OpenAI Says Its Jalapeno Chip Beats Nvidia GB300

OpenAI has just unveiled benchmark data for "Jalapeno," its custom inference chip developed with Broadcom. And the takeaway is bold: they claim it beats Nvidia’s Blackwell Ultra-based GB300 on both speed and power efficiency. To give you an idea of the gap, Jalapeno’s package power is rated at 700 watts compared to the 1,400 watts required by the GB300. That’s a massive efficiency advantage even before you look at the performance numbers.

Using InferenceX—a public benchmark from SemiAnalysis—Jalapeno was tested against rival systems across three massive models: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T. Across the board, Jalapeno produced 1.5 to 1.9 times more AI work per watt at peak throughput, with latency cuts between 1.7 and 3.6 times. For highly interactive workloads, like those used for AI agents, the advantage grew even further—up to 4.1 times higher performance. On the 1-trillion-parameter Kimi K2.5 model, it posted 1.5 times better performance per watt and 3.4 times lower latency.

OpenAI normalized these results using each chip’s published power rating, and while Jalapeno is rated at 700 watts, they noted that actual measured power draw stayed at or below 550 watts during testing. The architecture targets the two specific phases of inference: compute-heavy "prefill" and memory-dependent "decode." By keeping model state and cache data local within a single connected system, they’ve cut out the communication delays that usually slow things down when data moves between chips.

What’s really fascinating is that AI itself helped build this silicon. OpenAI used its own models to design Jalapeno, shortening the concept-to-tapeout path to just nine months. Their engineers also used the Codex tool to optimize open-weight models, with AI-generated code running up to 1.8 times faster than versions written by humans. OpenAI plans to start deploying Jalapeno by the end of the year as part of a multi-generation roadmap. They aren't ditching Nvidia yet—they’ll continue to buy from them to meet demand—and notably, they didn't test Jalapeno against Nvidia's newer Vera Rubin chips. But with Jalapeno built specifically for inference, the message is clear: OpenAI is building its own path to scale.