OpenAI announced that its new custom AI chip, called Jalapeno, outperformed Nvidia’s current-generation GB300 processor in internal testing, marking a notable milestone in the ChatGPT maker’s push to build its own AI infrastructure rather than relying entirely on outside chip suppliers. In benchmark testing, Jalapeno led in two specific categories, the amount of AI work it could process per unit of power consumed, and the speed at which it returned responses, according to OpenAI’s chip chief, who discussed the results in an interview and presented them publicly at the Hot Chips conference at Stanford University.
Jalapeno was developed in partnership with Broadcom, which builds custom chips for a range of major technology clients, and the two companies have touted the unusually short development timeline that brought the chip from concept to testing. OpenAI plans to begin using the chips to support its AI models later this year, running the low-voltage, 700-watt processor specifically to reduce power costs across its rapidly expanding data center footprint, power representing one of the largest ongoing expenses in operating AI infrastructure at scale.
Several important caveats temper how much weight investors should place on this result. Jalapeno was not tested against Nvidia’s newest chip generation, Vera Rubin, which only recently began shipping and represents Nvidia’s current cutting edge rather than its prior-generation GB300. Jalapeno is also not designed to train AI models at all, an area where Nvidia’s technology remains dominant. Instead, Jalapeno is built specifically for inference, the process of running an already-trained model to generate responses and complete tasks, a narrower but still commercially significant slice of the overall AI compute market.
OpenAI’s own chip chief was notably candid about the limits of this milestone, describing Nvidia as a genuinely strong partner that OpenAI will continue to rely on heavily going forward, a reminder that this announcement reflects supplier diversification rather than any intention to replace Nvidia outright. That diversification effort is broader than just Jalapeno. OpenAI already uses chips from Cerebras Systems for some of its smaller models, a company whose own record-breaking Nasdaq debut we covered earlier this summer, though OpenAI’s chip chief noted that architecture is best suited to smaller models, while Jalapeno is designed to handle considerably larger ones. Beyond Jalapeno, competing custom chip startups are pursuing similar goals, including Etched, which recently raised funding at a $21 billion valuation, and MatX, founded by former members of Google’s internal silicon design team.
For investors tracking the AI infrastructure ecosystem, this development is best understood as confirmation of a trend already well underway rather than a singular disruption. Every major AI company, from Google’s long-running TPU program to Amazon’s Trainium chips to Microsoft’s own custom silicon efforts, is pursuing some version of reduced dependency on any single chip supplier, and OpenAI’s Jalapeno simply extends that pattern to the company sitting at the center of the current AI boom. That dynamic creates real, sustained demand for the broader ecosystem of smaller specialized companies supporting custom chip development, including semiconductor design and IP licensing firms, advanced packaging providers, and specialized testing and validation companies that benefit regardless of which individual chip architecture ultimately wins the most market share.
Nvidia’s stock showed little reaction to the news, a reasonable response given the caveats involved. But the steady, accelerating march toward diversified AI chip supply chains remains one of the more durable structural themes shaping opportunity across the smaller companies that make up that supply chain.
