OpenAI and Broadcom have unveiled Jalapeno, the first OpenAI-designed processor, and the pitch is blunt: better performance per watt than current leading chips. That is a big claim in a market where Nvidia still sets the pace for AI hardware, and it points to where the real fight is heading – not just raw speed, but how efficiently a chip can keep large language models running at scale.
OpenAI says it built Jalapeno from scratch with Broadcom and Celestica, using what it learned from LLMs and inference workloads. The company has already put engineering samples into lab work on target frequency and power, including GPT-5.3-Codex-Spark. That kind of in-house chip effort usually starts as a hedge against supplier dependence, then turns into leverage when the bills for training and serving models get truly absurd.
What OpenAI says Jalapeno is built for
OpenAI says Jalapeno was designed to be flexible across LLMs, with an architecture shaped around inference needs for current and future AI models. The chip reduces data movement and balances compute, memory, and networking resources so the system gets closer to theoretical peak performance in real use. In AI hardware, that last part matters more than the marketing copy: a chip can look great on paper and still waste cycles shuffling data around like it missed the train.
Broadcom’s role is also telling. Its Tomahawk networking chips are part of the platform push, which suggests OpenAI is not just chasing a faster processor but a full-stack setup that can scale into production. That is the same direction other AI giants have taken, with custom silicon becoming a familiar response to Nvidia’s dominance and the power constraints that come with stuffing more model capacity into existing data centers.
The performance claim is still preliminary
OpenAI says the final performance has not been measured yet, but early tests indicate Jalapeno should deliver significantly better performance per watt than current state-of-the-art technology. A fuller technical report is due in the coming months. For now, the key detail is not that OpenAI says it has built a miracle chip; it is that the company wants ownership of the bottleneck, and inference is where the economics of AI are becoming painfully obvious.
- First OpenAI processor, developed with Broadcom and Celestica
- Designed for LLM inference and future model flexibility
- Engineering samples are already running lab workloads
- OpenAI says early tests show much better performance per watt
Why this matters beyond one chip
If Jalapeno works as advertised, OpenAI gets more than a faster accelerator. It gets bargaining power, tighter control over model deployment, and less reliance on off-the-shelf hardware that every rival can buy. The broader pattern is easy to see: AI companies are moving from software-only bets into custom silicon because the cost of scaling frontier models is now a hardware problem dressed up as a product strategy.
The open question is whether Jalapeno becomes a meaningful volume chip or just a strategic proof point. Either way, the message to Nvidia is obvious: the biggest AI customers no longer want to rent the whole stack forever. They want to own the boring, expensive plumbing too.