In the first week of July 2026, the AI infrastructure market experienced a dramatic wave of volatility.
On July 1 (Eastern Time), Bloomberg reported that Meta is planning to launch a cloud infrastructure business, aiming to sell AI compute power and model access to external customers. The news triggered sharp declines in CoreWeave (CRWV) and Nebius (NBIS) stock prices. CoreWeave dropped about 13.9% that day, closing at $85.69, while Nebius fell roughly 17% to $229.18. The sell-off intensified the following day—on July 2, CoreWeave fell another 4.6% to close at $81.75, and Nebius dropped 5.92% to $215.62. Together, the two companies lost billions of dollars in market value.
The market’s logic seemed straightforward: Meta, previously Neocloud’s largest customer, is now a potential competitor—AI compute supply is about to become excessive—Neocloud’s valuation needs to be reassessed.
But every link in this logic chain deserves a second look. On July 2, semiconductor research firm SemiAnalysis released a report with a nearly opposite assessment: "Meta’s data center and compute procurement will accelerate, not slow down. Capital expenditures in 2027 will be astonishingly high."
Let’s break down the structural logic behind this controversy, starting from the initial event, to answer a fundamental question: Is the market sell-off driven by facts, or by a misunderstanding?
The Spark of Market Controversy—Bloomberg’s Report and Its Ripple Effect
On July 1, Bloomberg, citing sources, reported that Meta is planning to launch a cloud infrastructure business under the project codename "Meta Compute." The report revealed two main directions: first, offering model hosting and access similar to AWS Bedrock; second, directly renting out raw compute power like CoreWeave.
This news quickly spread through the AI infrastructure sector. The market’s core concern: if Meta shifts from buyer to seller, Neocloud companies reliant on Meta’s orders would face a double blow—reduced demand and a new competitor with $100 billion in capital expenditure capacity.
BNP Paribas analysts noted in a subsequent report that, although demand remains strong and both companies’ GPU capacity is fully sold, short-term pricing is still dominated by concerns over "new entrants." Rosenblatt Securities took the opposite stance; analysts John McPeake and Tanu Chauhan wrote in a post-market report on July 2 that the sell-off "created a buying opportunity," emphasizing that Meta’s contract with CoreWeave likely lacks authorization to sublease compute power to third parties.
The coexistence of these two perspectives highlights a key issue: the market has yet to reach consensus on the real impact of Meta "selling compute power."
The Reality—Meta Is Still Accelerating AI Compute Procurement
SemiAnalysis countered with a straightforward data point: in the first half of 2026, Meta signed contracts for over 5GW of data center capacity in cloud services and hosting—excluding all progress on self-built projects.
What does 5GW mean? For comparison, Meta’s two largest data center campuses under construction have a combined capacity of 2.5GW. SemiAnalysis directly refuted the prevailing narrative that "only 5GW of data center projects are under construction in the US, half of which are delayed"—"Meta’s two campuses alone account for half that figure."
If a company were truly scaling back AI investment and anticipating a compute surplus, it wouldn’t sign over 5GW of external capacity in half a year while simultaneously building two 2.5GW campuses. There’s a clear logical contradiction in these actions.
SemiAnalysis’s core conclusion: Meta isn’t reducing third-party procurement—it’s using Neocloud providers to secure capacity faster. Since early 2024, Meta has signed nearly 10GW in contracts, with most new capacity still sourced from third parties. For suppliers like CoreWeave and Nebius, Meta’s orders could actually increase remaining performance obligations (RPO).
The True Purpose of 5GW Compute—Not a Single Business, but a "Flexible Compute Pool"
The root of the market’s misunderstanding lies in equating "Meta might sell compute" with "Meta’s compute has only one use." SemiAnalysis offers a different framework: Meta’s new capacity resembles a "flexible compute pool" that can be dynamically allocated across multiple high-value directions.
