Meta cloud computing AI capacity: what the cloud push means for AI builders

Meta is building a cloud business to sell excess AI compute capacity, potentially turning its massive infrastructure investment into a new revenue stream. For AI builders, this could mean more competition in model hosting and GPU access, but significant strategic details remain unresolved.

What happened

Meta is developing a cloud business to monetize its surplus AI computing power, according to CNBC and Bloomberg reports. The company will sell excess capacity to external customers, though it is still debating whether to offer access to hosted AI models or raw computing power. Meta stock rose roughly 8% on the news, while shares of AI-focused cloud providers CoreWeave and Nebius Group both dropped, with CoreWeave declining 13% and Nebius falling 15%.

Mark Zuckerberg first signaled the possibility during Meta's Q3 2025 earnings call and addressed it again in May 2026 at the company's annual shareholder meeting. "It's definitely on the table," Zuckerberg told investors, adding that if Meta overbuilds AI infrastructure, "then that is an option that we have."

The move follows Meta's plan to spend up to $145 billion on capex in 2026, primarily on data centers and GPUs for AI workloads. Meta is following a path already taken by SpaceX, which started selling excess compute this year and has inked deals worth $1.25 billion per month from Anthropic and $920 million per month from Google.

Why AI builders should care

Meta's entry into AI cloud services could increase supply of AI compute and model hosting, potentially affecting pricing and options for AI builders. Meta operates massive infrastructure for its own platforms, and even a fraction of excess capacity could rival that of dedicated GPU cloud providers. For founders and developers currently relying on CoreWeave, AWS, Azure, or Google Cloud for AI workloads, Meta could become a viable alternative for hosting and compute.

The impact was immediate: CoreWeave and Nebius stock dropped sharply, indicating that the market expects Meta's scale to disrupt current pricing dynamics. If Meta offers raw compute access at competitive rates, it could lower GPU costs for AI startups and model developers.

Practical implications

Two major unknowns will shape Meta's cloud strategy:

• Model access vs. raw compute: Meta is debating whether to sell access to its hosted AI models or raw computing power. For builders, raw compute is more flexible for custom training and deployment, but model access is simpler for inference workloads.
• Pricing and availability: No pricing details have been announced. Meta's infrastructure, built for massive internal workloads, could offer lower prices than specialist GPU cloud providers, but margins and access tiers remain unclear.

Meta's internal AI model strategy also matters here. The company launched Muse Spark in April 2026 under Alexandr Wang, whom Meta brought in from Scale AI for $14 billion. Meta positioned Muse Spark as a "powerful foundation" model, not a state-of-the-art offering. If Meta offers hosted access to Muse Spark or future models, it could compete with Anthropic, OpenAI, and Google on model hosting as well as compute.

Caveats

This move is still under development. The CNBC report notes that Meta is "debating" the exact offering, and a representative did not immediately respond to requests for comment. Concrete product details, pricing, and launch timeline remain unannounced. Meta has historically struggled to sell services beyond advertising, as covered in a separate CNBC analysis. The cloud market, especially for AI compute, is already fiercely competitive and capital-intensive. Meta's ability to successfully compete with established hyperscalers and nimble neocloud providers remains to be seen.

FAQs

What is Meta building in the cloud for AI compute?

According to Bloomberg and CNBC, Meta is building a new cloud business that could sell excess AI computing capacity to outside customers. Meta is deciding between two models: selling access to AI models hosted on its infrastructure or offering raw computing power. Mark Zuckerberg first signaled this was possible during Meta's Q3 2025 earnings call and again in May 2026, saying the option is "definitely on the table."

How will Meta monetize excess AI capacity?

Meta could monetize excess AI capacity by either offering hosted AI model access or selling raw computing power as a new revenue stream. This follows over $145 billion in planned 2026 capex for data centers and GPUs. SpaceX already sells excess compute, with reported deals paying $1.25 billion per month from Anthropic and $920 million from Google. No pricing or access details for Meta's offering have been announced yet.

Who are Meta's cloud competitors in AI compute and hosting?

Meta would enter a market dominated by Amazon, Microsoft, Google, and CoreWeave. The news immediately impacted competitors: CoreWeave stock fell 13% and Nebius Group dropped 15%. Meta's cloud scale from its existing infrastructure investment could challenge both hyperscalers and neocloud providers.

What is Muse Spark and how does it relate to Meta's cloud strategy?

Muse Spark, launched in April 2026, is Meta's first model under Alexandr Wang, whom Meta acquired from Scale AI for $14 billion. Meta positioned Muse Spark as a "powerful foundation" model rather than a state-of-the-art offering. The model underpins Meta's broader AI infrastructure push, which the new cloud business would monetize.