OpenAI unveils first custom chip, Jalapeño, built with Broadcom

OpenAI and Broadcom on June 24 unveiled Jalapeño, OpenAI's first custom-designed artificial intelligence chip, an accelerator built specifically to run large language models more cheaply and efficiently than the general-purpose graphics processors that dominate the market today. The announcement marks OpenAI's entry into custom silicon and the first product in what the two companies describe as a multi-generation hardware platform they are building together.

Why it matters

Jalapeño is OpenAI's most direct move yet to loosen Nvidia's grip on the hardware that powers its models, and the economics it claims — roughly half the cost per unit of inference — go to the heart of whether frontier AI can ever be profitable. Inference, the work of actually running models for users, is where the bulk of compute spending now goes, so a chip purpose-built for it rather than adapted from general-purpose GPUs could reshape the cost structure of the entire industry if the performance claims hold at scale. The nine-month development cycle, which OpenAI says was accelerated using its own models, is itself a marker of how AI is starting to compress the timelines of the chip industry that feeds it — and a warning shot to Nvidia that its largest customers are increasingly willing and able to build around it.

Unlike the GPUs that OpenAI and its rivals have relied on, which were originally designed for graphics and adapted to AI, Jalapeño was architected from the ground up for inference — the task of serving a trained model's responses to users. The companies said the design was optimized around the specific computational kernels, memory movement, networking and serving patterns that matter most for frontier models.

Broadcom Chief Executive Hock Tan said early testing showed the chip delivering cost savings of roughly 50 percent compared with typical AI GPUs, and performance per watt 'substantially better' than the current state of the art. For a company whose compute bills run into the billions, halving the cost of running its models would materially change the economics of its business.

Perhaps as striking as the chip itself was the speed of its development. OpenAI and Broadcom said Jalapeño went from initial design to manufacturing tape-out in about nine months, a pace they characterized as among the fastest ever achieved for a high-performance, advanced-node semiconductor. The companies said the timeline was compressed in part by using OpenAI's own models to accelerate parts of the design and optimization work.

The chip is a large, reticle-sized application-specific integrated circuit, or ASIC — a class of processor tailored to a narrow set of tasks rather than built for general use. That specialization is what allows it to wring more efficiency out of inference workloads, at the cost of the flexibility that has made Nvidia's GPUs the default choice across the industry.

The companies are targeting initial deployment of Jalapeño by the end of 2026, with the rollout expanding in subsequent years. Early deployments are slated to include gigawatt-scale data centers operated with Microsoft and other partners, underscoring the enormous power and infrastructure footprint that frontier AI now demands.

Jalapeño lands amid a broader scramble by the largest AI companies to reduce their dependence on Nvidia, whose chips have commanded premium prices and long waiting lists throughout the AI boom. By co-developing its own accelerator, OpenAI joins a growing list of technology giants designing custom silicon — a shift that, if the performance claims bear out, could gradually erode the pricing power that has made Nvidia one of the most valuable companies in the world.

OpenAI unveils first custom chip, Jalapeño, built with Broadcom

Sources