OpenAI and Broadcom Announce Chip Designed for LLM Inference at Scale

Ronni Holmvig Strøm · 2026-06-25

On Wednesday, OpenAI and Broadcom unveiled Jalapeño, a custom ASIC built for large language model inference. It is OpenAI's first in-house chip, designed in partnership with Broadcom, and it

On Wednesday, OpenAI and Broadcom unveiled Jalapeño, a custom ASIC built for large language model inference. It is OpenAI's first in-house chip, designed in partnership with Broadcom, and it positions the company to serve its models on silicon it controls rather than on Nvidia GPUs alone. A physical sample shipped this week, with deployment planned over the next two years. The chip itself will get the headlines.

What Jalapeño Is

Jalapeño is an application-specific integrated circuit, hardware designed for one job rather than for general-purpose computing. It trades the flexibility of a GPU for lower cost and a layout tuned to inference, the serving stage where a deployed model generates tokens in response to a user.

OpenAI describes a large compute chiplet paired with high-bandwidth memory rather than cheaper DRAM. According to the company, the design targets high throughput and low latency together, the combination that reasoning models and agentic workloads require at serving time. OpenAI says the chip delivers "performance per watt substantially better than current state-of-the-art" and has promised a technical report in the coming months. It has not published performance figures.

The partnership pairs Broadcom's chip design and manufacturing relationships with OpenAI's specification of what its own models need to run. Broadcom has built similar custom accelerators for other large technology companies.

The Development Timeline

OpenAI says the chip went from concept to a physical sample in nine months. ASIC programs at this scale are typically measured in years. Greg Brockman told CNBC the chip was designed end to end with help from OpenAI's own models, and that "the degree to which our models have been able to accelerate it was very surprising to us."

The company has set out a specific schedule for deployment. Broadcom's Hock Tan described small prototype deployment in late 2026, a ramp through 2027, and full production in the first half of 2028.

The Demand Context

Both companies framed the chip against sustained demand for compute. Brockman told CNBC that OpenAI "cannot get compute fast enough." Tan called demand across Broadcom's customers "simply insatiable" through 2028.

The economics of running current models have shifted the pressure toward inference. A reasoning model that generates intermediate steps before answering, and an agent that runs for minutes against external tools, both consume most of their compute at serving time rather than during training. Custom inference silicon is one route to lowering the cost of that compute, and reduces OpenAI's reliance on a single supplier.