OpenAI President Greg Brockman: 'We will remain in this compute shortage no matter what.'

OpenAI president Greg Brockman believes the AI industry will continue to deal with computing capacity constraints for the foreseeable future. "My prediction is very much that we will remain in this compute shortage no matter what," Brockman said during a roundtable discussion in New York City on Thursday. "Right now we have to make hard decisions on what models we actually train, what products we actually scale," he explained.

AI labs like OpenAI and Anthropic, along with Big Tech companies including Microsoft, Amazon, and Meta, are struggling to meet demand for their AI services as they juggle training new models and serving software to customers. That requires a massive amount of computing power from graphics chips like Nvidia's. Despite collectively spending hundreds of billions of dollars on data centers, the industry can't keep up.

During Google's Q2 earnings call on Wednesday, CFO Anat Ashkenazi said that despite spending heavily on its own data centers, the company will also contract with third parties to access computing capacity in Q3. Google signed a deal in June to rent AI capacity from SpaceX for $920 million per month.

Brockman noted that he sometimes observes how quickly OpenAI is ramping up capacity and assumes the company will have a surplus of compute, only for demand to increase further as OpenAI develops new models and new ways to use them. "I think that that is going to be how it will feel because we are shifting to this compute-powered economy," he explained.

The executive also said he believes customers are beginning to see a return on their investments in AI as they use the software to improve productivity. He dismissed the concept of workers and companies "token maxxing," or using as much AI computing as possible simply to show capability. "I think people are starting to see that return on investment is there if you use it right," Brockman said. "If you just token max, then yeah, of course, you're just going to get max tokens. But if you value max, if you really try to apply these technologies to problems that matter, then you get the return."