OpenAI Claims an AI Chip Breakthrough. Broadcom Collaboration Jalapeño Beats Nvidia’s GB300 in Tests.

OpenAI says its first custom artificial intelligence chip can beat some of Nvidia's leading processors on speed and power efficiency, a potentially significant step in the ChatGPT maker's effort to reduce the enormous cost of running AI models.

The new processor, called Jalapeño, outperformed Nvidia's GB300 in the amount of AI work completed per unit of power and the speed at which the system returned responses, according to a Bloomberg report published Tuesday.

OpenAI hardware chief Richard Ho said the GB300 had been the leading option on the public benchmarking system used for the comparison. OpenAI plans to begin deploying Jalapeño later this year as part of a broader effort to build more of the infrastructure powering ChatGPT and its other AI products.

The chip was developed with Broadcom, which partnered with OpenAI on a multigeneration custom accelerator program announced last year. OpenAI said in June that the first Jalapeño design went from initial development to manufacturing tape-out in just nine months.

The results matter because running AI models has become one of the technology industry's biggest expenses. Data centers require enormous quantities of electricity, and companies including OpenAI are spending heavily to secure enough computing capacity to meet growing demand.

Jalapeño is designed specifically for inference, the stage after an AI model has been trained when it responds to users, generates content, or carries out tasks. OpenAI says the chip can deliver both high throughput, allowing it to serve large numbers of users efficiently, and low latency, which translates into faster responses.

"In the lab, Jalapeno is showing performance both in the high-throughput domain, meaning it will be able to serve a lot of customers more cheaply, as well as the low-latency domain," Ho told Bloomberg. OpenAI's newly released benchmark results put numbers behind those claims.

Across GPT-OSS 120B, DeepSeek R1 and Moonshot AI's Kimi K2.5, the company said Jalapeño delivered 1.5 to 1.9 times more AI work per watt at peak throughput and between 1.7 and 3.6 times lower end-to-end latency than the comparison systems. For highly interactive workloads, OpenAI reported performance improvements of 2.1 to 4.1 times.

Power consumption is another major part of the pitch. Jalapeño is rated at 700 watts, while measurements presented at the Hot Chips conference showed sustained power at or below 550 watts for the workloads tested. OpenAI believes those efficiencies could lower the cost of operating its rapidly expanding data center infrastructure.

But there are limits to the Nvidia comparison. Jalapeño was not tested against Nvidia's newer Vera Rubin generation, which has begun shipping. It also is not intended to train AI models, an area where Nvidia remains crucial to OpenAI. Instead, Jalapeño is optimized for serving models after training is complete.

That means OpenAI is not preparing to abandon Nvidia. Ho told Bloomberg that Nvidia remains a key partner and that OpenAI expects to continue buying large quantities of its processors as demand for computing power grows.

The bigger change is that OpenAI is increasingly designing its own technology alongside the chips it buys from outside suppliers. The company says Jalapeño was created using knowledge from the models, software, networking and serving systems it operates, with OpenAI's own AI models also helping accelerate the chip's development.

A second-generation processor is already well into development, with tape-out expected in the coming months, while OpenAI has begun conceptual work on a third generation. "We're at a cost level and a power level that will bring the infrastructure cost down," Ho told Bloomberg. "This is step one."