OpenAI and Broadcom Unveil "Jalapeño," a Custom LLM Inference Chip to Reduce Reliance on Nvidia

From blank-slate design to silicon in nine months

OpenAI and Broadcom on June 24, 2026, unveiled Jalapeño, OpenAI's first custom-built AI chip — an application-specific integrated circuit designed from the ground up for large language model inferenceopenai +1. The announcement, made eight months after the partnership was publicly disclosed, marks OpenAI's most aggressive push yet to control its own hardware stack and reduce its dependence on Nvidia's GPUscnbc +1. Engineering samples were physically delivered to OpenAI CEO Sam Altman and President Greg Brockman at the event, and early tests show Jalapeño already running production workloads — including GPT-5.3-Codex-Spark — in a lab environmentopenai +1.

OpenAI and Broadcom moved from initial schematics to a completed manufacturing tape-out in just nine months, a pace the companies describe as the fastest ASIC development cycle ever achieved in high-performance advanced semiconductorsopenai +1. The speed was aided by OpenAI's own AI models, which were used to accelerate parts of the chip design and optimization processtechcrunch +1. "The degree to which our models have been able to accelerate it was very surprising to us," Brockman told CNBCcnbc.

Purpose-built for inference, not general compute

Unlike Nvidia's general-purpose GPUs, Jalapeño is an ASIC — narrower in scope but purpose-fit for the specific workloads behind ChatGPT, Codex, and OpenAI's APItechcrunch +1. The architecture minimizes data movement and balances compute, memory, and networking resources so real-world performance runs closer to theoretical peak efficiencyopenai. Early tests indicate performance-per-watt "substantially better than current state-of-the-art," and Bloomberg reported the chip is expected to cut inference costs by roughly 50%venturebeat +1.

Broadcom is contributing silicon implementation and Tomahawk networking technology, while manufacturing partner Celestica handles board, rack, and system integrationopenai +1. The goal is deployment at gigawatt-scale data centers alongside Microsoft and other partners by the end of 2026, with Broadcom CEO Hock Tan projecting a full ramp in the first half of 2028cnbc.

Closing the gap with hyperscaler rivals

The launch narrows a structural gap that has long disadvantaged OpenAI relative to cloud giants like Google and Amazon, which have run proprietary AI silicon — Google's Tensor Processing Units and Amazon's Trainium — to serve AI workloads at lower costventurebeat +1. OpenAI's 2025 financials underscore the urgency: the company posted $13.07 billion in revenue against $34 billion in total operating expenses, with compute costs including more than $10.59 billion paid to Microsoft for cloud infrastructureventurebeat.

Jalapeño does not replace Nvidia in OpenAI's stack. Nvidia finalized a $30 billion investment in OpenAI as part of a $110 billion funding round earlier in 2026, and its hardware will remain central to model trainingventurebeat +1. OpenAI also holds deals with AMD, Cerebras, and AWS for additional computecnbc. What Jalapeño changes is OpenAI's leverage at the inference layer — the point where AI actually reaches users — cutting per-query costs and shoring up a credible path toward profitability ahead of a widely anticipated IPOventurebeat.