The first direction is frontier model training. Meta Superintelligence Labs (MSL) remains the largest destination for incremental compute. SemiAnalysis makes it clear that Meta hasn’t abandoned frontier model training—the team is "excited" about its progress. This is the most direct capital expenditure narrative: to catch up with OpenAI and Anthropic, Meta needs massive training clusters, talent, and room for experimentation.
The second direction is ad recommendation systems. SemiAnalysis believes Meta aims to increase the complexity of its ad recommendation systems by more than 10x. Meta’s official financials show ad impressions grew 19% year-over-year in Q1 2026, with average price per ad up 12%. Meta Engineering previously reported that the GEM training stack improved effective training FLOPs by 23x, MFU by about 1.43x, and GPU scale by 16x. Doubling GEM training GPUs led to a 5% lift in Instagram and 3% in Facebook Feed ad conversion rates. This path is easier for investors to understand: if more compute boosts ad conversion, it’s not just "burning money on GPUs"—it’s part of ad revenue and pricing power.
The third direction is a model services platform. SemiAnalysis exclusively revealed that Meta is in final negotiations with Anthropic for private deployment rights to Claude, similar to how Amazon accesses Claude through Bedrock, but running inside Meta’s own data centers. This means Meta could offer not only its own models but also bundle Claude into its compute and platforms for external clients.
The fourth direction is large-scale, short-term, high-premium on-demand compute trading, similar to SpaceX. This is the most striking assessment in the report. SemiAnalysis estimates that SpaceX’s annualized revenue per GW from Anthropic is about $3.1 billion—2.6 times the typical Neocloud five-year IaaS average; the Google deal is even higher, at $4.8 billion/GW/year, or 4x. If Meta allocates just 200MW for such external deals, the report projects annualized revenue could exceed $10 billion. That scale could fundamentally change the market’s intuition about "Meta selling compute"—it may not be low-margin subleasing, but rather monetizing rapid data center buildouts by selling time windows to top-tier clients desperate for compute.
The Real Market Blind Spot—Compute Isn’t "Oversupplied," but "Structurally Scarce"
The deeper issue in this debate is that the market’s standard for judging "compute oversupply" may itself be flawed.
"Oversupply" can’t be measured by GW capacity alone. The real bottlenecks in AI data centers are rarely on-paper power, but usable GPUs, network, delivery speed, customer migration cost, and contract flexibility. Power doesn’t equal available GPUs, and a server room doesn’t equal deliverable compute—delivery speed itself is becoming a core competitive edge.
Morgan Stanley’s models show Meta will add about 2GW and 3.5GW of owned IT capacity in 2026 and 2027, respectively. For comparison, Amazon and Google are expected to add 5GW and over 9GW of IT capacity in 2027. In the context of industry-wide expansion, Meta’s 5GW is hardly a case for "oversupply."
SemiAnalysis concludes that the market’s misjudgment stems from focusing solely on "selling compute" without understanding why Meta is confident enough to keep expanding. If Meta simply subleased GPUs, becoming a bare-metal IaaS provider with ~30% gross margin, concerns about Neocloud’s valuation would be justified. But Meta’s new capacity is headed for far more complex uses than just "GPU subleasing."
Is the Neocloud Sell-Off Justified?—The Real Risk Is Concentration, Not Vanishing Demand
For CoreWeave and Nebius, market concerns aren’t entirely unfounded.
CoreWeave’s contract value with Meta has reached $35.2 billion—$14.2 billion signed in September 2024, running through 2031, and another $21 billion in April 2026, extending to 2032. Nebius’s contract with Meta could reach $27 billion. In its Q1 2026 shareholder letter, Nebius stated that its second major Meta contract exceeds 3.5GW in capacity.
Such customer concentration is inherently risky. When a single client accounts for such massive long-term contracts, any signal of a strategic shift can trigger a major repricing of revenue expectations.
But has Meta’s strategy really changed?
Current contracts show Meta is still ramping up third-party Neocloud usage. As long as Meta believes compute can be absorbed by MSL, ad systems, model services, or short-term high-premium deals, it makes sense to let Neocloud build clusters first, rather than wait for in-house projects to come online. Meta is willing to pay a premium for speed—this is precisely why third-party suppliers still matter.
