Meta secured more than 5 gigawatts of cloud and hosting capacity in the first half of 2026, according to a July 2 report from research firm SemiAnalysis, a finding that challenges recent market fears that the company may be pulling back from third-party AI infrastructure providers or preparing to flood the market with excess computing power.
The new capacity is separate from Meta’s own fast-growing data center footprint. SemiAnalysis said the company has signed nearly 10 gigawatts of cloud and hosting contracts since early 2024, with more than half of its recent capacity additions coming through outside hosting partners rather than self-built facilities.
The report landed at a sensitive moment for the AI infrastructure market. Shares of specialized cloud providers, often called neocloud companies, had sold off after reports suggested Meta could start selling surplus computing capacity to other companies. CoreWeave fell 13.9%, while Nebius Group dropped 17%, as traders worried that one of the sector’s biggest customers might become a major competitor.
SemiAnalysis presented a different view. Rather than reducing its reliance on outside vendors, the firm said Meta is still accelerating external purchases to gain faster access to large-scale GPU capacity. The report described third-party hosting as a key part of Meta’s strategy to shorten deployment timelines while its own large campuses are still under construction.
The tension between those two interpretations has become one of the most important questions in the AI infrastructure trade: is Meta building too much compute and preparing to resell the excess, or is it still struggling to secure enough capacity for its rapidly expanding artificial intelligence ambitions?
For now, the SemiAnalysis report suggests the answer may be both. Meta appears to be buying as much capacity as it can, while also preserving the flexibility to shift that computing power among internal AI research, advertising systems, enterprise services, and short-term external leases.
Meta is still buying from outside providers
The central finding from SemiAnalysis is that Meta’s demand for third-party infrastructure remains strong. The firm said Meta secured more than 5 gigawatts of external cloud and hosting capacity during the first half of 2026 alone. That figure does not include Meta-owned data centers, which are also expanding sharply.
Since early 2024, Meta has signed close to 10 gigawatts of cloud and hosting contracts, according to the report. That level of procurement places the company among the largest buyers of AI infrastructure in the world and reinforces how aggressively major technology firms are racing to acquire power access, GPUs, networking equipment, and data center space.
The distinction between self-built capacity and third-party hosted capacity is important. Building data centers can take years, especially when power availability, permitting, grid connections, cooling, and equipment delivery are all constraints. By using outside providers, Meta can bring capacity online faster, even if the cost per unit is higher.
SemiAnalysis said more than half of Meta’s new capacity contracts rely on third-party hosting rather than self-build projects. That means outside providers remain deeply embedded in Meta’s AI infrastructure plans, at least in the near term.
CoreWeave and Nebius are among the companies with major exposure to this spending cycle. SemiAnalysis said CoreWeave holds contracts valued at about $21 billion, while Nebius has contracts worth up to $27 billion. The report added that Meta’s orders could increase the remaining performance obligations of these providers, giving them more contracted revenue visibility.
That point matters because neocloud companies have drawn intense scrutiny over their rapid expansion, heavy borrowing, and reliance on a small number of very large customers. If Meta continues to expand its commitments, it could support the growth case for those companies. If Meta shifts away or renegotiates terms, it could create major financial pressure.
Why the sell-off happened
The market reaction before the SemiAnalysis report was sharp because the cloud infrastructure story is highly sensitive to small changes in perceived demand. AI data centers require enormous upfront spending, and many specialized providers have financed their growth through debt, equipment leasing, and long-term customer contracts.
Reports that Meta might sell excess compute capacity raised the possibility that a large customer could also become a powerful supplier. That would change the competitive balance for neocloud providers, whose business models depend on renting access to high-end GPUs and specialized AI clusters.
The fear was simple: if Meta has so much capacity that it can rent some of it to external customers, then companies such as CoreWeave and Nebius could face new competition from one of the largest and best-capitalized technology firms in the world.
That concern explains the steep share-price moves. Traders were not only reacting to the possibility of Meta selling compute. They were also reassessing the broader risk of oversupply in AI infrastructure, especially after a period in which GPU rental prices remained high because demand consistently exceeded available supply.
SemiAnalysis pushed back against the idea that Meta’s external purchases are slowing. The firm said the company is continuing to buy from third parties and using those providers to shorten delivery times. Still, the report did not dismiss the possibility that Meta may monetize some capacity when it makes economic sense.
Instead, SemiAnalysis framed Meta’s expanding infrastructure base as an “optional compute pool.” In that model, Meta can redirect capacity depending on which use case offers the highest strategic or financial return.
Four possible uses for Meta’s compute pool
SemiAnalysis said Meta’s capacity can be moved among four main uses: the Meta Superintelligence Labs training cluster, advertising recommendation systems, private model-as-a-service deployments such as Claude instances, and high-margin short-term rentals to external customers.
