Sachin Katti: The former academic behind OpenAI's trn-dollar AI buildout

By Dina Bass and Shirin Ghaffary

 

For months, OpenAI’s Sam Altman has been talking up plans to spend trillions of dollars on an unprecedented buildout of data centers stocked with advanced chips to power AI.

 

The task of turning that ambition into a reality now falls in part on Sachin Katti, a recent hire at OpenAI who spent much of his career in academia and likes to say he didn’t have a real job until he started working at Intel Corp. more than four years ago.

 

Katti, previously Intel’s chief technology and AI officer, joined the ChatGPT maker in November to serve as its head of industrial compute. In that role, Katti reports to OpenAI President Greg Brockman and oversees its industrial partnerships, engineering, and delivery and operations teams.

 

 

Katti, who spent more than 15 years as a Stanford professor, now works on finding additional data center capacity and lining up ways to get more components like AI chips and memory. And he’s doing so at a time when data center operators are contending with power grid constraints, memory chip shortages and growing pushback from local communities.

 

“We are very paranoid about the surprises that might be ahead of us,” he said in an interview with Bloomberg News, referring to the risks of everything from supply chain shortages to geopolitics. At the moment, he said, OpenAI is happy with its progress on infrastructure. 

 

To ensure it stays happy, Katti and the company are pushing aggressively to broaden the mix of suppliers OpenAI relies on and the geographic regions where it has a data center footprint, with the aim of moving as quickly as possible in its infrastructure expansion.

 

OpenAI previously said its computing capacity more than tripled in 2025 to roughly 1.9 gigawatts. By comparison, a single gigawatt is enough to power about 750,000 US homes at any given time. Altman, for his part, has said his goal is to eventually “produce a gigawatt of new AI infrastructure every week.”

 

To get there, OpenAI is working with partners on data centers in several US states, as well as in Europe, the Middle East and South America. It recently opted not to lease an expansion to its massive Stargate-branded data center in Abilene, Texas, built by Crusoe. Part of the reason, the company said, is because it had better options in other states like Wisconsin that enabled a more rapid expansion of capacity using the newest Nvidia Corp. chips.

 

OpenAI has also struck multibillion-dollar agreements with multiple chipmakers, including Cerebras Systems Inc. for running AI models after they’ve been trained, and Amazon Web Services for extra supply. “One of the attractive things about that is, frankly, the supply diversification,” he said of a recent deal to tap 2 gigawatts of capacity from Amazon.com Inc.’s Trainium chips. 

 

The AI developer is also making its own chips through a partnership with Broadcom Inc. OpenAI’s chip is on schedule to be deployed widely next year, Katti said, confirming recent comments from Broadcom Chief Executive Officer Hock Tan. Katti also said Tan’s prediction that deployments of the processor would reach 1 gigawatt of capacity next year is “in the right zip code.”

 

Still, OpenAI continues to rely on market leader Nvidia for the hardware used to train its cutting-edge AI models. OpenAI plans to deploy about 1 gigawatt of computing using Nvidia’s forthcoming Vera Rubin AI accelerators later this year. “It’s at that scale, probably even larger over time,” he said.

 

In the chip wafer and memory area, OpenAI is trying to help convince manufacturers to invest in new plants by assuring them there will be strong demand for years to come. Memory makers are wary of over-expanding and getting burned by a glut as happened to Micron Technology Inc. and SK Hynix Inc. after the pandemic ended.

 

“They don’t want to go into a boom-and-bust cycle,” Katti said. “They are very conservative on how they want to invest, and understandably so.”

 

Katti and the company are also keeping their eyes open for good acquisition opportunities that could further speed up their infrastructure efforts. 

 

“If you think about the physical world, it’s still largely unchanged in terms of how quickly things can get built,” he said. “It’s a three-year cycle. But our needs are changing every three months.”