OpenAI and Broadcom Reveal Jalapeno: A New AI Chip

A compact chip arrived at OpenAI this week and it felt like more than hardware. It was a statement. Named Jalapeno, the new ASIC marks OpenAI's first in-house chip effort, built in collaboration with Broadcom to accelerate AI inference for the company that brought ChatGPT to the world.

OpenAI calls the design an Intelligence Processor and describes it as the first accelerator in a platform the company is assembling to run advanced models faster, more reliably, and at greater scale. Broadcom will manufacture the silicon. OpenAI says it designed Jalapeno in roughly nine months and also developed large portions of the surrounding compute systems where these chips will run.

Why this matters for AI compute

GPUs from Nvidia have been the backbone of modern AI. They are flexible and powerful. But they come with high price tags and supply bottlenecks. ASICs trade some flexibility for efficiency and cost savings. That makes them attractive when a workload can be tightly optimized. Jalapeno is OpenAI's bet that carefully tuned hardware can deliver more compute per watt and help the company scale access to advanced models.

Is OpenAI breaking up with Nvidia? Not exactly. This is diversification, driven by demand that keeps accelerating. Earlier this year OpenAI struck deals to use Amazon's Trainium chips and signed partnerships with AMD and Cerebras. Now, with Broadcom producing Jalapeno, OpenAI aims to add another reliable route to raw compute as it builds racks of custom hardware that, at full ambition, could require around ten gigawatts of power.

There are practical milestones too. A physical sample of Jalapeno was delivered on Wednesday. The companies say they expect initial deployments by the end of 2026 and broader rollouts in the years that follow. If timelines hold, we could be watching the start of a sizable shift in how large language models are served in production.

Broadcom felt the news. Its stock rose about two percent after the announcement. The chipmaker has been one of the AI market's winners, with shares up roughly ten percent year to date and nearly sevenfold since late 2022, as major cloud and AI labs commission custom silicon.

There are trade-offs. ASICs are less flexible than general-purpose GPUs, so they work best when paired with a precise software and systems stack. OpenAI's pitch is that by designing more of that stack itself, it can squeeze better efficiency and extend access to powerful AI tools. In short: build the engine, tune the car, and you can drive farther for less.

The real test will be performance on real workloads and how quickly OpenAI can scale production with Broadcom. For now, Jalapeno is a bold move toward hardware independence and a reminder that the AI supply chain is evolving from off-the-shelf parts to custom platforms engineered for a new class of software.