OpenAI, Broadcom unveil Jalapeño LLM inference chip; gigawatt-scale deployment targeted by end-2026

OpenAI and Broadcom unveiled Jalapeño on June 24, 2026—a custom LLM inference accelerator co-designed in nine months and already running GPT-5.3-Codex-Spark in lab testing. The chip is described as a "blank-slate design for modern LLM inference, not a general-purpose accelerator adapted from earlier AI workloads." Jalapeño targets gigawatt-scale deployment by end-2026 as part of a broader compute platform rather than a standalone component; the surrounding ecosystem (boards, racks, networking, connectivity, power delivery) is integral to the design philosophy.

The announcement follows OpenAI's Titan project (co-developed with Broadcom and TSMC, targeting H2 2026 mass production on 3nm process, with Samsung supplying HBM4 memory), and marks a strategic hedge against NVIDIA dependence. OpenAI, Google, Apple, SpaceX, Amazon, Microsoft, and Meta are all pushing custom silicon in 2026 because AI infrastructure has become too expensive, strategically important, and supply-constrained to leave entirely in Nvidia's hands. This is not a post-Nvidia era; it is a "post-naïve era" where large infrastructure builders now view compute strategy as inseparable from margin strategy and supply-chain control.

Broadcom is positioned as the "quiet architect" of this custom chip alternative: rather than selling accelerators through retail channels, it partners with hyperscalers to turn predictable, high-volume workloads into bespoke silicon. The resulting chips remain invisible to end-users but reshape the economics of the services they touch. For OpenAI, owning inference silicon is essential to reducing the per-token cost of serving frontier models at scale and maintaining operational independence from GPU supply constraints. Architects should expect custom silicon to capture an increasing fraction of production inference workloads over 2026–2027.