OpenAI unveils its first custom chip, built by Broadcom

OpenAI has a new chip, and it comes with a spicy name: Jalapeño. Built in partnership with Broadcom, this is the company's first stab at custom silicon, and it's aimed squarely at inference — the part of AI where a trained model actually answers your prompt. OpenAI says its own models helped design the thing, which is either a nice bit of symbolism or a genuinely useful engineering shortcut, depending on how cynical you're feeling.

The headline claim is better performance-per-watt than current top-tier hardware, though that's based on early testing, not shipped product. Still, the strategic logic is obvious. OpenAI burns an enormous amount of money running inference for things like real-time coding models, and even modest efficiency gains at that scale translate into real savings. Training frontier models will almost certainly stay on Nvidia GPUs for now — that workload is too demanding and too entrenched to swap out yet. But inference is a different beast, and it's one OpenAI clearly wants to own more of.

This isn't a novel idea industry-wide. Google has TPUs, Amazon has Trainium and Inferentia, and both built those chips for exactly the same reason: reduce reliance on Nvidia and tailor hardware to your own workloads instead of buying whatever's on the general market. Greg Brockman said as much on OpenAI's podcast, describing the company's approach as hunting for underserved workloads it understands deeply enough to accelerate directly. That's a polite way of saying Nvidia's chips, brilliant as they are, weren't built with OpenAI's specific inference patterns in mind.

What's more interesting than the chip itself is what it signals about OpenAI's ambitions. The company isn't just shipping models anymore — it's building the data centers, the networking, the scheduling systems, and now the silicon underneath all of it. OpenAI framed this explicitly in its announcement, describing a full-stack approach where every layer, from chip architecture to product experience, gets optimized toward the same goal. That's a very different company than the one that shipped GPT-3 as an API a few years back. It's starting to look a lot more like an infrastructure company that happens to also make chatbots.