When Mark Zuckerberg gets a bold new business idea, he likes to throw money at it. Last summer, he dropped $14.3 billion for a 49% stake in Scale AI, allowing him to poach its wunderkind founder Alexandr Wang to lead a new project to build artificial-intelligence systems that surpass human intelligence. He doled out $300 million pay packages to poach top scientists to build an all-star team called Superintelligence Labs.
Maybe it was the clashing egos or Meta Platforms Inc.’s chronic inability to develop an original idea that will ever be as lucrative as social media — see the company’s painfully expensive investments in the metaverse and crypto — but several of its rock star researchers have since fled to other labs, and its new Muse Spark AI system hasn’t matched the top benchmarks of frontier models from OpenAI and Anthropic PBC. Meta itself acknowledged gaps on coding and agents, and its attempts at building open-source AI models under the brand name Llama have also floundered.
Zuckerberg isn’t quite throwing in the towel, but his planned pivot to becoming a cloud business provider, recently reported by Bloomberg News, suggests a grudging acceptance that he won’t get the money or prestige he craves from building AI models. He’s now exploring plans to sell computing power to rival AI labs — an unsexy, strategic pivot to infrastructure that could become a lucrative business line, so long as the AI boom doesn't go bust.
AI builders continue to need computing power, but constrained supply makes it expensive. Equipment shortages, permitting barriers and difficulties connecting to electricity grids have snarled tech company efforts to build new data centers. Blackstone Inc.’s QTS recently abandoned plans to build part of a 2,100-acre data center in Virginia, after local residents protested for years to kill the project.
But Meta already has a stable of more than 30 data centers commanding 4.8 gigawatts of North American capacity, which makes it fourth among the major operators after Amazon.com Inc.’s 10.6 GW, Microsoft Corp.’s 5.5 GW and Google’s 5.2 GW, according to a recent analysis by investment firm Jefferies Group. Any investor in Meta has a painful memory of Zuckerberg’s questionable plan to spend billions of dollars to boost that capacity — including between $115 billion and $135 billion in capital expenditure for 2026 — with an aim to get to more than 10 GW by the end of this year.
Using computing power for your own needs and selling the spare capacity echoes some of the logic behind Amazon’s launch of its juggernaut AWS cloud business two decades ago, one of the most successful moves in business history. AWS currently represents about 17% of Amazon's revenue but contributes the majority of its operating profit, generating roughly $45.6 billion in operating income in 2025, at margins of around 35%. Amazon is essentially a cloud company with a store attached.
But a closer parallel to Zuckerberg is a man he has long sought to emulate: Elon Musk. The Space Exploration Technologies Corp. billionaire (and occasional trillionaire) has begun successfully leasing computers for running AI to Anthropic and Alphabet Inc.’s Google. While Zuckerberg benefits from the buildout of server farms to run his social media empire, Musk was simply good at ginning up his data centers Colossus 1 and Colossus 2 at lightning speed, trucking in enormous mobile natural-gas turbines to create power on-site rather than waiting months or years for permission to siphon it from the city grid. Now, Anthropic alone is set to contribute $15 billion in annual revenue to SpaceX as Musk’s new tenant.
Still, Musk has had to admit defeat in building frontier AI, with his Grok model falling behind on capability benchmarks and barely taking market share in selling AI to companies.
Musk and Zuckerberg have the ignoble honor of establishing “The Tech Bros Who Failed at AI” club, but doing so comes with a galaxy of benefits because of their pivot to infrastructure. Tech analyst Richard Windsor at research firm Radio Free Mobile suggests Meta could sell a thin slice of spare data center capacity in 2026 and scale up to five gigawatts’ worth by 2030, priced at roughly what it costs to build the facilities in the first place. Windsor suggests the new business could add around $70 to Meta's share price, currently around $583, though that uplift rests on how much capacity Meta will actually free up — still very much unknown — and what customers will pay for it — another uncertain variable.
Meta has long needed to diversify itself away from the cyclical online advertising business from which it derives about 98% of its revenue, especially as it faces new challenges from chatbots luring people away from websites and Google search results. A compute-leasing business seems the first credible attempt since its disastrous efforts with the metaverse and crypto. It also offers both Zuckerberg and Musk a hedge against failing to build AI while capitalizing on a valuable fixed asset — AI chips and servers — that might otherwise be sitting idle.
It may not matter if Zuckerberg’s pivot is a shrewd strategic move that was in his back pocket, or a consolation prize. With compute still in high demand, he’s well positioned to reap the benefits of the shift.
A message from Advisor Perspectives and VettaFi: Discover something new! Click here to register for our upcoming webcasts.
Bloomberg News provided this article. For more articles like this please visit
bloomberg.com.
Read more articles by Parmy Olson
