OpenAI says its Jalapeño chip delivers faster AI responses in tests

In short: OpenAI says its new Jalapeño computer chip ran AI responses faster and used less energy than top Nvidia chips in a public benchmark test.

What happened

OpenAI says it has new performance results for Jalapeño, an AI chip it first introduced in June. A chip is the hardware inside servers that does the heavy lifting, like the engine in a car.

OpenAI says Jalapeño is designed for “AI inference,” which means running a trained AI model to produce an answer. Put simply, it is the part of the job where a chatbot takes what it already learned and writes back to you.

In a briefing with reporters, OpenAI hardware vice president Richard Ho said Jalapeño aims to avoid a common trade-off. Many systems can either respond quickly (low “latency,” meaning less waiting) or handle lots of requests at once (high “throughput,” meaning more work in the same time), but not both.

To measure performance, OpenAI used a benchmarking platform called InferenceX. A benchmark is like a standardized test that lets you compare different systems under similar conditions.

OpenAI says Jalapeño delivered about 1.5 to 1.9 times more work per watt than the best results recorded using Nvidia’s GB200 or GB300 chips, across three AI models: GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T. It also reported 1.7 to 3.6 times lower end-to-end latency, meaning faster full responses.

Jalapeño is an ASIC (an “application-specific integrated circuit,” meaning a chip built for one main job, like a kitchen gadget that only makes waffles). OpenAI made it in partnership with Broadcom.

Why it matters

If these results hold up in real-world use, faster and more energy-efficient chips could mean AI services that feel more responsive, cost less to run, and stay available as more people use them.

Source: The Verge AI