Meta Compute Turns AI Capacity Into A Cloud Option

Meta is preparing a cloud-infrastructure business that would give external customers access to AI models and computing capacity, creating a potential challenge to Amazon Web Services, Microsoft Azure and Google Cloud.

The planned effort centres on an internal group called Meta Compute.

Meta has put the group around AI-infrastructure buildout and operations, with Santosh Janardhan, Daniel Goss and Dina Powell McCormick named in the leadership structure.

The service model under consideration has two parts.

One route would let outside developers pay to run queries against AI models on infrastructure owned and operated by Meta.

Another route would rent raw GPU capacity directly to customers.

Meta's proprietary generative-AI model Muse Spark is one of the models named for the query-access option.

The model has already taken on work that had initially been handled by Google's Gemini inside Meta.

Zuckerberg Says External Compute Sales Are On The Table

Mark Zuckerberg told shareholders in May that selling compute access was being considered and said the idea was “definitely on the table”.

Zuckerberg said outside companies ask Meta almost every week about API service access or compute capacity that they could buy at a premium to Meta's own purchase cost.

He explained that no external deals had been made because Meta's internal demand had consumed available capacity.

The commercial trigger is overbuild risk.

Zuckerberg said that if Meta builds more capacity than it needs internally, the company will sell the extra capacity externally.

That would turn part of Meta's AI infrastructure programme from a cost centre into a possible cloud product.

The disclosed plan still leaves open whether Meta can expose that capacity to developers with the reliability, pricing and product packaging expected in cloud services.

AI Infrastructure Spend Creates Revenue Pressure

Meta listed projected spending of as much as US$145bn on AI infrastructure this year.

The company is also investing heavily in superintelligence infrastructure and has reached large capacity agreements with CoreWeave, Google and Oracle.

Those commitments explain why investors are questioning how the infrastructure programme will translate into revenue.

A Meta Compute service would give the company a direct route to monetise spare AI capacity, while keeping priority access for its own model work.

The supply backdrop is already visible inside Meta.

Google said it could not meet Meta's demand for AI compute, and the resulting Gemini access limits delayed some internal AI efforts.

Meta then asked employees to reduce AI token consumption while more work shifted to Muse Spark.

That internal shift gives Meta Compute a second purpose: it supports Meta's own model development while also creating a possible external service if the company builds surplus capacity.

Muse Spark Still Lacks A Developer Release Date

Meta disclosed Muse Spark in April, but the model has not yet been released to developers.

Meta has not disclosed a confirmed Muse Spark developer release date, customer names, pricing, service-level commitments, GPU inventory, regional availability, or signed external compute contracts.