A week after OpenAI and Broadcom unveiled their jointly designed AI chip, code-named Jalapeño, the announcement is still moving markets and prompting questions about whether the custom-silicon push can break Nvidia’s grip on the training business. Broadcom shares have risen since the launch, a sign that investors are taking the threat to Nvidia’s near-monopoly seriously.
The chip is designed as a general-purpose ASIC, a custom accelerator that can be adapted to run a range of large language models rather than a single architecture. That positioning is deliberate: it contrasts with the tightly coupled approach Nvidia has built, in which its hardware, software and networking are designed to work best together and to be hard to replace. OpenAI and Broadcom are betting that enterprises and model developers will want a chip that does not lock them into one vendor’s stack.
The companies have disclosed few technical details, and people familiar with the project say the chip has been in development for more than a year, with a team drawn from both companies working on the design in Silicon Valley. Broadcom brings the engineering discipline of its custom-chip business, which has produced accelerators for Google’s TPU line and Meta’s in-house chips, while OpenAI brings the software and the demand: the company controls some of the largest model-training workloads in the world and can guarantee the chip a customer.
The economics are the point. Nvidia’s accelerators command premium prices, and the high-bandwidth memory they require has been a bottleneck and a cost driver across the industry. A custom chip designed for OpenAI’s specific workloads can be optimized for cost and efficiency in ways that general-purpose hardware cannot, and Broadcom’s manufacturing partnerships give the project access to leading-edge capacity. If the chip delivers on its cost targets, it will put pressure on the prices Nvidia can charge for the rest of the industry.
The market reaction has been instructive. Broadcom’s stock rose in the days after the announcement, with analysts citing the OpenAI relationship as evidence that the company’s custom-silicon franchise has room to grow. Nvidia’s shares barely moved, which some investors read as confidence and others as complacency. The divergence has become a talking point among semiconductor investors, who have spent the past two years debating whether Nvidia’s dominance is structural or simply the default.
History cuts both ways. Custom chips have failed against Nvidia before: Google’s TPU has not dented the market share of the company’s accelerators, and Amazon’s Trainium has carved out a niche without becoming a general alternative. Yet the market has shifted in the custom direction this year, as every major cloud provider has either built its own silicon or commissioned it from Broadcom or Marvell. The question is no longer whether custom chips will exist; it is whether they will scale beyond their sponsors.
For OpenAI, the chip is part of a broader strategy of vertical integration. The company has been building its own data centers, signing long-term power agreements and designing its own hardware, a pattern that reduces its dependence on any single supplier, including Nvidia. The Jalapeño project does not mean OpenAI will abandon Nvidia entirely, people familiar with its plans say, but it gives the company a negotiating position it did not have before, and a hedge against the allocation problems that have defined the AI supply chain.
For Broadcom, the deal is a validation of its custom-silicon strategy. The company has positioned itself as the foundry of choice for companies that want their own chips without building a semiconductor division from scratch, and OpenAI is the most prominent customer yet. Analysts estimate the project could add meaningfully to Broadcom’s AI revenue if it reaches volume production, and the company has signaled it expects more such deals as the industry’s custom-chip ambitions grow.
The near-term test is production. A chip announced is not a chip shipped, and the timeline for Jalapeño’s deployment has not been disclosed. OpenAI will need the chips in quantity to train its next-generation models, and any delay would undercut the project’s credibility. The company’s hardware team has been hiring aggressively, according to people who track its recruiting, a sign that the project is funded for the long run.
For Nvidia, the challenge is not any single chip but the pattern they represent. Every custom accelerator announced by a major customer is a vote for the idea that the AI hardware market will eventually fragment, and Broadcom’s pipeline suggests the votes are accumulating. Nvidia’s answer has been to move faster on its own roadmap and to make its platforms harder to leave; the strength of that answer will determine whether Jalapeño is remembered as a curiosity or as the beginning of the end of an era of dominance.
