OpenAI's Jalapeño Chip Beats Nvidia in Power Efficiency and Throughput Tests

OpenAI has published its first public benchmark results for Jalapeño, the custom AI inference chip it built in partnership with Broadcom, and the numbers show it outperforming Nvidia's Blackwell-class GB200 and GB300 GPUs on power efficiency and latency. According to detailed results shared with chip analysis firm SemiAnalysis and TechCrunch at the Hot Chips conference, Jalapeño delivered 1.5x to 1.9x higher throughput per kilowatt and 1.7x to 3.6x lower latency than NVDA's GB200/GB300 systems, running on a 700-watt ASIC compared with roughly 1,400 watts for Nvidia's flagship GPU.

OpenAI Head of Hardware Richard Ho told TechCrunch the results show "a very, very significant performance advance over state of the art," and the chip's rapid path to market, about 16 months from initial team hiring to manufacturing tape-out, underscores how aggressively OpenAI is moving into custom silicon as it looks to lower the cost of running its own models rather than depend solely on merchant GPUs. Jalapeño is part of what OpenAI describes as a "multigenerational platform" combining custom chips, models, and memory, co-developed with AVGO.

On SemiAnalysis's InferenceX benchmark suite, Jalapeño hit over 700 tokens per second per user on the DeepSeek R1 model and roughly 1,400 tokens per second per user on GPT-OSS. SemiAnalysis wrote that OpenAI is "beating every Nvidia, AMD, and Google chip we have been able to test on multiple top open source models," though it gave no numeric comparisons against those AMD or Google chips. SemiAnalysis also cautioned that all of the figures were supplied by OpenAI rather than independently measured end-to-end, and that the Blackwell-class comparison is "somewhat incomplete and unfair" since Nvidia's newer Vera Rubin platform, the more appropriate rival, has not yet been benchmarked against Jalapeño.

The deployment timeline is more gradual than a simple "later this year" framing suggests: OpenAI plans only small production volumes by the end of 2026, with broader deployment ramping through 2027. If the efficiency gains hold up at that scale, Jalapeño could pressure Nvidia's position in the inference market and push other AI labs further toward custom silicon, but the real test will come once independent benchmarks and head-to-head Rubin comparisons are available.

Related Stocks

• NVDA
• AVGO
• GOOGL
• OSS
• CRWV
• CBRS
• NOC
• PDD
• META
• TSM