OpenAI Unveils Jalapeño, Its First Custom AI Chip with Broadcom
What happened: OpenAI announced Jalapeño, its first custom inference ASIC, co-developed with Broadcom in just nine months from design to tape-out.
What happened: OpenAI announced Jalapeño, its first custom inference ASIC, co-developed with Broadcom in just nine months from design to tape-out. The chip is optimized for large language model (LLM) inference and agentic AI workloads, and is already running internal models such as GPT-5.3-Codex-Spark. The initial data center deployment is slated for late 2026, with Microsoft and other partners integrating the hardware. OpenAI claims Jalapeño offers substantially better performance per watt and up to 50% lower inference costs compared to current state-of-the-art chips, though no third-party benchmarks have been released.
Why it matters: Jalapeño represents a strategic move by OpenAI to control its AI infrastructure stack and reduce reliance on Nvidia and AMD for inference workloads. The rapid development cycle and partnership with Broadcom and Celestica signal a new phase of vertical integration in the AI hardware race. However, the chip is inference-only, so OpenAI will still depend on external suppliers for model training. The cost and power efficiency claims, if validated, could reshape economics for large-scale AI deployment, but independent verification is still pending.
Source: Decrypt