OpenAI Claims Jalapeño Inference Chip Beats Nvidia Blackwell in First Eval
The AMW Read
First Jalapeño-vs-Blackwell inference claim updates OpenAI’s custom-silicon case study and silicon competition, but unscaled results and no Rubin comparison keep impact at segment level.
OpenAI Claims Jalapeño Inference Chip Beats Nvidia Blackwell in First Eval
OpenAI released an evaluation of Jalapeño, its Broadcom-developed AI server chip, claiming the part is faster and more powerful than Nvidia's Blackwell hardware for inference. The ASIC is described as optimized to run existing models rather than a training-only design. The same reporting notes the chip has not yet been tested at scale and has not been compared against Nvidia's newer Rubin platform. Per the AI Market Watch index, OpenAI is tracked in Foundation Models / LLMs with about $199.6B in total funding across an index of roughly 5,000 companies rather than a full market census.
The claim lands on a load-bearing industry question: whether a frontier lab can build inference silicon that cuts Nvidia dependence on the serving path where latency and work-per-watt set operating cost. Jalapeño is inference-focused and co-developed with Broadcom, so the competitive surface is production decode and agent serving, not training cluster leadership. Prior OpenAI Jalapeño coverage already flagged early efficiency claims and a limited late-2026 rollout; this evaluation extends that arc by putting a direct Blackwell comparison on the record while leaving scale and Rubin as open tests. First-generation custom ASICs have often looked strong in lab settings and then struggled once multi-tenant, rack-scale traffic arrives.
For builders and investors, the decision point is production proof, not the first eval headline. Until Jalapeño runs at fleet scale with stable software and economics, Nvidia remains the default allocation partner even for labs shipping their own chips. Watch whether subsequent disclosures close the scale and Rubin gaps, and whether serving cost or latency for OpenAI products actually moves as silicon ramps.