This flexibility is central to understanding Meta’s strategy. The company is not simply building a fixed pool of servers for one product line. It is creating a large base of compute that can support different workloads as priorities shift.
Meta Superintelligence Labs, or MSL, is the most strategically visible use. The division is central to Meta’s attempt to compete with companies such as OpenAI and Anthropic in frontier AI model development. Training advanced models requires enormous clusters of GPUs working together for long periods, making raw compute access a critical input.
Advertising systems are another major outlet. Meta’s core business still depends on matching ads to users across platforms such as Facebook and Instagram. Better recommendation and conversion models can directly improve financial results, turning AI infrastructure into a revenue engine rather than only a research expense.
Private model-as-a-service deployments represent a third possible use. These arrangements could include dedicated or private AI model instances for enterprise customers, including deployments based on models such as Claude. Such services may appeal to companies that want advanced AI capabilities but need more control, privacy, or performance guarantees than standard public cloud products provide.
The fourth use is short-term external rental. This is the part of the story that unsettled traders in neocloud stocks. If Meta can rent high-end compute at premium prices during periods when it does not need all of the capacity internally, it may be able to recover costs faster and improve returns on its capital spending.
Advertising gives Meta a steady internal outlet
One reason the oversupply argument remains uncertain is that Meta has already shown it can use additional GPUs to improve its advertising business.
SemiAnalysis cited Meta’s disclosure that doubling the number of GPUs used by the GEM training stack improved ad-conversion rates by 5% on Instagram feeds and 3% on Facebook feeds. For a company with Meta’s scale, even modest improvements in advertising efficiency can translate into large revenue gains.
In the first quarter of 2026, Meta reported that ad impressions rose 19% from a year earlier, while average prices increased 12%. Those figures suggest that the company’s advertising machine remains a powerful internal consumer of compute.
This matters because AI infrastructure spending is often discussed in terms of experimental research or future products. Meta’s case is different. A significant portion of additional compute may feed directly into an already profitable advertising platform.
Better ranking models, recommendation engines, ad targeting systems, and conversion prediction tools can all increase the value of Meta’s ad inventory. If added GPU capacity improves ad performance, Meta has a strong reason to keep using much of that capacity internally rather than selling it to the market.
That does not eliminate the possibility of short-term rentals. It does, however, suggest that Meta’s compute needs are not limited to frontier model training. The company has several internal systems that can absorb large amounts of additional capacity if the economics are attractive.
Short-term rentals could be highly profitable
SemiAnalysis estimated that short-duration external contracts can generate much higher revenue per watt than conventional cloud leases. The firm modeled transactions similar to those between SpaceX and Anthropic at annualized revenue of about $3.1 billion per gigawatt. That is roughly 2.6 times the typical five-year infrastructure-as-a-service rate.
The report also modeled deals with Google at about $4.8 billion per gigawatt. These figures highlight why Meta may want the option to rent capacity externally even if it remains a major internal consumer of AI compute.
According to SemiAnalysis, if Meta allocated just 200 megawatts to high-priced arrangements, yearly revenue could exceed $10 billion. That scenario would give the company a potentially fast payback path on newly installed infrastructure.
However, the report also noted an important limitation: these contracts are often flexible and not guaranteed. Short-term rental opportunities may be lucrative when demand is tight, but they may not provide the same predictable revenue base as long-term infrastructure contracts.
That makes the strategy attractive but potentially volatile. Meta could earn premium rates when the market is short of compute, but those returns could fall if more capacity comes online across the industry or if AI companies slow their model-training schedules.
For neocloud providers, this is one of the biggest risks. A large amount of flexible capacity entering the rental market could change pricing dynamics, even if Meta is not trying to build a full-scale cloud business in the traditional sense.
Meta’s own data center buildout is also expanding
Meta’s third-party contracts are only part of the broader infrastructure expansion. SemiAnalysis said the company has two major campuses under development that together represent about 2.5 gigawatts of on-site capacity.
When those self-built projects are combined with cloud and hosting contracts, Meta’s total available power could be far above earlier estimates that placed its nationwide data center projects at about 5 gigawatts.
This scale reflects the wider transformation of AI infrastructure into one of the most important capital spending categories in the technology sector. Compute is no longer just a support function for software companies. It is becoming a strategic asset, similar to energy access, chip supply, and network capacity.
Meta’s capital expenditure plan has reportedly expanded to a range of $125 billion to $145 billion for 2026. Some market watchers see potential cloud sales as an attempt to generate returns on that massive spending program. Others argue that Meta still faces real capacity constraints and will need outside partners for years.
The company’s second-quarter earnings report, scheduled for July 29, is likely to receive close attention. Traders will look for updates on spending guidance, infrastructure commitments, AI product progress, and any details about a possible cloud business initiative.
Model development explains the scale of demand
The strongest argument for Meta’s continued compute appetite is the intensity of the AI model race.
