메타, AI 클라우드 시장 진입: Neocloud, CoreWeave·Nebius 추락 속 라이벌로 부상

Meta Platforms is developing plans to sell its surplus AI computing capacity to outside customers, Bloomberg News reported Wednesday, a move that would position the social media company as a direct competitor to Amazon Web Services, Microsoft Azure, and Google Cloud while simultaneously threatening two of its largest infrastructure suppliers. Meta shares surged more than 10% on the news, adding roughly $149 billion in market value to a stock that had fallen nearly 15% against the S&P 500 this year. For developers, cloud buyers, and anyone tracking the trajectory of AI infrastructure, the report delivers a concrete answer to a question Wall Street has been asking all year: what will Meta do with its projected $125 billion to $145 billion in capital expenditure?

The initiative operates under the internal banner Meta Compute, which oversees the construction and management of the company’s AI infrastructure. Three executives lead the effort: Santosh Janardhan, head of infrastructure; Daniel Gross, who oversees Meta Superintelligence Labs; and Dina Powell McCormick, the company’s president. This leadership trio—an infrastructure chief, an AI research head, and a senior business executive—signals a structured push into the enterprise market rather than a speculative side project.

Two distinct offerings are under consideration, according to individuals familiar with the matter who requested anonymity due to the plans’ ongoing development.

The first is a hosted AI model service. Developers would pay to run queries against AI models—including Muse Spark, Meta’s proprietary closed-weight foundation model launched on April 8—hosted on Meta’s infrastructure. Access would be restricted to Meta’s platform via an API mechanism similar to Amazon Web Services’ Bedrock, which grants developers access to foundation models without requiring them to provision or manage hardware. The second offering is raw compute capacity: leasing bare GPU cycles to third parties following the neocloud model pioneered by firms like CoreWeave.

Meta declined to comment. The plans remain in development and could change.

The two product offerings map to fundamentally different cloud business models, and understanding the distinction is critical for any developer evaluating Meta as a future vendor.

Hosted AI model services operate through an abstraction layer. When a developer submits a prompt, the request hits an API endpoint; the cloud provider manages GPU allocation, load balancing, model weight handling, inference optimization, and failover. Under the hood, AWS Bedrock leverages PyTorch compilation, CUDA graph optimizations, and kernel fusion to reduce inference latency while keeping the hardware invisible to the customer. Billing is structured per token or per invocation, meaning the customer’s engineering team writes application code without ever interacting directly with a GPU.

Raw compute rental—the neocloud model—operates in direct contrast. Customers receive bare-metal access to GPU hardware, typically NVIDIA H100s or Blackwell-generation chips, with no virtualization overhead. Inter-GPU bandwidth within a cluster utilizes NVLink-4, delivering 900 gigabytes per second of intra-node throughput, alongside InfiniBand for ultra-low-latency cross-rack communication. Customers deploy their own software stacks directly onto the hardware. Pricing is set per GPU-hour; CoreWeave has charged approximately $1.39 per hour for an H100, compared to roughly $3.67 per hour on Azure for equivalent performance. Commercial terms typically involve multi-year take-or-pay contracts, requiring customers to commit to paying for a fixed number of GPU-hours at a fixed rate regardless of actual utilization.

If Meta enters this market, its structural advantage lies in scale. The company is allocating $125 billion to $145 billion toward AI infrastructure in 2026 alone, including a 2,250-acre hyperscale campus in Louisiana and a one-gigawatt data center under construction in the American Midwest. At this magnitude, Meta could launch a GPU rental fleet rivaling or exceeding the current operations of CoreWeave or Nebius, all without relying on the GPU-collateralized debt that neocloud providers depend on to finance their hardware.

Meta’s surplus capacity is structural, not accidental. AI training workloads are inherently bursty: running a frontier model consumes GPUs at near 100% utilization for days or weeks, after which the cluster drops to 30%–50% utilization during the inference serving period between training runs. Given Meta’s capex scale—a buildout designed to meet peak demand for training future iterations of Muse Spark and its successors—the company has constructed infrastructure sized for maximum training load. Consequently, a significant block of GPU capacity sits underutilized whenever active training runs are not in progress. This surplus is precisely what Meta Compute aims to monetize.

The precedent was established by SpaceX’s xAI unit, which built its Colossus 1 data center in Memphis to train the Grok AI model but found itself with excess capacity. In May, Anthropic signed a deal to lease all of Colossus 1’s compute at $1.25 billion per month through May 2029; in June, Google agreed to pay $920 million per month for capacity at Colossus 2. The SpaceX parallel is significant because it established a market benchmark for hyperscale GPU capacity leased to major AI developers—and Meta is closely monitoring these terms.

The sharpest market reaction Wednesday centered not on Meta’s gains, but on the damage inflicted on neocloud providers. CoreWeave shares fell roughly 13% to 15%; Nebius dropped by a similar margin. IREN declined approximately 6.5%.

The concern is structural. Neocloud companies built their businesses around a scarcity condition: AI demand outpaced traditional cloud providers’ ability to expand GPU capacity, creating a window for specialized GPU-rental firms. CoreWeave capitalized on this opportunity partly by securing a $21 billion capacity agreement with Meta—a deal announced on April 9 and running through December 2032. Nebius signed a separate agreement worth up to $27 billion with Meta over five years, with capacity commitments beginning in 2027. These contracts now present a double-edged dynamic. Meta has historically been one of the neoclouds’ most important customers. If Meta builds sufficient internal capacity to meet its own needs and then sells the remainder externally, it simultaneously removes a key source of demand and introduces a new source of supply into the same market.

D.A. Davidson managing director Gil Luria summarized the risk directly: “The impact of adding Meta’s capacity to the market is more likely to be on neoclouds than the big hyperscalers. Those companies like CoreWeave and Nebius rely on Meta for their growth and Meta may not need them anymore.”

The financing implications are substantial. The neocloud business model depends on multi-year take-or-pay contracts to secure collateral for GPU-backed debt. CoreWeave’s $8.5 billion debt facility—the first investment-grade rated GPU-backed financing, priced in March 2026—was borrowed against exactly this type of contracted cash flow. If Meta’s agreements with CoreWeave and Nebius face renegotiation or non-renewal in coming years, the collateral underpinning that debt structure will weaken significantly.

Meta’s push into cloud infrastructure also reflects a competitive lesson from earlier this year. According to Financial Times reporting, Google restricted Meta’s access to its Gemini AI models around March 2026 after failing to supply the full computing capacity Meta requested. The shortfall forced Meta to instruct employees to ration AI token consumption and accelerated the company’s shift toward reliance on Muse Spark—its own closed-weight model—to handle content moderation, scam detection, and internal coding workflows.

That episode underscores why a company at Meta’s scale cannot simply purchase its way to sufficient AI compute: even Google, one of the world’s largest data center operators, ran short. Owning the infrastructure—and controlling access to it—is the only reliable hedge. A cloud business effectively turns that defensive hedge into a revenue stream.

Can Meta actually win in cloud? The incumbent providers—AWS, Azure, and Google Cloud—have spent decades building the software platforms, enterprise sales capabilities, developer ecosystems, and customer support infrastructure that sit atop the hardware. Meta possesses data centers and GPUs, but it currently lacks a published pricing model, an announced launch date, a disclosed customer pipeline, or an enterprise sales organization tailored to external cloud customers.

Analysts note that the competitive threat to established cloud providers remains limited in the near term. AWS, Azure, and Google Cloud are deeply embedded in enterprise software stacks through identity management, database