OpenAI and Broadcom drop Jalapeño, custom silicon for LLM inference

Training frontier models gets the headlines, but inference is where the AI bills actually pile up.

OpenAI is tackling that balance sheet directly. The company just unveiled Jalapeño, its first custom AI processor designed specifically for large language model inference, built in deep partnership with Broadcom.

Broadcom CEO Hock Tan and President Charlie Kawwas personally delivered the first silicon to Sam Altman and Greg Brockman, marking the start of a multi-generation hardware platform.

Built with AI in nine months flat

In traditional semiconductor timelines, taking a high-performance advanced ASIC from a clean sheet to physical tape-out takes years. OpenAI and Broadcom wrapped it up in just nine months.

OpenAI claims this is the fastest ASIC development cycle ever achieved in modern semiconductors. While Broadcom provided silicon implementation and Celestica handled board and rack integration, OpenAI also used its own AI models to accelerate parts of the circuit design and architecture optimization.

Lab samples are already running frontier workloads at target production frequency and power. OpenAI confirmed that early benchmarks include running GPT-5.3-Codex-Spark in the lab.

While complete numbers will arrive in an upcoming white paper, OpenAI says Jalapeño delivers performance per watt substantially better than current market leaders by matching memory bandwidth and Tomahawk networking to actual LLM runtime bottlenecks.

The great inference hedge

Jalapeño is not a general-purpose GPU designed to handle graphics, physics, or legacy machine learning. It is an inference engine built entirely around the serving patterns of ChatGPT, Codex, and OpenAI's upcoming agentic workloads.

OpenAI plans to deploy the processor at gigawatt scale with data center partners, including Microsoft, starting before the end of 2026.

As TechCrunch pointed out following the announcement, this move puts OpenAI alongside Google and Apple in hedging against single-supplier risk. Nobody is cutting ties with Nvidia tomorrow, but relying solely on one chipmaker for long-term survival is no longer an option.

Why it matters

This completes OpenAI's transition into a true full-stack computing giant. Controlling the models, the orchestration layers, and the underlying silicon allows OpenAI to squeeze out server inefficiencies that general-purpose hardware cannot address. For developers and users, that means cheaper API calls, snappier ChatGPT responses, and better margins for agent workflows that make dozens of iterative calls.

For the broader industry, it is another clear signal that Nvidia's near-monopoly is shifting. While training clusters remain heavily green, inference volume is where the true scale lives. With Google leaning into TPUs, Amazon backing Inferentia, and now OpenAI rolling its own silicon, the battle for the inference server rack is wide open.

We think the development speed alone is a fascinating proof point for using AI to design AI hardware. However, building an engineering sample in a lab is very different from deploying gigawatts of custom silicon alongside partners like Microsoft and Celestica. If execution stumbles or advanced packaging capacity tightens in 2026, OpenAI may still find itself waiting in the same hardware queues it is trying to escape.