This page may contain stale information. Last updated: 2026-06-24

Definition

Jalapeño is openai’s first custom LLM inference accelerator, co-developed with broadcom. OpenAI calls it an “Intelligence Processor” — purpose-built for inference (not training), architected around OpenAI’s LLM serving stack.

Key Specifications (Announced June 24, 2026)

  • Development cycle: 9 months design to tape-out — claimed fastest ASIC cycle in high-performance semiconductors
  • AI-assisted design: OpenAI models accelerated parts of chip design and optimization
  • Early testing: Substantially better performance-per-watt vs state-of-the-art; ~50% cost savings vs typical GPUs (Hock Tan, Broadcom CEO)
  • Lab workload: GPT-5.3-Codex-Spark running on engineering samples
  • Deployment: Microsoft and partner data centers targeted end of 2026; multi-generation roadmap with 2028 follow-on
  • Networking: Broadcom Tomahawk silicon for scale-out

Performance claims are preliminary — OpenAI states final benchmarks pending. Jalapeño is inference-only; training likely remains on nvidia GPUs.

Sources