Meta Builds Cloud Business to Sell AI Compute

If Meta follows through, it would add a fourth major seller of large-scale AI compute and hosted models alongside AWS, Azure and Google Cloud - and it would do so via two distinct tracks rather than one. For practitioners managing training and inference budgets, the more consequential track is not the neocloud-style raw-capacity resale (a crowded space CoreWeave, Nebius and others already occupy) but the Bedrock-style offering: paid access to Meta's own frontier models, including its Muse Spark line, running on infrastructure Meta already operates at scale. That would give buyers a fourth credible source of frontier-model API access with different pricing and licensing terms than the incumbent clouds.

Citing people familiar with the matter, Bloomberg (Riley Griffin and Kurt Wagner) reported Meta is building the cloud business under an internal initiative called Meta Compute, led by Santosh Janardhan (Meta's head of infrastructure), Daniel Gross (a leader inside Meta Superintelligence Labs) and Meta President Dina Powell McCormick. Two offerings are under consideration: selling access to AI models - including Meta's own Muse Spark models - hosted on its existing infrastructure, similar to AWS's Bedrock; and selling "raw" computing capacity, similar to neocloud businesses like CoreWeave. A Meta spokesperson declined to comment, and Bloomberg noted the plans are still in development and could change. Reuters independently reported the story (citing Bloomberg) and confirmed the premarket stock move but said it could not independently verify the underlying details.

Bloomberg reported Meta shares rose as much as 8.6% in premarket trading before paring gains; Reuters and other outlets, checking at a different point in premarket trading, cited a more modest gain near 6% - the kind of spread that is normal for fast-moving premarket quotes rather than a factual conflict. CoreWeave, viewed as a potential competitor, fell as much as 9.7% on the report, per Barron's.

The quote outlets have circulated - "It's definitely on the table" - comes from Meta's May shareholder call, where Zuckerberg elaborated further: "Almost every week there are different companies that come to us from the outside asking us to both stand up an API service or asking if we have compute that they could buy from us at some premium to what we've bought it at... We haven't done that yet because we think we have a use for the compute. But obviously if we get to a point where we feel that we have overbuilt, then that is an option that we have."

This fits a broader pattern of AI-infrastructure-heavy companies turning capex into a second revenue line. Elon Musk's xAI (via SpaceX) has already been renting capacity from its Memphis data center to Anthropic and has a compute arrangement with Google - a business Bloomberg Intelligence estimates could generate over $50 billion for xAI by 2028 and $100 billion by 2030. Meta's own infrastructure buildout includes existing compute deals with CoreWeave, Google and Oracle, underscoring how tightly coupled the major AI labs and hyperscalers have become on the supply side.

The practical value of either Meta track depends on product details not yet published: billing units (token, GPU-hour, spot vs. reserved), isolation and tenancy guarantees, model licensing terms, and whether Meta targets training or inference workloads. Teams evaluating vendor concentration risk should watch for Meta's first official product announcement rather than react to the rumor itself - Bloomberg, Reuters and Barron's all note the plans remain unconfirmed by Meta and subject to change.

An official Meta Compute product announcement or developer documentation; published pricing/billing units and SLAs; named early customers or pilots; whether Meta prioritizes the Bedrock-style model-access track or the raw-capacity track first; and any expansion of compute-leasing arrangements elsewhere in the industry (xAI-Anthropic being the current template).

The report is significant because it points to a widening pool of large-scale AI compute and frontier-model API supply from a company that has spent hundreds of billions on infrastructure it has not yet needed to fully utilize externally - but as of this reporting, Meta has not confirmed the plan, and no pricing, timeline or product name beyond the internal "Meta Compute" label has been made public.