The real risk for Neocloud isn’t that Meta will stop buying, but that Meta’s procurement structure could change—from "long-term lock-in" to "flexible scheduling." If Meta increasingly adopts SpaceX-style short-term, high-premium deals, the value of Neocloud’s long-term contracts may be reassessed. This is a contract structure issue, not a demand disappearance issue.
Conclusion
Was the Neocloud sell-off triggered by Meta’s "cloud business plans" a market misjudgment? According to the SemiAnalysis framework, the market made at least three key errors.
First, it misread "compute scheduling" as "supply competition." The fact that Meta added 5GW of capacity shows it’s expanding AI infrastructure at unprecedented speed. A shrinking buyer doesn’t sign over 5GW in external contracts in six months.
Second, it mistook "one possibility" for "the only narrative." Meta might sell compute externally, but that’s just one of four possible uses for the 5GW. MSL training, ad optimization, model service platforms, and high-premium compute trading—all are fundamentally different from "low-margin GPU subleasing."
Third, it confused "structural change" with "vanishing demand." Neocloud’s risks are real—client concentration, contract flexibility, financing costs—but these are not the same as "Meta stopped buying." Meta is still procuring, but the methods and structures may be evolving.
Of course, this analysis requires caution. 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 play a role. If Meta ends up signing a large number of long-term compute deals with little exit flexibility, and falls behind in model development, over 5GW of new external compute could rapidly become a capital expenditure burden.
But for now, equating "Meta might sell compute" with "AI compute is about to be oversupplied and Neocloud is losing value"—lacks sufficient support both in data and logic.
The AI infrastructure race is shifting from "who can build more" to "who can schedule more efficiently." Meta’s 5GW compute expansion is less a signal of "oversupply," and more a sign that a tech giant is redefining its role in this race—from AI application company to orchestrator of AI infrastructure.
FAQ
Q1: What exactly does Meta’s 5GW of compute refer to?
5GW (gigawatts) measures the IT capacity of data centers, representing the total power consumption of servers, GPUs, and other computing equipment. According to SemiAnalysis, Meta signed for over 5GW in cloud leasing and hosting capacity in the first half of 2026, not including in-house projects. For reference, Meta’s two largest campuses under construction have a combined capacity of about 2.5GW.
Q2: Why did CoreWeave and Nebius plunge?
After Bloomberg’s July 1 report that Meta plans to launch a cloud infrastructure business, the market worried Meta would shift from major customer to direct competitor. As of the July 3 close (Beijing time), CoreWeave was at $81.75, down 4.60%; Nebius at $215.62, down 5.92%. The two companies have over $60 billion in long-term contracts with Meta, so the narrative of "customer turning competitor" directly triggered a valuation reset.
Q3: Why does SemiAnalysis believe the market misjudged?
SemiAnalysis argues the market mistook "compute scheduling" for "supply competition." The fact that Meta signed 5GW of external capacity in six months shows it’s still accelerating expansion. The report notes that new compute has four high-value destinations—MSL training, ad recommendation systems, model service platforms, and high-premium short-term trading—not just low-margin subleasing.
Q4: What is the real risk for the Neocloud industry?
Neocloud’s real risk isn’t disappearing demand, but changing client structure. Hyperscale customers like Meta and Microsoft account for most of CoreWeave and Nebius’s long-term contracts. If these clients increasingly prefer short-term, flexible deals over long-term lock-ins, Neocloud’s revenue visibility will weaken. Concentration risk is real, but the narrative of vanishing demand lacks evidence.
Q5: Will there really be an AI compute glut?
Current data doesn’t support the "compute oversupply" narrative. SemiAnalysis points out that market discussions often ignore a key fact: power doesn’t equal available GPUs, and server rooms don’t equal deliverable compute. Delivery speed is becoming a core competitive factor. Morgan Stanley models show Meta will add about 2GW and 3.5GW of owned IT capacity in 2026 and 2027, while Amazon and Google are expanding even faster over the same period.
