OpenAI just built its own chip. Nvidia's monopoly on inference just cracked.

OpenAI and Broadcom just unveiled Jalapeño.

OpenAI's first custom inference chip. Designed from scratch in nine months. Optimized exclusively for LLM workloads.

This isn't a research prototype. Engineering samples are already running GPT-5.3-Codex-Spark at production frequency and power. Early testing shows performance per watt "substantially better than current state-of-the-art."

Read that again. The company that buys more Nvidia GPUs than almost anyone just built silicon that outperforms them on the workload that matters most.

Hock Tan told CNBC demand from his six largest customers is "simply insatiable" — not just for 2026, but through 2028. OpenAI cannot get compute fast enough. So they stopped waiting.

Initial deployment begins end of 2026. Full gigawatt-scale rollout in 2027. This is a multi-generation roadmap, not a one-off experiment.

Every enterprise CFO running inference on OpenAI's API just got a structural cost reduction baked into their 2027 budget. Every Nvidia dependency analysis just got more complicated. Every chip vendor betting on CUDA lock-in just lost leverage.

If your AI strategy assumes inference costs stay flat or rise, rewrite the model. The economics just shifted.

SOURCE: https://www.cnbc.com/2026/06/24/openai-and-broadcom-reveal-jalapeno-first-ai-chip-in-partnership.html
VERIFIED: OpenAI official announcement, CNBC interview with Greg Brockman and Hock Tan, Pulse2, BetaNews
SIGNAL: OpenAI building custom inference silicon breaks the Nvidia dependency assumption every enterprise AI budget was built on. Inference costs are about to structurally decline.