Meta is building a cloud division through which it would sell excess computing power for artificial intelligence. This was reported by Bloomberg, citing people familiar with the matter. Mark Zuckerberg’s company wants to make money from the infrastructure into which it has poured tens of billions of dollars while taking on three established rivals that have dominated the industry for a decade.
The project is expected to be called Meta Compute. According to Bloomberg, it is being led by infrastructure chief Santosh Janardhan together with Daniel Gross of Meta Superintelligence Labs and President Dina Powell McCormick. The company is considering two ways to offer the computing power.
Two options and the competition
The first option would give developers access to Meta’s models running on its own infrastructure, including the closed Muse Spark model. This would be similar to Amazon Web Services’ (AWS) Bedrock service. The second option is to sell raw computing power in much the same way as so-called neoclouds such as CoreWeave.
Whichever path Meta chooses, it will put itself in direct competition with AWS, Google Cloud, and Microsoft Azure. This trio dominates the enterprise market for AI computing power. Their combined contracted future business exceeds one trillion dollars, meaning Meta faces three deeply entrenched rivals.
More interesting than the battle with hyperscalers, however, is who Meta threatens the most. The company itself was one of the neoclouds’ most important customers. In April, it expanded its contract with CoreWeave to $21 billion and signed contracts with Nebius worth up to another $27 billion. Altogether, it committed roughly $48 billion to renting third-party graphics cards because its own construction could not keep pace with demand.
Now the customer could become a competitor. “Adding Meta’s capacity to the market is more likely to affect neoclouds than the major hyperscalers. Companies such as CoreWeave and Nebius have built their growth around Meta, and Meta may no longer need them,” reported Reuters.
Zuckerberg hinted at it in the spring
At a meeting with shareholders in May, Zuckerberg said that entering the cloud business was “definitely on the table.” He added that companies approach Meta “almost every week” seeking to buy access to its models or spare computing capacity.
Meta’s CEO described the entire idea as insurance against the company ultimately building more than it can use itself. “If we get to a point where we feel we’ve overbuilt, then this is an option we have,” he said. Selling the surplus would thus turn the looming problem of overinvestment into a revenue-generating activity.
That threat is not insignificant. In April, Meta raised its estimate for this year’s capital expenditures to between $125 billion and $145 billion. It cited more expensive components and competition for land, energy, and construction work. In the same week, Zuckerberg told employees that roughly 8,000 planned layoffs were a direct consequence of the infrastructure budget.
Meta and its data centers
In February, Meta signed a $100 billion agreement with AMD for 6 GW of computing power. Its contracts for graphics cards from AMD and Nvidia are worth around $110 billion combined. The company is also developing its own MTIA chips for inference and has signed a multibillion-dollar deal with Amazon for Graviton processors to address a shortage of standard CPUs.
This also includes enormous data centers. The Prometheus and Hyperion campuses are designed to scale to 1 GW and eventually as much as 5 GW. Capacity on this scale comes in large, indivisible blocks timed according to demand forecasts. This is precisely why a company that until recently paid neoclouds tens of billions for time on graphics cards could suddenly find itself with excess computing power that it would be profitable to sell.
A similar scenario has already played out at SpaceX. It leased the entire capacity of the Colossus 1 data center in Memphis, more than 300 MW, to Anthropic for roughly $1.25 billion per month. It subsequently agreed to lease capacity to Google as well for nearly one billion dollars per month. According to an estimate by Bloomberg Intelligence, these agreements could generate more than $50 billion by 2028.
Meta’s market position
The Reality Labs division alone deepened its losses by more than four billion dollars in the first quarter of this year. Meta’s shares were down roughly 15 percent over the past year and underperformed the S&P 500 index.
Moreover, enterprise cloud is a service industry. That means service contracts with availability guarantees, sales engineers, and lengthy purchasing cycles—in other words, things Meta has never operated before. The company does own MTIA chips, one of the largest fleets of Nvidia graphics cards, the Hyperion campus, and Llama, the most downloaded family of open models. A paid interface where developers can rent time on Meta’s chips to run its models is a logical product. So is leasing raw capacity to customers priced out elsewhere. But turning that into a functioning business is something else entirely.
Meta’s shares jumped more than 10 percent after the announcement, their largest one-day gain in more than five months. CoreWeave, by contrast, fell 10.8 percent and Nebius dropped 12.4 percent. The market thus made it clear who it believes has the most to lose.
Zuckerberg will have an opportunity to give the project a more concrete shape during the second-quarter earnings call. Until then, Meta Compute remains primarily an insurance policy against a budget that continues to swell every quarter.
Sources: tomshardware.com and finance.yahoo.com
