OpenAI's first piece of custom silicon, an inference accelerator code-named Jalapeño, has posted benchmark results that beat Nvidia's Blackwell systems — and, on some measures, Nvidia's newer Rubin platform — on AI work delivered per watt, according to figures OpenAI shared with the chip-analysis firm SemiAnalysis. The results mark the most concrete evidence yet that a major AI lab is building hardware that could reduce its dependence on Nvidia, though the chip remains in early testing and is not yet in production.
SemiAnalysis, whose InferenceX benchmark suite was used for the tests, said it verified some of the runs in person at OpenAI's facilities. The firm's founder, Dylan Patel, was quoted as saying OpenAI is "beating Nvidia Blackwell and even Rubin," though SemiAnalysis was careful to note it had not run the full benchmark suite independently and had not reviewed results from more demanding, real-world "AgentX" workloads. Across three tested models — GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T — OpenAI reported Jalapeño delivering 1.5 to 1.9 times more throughput per watt than comparison systems at peak load, with latency reductions ranging from 1.7 to 3.6 times.
An important asterisk
The comparison carries real caveats that both OpenAI and outside analysts have acknowledged. Jalapeño uses newer HBM4 memory that Nvidia's Blackwell chips do not, and the rival systems in the test were not configured with optimizations such as multi-token prediction that would likely narrow the gap. SemiAnalysis's own writeup frames the results as encouraging but preliminary, noting the tests used simplified context lengths rather than the long, multi-turn conversations typical of production AI agents. In other words: a real result, but not yet an apples-to-apples verdict on whether Jalapeño can outperform Nvidia's shipping hardware at scale.
Jalapeño is designed and fabricated in partnership with Broadcom, using TSMC's N3P manufacturing process. Design work began in mid-2024, and the chip's design was finalized for fabrication in November 2025. It remains at the engineering-sample stage — the phase before mass production — while Nvidia's Rubin systems are already shipping to customers, according to separate reporting on the chip's development timeline. Production is expected to ramp gradually through 2027, with the bulk of output not arriving until late in the year.
The chip is inference-only, meaning it runs trained AI models rather than training new ones, and OpenAI has said it intends to use Jalapeño internally to run its own products rather than sell it to outside customers. If the efficiency gains hold up under fuller testing, it would give OpenAI a tool to lower the enormous power costs of running models like ChatGPT at scale — and add OpenAI to a short list of AI labs, alongside Google and Amazon, now designing their own chips rather than relying solely on Nvidia.
