OpenAI and Broadcom Claim New AI Chip Outperforms Nvidia GB300

What Happened

OpenAI and Broadcom have announced a custom AI processor called Jalapeño that the companies say outperformed Nvidia's GB300 chip in benchmark testing. The processor completed more AI work per unit of power consumed and delivered faster throughput than Nvidia's current high-end data center GPU, according to claims published Monday.

The Chip and Its Benchmarks

Jalapeño is a custom application-specific integrated circuit, or ASIC, developed through a collaboration between OpenAI and Broadcom. According to the International Business Times report citing the announcement, the chip exceeded the Nvidia GB300 on two specific metrics: performance per watt, a measure of energy efficiency, and overall system throughput, which refers to the volume of AI inference or training work completed within a given time period.

The GB300 is Nvidia's latest generation data center accelerator, part of the Blackwell Ultra architecture. It has been broadly positioned by Nvidia as its most capable chip for large-scale AI workloads. OpenAI's claim that an in-house design has surpassed it in key metrics marks a notable development in the competitive landscape for AI silicon.

Specific benchmark figures, testing methodologies, and independent verification details were not included in the wire report at time of publication.

Background

OpenAI has been publicly developing its own silicon strategy for several years. The company has relied heavily on Nvidia hardware to train and run its models, including the GPT series and the systems underlying ChatGPT. Nvidia's dominance in AI accelerator hardware has made it one of the most valuable companies in the world, with its chips used by virtually every major AI laboratory and cloud provider.

Broadcom is one of the world's largest semiconductor companies and has established itself as a leading designer and manufacturer of custom AI accelerators for major technology clients. The company produces tensor processing units for Google and has custom chip engagements with several other hyperscalers. Broadcom does not manufacture chips directly but works with foundries, typically Taiwan Semiconductor Manufacturing Company, to produce finished silicon.

The partnership between OpenAI and Broadcom to develop Jalapeño follows a broader industry pattern of large AI consumers designing proprietary chips to reduce dependence on Nvidia and potentially lower the cost of running AI workloads at scale.

What It Means in Practice

If the performance claims hold under independent scrutiny, the Jalapeño chip could allow OpenAI to run its AI infrastructure more cheaply per unit of computation. Energy costs are a major operational expense for large language model training and inference. A chip that delivers more computation per watt directly reduces the electricity cost of running AI services.

The development also adds OpenAI to a list of companies that includes Google, Amazon, and Microsoft, all of which have developed or are developing custom AI silicon to reduce purchasing dependence on Nvidia. Each of those companies continues to purchase Nvidia hardware alongside their proprietary chips.

Nvidia's stock has been closely watched by investors ahead of the company's upcoming earnings release. The company has disclosed an expanding investment strategy beyond chip hardware in recent weeks.

Broadcom's role as the design and manufacturing partner positions it as a potential beneficiary of any scaling of Jalapeño production. The financial terms of the OpenAI-Broadcom arrangement were not disclosed in available reports.

What Comes Next

OpenAI has not announced a public deployment timeline for Jalapeño, and independent benchmark validation from third-party researchers or institutions has not been published as of this report.