In July 2026, a news story about Meta sent shockwaves through global capital markets. The social networking giant was preparing to launch a new cloud infrastructure business called "Meta Compute," aiming to sell AI computing power and model access to external customers. The business model has two main tracks: first, hosting open-source models like Llama for developers to access—mirroring Amazon AWS’s Bedrock service; second, directly leasing raw GPU clusters, entering the territory of Neocloud, the emerging AI cloud service providers.
The capital markets reacted swiftly and in divergent ways. On the day the news broke, Meta’s stock surged 8.8%, adding roughly $127 billion in market value in a single day. Meanwhile, CoreWeave plunged 13.92%, and Nebius plummeted 17.01%. In just a few hours, one news item redistributed several companies’ market caps.
At the heart of this upheaval lies a fundamental question: When one of the world’s largest buyers of AI compute starts considering selling its capacity, is the competitive logic of the AI infrastructure market undergoing a fundamental shift? This article analyzes Meta’s capital expenditure scale, business model choices, and the competitive impact on Neocloud companies from three perspectives.
Meta’s Compute Landscape: From $65 Billion to $145 Billion
To understand the significance of Meta entering the compute leasing market, we first need to clarify the scale of its AI infrastructure.
Meta’s capital expenditures on AI infrastructure have grown exponentially. In 2023, it reached $28.1 billion; in 2024, $39 billion; and in 2025, it jumped to $72.2 billion. For 2026, the company’s guidance is a staggering $125 billion to $145 billion. For context, Sweden’s entire GDP in 2024 was about $604 billion—Meta’s annual capital spending is nearly a quarter of Sweden’s total economic output.
This rapid growth reflects fast-moving strategic iterations. Zuckerberg’s capital expenditure forecast for early 2025 was just $65 billion, which was raised to $72 billion by May, and then to $115–$145 billion in January 2026. In just one year, the guidance more than doubled.
On the infrastructure deployment front, Meta plans to deploy 7 GW of compute infrastructure this year. In the first half, it added 1 GW, expects to add another 5.5 GW by year-end, and plans to double capacity again next year. According to SemiAnalysis, in just the first half of 2026, Meta signed over 5 GW of IT capacity in cloud and colocation services, not counting its simultaneously expanding self-built data centers.
Jensen Huang commented, "No one deploys AI at Meta’s scale."
From Largest Buyer to Potential Rival: The Two Paths of Meta Compute
Meta’s management has repeatedly conveyed a core logic in earnings calls: If AI market demand falls short of expectations, the computing capacity already built can be flexibly repurposed—including leasing it externally. Meta Compute is the realization of this "other use."
Currently disclosed information shows Meta is considering two main service models.
The first is a managed API and model access service, where Meta operates the data centers and chips, and developers access AI models—including Muse Spark—via API. This business model is similar to AWS’s Bedrock platform.
The second is direct leasing of "bare metal" compute, allowing enterprises to rent GPUs and compute resources—a model similar to CoreWeave and other Neocloud companies.
According to Huatai Securities, Meta’s move is not a sign of "excess compute," but rather an optimization of compute assets—a "swap out the old for the new." Meta will concentrate new-generation chips like GB and Rubin on training and iterating frontier models, while releasing older H-series clusters for inference or external leasing to optimize asset utilization. By 2026, Meta’s data centers will have 5 GW of available capacity.
This monetization model has already seen preliminary validation with xAI. xAI’s leasing agreements with Anthropic and Google (Colossus 1 and 2, with monthly rents of $1.25 billion and $920 million, respectively) have demonstrated the cash flow potential of compute leasing.
Neocloud vs. Hyperscalers: Two Distinct Competitive Logics
Meta’s entry triggered intense market reactions because of its unique position—it’s both a new player among hyperscale cloud providers and the largest customer in the Neocloud space.
The current AI compute market has two main camps. One is the hyperscale cloud providers, including Microsoft, Amazon, Meta, and Alphabet—large tech firms with profitable business units to subsidize capital expenditures. The other is Neocloud companies like CoreWeave and Nebius, which build specialized GPU data centers and sell compute capacity.
The two differ fundamentally in business models. Hyperscalers have mature cloud ecosystems, generating revenue by leasing compute, storage, and software services. Neocloud companies focus on cost-effective GPU services, with their advantage lying in specialization—from hardware and networking to integrated AI cloud scheduling stacks.
But Neoclouds’ vulnerability is also clear: they are highly dependent on a few large customers. The CoreWeave–Meta relationship is especially massive. The two signed a $14.2 billion agreement in September 2025, followed by an additional $21 billion in April 2026, bringing total contract value through December 2032 to over $35 billion. A single customer accounts for a substantial portion of CoreWeave’s $99.4 billion in backlogged revenue orders.
When the largest customer starts building and selling its own compute, the risks of vertical integration become apparent.
