OpenAI has shown off the first samples of Jalapeno, its in-house chip for running already-trained AI models, and says testing has begun. Built with Broadcom, the silicon is meant to give OpenAI tighter control over the hardware behind ChatGPT-style products while trimming the eye-watering cost of ChatGPT inference, the part of AI that happens after training is done and the model actually answers users.
The company says the custom approach should improve efficiency by cutting the amount of data that has to move around, which is often where generic GPUs start sweating. Broadcom chief Hock Tan said the chip could use up to 50% less resources than typical graphics processors, a claim that matters because the AI industry is increasingly judged not just on model quality, but on how cheaply it can keep the lights on.
What OpenAI says Jalapeno is built to do
According to OpenAI and Broadcom, Jalapeno was developed from scratch in a short timeline and is tuned for inference rather than training. That is a sensible bet: training gets the headlines, but inference is the bill that keeps arriving every month.
The chip is designed around the specific demands of compute, memory, and networking in high-end AI systems. Final versions are slated for large Microsoft data centers and other OpenAI partners starting at the end of this year, with a next Jalapeno version planned for 2028.
Broadcom, Microsoft and the custom silicon race
OpenAI first disclosed the Broadcom tie-up in October, part of a broader push among AI companies to reduce reliance on Nvidia’s expensive, power-hungry accelerators. Google has its own tensor chips, Amazon has been developing custom silicon for years, and now OpenAI is effectively joining the club it helped make fashionable.
Tan said the companies may beat his earlier forecast for deploying 1.3 GW of AI chips next year, because demand is running hot. That is corporate optimism with a power bill attached: the more inference OpenAI can do on its own hardware, the better its margins look, at least in theory.
The money, the financing and the risk
OpenAI has already raised $122 billion this year to fund chips, data centers, and hiring, a staggering sum that underlines how capital-intensive the AI race has become. Broadcom, meanwhile, said it built a financing mechanism with Apollo Global Management and Blackstone, while the exact payment structure with OpenAI is still being finalized.
That leaves both sides exposed to the same reality: custom hardware can lower costs later, but it burns cash first. Broadcom’s stock rose 1.6% to $386.25 after the announcement, and is up almost 10% this year, so investors are clearly willing to pay for the promise.
What happens after Jalapeno
The real test is whether Jalapeno can move from a promising sample to a workhorse inside OpenAI’s most crowded products. If it does, expect more custom chips, more partner financing, and less patience for off-the-shelf GPUs than the industry would like to admit.