OpenAI just stepped into the silicon business. On Wednesday, the company unveiled its first custom-built inference processor, designed and manufactured with Broadcom, according to TechCrunch AI. The chip is called Jalapeño, and OpenAI says it was built specifically for the demands of its own inference systems. In a twist worth noting, the company says its own AI models helped design the chip.
This is significant because it marks OpenAI’s move from buying hardware to building it. The Broadcom partnership was announced back in October, but rumors of an in-house chip had circulated for far longer, all pointing to the same goal: cutting OpenAI’s heavy reliance on Nvidia’s GPUs.
What Jalapeño Actually Does
Jalapeño is built for inference, not training. That distinction matters. Inference is the work of running a finished AI model when you type a prompt and wait for an answer. Training is the much heavier job of building the model in the first place.
The chip is still in testing, but OpenAI reports early results show much better performance-per-watt than current top alternatives, as TechCrunch AI details. Translation: more output for less electricity. In the announcement, OpenAI leaned hard on the chip’s low operating cost when running real-time coding models.
A few things stand out:
- Inference only, for now. Heavier tasks like pre-training will likely stay on Nvidia hardware.
- Cost is the point. Even small cuts to inference costs can move OpenAI’s bottom line in a big way.
- Self-designed. OpenAI’s models assisted in developing the silicon, a sign of where this loop is heading.
Why This Matters for the Industry
OpenAI isn’t first here, and it knows it. Google and Amazon have both built their own custom chips, usually called “AI accelerators,” which are pieces of silicon designed to speed up machine learning work. Google has its TPUs. Amazon has Trainium and Inferentia. OpenAI is now joining that club, and the message to Nvidia is clear: the biggest AI labs want to own more of their stack.
The economics explain the push. Running these models at scale is expensive, and inference is the cost that repeats every single time a user hits enter. President Greg Brockman framed the strategy on OpenAI’s in-house podcast shortly after the Broadcom deal went public. “We have a deep understanding of the workload,” Brockman said. “We’ve really been looking for specific workloads that are underserved, [and asking] how can we build something that will be able to accelerate what’s possible?”
What stands out to me is the vertical ambition. OpenAI put it plainly in the announcement: it is “designing the infrastructure underneath” its models, including “chip architecture, kernels, memory systems, networking, scheduling, deployment systems, and product experience.” The company argues that operating across the full stack lets it optimize every layer toward one goal: “making its models faster, more reliable, and more affordable for users.”
The Bigger Picture
Optimizing inference may turn out to be one of the deciding factors in the economics of AI. As models get baked into more products, the cost of serving them becomes the cost of the business. OpenAI is already building agentic products like Codex, the models behind them, and the data centers to run them. Custom chips are the logical next layer down.
For practitioners, here’s what to watch:
- Pricing pressure. If inference gets cheaper for OpenAI, that could eventually show up in API and product pricing.
- Nvidia’s grip. Don’t expect it to break overnight. Training still runs on Nvidia, and that’s where much of the spend lives.
- The custom-silicon trend. Three of the largest AI players now build their own chips. Expect others to follow.
Jalapeño is still being tested, so the real proof will come when it runs production workloads at scale. But the direction is set. OpenAI wants to control the hardware its models run on, and it’s willing to build the chips to do it. You can find the full details in the original report at TechCrunch AI.