Meta 5GW AI Computing Power Expansion Panorama: Why the "Cloud Business Sell-off" Could Be a Market Misjudgment?

In the first week of July 2026, the AI infrastructure market experienced a dramatic emotional shock.

On July 1st Eastern Time, Bloomberg reported that Meta is formulating plans to launch a cloud infrastructure business, selling AI computing power and model access to external customers. As soon as the news broke, CoreWeave (CRWV) and Nebius (NBIS) saw significant drops in their stock prices. CoreWeave fell about 13.9% that day, closing at $85.69; Nebius fell about 17%, closing at $229.18. The sell-off intensified the next day—on July 2nd, CoreWeave dropped another 4.6%, closing at $81.75; Nebius fell about 5.92%, closing at $215.62. The combined market value of the two companies evaporated by tens of billions of dollars.

The market's logical chain appears clear: Meta goes from being Neocloud's biggest customer to a potential competitor—AI computing supply is about to become oversupplied—Neocloud valuations need to be reassessed.

But every link in this logical chain deserves a fresh look. On July 2nd, semiconductor research firm SemiAnalysis released a report offering an almost completely opposite judgment: "Meta's data center and computing power procurement will accelerate, not slow down. Capital expenditures in 2027 will be shockingly high."

Starting from the triggering event, let's deconstruct the structural logic behind this controversy layer by layer, and answer one core question: Is the market selling off facts or misinterpretations?

The Starting Point of the Market Controversy—The Chain Reaction from the Bloomberg Report

On July 1st, Bloomberg cited sources reporting that Meta is making plans to launch a cloud infrastructure business, project codenamed "Meta Compute." The report indicated two possible directions: one is to offer model hosting access similar to AWS Bedrock, and the other is to directly lease raw computing power like CoreWeave.

This news quickly spread across the AI infrastructure sector. The core logic behind market concerns is: if Meta shifts from buyer to seller, Neocloud companies that rely on Meta's orders will face a double blow—reduced demand while competing against a rival with a hundred-billion-dollar capital expenditure capacity.

Analysts at BNP Paribas noted in a subsequent report that although demand remains strong and both companies' GPU capacity is sold out, the market's wariness of "new entrants" still dominated short-term pricing. Rosenblatt Securities held the opposite view; analysts John McPeake and Tanu Chauhan said in a post-market report on July 2nd that the sell-off "provides a buying opportunity," emphasizing that Meta's contracts with CoreWeave likely lack authorization to sublease computing power to third parties.

The coexistence of these two views itself indicates a problem: the market is far from reaching a consensus on the substantive impact of Meta "selling computing power."

The Real Situation—Meta Is Still Accelerating AI Computing Power Procurement

SemiAnalysis's report provides the most direct counter-evidence: in the first half of 2026, Meta has already signed over 5GW of data center capacity in cloud services and colocation—and this does not include the entire progress of its self-built projects.

What does 5GW mean? For comparison, Meta's two largest data center campuses currently under construction have a combined capacity of 2.5GW. The SemiAnalysis report directly refutes the earlier market narrative that "only 5GW of data centers are under construction in the US, with half of the projects delayed"—"Just Meta's two campuses alone equal half of that number."

If a company were truly scaling back AI investments and believing computing power would soon be oversupplied, it wouldn't sign over 5GW of external capacity within half a year while simultaneously advancing two self-built campuses at the 2.5GW level—these actions show an obvious logical contradiction.

SemiAnalysis's core conclusion is: Meta is not reducing external procurement; it is using third-party Neocloud providers to get capacity faster. Since early 2024, Meta has signed nearly 10GW in contracts, with most new capacity still coming through third parties. For suppliers like CoreWeave and Nebius, Meta's orders may actually continue to boost their Remaining Performance Obligations (RPO).

The Real Use of 5GW Computing Power—Not a Single Business, but an "Optional Computing Pool"

The root of the market's misinterpretation lies in understanding "Meta may sell computing power" as "Meta's computing power has only one destination." The SemiAnalysis report offers a completely different framework: Meta's new capacity is more like an "optional computing pool" that can be flexibly allocated across multiple high-value directions.

The first direction is frontier model training. Meta Superintelligence Labs (MSL) remains the biggest destination for incremental computing power. SemiAnalysis explicitly states that Meta has not given up training frontier models; the team is currently "excited" about its progress. This is the most direct narrative for capex: Meta needs sufficiently large training clusters, talent, and room for trial and error to catch up with OpenAI and Anthropic.

