OpenAI Partners With Broadcom in Third Major Chip Deal
OpenAI has signed its third major semiconductor partnership, teaming up with Broadcom to co-design custom AI chips and purchase 10 gigawatts of computing capacity. This landmark deal brings OpenAI’s total secured computing power to over 26 gigawatts—equivalent to 26 nuclear reactors—as the company races to build infrastructure for services like ChatGPT.
Key Takeaways
- OpenAI’s third major chip deal with Broadcom includes custom co-designed processors
- Total computing capacity now exceeds 26 gigawatts across all partnerships
- Potential additional spending of $350-500 billion on top of existing $1 trillion commitments
- First time OpenAI has produced its own AI chips specifically for inference
Massive Chip Acquisition Spree
According to Financial Times reports, the Broadcom partnership follows recent deals with Nvidia and AMD. In September, OpenAI committed to 10 gigawatts of Nvidia chips, followed by 6 gigawatts from AMD last week.
The company also signed a $300 billion data center agreement with Oracle spanning five years. Combined with previous commitments, OpenAI could spend an additional $350-500 billion beyond the roughly $1 trillion already allocated for chips and data centers.
Custom Chip Development
OpenAI has collaborated with Broadcom for 18 months to develop custom processors specifically optimized for running its AI models. This marks the first time the Microsoft-backed startup has produced its own AI chips.
This will give us a gigantic amount of computing infrastructure. The race to develop AI infrastructure is the biggest joint industrial project in human history.
— Sam Altman, OpenAI CEO
The custom chips are designed specifically for inference—the process where AI models respond to user requests. Deploying this massive computing capacity will require building extensive new data center infrastructure.
Sam Altman announced the partnership in a recent podcast, emphasizing the scale of investment needed to support advanced AI systems like ChatGPT.
