OpenAI Claims New Jalapeno Chips Outperform Nvidia Processors In Tests

OpenAI has said its internally developed Jalapeno AI processor outperformed Nvidia Corp’s GB300 chip in tests measuring power efficiency and response speed, marking a step forward in the ChatGPT maker’s efforts to build more of its computing infrastructure in-house.

OpenAI chip chief Richard Ho said Jalapeno led the GB300, which was the top-performing option on the public benchmarking system used for the comparison, as per a report by Bloomberg (BBG).

The chip was evaluated on the amount of AI work it could complete for a given amount of power and how quickly it could return responses.

OpenAI plans to begin deploying Jalapeno for its AI models later this year, as per BBG.

Ho said the processor could substantially lower the company’s infrastructure costs if production can be scaled and the expected savings materialise.

“The chip is really good — it should drop it by a lot,” Ho said on the Bloomberg Tech show on Tuesday.

Chip Focuses On AI Inference

Jalapeno has been designed primarily for inference, the stage at which trained AI models process prompts and generate responses, rather than for training models.

The chip operates at 700 watts, which Ho said could help reduce electricity costs across OpenAI’s data centres, as per the report.

Jalapeno was not compared with Nvidia’s latest Vera Rubin generation, which has only recently begun shipping, the Bloomberg report said.

OpenAI tested Jalapeno using its own smaller open-source model, along with models from DeepSeek and Moonshot AI.

The strongest gains were recorded on Moonshot’s Kimi model, while internal testing also showed promising results on some larger OpenAI models that have not yet been released.

Nvidia remains an important supplier to OpenAI, with Ho saying the company will continue to require substantial quantities of Nvidia processors.

OpenAI Developing Next Chip

OpenAI developed Jalapeno with Broadcom Inc. OpenAI has already started work on a second-generation chip, which Ho expects to reach the tape-out stage in the coming months.

The company has also begun developing concepts for a third-generation processor as it seeks to contain the cost of building AI infrastructure globally.

"We’re at a cost level and a power level that will bring the infrastructure cost down,” Ho said, adding, "This is step one."