Recent internal updates suggest the company is pushing hard to close the gap with leading AI labs. Alexandr Wang, head of Meta’s superintelligence division, told employees in a town hall that an upcoming model code-named “Watermelon” has reached performance comparable to OpenAI’s flagship GPT-5.5 model on key benchmarks.
If accurate, that claim would help explain why Meta continues to secure such large amounts of hardware. Training and refining frontier models requires repeated experiments, large datasets, inference testing, safety work, and post-training runs. Each stage consumes substantial compute.
In that context, Meta’s purchases may not look excessive. They may instead represent the cost of staying competitive in a field where access to GPUs can determine how quickly a company can test ideas and improve models.
At the same time, internal progress appears uneven. Mark Zuckerberg recently acknowledged in an internal meeting that work on AI agents had not advanced as quickly as he had hoped over the past several months.
That admission adds complexity. If Meta’s model training efforts require huge compute but its product applications are slower to mature, the company could temporarily have more usable capacity than its internal teams can absorb. In that case, external rentals would be a practical way to offset costs without abandoning the broader buildout.
The risk for neocloud providers
SemiAnalysis argued that the biggest risks for neocloud suppliers are not necessarily demand collapse, but concentration and contract structure.
Companies such as CoreWeave and Nebius have grown rapidly by serving large AI customers. That growth can be attractive when contracts are expanding, but it can also create exposure if a few customers control a large share of revenue.
Their valuations depend on several variables, including financing costs, GPU depreciation, power availability, utilization rates, and the flexibility of customer agreements. If large clients can adjust reserved capacity without fully paying for unused resources, the providers could face stranded assets.
Debt is another concern. Nebius saw total liabilities rise by 1,040% year over year, according to the information cited in the market discussion. CoreWeave reported quarterly interest expenses of $536 million against more than $50 billion in liabilities.
These figures show why traders reacted so strongly to the idea of Meta becoming a compute seller. Highly leveraged infrastructure providers need stable demand, high utilization, and strong pricing to justify expansion. A shift in strategy by a major customer can quickly change the risk profile.
Still, the SemiAnalysis report suggests that Meta remains a major buyer, not merely a future competitor. That distinction may ease some fears, but it does not remove the structural vulnerability of providers that depend heavily on a small group of hyperscale and AI lab customers.
Pricing will be the key signal
One of the clearest indicators to watch will be GPU rental pricing. So far, prices have remained elevated because demand has consistently outpaced supply. AI labs, enterprises, cloud companies, and application developers all need access to high-performance chips, and supply chains remain tight for the most advanced systems.
If Meta begins leasing meaningful amounts of capacity into the market, pricing could become the first visible sign of whether the supply-demand balance is changing.
A small amount of premium short-term rental activity may not disrupt the market. It could simply reflect normal optimization by a company with flexible infrastructure rights. But a large and sustained push by Meta to sell compute would be different. It could pressure neocloud providers, lower rental rates, and reduce the cost of AI development for other enterprises.
The effect would depend on timing, scale, and contract terms. If Meta leases only temporary excess capacity during gaps between internal workloads, market impact may be limited. If it builds a structured external cloud offering with large blocks of available compute, the consequences could be much larger.
Meta’s strategy remains flexible
The most important takeaway from the SemiAnalysis report is that Meta’s AI infrastructure strategy is not one-dimensional. The company is building its own facilities, signing major third-party hosting contracts, training frontier models, improving advertising systems, exploring enterprise AI services, and potentially renting capacity when returns are attractive.
That flexibility may be intentional. In a fast-changing AI market, it is difficult to know which workload will deliver the best return two years from now. By building a broad compute pool, Meta can shift capacity as technology, products, and demand change.
The risk is that flexibility does not eliminate cost. If Meta’s frontier AI work under MSL fails to keep pace with competitors such as OpenAI and Anthropic, and if internal applications do not absorb the capacity quickly enough, external compute could become a source of capital expenditure pressure rather than strategic advantage.
Contracts with limited flexibility would magnify that risk. If Meta is locked into capacity it cannot fully use or resell profitably, the financial burden could grow. If the contracts are flexible and demand remains strong, the company may be able to turn excess capacity into a revenue opportunity.
For traders, that is why the story remains unsettled. SemiAnalysis has provided evidence that Meta is still expanding purchases from third-party providers, which counters the simplest oversupply narrative. But the possibility that Meta could rent out some of its capacity remains real, and even a partial move could reshape expectations across the neocloud sector.
Whether Meta’s more than 5 gigawatts of added external capacity becomes a competitive weapon, a profit center, or a cost burden will depend on how effectively the company uses it through 2027. Its ability to balance model training, advertising improvements, enterprise services, and short-term rentals will determine whether the AI infrastructure boom translates into durable operating gains or another round of pressure on an already stretched data center market.
Want to see how AI transforms trading like Meta’s cloud strategy? Explore AI copy trading for smarter market decisions.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
