OpenAI and Broadcom ship first custom inference chip

*OpenAI and Broadcom introduced Jalapeño, a processor built specifically for large-scale LLM inference workloads.*

OpenAI and Broadcom have announced Jalapeño, a custom silicon design aimed at running large language model inference more efficiently at scale. The move places the AI company directly into hardware development at a time when inference demand is outpacing available capacity.

What changed

The two companies described Jalapeño as an “Intelligence Processor” optimized for the inference stage of LLM workloads. Prior systems relied on general-purpose GPUs or earlier accelerators. The new chip targets improvements in performance per watt and overall throughput when serving production models.

Technical details

OpenAI’s announcement states the part was developed with Broadcom to raise efficiency and scale across its AI systems. No clock speeds, transistor counts, or power figures were released. The partners positioned the device as the first in a line of custom inference silicon rather than a one-off prototype.

Industry context

Ars Technica noted the announcement arrives amid broader competition to secure enough compute for growing model usage. Neowin framed the effort as OpenAI’s entry into hardware territory long dominated by established chip vendors. None of the sources provided third-party benchmarks or deployment timelines.

Why it matters

OpenAI now controls part of the stack that converts trained models into production responses. That reduces reliance on third-party accelerators for its highest-volume workloads and gives the company a direct lever on cost and latency. Whether the same design reaches outside customers remains open; the initial focus stays internal scale.

---

Sources:

• Ars Technica
• OpenAI
• Neowin

{
  "excerpt": "OpenAI and Broadcom unveiled Jalapeño, their first custom chip built for large-scale LLM inference.",
  "suggestedSection": "ai",
  "suggestedTags": ["openai", "broadcom", "ai-chip", "inference"],
  "imagePrompt": "Abstract silicon dies and wafer fragments rest on a matte surface, faint traces of etched circuitry catching low side light. The arrangement suggests scale and precision without showing any device or brand. muted color palette, cinematic lighting, 16:9"
}