OpenAI's first custom silicon, 'Jalapeño,' targets inference costs and Nvidia reliance

OpenAI has revealed Jalapeño, its first in-house processor, co-designed and manufactured with Broadcom and aimed squarely at inference — the job of running already-trained models against user requests rather than training new ones. The company says its own models helped design the chip, and that early testing shows markedly better performance-per-watt than today’s leading alternatives. Heavier work like pre-training is still expected to run on Nvidia GPUs.

The move follows a partnership announced in October and fits a broader industry pattern: Google and Amazon have both built their own AI accelerators to cut dependence on Nvidia. OpenAI president Greg Brockman framed the effort as exploiting the company’s deep knowledge of its own workloads to accelerate tasks that general-purpose hardware underserves. OpenAI highlighted the chip’s low operating cost when running real-time coding models — a pointed detail given products like Codex.

The strategic logic is economic. Inference, not training, is where the recurring cost of serving millions of users accumulates, so even modest per-query savings compound at scale. By owning more of the stack — chip architecture, kernels, memory, networking, scheduling, and the products on top — OpenAI can tune every layer toward the same goal of cheaper, faster, more reliable model serving, and lean less on a single dominant supplier.