OpenAI's New Chip Beats Nvidia's GB300 in Key Tests

OpenAI published benchmark results claiming its custom inference chip, developed with Broadcom and codenamed Jalapeño, outperformed Nvidia's GB300 processor on two specific measures: the amount of AI work handled per unit of power, and the speed of returning a response, according to Bloomberg's reporting on the results, confirmed by OpenAI chip chief Richard Ho.

Jalapeño is an inference chip, meaning it's built specifically to run and serve already-trained AI models rather than train new ones from scratch. Ho said the chip achieved its results while operating at 700 watts, a detail that matters because power consumption is one of the largest ongoing operating costs for AI data centers, according to Yellow's reporting on the announcement. Ho explained the practical implication for OpenAI's customers directly: Jalapeño's combination of high throughput and low latency could let customers choose models that prioritize either lower operating costs or faster response times, depending on their specific needs.

OpenAI was notably transparent about the limits of this comparison rather than presenting it as an unqualified win. Jalapeño was not tested against Nvidia's newer Vera Rubin generation, which has only just begun shipping, and the comparison instead measured against the GB300, Nvidia's previous leading chip on the public benchmarking system used for the test, according to The Next Web's reporting on the disclosure. The chip also cannot train models at all, meaning it addresses only one half of the AI compute equation.

Jalapeño vs. GB300: What Was Actually Tested

The specific models used in testing are worth understanding, since they weren't exclusively OpenAI's own systems. The tests ran a small open model built by OpenAI itself, alongside third-party models from DeepSeek and Moonshot, with the widest performance advantage showing up specifically on Kimi, the largest model tested, according to The Next Web's reporting. That detail connects directly to the momentum we've tracked closely in our coverage of Kimi K3's rapid rise to rival Claude and ChatGPT, underscoring how central Chinese open-weight models have become even in benchmark testing conducted by a major Western AI lab.

Ho was careful to frame this development as complementary to OpenAI's existing Nvidia relationship, not a replacement for it. "Nvidia is a really good partner, and we continue to need a lot of Nvidia," Ho said, according to The Next Web's reporting. That framing matters given the roadmap already in motion: a second-generation Jalapeño chip tapes out in the coming months, with a third generation already in concept, according to Axios's reporting on the announcement.

OpenAI's move fits a genuinely widespread industry pattern of major AI buyers developing custom chips specifically to reduce dependence on Nvidia for inference workloads. Anthropic is separately designing its own silicon, and Google has been in talks with Marvell about custom inference chips, according to The Next Web's reporting. This connects directly to the broader chip diversification trend we've tracked closely, including AMD's Helios platform landing Microsoft as its first major customer and Nvidia's own recent 15% price hike on AI servers, a cost pressure that's making alternatives to Nvidia's premium pricing increasingly attractive across the industry.

This announcement is worth understanding for any business evaluating AI infrastructure costs, since inference, not training, represents the ongoing operational expense most companies actually pay repeatedly every time an AI model responds to a request. If custom inference chips like Jalapeño genuinely deliver meaningful power efficiency gains at scale, that cost savings could eventually flow through to lower prices for AI-powered products and services built on top of OpenAI's infrastructure.

For businesses tracking the broader AI hardware landscape, this is a signal that Nvidia's dominance, while still substantial, faces genuine, credible pressure from multiple directions simultaneously, custom silicon from its own largest customers, competing chip makers like AMD, and rising memory costs squeezing its own pricing.

What is OpenAI's Jalapeño chip?
Jalapeño is OpenAI's custom AI inference chip, developed with Broadcom, designed to run and serve already-trained AI models rather than train new ones, with OpenAI reporting it outperforms Nvidia's GB300 on power efficiency and response speed.

Did OpenAI compare Jalapeño against Nvidia's newest chip?
No. The comparison was made against Nvidia's GB300, not the newer Vera Rubin generation, which had only just begun shipping at the time of testing, a limitation OpenAI itself disclosed.

Is OpenAI trying to stop using Nvidia chips?
No. OpenAI chip chief Richard Ho said Nvidia remains "a really good partner" and that OpenAI "continues to need a lot of Nvidia," framing Jalapeño as a complementary option rather than a full replacement.

OpenAI published benchmark results showing its custom Jalapeño inference chip, developed with Broadcom, outperformed Nvidia's GB300 on power efficiency and response speed, though the chip was not tested against Nvidia's newer Vera Rubin hardware and cannot train AI models. Testing included OpenAI's own model alongside third-party systems from DeepSeek and Moonshot, with the widest performance advantage appearing on Moonshot's Kimi model. OpenAI's chip chief emphasized the company still relies heavily on Nvidia, framing the custom chip as complementary rather than a replacement, as a second-generation Jalapeño is already set to tape out in the coming months.