Meta Launches AI Cloud to Sell Excess Compute — META Surges 10%, CoreWeave Falls 10.8%
Meta is building a cloud computing business to sell excess AI capacity to outside developers and enterprises, Bloomberg reported Tuesday. The company confirmed the plans Wednesday, sending its shares up more than 10% while AI cloud infrastructure competitors CoreWeave and Nebius each fell sharply on the implications.
The structure would let customers access Meta-hosted AI models — primarily Llama variants — through a Bedrock-style API without managing hardware. Meta may also offer raw compute, moving it directly into competition with the neo-cloud providers it has been renting from.
The Numbers
| Company | Move | Change |
|---|---|---|
| META | Announced cloud business | +10% |
| CoreWeave (CRWV) | Major Meta customer turned competitor | -10.8% |
| Nebius | GPU cloud exposed to same risk | -12.4% |
Meta has signed $21B in CoreWeave AI cloud capacity through 2032. That contract now looks different: Meta may both reduce future purchasing and compete for the same external customers CoreWeave was banking on.
Why It Happened
Meta spent an estimated $150B on AI infrastructure in 2025 and has committed to further expansion. The problem is utilisation. AI workloads are spiky — training runs consume enormous capacity for weeks, then stop. Between runs, hundreds of thousands of GPUs sit partially idle.
Zuckerberg acknowledged the dynamic publicly before the Bloomberg report, noting that outside companies ask Meta for compute access “almost every week.” The new business converts that informal demand into a revenue line.
The model mirrors AWS’s origin: Amazon built cloud computing for its own e-commerce workloads, found spare capacity, and turned it into a $100B+ business. Meta is attempting the same playbook, but entering a market that already has AWS, Azure, Google Cloud, CoreWeave, and a dozen smaller GPU clouds.
What Meta Is Actually Building
The initial product is closer to managed model inference than raw cloud. Developers would call Meta’s hosted Llama models through an API, not provision raw GPU instances. Raw compute access may follow, but model inference is the first motion.
This distinction matters for CoreWeave’s risk exposure. CoreWeave sells raw GPU infrastructure — H100 and GB200 clusters — to enterprises and labs. If Meta only offers managed Llama inference, CoreWeave loses a strategic Meta customer (as Meta self-serves model workloads) but not necessarily the customers who want custom model training on bare metal.
The raw compute option — if Meta pursues it — changes that calculus entirely.
The Structural Problem
Cloud is not just racks and chips. Customers expect billing, uptime SLAs, multi-region availability, security certifications, enterprise support, migration tooling, and stable developer APIs. AWS and Azure have spent decades building those layers. Meta has world-class infrastructure engineering but no cloud services business.
The bet is that Meta can stand up a viable compute-as-a-service product before its excess capacity pressure becomes an expense problem, not an opportunity. With $150B in annual capex and AI revenue still building, the margin for error is tight.
CoreWeave and Nebius have first-mover advantages in the neo-cloud tier. But they are also heavily exposed: CoreWeave has $99.4B in backlog as of Q1 2026, a significant portion of which is anchor hyperscaler contracts that could shift if those hyperscalers start self-serving their compute needs.
Meta is not the last company that will do this. The AI infrastructure arms race has overbuilt. The question is who monetises the surplus first.