SILICON VALLEY — "One of the biggest constraints is that we haven't been able to roll this out globally, because the service requires significant computing resources to run."
That was Chief Executive Mark Zuckerberg's answer on a conference call held after Meta released third-quarter earnings on the 21st, local time. An investor had asked what was limiting the growth of the user base for Muse, the artificial intelligence agent launched on Sept. 8. Put plainly, the question was why Meta has kept Muse available only in the United States and Canada for nearly two months. From an investor's standpoint, it was a fair thing to wonder: Muse has been a hit, passing 10 million sign-ups within 40 days of launch, yet the service still runs only in the North American market.
The answer was a shortage of data centers. Running a personal AI agent around the clock, every day of the year, requires an enormous amount of computing resources. It takes a vast number of AI chips, including graphics processing units and central processing units, along with proportionally enormous costs and electricity. Linking the service to the social media applications Meta owns — Facebook, Instagram, WhatsApp and Threads — could maximize revenue, but the computing capacity Meta has secured so far could not support it. Dots, the competing model from OpenAI, is available only on plans costing more than $100 a month, while Muse is free, leaving it especially short of capacity.
The mood has shifted completely from early July, when reports said Meta was weighing a plan to sell its spare computing capacity to outside buyers. "The 10 million download figure is even more impressive given that it was achieved in the U.S. and Canada alone," Zuckerberg said, adding that the company could have attracted far more users had it launched in more regions. Competition is intensifying to the point that even Anthropic has unveiled a personal AI agent, and Meta's failure to secure enough data centers is holding it back.
"Running Muse requires significant computing, but we expect to improve cost efficiency by more than tenfold next year," Zuckerberg said. "That will allow us to secure profitability as revenue scales up." He projected that "revenue per user for Muse will be higher than for our existing family of apps — on the order of several hundred dollars a year." Truist Securities estimated that Muse could generate $28.5 billion in incremental revenue by 2030.
Capacity shortages are a shared worry among hyperscalers that run large-scale computing infrastructure. Bloomberg has projected that the four largest hyperscalers — Amazon, Microsoft, Google and Meta — will invest $2.4 trillion (3,184.8 trillion won) in AI infrastructure over the coming years to strengthen their data center capabilities. Meta's move on the 26th to raise money for AI investment through a record $95 billion share issue was also a decision aimed at expanding infrastructure. Amazon Chief Executive Andy Jassy said on the 23rd, in a podcast interview with Ben Horowitz, co-founder of venture capital firm Andreessen Horowitz, that leasing contracts for Amazon Web Services computing capacity are nearly all booked through 2028.
Bloomberg reported last month, citing sources, that Microsoft plans to more than triple its data center capacity from 12 gigawatts now to more than 38 gigawatts by 2032. That is more than the entire state of New York uses during its hours of peak electricity demand. Two years ago, Microsoft paused construction over concerns about a data center glut, leaving it unable to respond adequately to explosive demand for AI development. Computing shortages are also why its Xbox video game console unit had no choice but to cap how long subscribers can play and to warn that users exceeding those limits would be charged.
The picture is no different at the largest AI startups. Connection errors in the latest version of Claude, Anthropic's AI model, in the first half of this year stemmed from a shortage of computing resources. The same reason lay behind OpenAI's decision to suspend its video generation service to concentrate on developing AI coding tools.
In May, Anthropic signed a contract with SpaceX, the company led by Elon Musk, to lease more than 300 megawatts of computing capacity. The deal stunned the industry when it emerged that it could be worth as much as $45 billion over several years. In late February, when Anthropic was clashing with the U.S. Defense Department over its refusal to allow military use of AI, Musk had savaged Anthropic Chief Executive Dario Amodei, saying the company "hates Western civilization." Despite the friction between them, Amodei reached out to his rival to resolve the computing shortage, and Musk publicly praised him, saying he was "impressed by Anthropic's management." For Anthropic, setting personal feelings aside and clearing a business bottleneck was the more urgent task.
Data center shortages can also plant the seeds of internal conflict at AI companies. In August, word emerged that Demis Hassabis, the chief executive of Google DeepMind known as the father of AlphaGo, was stepping back from frontline management and that Google co-founder Sergey Brin was becoming indirectly involved in AI development. The news drew a range of interpretations, one of which was that Hassabis and Brin were at odds. The Wall Street Journal reported that Brin's intervention proved the trigger amid a tense tug-of-war over how to allocate limited computing resources, and that he had voiced frustration that Hassabis was not devoting enough computing power to long-term research.
