Key Points
- OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom chip, on June 24
- The processor is built for inference, not training, and tuned for low operating cost
- It went from design to tape-out in nine months, sped along by OpenAI's own models
- Initial data-center deployment is targeted for late 2026
OpenAI now has its own silicon. On June 24, the company and Broadcom unveiled Jalapeño, OpenAI's first custom processor — an accelerator built specifically for inference, the work of running a model after it has been trained. Pre-training, OpenAI says, will keep running on Nvidia hardware; Jalapeño is aimed squarely at the cheaper, higher-volume half of the workload, where serving billions of queries a day turns electricity and chip cost into the binding constraint.
Broadcom is running the same design-partner playbook with Meta, whose Iris chip enters production in September 2026.
What Jalapeño is
OpenAI calls it an "Intelligence Processor," architected around its own view of how large-language-model inference should run. The company says early testing shows performance per watt "substantially better than current state-of-the-art," with engineering samples already running workloads in the lab, including its GPT-5.3-Codex-Spark coding model. The chip went from initial design to manufacturing tape-out in roughly nine months — fast for custom silicon — and OpenAI used its own models to accelerate parts of the design and optimization. "We have a deep understanding of the workload," OpenAI president Greg Brockman said. "We've really been looking for specific workloads that are underserved."
The full-stack play
Jalapeño is the first concrete output of the strategic collaboration OpenAI and Broadcom announced in October 2025 to deploy 10 gigawatts of OpenAI-designed accelerators. Sam Altman's company is now following the path Google blazed with its TPU and Amazon with Trainium: design the inference chip in-house, lean on a partner like Broadcom for the silicon implementation, and pull the most expensive, supply-constrained part of the stack out of Nvidia's hands.
The in-house silicon push isn't OpenAI's alone: rival Anthropic is now in talks with Samsung to develop its own custom AI chip, widening the race to design around Nvidia.
The detail worth sitting with is that OpenAI used its own models to help design the chip that will run its models — the first visible loop of AI compounding on its own hardware roadmap. A nine-month tape-out for a frontier-class accelerator is the kind of cycle time that, if it holds, reshapes who can credibly build custom silicon at all, and turns "we depend on Nvidia" from a strategic fact into a temporary one. That dependence is exactly what's now being financed at scale: Nvidia is reportedly in talks to guarantee $250 billion for OpenAI's planned Ohio data center, underwriting the very infrastructure Jalapeño is meant to make less Nvidia-dependent.
The same repricing is now hitting the inference layer directly: model-serving company Baseten raised a $1.5 billion Series F at a valuation of up to $13 billion, a bet on the half of the AI bill that recurs every time a model runs — exactly the workload Jalapeño is built to serve.
Source: TechCrunch, CNBC
Comments