The second direction is the advertising recommendation system. SemiAnalysis believes Meta is confident it can increase the complexity of its ad recommendation system by more than 10 times. Meta's official financial reports show that in Q1 2026, ad impressions grew 19% year-over-year, and average unit price grew 12%. Meta Engineering previously stated that GEM-related training stacks achieved a 23x increase in effective training FLOPs, about 1.43x improvement in MFU, and a 16x expansion in GPU scale; after doubling the GPUs for GEM training, Instagram and Facebook Feed ad conversion rates improved by 5% and 3% respectively. This path is easier for investors to understand: if more computing power can improve ad conversion rates, it's not simply "burning money on GPUs" but part of ad revenue and pricing power.

The third direction is a model service platform. SemiAnalysis exclusively reveals that Meta is in the final stages of negotiations to sign an agreement with Anthropic to obtain private deployment rights for Claude, similar to how Amazon gets Claude through Bedrock, with the difference being that it would run in Meta's own data centers. This means Meta could not only sell its own models in the future but also bundle Claude into its computing power and platform for external services.

The fourth direction is large-scale, short-term, high-premium on-demand computing transactions similar to SpaceX. This is also the most impactful judgment in the report. SemiAnalysis estimates that SpaceX's transaction with Anthropic yields annualized revenue of about $3.1 billion per GW, which is 2.6 times the typical five-year IaaS average price from Neoclouds; the transaction with Google is even higher, at about $4.8 billion per GW per year, equivalent to 4 times. If Meta allocates only 200MW for similar external transactions, based on the calculations from the report's public pages, annualized revenue could exceed $10 billion. This scale is enough to change the market's intuition about "Meta selling computing power externally"—it might not necessarily be low-margin subleasing but could be selling time windows to capital-starved top-tier clients using quickly deployable data center capacity.

What the Market Truly Overlooks—Computing Power Is Not "Oversupplied" but "Structurally Scarce"

The deeper issue in this controversy is that the market's criteria for judging "computing oversupply" might themselves be flawed.

"Capacity oversupply" cannot be judged solely by total GW. What is truly scarce in AI data centers is often not paper electricity but available GPUs, networking, data center delivery speed, customer migration costs, and contract flexibility. Electricity does not equal available GPUs; data center space does not equal deliverable computing power—delivery speed itself is becoming a core competitive advantage.

Morgan Stanley's model shows Meta will add about 2GW and 3.5GW of self-operated IT capacity in 2026 and 2027 respectively. For comparison, Amazon and Google's new IT capacity in 2027 could be as high as 5GW and over 9GW respectively. On the scale of the entire industry's expansion, Meta's 5GW does not constitute a reason for "oversupply."

The judgment in the SemiAnalysis report is: the market misjudges because it only sees the action of "selling computing power" without understanding why Meta has the confidence to keep expanding. If Meta were simply subleasing GPUs to become a bare-metal IaaS provider with about 30% gross margin, then the market's concerns about Neocloud valuations would be justified. But the destinations of Meta's new capacity are far more complex than "subleasing GPUs."

Is the Neocloud Sell-Off Reasonable?—Risk Comes from Concentration, Not Disappearing Demand

For CoreWeave and Nebius, the market's concerns are not entirely without basis.

CoreWeave's contract with Meta has reached $35.2 billion—a $14.2 billion contract signed in September 2024 extending to 2031, and an additional $21 billion contract signed in April 2026 extending to 2032. Nebius's contract with Meta could be up to $27 billion. In its Q1 2026 shareholder letter, Nebius stated that its second major contract with Meta involves over 3.5GW of capacity.

Such a high concentration of clients is itself a risk. When one client contributes such massive long-term contracts, any signal of change in that client's strategic direction will trigger a repricing of revenue expectations.

But the question is: has Meta's strategic direction really changed?

Looking at existing contracts, Meta is still accelerating its use of third-party Neocloud services. As long as Meta believes computing power can be absorbed by MSL, the ad system, model services, or short-term premium transactions, it has reason to let Neocloud build clusters first, rather than waiting for its self-built projects to be delivered slowly. Meta is willing to pay a premium for speed—this is precisely why third-party suppliers still have value.