Three Scenarios for Market Impact
Based on the above, Meta’s entry into the AI compute leasing market could have three main impacts.
First, compute prices may come under pressure. Morgan Stanley estimates that if Meta leases 250 MW of compute for a year at $40 per watt, it could add about $2.97 per share in earnings in 2028—an 8% profit boost. If leasing scales up to 1,000 MW, the increase could reach 33%. Theoretically, a surge in supply would push rental prices down. However, compute rental prices are still rising—industry insiders report that prices continued to climb in early July 2026—so the supply-demand relationship has not fundamentally reversed.
Second, Neocloud profit margins may be squeezed. If Meta simply subleases GPUs, becoming a bare-metal IaaS provider with roughly 30% gross margin, market concerns about Neocloud valuations are justified. When a major customer becomes a supplier, original vendors lose bargaining power. CoreWeave’s stock dropped more than a third in three months, and Nebius fell 37.3% from its 2026 peak, reflecting market pricing of this risk.
Third, industry competition may accelerate and fragment. Some analysts believe the market is overreacting. SemiAnalysis reports that Meta is not reducing external procurement, but using third-party Neoclouds to secure capacity faster. Since early 2024, Meta has signed nearly 10 GW in contracts, with most new capacity still coming from third parties. For suppliers like CoreWeave and Nebius, Meta’s orders could actually boost their remaining obligations.
Bank of America analysts also note a shortage of dedicated AI data center supply. By 2030, data center power demand may exceed new capacity by 100 GW, leaving room for multiple suppliers.
Data Check: Latest Market Performance
As of the close on July 24, 2026, relevant stocks performed as follows:
- Meta: $595.19, down 1.80% for the day
- Nebius (NBIS): $187.77, down 15.02% for the day
- CoreWeave (CRWV): $71.88, down 11.37% for the day
Neocloud stocks fell much more sharply than Meta itself, indicating that the market is still digesting structural concerns over the "customer turned competitor" risk.
From a fundamentals perspective, Nebius reaffirmed its full-year 2026 guidance: annualized run-rate revenue of $7–9 billion, group revenue of $3–3.4 billion, and group adjusted EBITDA margin of about 40%. The company raised its 2026 capital expenditure guidance to $20–25 billion. CoreWeave reaffirmed its full-year 2026 revenue guidance at $12–13 billion.
Both companies maintain high growth expectations, but the uncertainty brought by Meta’s entry is recalibrating market valuation logic.
Conclusion
Meta’s move into the AI compute leasing market is fundamentally a result of value shifting from the "model layer" to the "infrastructure layer" in the AI race. When a company invests over $100 billion annually in AI infrastructure, seeking diversified returns for those assets becomes a rational business decision.
For Neocloud companies, the challenge is real—the largest customer has become a direct competitor, a structural shift that cannot be ignored. But opportunities remain: long-term demand for AI compute is still strong, and there is a massive gap in dedicated AI infrastructure supply. CoreWeave and Nebius have built specialized capabilities in GPU-optimized data centers and integrated scheduling stacks—skills not easily replicated by hyperscalers.
Current market pricing reflects a discount for uncertainty. Ultimately, the outcome will depend on two variables: whether incremental demand for AI compute can absorb the new supply from Meta, and whether Neocloud companies can carve out differentiated survival spaces after the giants enter.
FAQ
Q: What is Meta Compute?
Meta Compute is Meta’s planned cloud infrastructure business, designed to sell AI compute power and model access to external customers. Its business model includes hosting AI models for developer access (similar to AWS Bedrock) and direct leasing of raw GPU clusters (similar to the Neocloud model).
Q: What does Meta’s entry into compute leasing mean for Nebius and CoreWeave?
Meta is one of CoreWeave’s largest customers, with contracts totaling over $35 billion. Meta’s self-built compute and external leasing means it moves from "largest buyer" to "direct competitor," potentially weakening Neocloud companies’ bargaining power and profit margins.
Q: Is there really an oversupply of AI compute?
There is no widespread oversupply at present. Industry data shows effective utilization rates of intelligent compute clusters average less than 20%, but there is a roughly 40% shortage of high-end compute needed for large model training. Meta’s compute leasing is more about optimizing existing assets than signaling weak demand.
Q: Can Neocloud companies survive after Meta enters the market?
Analysts believe there is still room for multiple suppliers. Bank of America expects data center power demand to exceed new capacity by 100 GW by 2030. Neocloud companies’ expertise in GPU-optimized data centers and integrated scheduling creates differentiated barriers.
Q: What is the long-term trend for the AI compute leasing market?
The AI compute leasing market is shifting from a "seller’s market" to multi-party competition. Hyperscale cloud providers (Microsoft, Amazon, Google, Meta) and specialized Neocloud companies (CoreWeave, Nebius) will form complex relationships of both cooperation and competition. Compute prices may rationalize, but long-term demand growth will continue to support market expansion.
