TheChek News Flow

When OpenAI in March released ChatGPT features that could turn photos into animated characters, a spike in usage forced the company to put temporary limits on the features.

It wasn’t the first time this had happened. Chief Financial Officer Sarah Friar told investors at a Goldman Sachs conference last week that the company has to regularly hold up launches of new features or AI models because of a shortage of compute capacity, or it purposely has to make some of its products work more slowly. OpenAI is “massively compute constrained,” she said.

The Takeaway

One solution: the company plans to spend about $100 billion renting backup servers from cloud providers over the next five years, executives have told some shareholders. That’s on top of the $350 billion the company has already projected to spend on server rentals from cloud providers this year through 2030.

The backup plan reflects CEO Sam Altman’s belief that access to servers will be a defining advantage in the AI race; he has told colleagues OpenAI will someday need more power than the entire U.S. power grid to run its servers.

He and his executives believe future product hits or artificial intelligence breakthroughs will cause unforeseen surges in OpenAI’s computing needs, according to several people who have spoken to them.

Public comments from Altman and Friar imply OpenAI is already using all the large clusters of Nvidia graphics processing units it has access to. If it has another viral hit that strains its computing capacity or doesn’t have enough servers to train a promising new AI model, it would be at risk of losing more business or users to competitors such as Google or Meta Platforms that may not face the same problems.

Including the backup servers, OpenAI is planning to spend about $85 billion a year on server rentals over the next five years, on average, according to the company’s private projections. To put that into perspective, customers of Amazon Web Services, Microsoft, Google Cloud and Oracle collectively spent roughly $200 billion renting servers from the four firms last year.

OpenAI’s $100 billion in backup server costs is one reason why its long-term cash burn projections total $115 billion through 2029, The Information earlier reported, and why it may need to raise that much capital. The company also projected it will start generating cash in 2030.

Despite the extra costs, there is a silver lining in the backup servers. OpenAI executives say the servers are “monetizable” because they will help the company gain additional revenues it hasn’t yet included in its projections, either by powering research breakthroughs or a surge of usage of its products, according to a person who spoke to the executives. OpenAI has projected its revenue will rise to roughly $200 billion in 2030 from $13 billion this year, largely on the strength of ChatGPT. That revenue would be greater than what Nvidia or Meta Platforms generated in the past 12 months.

Hoarding Concerns?

OpenAI could also lower its back-up server costs or monetize the servers in other ways, including by giving back the excess capacity to Oracle, Microsoft or other cloud firms that provided it—but only if the cloud providers agree. Or it could find new customers for the server clusters, though it isn’t clear which companies would be in a position to spend that kind of money (other than perhaps OpenAI’s top competitors) and whether the cloud providers would allow OpenAI to resell the capacity.

If OpenAI doesn’t use the excess servers or can’t find other customers for them, it could be seen as hoarding GPUs. Nvidia, whose hardware currently powers virtually all of OpenAI’s servers, would likely want to avoid that result because it could deprive other companies of the chips and prompt them to use rival AI chips.

Executives have projected that OpenAI’s spending on the back-up servers will peak at $40 billion in 2028. It isn’t clear why the projections show a marked decline in those costs in later years, but OpenAI has separately said it will eventually develop its own data centers and chips, which could make up the difference.

“Think of OpenAI becoming very much a full-stack company,” Friar said at the Goldman Sachs conference.

Silicon Valley founders and executives are familiar with viral products that cause server shortages—and their potential consequences. For instance, outages involving Friendster, an early social network that took off more than 20 years ago, opened the door to competitors.

OpenAI experienced several outages in the past year or so, sometimes lasting multiple hours, though it managed to maintain its dominance in the chatbot market. The company is closing in on 1 billion weekly active users by the end of the year, after hitting 700 million weekly active users in July.