OpenAI builds first chip with Broadcom
and TSMC, scales back foundry ambition
- OpenAI is working with Broadcom and TSMC to create its first custom AI chip, targeting release by 2026
- The company is using AMD chips alongside Nvidia’s to diversify supply and cut costs amid shortages
- OpenAI dropped plans to build its own chip factories due to high costs and long timelines
- It has built a team of experts, including former Google engineers, to lead in-house chip design efforts
- Rising compute costs and demand for AI infrastructure drive OpenAI’s strategy to optimize resources and expand partnerships
OpenAI, the company behind ChatGPT, is partnering with Broadcom and TSMC to develop its first custom AI chip while also using AMD chips alongside Nvidia’s. This move aims to meet growing infrastructure demands, lower costs, and secure a more diverse chip supply. While OpenAI considered building its own factories, it has shelved the idea due to high costs and long timelines, instead focusing on designing chips in-house.
OpenAI’s strategy mirrors larger tech firms like Amazon, Meta, and Google, blending internal development with external partnerships. By diversifying its chip suppliers and creating custom solutions, OpenAI is addressing the challenges of rising chip costs and shortages, which affect Nvidia-dominated markets.
Broadcom is helping OpenAI design an inference chip, essential for deploying AI applications. OpenAI has secured manufacturing capacity with TSMC to produce this chip by 2026, though the timeline might change. Meanwhile, AMD’s MI300X chips are being tested through Microsoft Azure, aiming to compete with Nvidia’s GPUs, which currently hold over 80% of the market.
To support these efforts, OpenAI has assembled a team of 20 chip experts, including engineers from Google’s Tensor Processing Unit (TPU) team. While committed to Nvidia for advanced GPUs like the Blackwell series, OpenAI remains cautious about straining its relationship with the dominant supplier.
Running AI systems is expensive, with OpenAI projecting a $5 billion loss this year, mainly due to high computing costs. The company is working to optimize resource use and reduce reliance on a single chip provider, reflecting broader shifts in the tech industry as AI demands grow.
Read the full article: OpenAI builds first chip with Broadcom and TSMC, scales back foundry ambition
