OpenAI has unveiled new performance results for its custom AI chip, Jalapeno, at the Hot Chips conference in Palo Alto, California, saying the processor can make AI responses faster while using less power.
The company said Jalapeno delivered between 1.5 and 1.9 times more AI work for every unit of power used, while reducing the time taken to generate a response by between 1.7 and 3.6 times across three AI models it tested.
For highly interactive workloads, performance was between 2.1 and 4.1 times higher, OpenAI said.
The announcement is part of OpenAI’s push to build more of the computing infrastructure needed to run its AI services in-house, as demand for AI continues to drive up the need for expensive chips and electricity.
OpenAI CEO Sam Altman summed up the announcement rather more simply on X, posting: “we made a chip and it is fast.”
What is Jalapeno?
Jalapeno is an inference chip. In simple terms, it is designed to run AI models and generate answers after those models have already been trained.
Think of AI training as teaching a student, while inference is the student actually answering the questions. Jalapeno is built for the second part.
That means the chip is intended for the enormous amount of computing that happens when users ask ChatGPT questions or when AI agents carry out tasks.
OpenAI said it designed Jalapeno specifically around the way modern AI models work, rather than adapting a general-purpose chip for the job.
Also Read: India’s AI data-centre push could drive 5% of global chip demand by 2030
Faster AI with less power
One of the biggest challenges in running AI services is balancing speed with efficiency.
A system can be designed to handle a large number of requests at once, but that does not necessarily mean each user gets a faster response. Optimising for very low response times can also require more computing power.
OpenAI says Jalapeno is designed to improve both.
The company tested the chip on three publicly available models — GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T — using InferenceX, a benchmark from SemiAnalysis that measures how quickly and efficiently AI systems serve requests.
On Kimi K2.5 1T, the largest model tested, OpenAI said Jalapeno delivered about 1.5 times higher performance per watt and 3.4 times lower end-to-end latency than the comparison system.
In plain English, that means
Read More