OpenAI and Broadcom Unveil Jalapeno Inference Chip

OpenAI and Broadcom (NASDAQ: AVGO) unveiled Jalapeno on June 24, 2026, calling it OpenAI's first "Intelligence Processor." Unlike a general-purpose AI accelerator adapted for LLM work, Jalapeno was designed from scratch around the memory movement, kernels, and serving patterns that matter most for large-language-model inference, informed by OpenAI's operational experience running ChatGPT, Codex, and its API. Broadcom handles silicon implementation and networking, including its Tomahawk switching silicon, while Celestica contributes board, rack, and system integration. Engineering samples are already running production workloads, including GPT-5.3-Codex-Spark, inside OpenAI's labs at target frequency and power.

OpenAI and Broadcom say Jalapeno went from initial design to manufacturing tape-out in nine months, which they describe as the fastest ASIC development cycle achieved in high-performance advanced semiconductors, aided by using OpenAI's own models to accelerate parts of the design and optimization process. Early testing shows performance-per-watt substantially better than current state-of-the-art alternatives, though OpenAI has not yet published detailed benchmark data; a technical report is promised in the coming months. TechCrunch notes the Broadcom partnership was first announced in October 2025 and that OpenAI's custom-chip ambitions have been rumored since early 2025 as a way to reduce dependence on Nvidia.

Training workloads are expected to remain on Nvidia hardware for now; Jalapeno targets inference, the step that runs pre-built models in response to user requests. Even modest reductions in inference cost matter given how much of OpenAI's compute spend goes to serving ChatGPT and Codex traffic, and the same dynamic, specialized inference silicon lowering per-query cost, is likely to spread across the industry as more labs reach OpenAI's scale.

Jalapeno is positioned as the first generation of a multi-generation compute platform, with initial deployment targeted for the end of 2026 and gigawatt-scale data centers planned with Microsoft and other partners. Watch for the promised technical report with concrete performance-per-watt numbers, and for whether Broadcom's other hyperscale customers respond to a competitor's custom silicon shipping into production.

OpenAI's move follows a now-familiar industry pattern: as frontier labs scale, they tend to shift from buying general-purpose GPUs to co-designing inference-specific silicon, as Google and Amazon did with TPUs and Trainium. That pattern is about industry economics broadly, not a specific claim about OpenAI's competitive intentions toward Nvidia beyond what the companies have publicly stated.