Meta is building a cloud business to sell excess AI computing capacity to outside customers, according to Bloomberg, confirmed by CNBC. The stock surged 10% on Wednesday. Shares of neocloud competitors CoreWeave and Nebius Group each fell roughly 12%.
What Meta Is Selling
The company is still deciding the scope. One option is selling raw computing power. The other is offering access to AI models hosted on Meta’s infrastructure, according to Bloomberg. Either path converts idle GPU capacity into recurring revenue.
Meta told investors in April it plans to spend up to $145 billion on capital expenditures in 2026 for data centers and GPU procurement, according to CNBC. That number had made investors uneasy. A cloud business that monetizes unused capacity changes the math: infrastructure spending becomes partially self-funding rather than a pure cost center.
The Competitive Fallout
The immediate market reaction tells the story. CoreWeave and Nebius, both neocloud companies that built their business on selling GPU compute to AI developers, each lost about 12% of their value within hours of the announcement, per CNBC.
Meta entering cloud compute puts it in direct competition with Amazon Web Services, Microsoft Azure, Google Cloud, CoreWeave, and the growing cohort of inference-focused providers like Groq. The difference: Meta already owns the infrastructure. It does not need to raise billions to build data centers because it already committed the capital for its own AI workloads. Excess capacity is a byproduct, not a fundraise.
Inference Demand Exceeds Supply
Demand for AI computing power has far outpaced supply since OpenAI launched ChatGPT in late 2022, according to CNBC. Model developers and enterprises deploying AI agents at scale need inference capacity, the compute required to actually run trained models on live workloads. That bottleneck has turned into a business opportunity for any company holding surplus GPUs.
The timing aligns with a broader shift in spending from model training to inference. JPMorgan revised its global AI capex forecast to $5.5 trillion through 2030 in its midyear outlook. The BIS warned last week that circular financing in AI infrastructure could create credit risk. Meta’s move reframes its own capex story: instead of burning cash to train models, it can generate cloud revenue from the same hardware.
The Agent Infrastructure Layer
For teams building autonomous agents, Meta’s cloud business creates another option for inference-on-demand. Agents executing multi-step workflows consume compute non-linearly, with each tool call, reasoning loop, and retry burning GPU cycles. A new large-scale provider entering the market puts downward pressure on inference pricing, which directly affects the economics of running agents in production.
Whether Meta captures meaningful cloud market share or simply forces incumbent pricing lower, the entry validates the thesis that inference compute is the next commodity market in AI infrastructure.