek R1 and Kimi K2.5 1T, "Jalapeño delivered 1.5 to 1.9 times more AI work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency than the comparison systems," and that for highly interactive workloads it delivered 2.1 to 4.1 times higher performance. On Kimi K2.5, the largest public model tested, OpenAI reports roughly 1.5 times higher peak performance per watt and 3.4 times lower end-to-end latency than the comparison system. The benchmark named is real. InferenceX, formerly InferenceMAX, is SemiAnalysis's open-source continuous inference benchmark research platform covering GB200 NVL72, GB300 NVL72, B200, MI355X and other accelerators. The decisive caveat comes from the benchmark operator itself. SemiAnalysis wrote that all numbers were provided to them by OpenAI, that they verified the InferenceX runs in person in the lab but did not run the full suite of InferenceX benchmarks and had not seen AgentX results, AgentX being their preferred suite for chip comparison because its long-context multi-turn characteristics reflect realistic production cache behaviour. SemiAnalysis further stated that the comparison to Blackwell is "somewhat incomplete and unfair" because Jalapeño competes against chips like Rubin that also use HBM4, that Vera Rubin systems are shipping to customers now while OpenAI has nothing beyond engineering samples, and that the models tested are not on the open frontier. Separately, however, SemiAnalysis concluded that even compared with Rubin, Jalapeño's single-token-prediction output throughput per megawatt surpasses the Vera Rubin multi-token-prediction figures Nvidia and CoreWeave published in July, and described the part as "beating every Nvidia, AMD, and Google chip we have been able to test". Comparator detail matters. OpenAI's own chart captions show two different Nvidia comparators: GPT-OSS-120B at nominal 8k/1k, STP, against a GB200 at 1,200 W package TDP, and DeepSeek R1 and Kimi K2.5 against a GB300 at 1,400 W package TDP, with Jalapeño at 700 W. Tom's Hardware notes that Jalapeño was not tested against Vera Rubin, does not train models, and that the major comparisons ran Jalapeño's single-token prediction against GB300 configurations doing the same even though Nvidia deployments commonly use multi-token prediction in production. Power normalisation used rated TDP, not measured draw. OpenAI revealed at Hot Chips that the processor is rated at 700 watts but measured sustained power stayed at or below 550 W on the workloads tested, and OpenAI normalised the benchmark results against each accelerator's published package TDP. Deployment status: OpenAI's first custom AI chip is expected to begin deployment in the company's computing infrastructure by the end of the year. Nvidia's response was dismissive rather than a technical rebuttal: Jensen Huang brushed off the chip, saying he remains confident in Nvidia's technology and ability to supply the AI industry even as customers develop their own processors.