# OpenAI says Jalapeno beats Nvidia Blackwell on inference speed per watt > Source: > Published: 2026-08-25 17:59:23+00:00 # OpenAI says Jalapeno beats Nvidia Blackwell on inference speed per watt **The first-party accelerator beat GB200 and GB300 systems in OpenAI-run tests, with deployment planned inside OpenAI's infrastructure by December.** By [Ryan Merket](/author/ryan-merket) ยท Published Primary source: [OpenAI on X](https://x.com/OpenAI/status/2092300846675505602) ## Why it matters Jalapeno gives OpenAI a first-party route to lower inference costs and faster agent responses while adding bargaining power in a compute market still dominated by Nvidia. [OpenAI (@OpenAI)](https://x.com/OpenAI/status/2092300846675505602) published the first measured results for Jalapeno on Tuesday, saying its custom inference chip delivered between 1.5 and 1.9 times more AI work per watt at peak throughput than Nvidia GB200 and GB300 systems across three public models. OpenAI also reported end-to-end latency between 1.7 and 3.6 times lower than the comparison systems. ([openai.com](https://openai.com/index/jalapeno-first-results/)) [https://x.com/OpenAI/status/2092300846675505602](https://x.com/OpenAI/status/2092300846675505602) The results move Jalapeno past the engineering-sample claims OpenAI and Broadcom made when they [unveiled the chip on June 24th](https://openai.com/index/openai-broadcom-jalapeno-inference-chip/). Broadcom handled silicon implementation and networking, while Celestica worked on boards, racks and system integration. OpenAI said the program went from its initial design to manufacturing tape-out in nine months, with OpenAI models used during design and optimization. ([openai.com](https://openai.com/index/openai-broadcom-jalapeno-inference-chip/)) OpenAI President and co-founder Greg Brockman framed Jalapeno at the June launch as part of a full-stack infrastructure strategy aimed at making compute cheaper and more abundant. Tuesday's numbers provide the first public evidence for that bet, although they remain OpenAI's measurements and the hardware has yet to enter production service. ### What OpenAI measured OpenAI tested Jalapeno using [InferenceX](https://inferencex.semianalysis.com/about), SemiAnalysis' open-source inference benchmark, against commercially available Nvidia systems. The tests covered OpenAI's [GPT-OSS 120B](/models/openai/gpt-oss-120b), [DeepSeek R1](/models/deepseek/deepseek-r1) 670B and Moonshot AI's [Kimi K2.5](/models/azure/kimi-k2.5) 1T, spanning high-throughput serving and low-latency interactive workloads. InferenceX measures metrics including throughput, latency, interactivity and energy consumption across fixed input and output lengths. ([inferencex.semianalysis.com](https://inferencex.semianalysis.com/about?utm_source=openai)) On GPT-OSS 120B, OpenAI reported about 1.9 times more peak throughput per kilowatt than an Nvidia GB200 system, with end-to-end latency of 1.03 seconds compared with 1.80 seconds. The Jalapeno system produced 85,448 mixed tokens per second per kilowatt, versus 44,960 for the GB200 configuration. ([openai.com](https://openai.com/index/jalapeno-first-results/)) Against a GB300 system running DeepSeek R1, Jalapeno produced about 1.7 times more peak throughput per kilowatt and cut end-to-end latency from 5.99 seconds to 1.65 seconds, according to OpenAI. On Kimi K2.5, Jalapeno delivered about 1.5 times the peak throughput per kilowatt and reduced latency from 5.31 seconds to 1.56 seconds. ([openai.com](https://openai.com/index/jalapeno-first-results/)) OpenAI normalized the comparisons using each accelerator's published package power rating. Jalapeno carries a 700-watt rating, while the tested GB200 and GB300 systems were rated at 1,200 watts and 1,400 watts, respectively. OpenAI said Jalapeno's measured sustained power stayed at or below 550 watts during the workloads, though it used the higher 700-watt rating in the normalized results. ([openai.com](https://openai.com/index/jalapeno-first-results/)) The benchmark harness is public. OpenAI produced the Jalapeno measurements, and production qualification, software maturation and broader model validation are still underway. That makes the planned deployment inside OpenAI's infrastructure the next consequential test: whether the performance survives production traffic, reliability requirements and rack-scale operations. ### A chip built around inference Jalapeno targets inference, the process of running an already-trained model to answer requests. OpenAI designed the accelerator around the competing demands of prompt processing and token generation. Prompt processing is compute-heavy, while generating tokens is constrained by memory bandwidth. Communication between chips can introduce another delay. OpenAI says Jalapeno keeps model state, including the KV cache used while producing responses, close to the relevant compute and memory resources. The design treats networking as part of the accelerator architecture, reducing the amount of data that must move between components during a request. That approach is intended to preserve both high throughput and low latency as workloads move between prompt processing and generation. ([openai.com](https://openai.com/index/jalapeno-first-results/)) AI also played a role after the hardware design was finished. OpenAI said Codex and an internal model helped bring three open-weight models to high performance on Jalapeno within two months. For selected attention and mixture-of-experts blocks in GPT-OSS, AI-generated implementations ran between 1.5 and 1.8 times faster than versions written by human experts. OpenAI limited that claim to those selected blocks, rather than the performance of a complete model. ([openai.com](https://openai.com/index/jalapeno-first-results/)) ### OpenAI wants leverage over inference costs OpenAI plans to begin deploying Jalapeno in its own compute infrastructure by the end of 2026. The immediate product effect would arrive through faster ChatGPT responses, more responsive Codex sessions and agents, and additional capacity during periods of high demand, according to OpenAI. ([openai.com](https://openai.com/index/jalapeno-first-results/)) The larger payoff is control over the cost of serving models. Inference is a recurring expense tied directly to usage, and agentic products can multiply that cost by issuing many model calls during a single task. A purpose-built chip that finishes each step faster and processes more work within the same power envelope gives OpenAI another way to improve service economics without waiting for a supplier's next accelerator cycle. Jalapeno adds that leverage alongside OpenAI's existing chip suppliers. OpenAI said it will continue deploying Nvidia accelerators and hardware from other partners for training and inference. Chief Financial Officer Sarah Friar described the custom chip as a first-party path within a broader infrastructure portfolio, giving OpenAI more choice when assigning workloads and negotiating for compute. ([openai.com](https://openai.com/index/the-full-stack-behind-abundant-intelligence/)) OpenAI is already working beyond the first chip. Jalapeno Gen 2 is deep in development, according to OpenAI, and Gen 3 is taking shape. The first generation now has benchmark numbers. The year-end deployment will determine whether those numbers translate into faster products and lower serving costs at OpenAI's operating scale. ([openai.com](https://openai.com/index/jalapeno-first-results/))