OpenAI's Jalapeño Chip Puts Nvidia's Inference Monopoly on Notice

1. OpenAI's Jalapeño Chip Puts Nvidia's Inference Monopoly on Notice

OpenAI announced Jalapeño on June 23, 2026: its first proprietary AI chip, designed in-house and brought to production in partnership with Broadcom. The chip targets LLM inference workloads specifically, with ChatGPT, Codex, and the OpenAI API named as the initial deployment surfaces. No pricing, node process, or throughput figures were disclosed in the announcement, but the framing was explicit: Jalapeño is purpose-built for the workloads OpenAI runs at scale today, with future agentic products named as the next target.

This is a direct shot at Nvidia's position in inference infrastructure. OpenAI currently runs on H100s and H200s, and Nvidia's pricing power over that supply chain has been a structural cost pressure for every frontier lab. By moving chip design in-house and routing manufacturing through Broadcom rather than TSMC-via-Nvidia, OpenAI joins Google (TPUs), Amazon (Trainium/Inferentia), and Microsoft (Maia) in treating custom silicon as a cost and control lever, not a vendor relationship. The Broadcom partnership matters here: Broadcom has co-designed custom AI accelerators for Google for years, giving OpenAI access to production-grade ASIC expertise without building a silicon team from scratch.

The pattern worth watching is what happens to the API pricing floor. If Jalapeño delivers meaningful inference cost reductions, OpenAI has room to compress API prices further, which would squeeze competitors like Anthropic and Mistral who still depend on third-party GPU clouds. The next signal to track: whether Jalapeño appears in OpenAI's data center buildout disclosures, and whether Microsoft, as OpenAI's primary cloud partner, gets preferential access or gets quietly routed around.

Source: @OpenAI on X