OpenAI has shared the first results from testing its new AI chip, called Jalapeño. The company built the chip to handle inference, which is the process of running AI models and generating responses for users.
OpenAI said the chip can complete more work while using less power. It also responds to requests faster than current hardware options on the market.
Most computer chips face a tradeoff. They can be fast, or they can be efficient, but rarely both at the same time. OpenAI says Jalapeño avoids this tradeoff by combining speed and efficiency in one design.
The company tested the chip across three different AI models. These included GPT-OSS 120B, DeepSeek R1, and Kimi K2.5. The models were developed both inside and outside OpenAI.
Across all three models, Jalapeño produced 1.5 to 1.9 times more AI work per watt at peak performance. It also delivered responses 1.7 to 3.6 times faster than comparison systems.
How The Chip Was Tested
OpenAI ran its tests using InferenceX, a public benchmark created by the research firm SemiAnalysis. The benchmark measures the full process of handling an AI request from start to finish.
The company compared Jalapeño against other AI systems already used commercially. Jalapeño is rated at 700 watts, but its actual power use stayed at or below 550 watts during testing.
OpenAI said the chip performed best on Kimi, the largest model in the test group. On that model, it delivered about 1.5 times more performance per watt and 3.4 times lower latency than the comparison system.
The chip's design focused on keeping data close together. This reduces the time chips spend waiting to move information between parts of a system.
AI models generate answers in two stages. The first stage reads the prompt, and the second stage writes the response word by word. OpenAI designed Jalapeño to work efficiently through both stages.
What Comes Next
OpenAI used its own AI models to help design Jalapeño. The company said this shortened the chip's development timeline to nine months from initial design to production.
The team also used an OpenAI coding tool called Codex to help write software for the chip. Using this tool, engineers added support for three additional AI models within two months.
OpenAI plans to begin using Jalapeño inside its own computing systems before the end of this year. The company called this the first generation of the chip, with two more generations already in development.
OpenAI said it will keep using chips from Nvidia and other hardware makers alongside Jalapeño. The company said meeting AI demand will require computing power from multiple sources, not one chip alone.
The company is now working on production testing and preparing to run Jalapeño at a larger scale. OpenAI said it will keep validating the chip's performance across more AI models in the coming months.
