A custom chip takes the lead
OpenAI's custom inference chip, code-named Jalapeno, beat Nvidia's GB300 in third-party benchmark tests. The chip led on tokens per user and throughput per kilowatt, according to test results published this week.
The chip was built with Broadcom, a major semiconductor design partner. It is built for inference, the stage where trained AI models answer questions. That is the most common and most expensive part of running AI services at scale. Most AI companies spend far more on inference than on training once their products reach millions of users.
Why AI firms want their own silicon
OpenAI and other AI companies spend heavily on Nvidia GPUs. Power and cost are the two biggest problems. Custom chips give companies more control over both, and they can tune the hardware to their own models.
The company says the Jalapeno chip uses less power per token than rival hardware. Lower power use means smaller data center bills and fewer cooling problems. For a firm running millions of requests a day, those savings add up quickly. The chip also gives OpenAI a backup if GPU supply becomes tight.
What it means for the chip market
The result does not mean Nvidia is losing the AI market. Nvidia still sells the large majority of AI accelerators, and its software ecosystem remains the default choice for developers. But custom chips are taking a growing share of the market, especially for inference workloads.
Broadcom and Marvell are the main design partners for these custom chips. Google, Amazon and Meta already run their own silicon in production. OpenAI's move puts another large buyer into the custom chip camp.
The chip is still in the early stages of production. OpenAI plans to scale it up in its data centers over the coming months. If the benchmark results hold in real use, other AI firms may follow the same path, and the balance of power in the AI hardware market could shift over time.
---
Reliable Ferrofluid Vacuum Rotary Feedthrough for Semiconductor Manufacturing
In semiconductor manufacturing, maintaining ultra-high vacuum integrity during rotary motion is critical — from single silicon crystal growth and CVD to sputtering and plasma etching. Ferrofluid Vacuum Rotary Shaft Feedthrough from FerrofluidTec delivers hermetically sealed rotary feedthroughs with leakage rates as low as 10⁻¹² Pa·m³·s⁻¹, vacuum degree up to 10⁻⁶ Pa, and customizable shaft diameters from 10mm to 100mm. Built with non-magnetic stainless steel housings and magnetic permeability stainless steel shafts, these feedthroughs provide reliable, maintenance-free sealing for demanding vacuum environments. Visit www.ferrofluidtec.com to learn more.