On July 1, 2026, Bloomberg reported that Meta is preparing a cloud computing business designed to sell its excess AI compute power to third parties, as well as access to its hosted models. The project, internally dubbed “Meta Compute” according to the agency, would put the owner of Facebook and Instagram in direct competition with Amazon Web Services (AWS), Microsoft Azure and Google Cloud. When asked, a Meta spokesperson declined to comment: the company neither confirmed nor denied the report, which remains unconfirmed. Markets, however, reacted immediately, and that may be where the real story lies: compute has become such a scarce and expensive resource that the mere prospect of reselling it was enough to reshape the stock market’s view of the company.
Two services under consideration, none confirmed
According to Bloomberg, the initiative would be led by Meta’s infrastructure chief, Santosh Janardhan, alongside Daniel Gross, from the Meta Superintelligence Labs, and the company’s president, Dina Powell McCormick. Two service models would be under study: the sale of “raw” compute capacity (GPU cycles rented to developers, as specialized providers already offer), and access to models hosted on Meta’s infrastructure, including its proprietary closed-weight model Muse Spark, in a manner similar to AWS’s Bedrock service.
These elements come from a single press report. The name “Meta Compute,” the service models and the timeline have not been confirmed by Meta, which has remained silent. These reported plans should therefore be treated cautiously, in the absence of any announced commercial offering. That caution is all the more warranted because the issue directly affects the group’s market valuation.
The stock market decided before Meta did
On the day of the report, Meta shares closed up 8.8%, a notable move for a stock that was then down about 15% for the year. By contrast, specialized AI infrastructure providers, the so-called “neoclouds,” sold off sharply: CoreWeave fell 13.9% and Nebius dropped 17% in the same session. The market’s logic is straightforward: if Meta starts selling its own compute, it will rent less from these providers and could even compete with them for customers. Value shifts toward whoever owns the data centers.
That sharp divergence in a single trading session says it all. Compute has moved out of the internal cost column: it is now an asset whose scarcity creates value, and whose mere prospect of being brought to market is enough to move tens of billions of dollars in market capitalization. Compute has become a strategic commodity.
Why Meta would have compute to sell
The starting point is an investment wall. In its first-quarter 2026 results, released on April 29, Meta raised its capital expenditure forecast for the year to a range of $125 billion to $145 billion, up from the $115 billion to $135 billion announced in January. By comparison, these expenses reached $72.2 billion for all of 2025: the trajectory amounts to a near doubling in one year. Meta officially attributes the April revision to higher component prices and, to a lesser extent, additional data center costs; the initial January forecast, meanwhile, was justified by efforts around Meta Superintelligence Labs. In both cases, the group does not isolate a figure dedicated solely to AI.
That money is financing data centers at an unprecedented scale. In July 2025, Mark Zuckerberg announced the construction of several “multi-gigawatt” compute clusters: the first, called Prometheus, was expected to come online in 2026; a second, Hyperion, is set to scale up to 5 gigawatts over several years. “Just one of them covers a significant portion of Manhattan’s footprint,” the executive said at the time. When you build at that scale to absorb model-training peaks, you are left, between cycles, with unused capacity. That reserve is what Meta would reportedly like to monetize.
The idea is not new. At the annual shareholder meeting on May 27, 2026, Zuckerberg described a cloud business as “clearly on the table,” adding that “almost every week, different companies come to us from the outside asking us either to set up an API service, or whether we have compute they could buy from us, at a premium to the price we bought it for.” He said they had not done so yet because there was no excess capacity, but that the option would exist “if we get to a point where we feel we’ve built too much.”
The paradox of a dependent giant
The shift highlights a contradiction at the heart of the AI economy. Meta is both a potential seller of compute and a major buyer. The group has significantly expanded its commitments to the neoclouds it would compete with: a new roughly $21 billion commitment with CoreWeave on April 9, 2026 (running through December 2032), on top of an earlier contract worth about $14.2 billion signed in 2025, bringing the total with that single provider to roughly $35 billion, and an agreement that could reach up to $27 billion with Nebius, concluded on March 16, 2026. The company is thus renting massive amounts of infrastructure from third parties while reportedly preparing to resell its own.
This dual role has practical limits. According to a Rosenblatt Securities analyst note released in the wake of the report, Meta would not have the right to resell to third parties the capacity it leases from CoreWeave. The master agreement between the two companies does include a restriction, which prevents the customer from renting out or distributing the services without CoreWeave’s prior written consent; but the claim that resale is impossible remains an analyst interpretation rather than a clause explicitly stating a ban. That nuance matters: it limits what Meta could actually commercialize in the near term, and it is a reminder that these contractual structures are anything but fixed.
A move that goes beyond Meta
Meta would not be the first company to turn excess capacity into revenue. Starting in May 2026, the logic had already been pioneered by SpaceX and xAI, which opened capacity in the Colossus data center to outside customers: Anthropic in May, then Google and Reflection AI in June. The convergence is striking: players that overinvested in infrastructure for their own needs are discovering that it is worth money when they are not using it, and are trying to turn it into a profit center. This compute economy did not emerge overnight: as early as 2025, AI players were striking multibillion-dollar deals to secure capacity from specialized providers, as illustrated by the $11.9 billion agreement between OpenAI and CoreWeave.
This shift is redrawing the balance of power in AI. For a long time, the race was about models; now it is moving toward those who own the data centers, the energy and the chips. For companies consuming AI, one more provider broadens choice and could put pressure on prices. But the trade-off is greater concentration: a handful of giants with the capital to build multi-gigawatt clusters will control both the models and the pipeline that runs them. The debate over dependency on infrastructure providers, already intense in relation to model access, is now extending to the most fundamental resource of all: compute itself.
At this stage, the project remains a reported and unconfirmed plan. The backdrop, however, is verifiable: $125 billion to $145 billion in capex, infrastructure contracts worth tens of billions, and a stock that jumps 8.8% on a simple rumor of resale. In the AI economy, compute has become the market, whether Meta enters it or not.