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

OpenAI unveiled Jalapeño yesterday.

Its first custom inference chip. Built with Broadcom. From design to tape-out in nine months.

Not a prototype. Engineering samples are running GPT-5 workloads right now.

Early testing shows performance-per-watt "substantially better" than current Nvidia hardware. Bloomberg reports it cuts inference costs by roughly 50%.

Here's what actually happened.

OpenAI didn't just design a chip. They built the full stack: architecture, kernels, memory systems, networking, scheduling, and deployment.

Broadcom handled silicon implementation. TSMC manufactures it. Celestica integrates the systems. Target: gigawatt-scale deployment by end of 2026.

This is the same playbook Google ran with TPUs and Amazon ran with Trainium. But OpenAI's version is different.

They trained their own models to help design the chip. Software and hardware co-developed from day one.

For enterprises running on OpenAI's API, this changes the economics permanently. Lower inference costs mean cheaper tokens. But it also means OpenAI controls the hardware their models run on.

If you're benchmarking inference costs against Nvidia today, those numbers have an expiration date.

Audit your compute contracts now. The pricing floor just moved.

SOURCE: https://openai.com/index/openai-broadcom-jalapeno-inference-chip

VERIFIED: OpenAI official blog, TechCrunch (June 24, 2026), Yahoo Finance, VentureBeat

SIGNAL: OpenAI joining Google, Amazon, and Meta in custom silicon means Nvidia's inference moat is narrowing. Enterprise compute pricing will shift as custom chips deploy at scale.