The real risk for Neocloud is not that Meta stops procuring, but that the structure of Meta's procurement changes—from "long-term lock-in" to "flexible deployment." If Meta increasingly adopts SpaceX-style short-term high-premium transaction models, the long-term contract value for Neocloud could be reassessed. This is a question about contract structure, not about disappearing demand.

Conclusion

The Neocloud sell-off triggered by Meta's "cloud business" news—is it a market misjudgment? Based on the framework provided by SemiAnalysis, the market made at least three layers of misinterpretation.

First, mistaking "computing power deployment" for "supply competition." The fact that Meta added 5GW of capacity in half a year shows that the company is still expanding its AI infrastructure at an unprecedented pace. A buyer that is scaling back would not sign over 5GW in external contracts within six months.

Second, mistaking "one possibility" for "the only narrative." Meta may indeed sell computing power externally, but that is just one of four destinations for 5GW of capacity. MSL training, ad recommendation optimization, model service platforms, and short-term high-premium computing transactions—each path is fundamentally different from "low-price GPU subleasing."

Third, mistaking "structural change" for "disappearing demand." The risks that Neocloud faces are real—client concentration, contract flexibility, financing costs—but these are different from "Meta stops buying." Meta is still buying; only the methods and structure of procurement may be changing.

Of course, this analysis needs to remain cautious. Whether MSL can catch up with OpenAI and Anthropic remains highly uncertain. Frontier model competition isn't solved by GPU count alone; data strategy, research teams, training stability, product distribution, and inference costs all affect results. If Meta ends up signing a large volume of long-term computing contracts without flexible exit arrangements, and frontier model catch-up goes poorly, over 5GW of new external computing power could more directly translate into capital expenditure pressure.

But for now, equating "Meta may sell computing power" with "AI computing is about to be oversupplied and Neocloud loses value"—at least at the data and logic levels—lacks sufficient support.

Competition in AI infrastructure is shifting from "who builds more" to "who deploys more efficiently." Meta's 5GW computing expansion is less a signal of the "supply glut" the market fears, and more a tech giant redefining its role in this race—from an AI application company to an AI infrastructure orchestrator.

FAQ

Q1: What exactly is Meta's 5GW of computing power?

5GW (gigawatts) is a unit of measurement for data center IT capacity, referring to the total power consumption capacity of servers, GPUs, and other computing equipment. The SemiAnalysis report indicates Meta has signed over 5GW in cloud leasing and colocation capacity in the first half of 2026, not including self-built projects. For reference, Meta's two largest campuses under construction have a combined capacity of about 2.5GW.

Q2: Why did CoreWeave and Nebius crash?

After Bloomberg's July 1st report that Meta planned to launch a cloud infrastructure business, the market feared Meta would transform from a major customer into a direct competitor. As of the close of Beijing time on July 3rd, CoreWeave stood at $81.75, down 4.60%; Nebius at $215.62, down 5.92%. The two companies have long-term contracts with Meta totaling over $60 billion. The narrative of "customer becomes competitor" directly triggered a valuation reassessment.

Q3: Why does SemiAnalysis believe the market misjudged?

SemiAnalysis believes the market mistook "computing deployment" for "supply competition." The fact that Meta signed 5GW in external capacity in half a year indicates it is still expanding aggressively. The report states new computing power has four high-value destinations—MSL training, ad recommendation systems, model service platforms, and short-term premium transactions—rather than only low-price subleasing.

Q4: What is the real risk for the Neocloud industry?

The real risk for Neocloud is not disappearing demand but changes in customer structure. Hyperscalers like Meta and Microsoft contribute most of CoreWeave and Nebius's long-term contracts. If these clients increasingly adopt short-term, flexible transaction models instead of long-term lock-in contracts, Neocloud's revenue visibility will be weakened. Concentration risk is real, but the narrative of disappearing demand lacks evidence.

Q5: Will AI computing really be oversupplied?

From current data, the narrative of "computing oversupply" lacks support. SemiAnalysis points out that market discussions of "oversupply" often overlook a key fact: electricity does not equal available GPUs, and data center space does not equal deliverable computing power. Delivery speed itself is becoming a core competitive advantage. Morgan Stanley's model shows Meta will add about 2GW and 3.5GW of self-operated capacity in 2026 and 2027, while Amazon and Google are expanding even more in the same period.