SemiAnalysis Refutes "Compute Oversupply Theory": Meta's Compute Expansion Far Beyond Imagination, Next Year's Capital Expenditure Will Be "Astonishingly High"
TechFlow Selected TechFlow Selected
SemiAnalysis Refutes "Compute Oversupply Theory": Meta's Compute Expansion Far Beyond Imagination, Next Year's Capital Expenditure Will Be "Astonishingly High"
Meta holds four major monetization trump cards: AI models, ad capacity expansion, API services, and high-premium short-term leasing. Every GW of computing power has a high-value outlet, ensuring Meta's computing power investment is positioned for both offense and defense.
Navigating Web3 tides with focused insightsNews of Meta "selling compute capacity" slammed the AI hardware sector, CoreWeave fell 13% in a single day, Nebius fell 15%, and the narrative of "compute oversupply" spread rapidly.
However, renowned semiconductor research institution SemiAnalysis believes the market's interpretation of Meta "renting compute capacity = cutting expenses" is wrong. On July 3, the institution released an interpretation report stating:
"We believe both interpretations are wrong; Meta's data center and compute capacity procurement will accelerate, not slow down. Capital expenditure in 2027 will be astonishingly high."
To substantiate this judgment, SemiAnalysis provided a set of specific figures: In the first half of 2026 alone, Meta has already contracted for over 5GW of data center capacity, covering cloud leasing and colocation facilities, not including the full progress of self-built projects.
Satellite/aerial images of Meta's two largest data center campuses currently under construction—the combined capacity under construction at these two campuses reaches 2.5GW.
The report also refuted another widely circulated market narrative—"Half of U.S. data center projects are delayed, only 5GW under construction nationwide." SemiAnalysis stated: Just Meta's two campuses already equal half of this figure, "these headlines are completely wrong."
Four Ways Out: Why Meta Can Continue to Bet on Compute Capacity
SemiAnalysis believes the market misjudged because it only saw the action of "selling compute capacity," without seeing why Meta has the confidence to continue expanding.
The report outlined four high-value monetization directions, each fundamentally different from the "bare metal IaaS" model of ordinary Neoclouds.
First, frontier AI models (MSL) remain the core.
SemiAnalysis clearly stated that Meta has not given up on training frontier models. Meta Superintelligence Labs (MSL) remains the largest destination for incremental compute capacity. The report stated that the team is currently "excited" about its progress and will subsequently release a dedicated in-depth report to assess MSL's chances of catching up to Anthropic and OpenAI.
Second, Advertising Recommendation System (RecSys): 10x expansion space.
SemiAnalysis believes Meta is confident it can expand the complexity of the advertising recommendation system by more than 10 times, thereby accelerating revenue growth. This requires investment in both inference and training compute capacity. Larger, more expensive RecSys models are already driving advertisers to pay higher prices while maintaining strong Return on Ad Spend (ROAS), also allowing users to spend more time on Meta-family apps, expanding monetizable ad inventory.
Third, Bedrock-like model API services.
SemiAnalysis exclusively disclosed that Meta is in the final negotiation stage of signing an agreement with Anthropic, will obtain private deployment rights for Claude, similar to how Amazon obtained Claude through Bedrock, with the difference being it runs within Meta's own data centers. This means Meta will not only sell its own models in the future but can also package Claude into its own compute capacity and platform to provide services externally.
SemiAnalysis listed three monetization paths:
- Internal use: Meta itself needs Claude tokens, and Anthropic's supply cannot keep up with demand; private instances can also provide stronger security and privacy protection.
- Selling Claude services externally: Similar to the Bedrock model, Meta controls the complete technology stack from CPU to GPU to network, with high security; but as a new entrant, establishing enterprise client relationships is a challenge.
- Vertical integration, building the application layer: Meta is one of the largest advertising platforms globally, with a path to build AI Agent products in the sales and marketing fields, deeply integrating frontier models.
SemiAnalysis also pointed out that distributing models to free social media users and the Meta hardware ecosystem (such as smart glasses) is also a potential option, and holds high strategic value for OpenAI and Anthropic, "they are likely willing to make concessions in exchange for the opportunity to access this distribution channel."
Third, "SpaceX-style" bulk compute capacity leasing. Musk created a new market, and Meta wants to come in and get a share.
This is one of the most impactful judgments in the report.
The compute capacity leasing agreement signed between SpaceX and Google shocked the entire AI infrastructure circle—its pricing is 4 times that of similar industry products, and the agreement pricing with Anthropic is also 3 times that of peers.
SemiAnalysis's AI Cloud TCO team tracks hundreds of GPU cloud transactions annually and holds the most complete GPU pricing database globally. Their conclusion is: "We have never seen an agreement of such scale and such short term. This contract is nominally three years, but both parties can cancel within 90 days—actually it is a 3-month contract with automatic renewal."
Why has no one done this before? Because very few companies can do this. Ordinary Neoclouds require multi-year contracts to cover financing costs and simply do not have the conditions to provide a 90-day cancellation option. The three major hyperscale cloud vendors (Microsoft, Amazon, Google) have the technical capability to do it, but each has higher-value long-term binding strategies—Microsoft took equity and IP in OpenAI, Amazon focuses on promoting Bedrock and Trainium, Google does TPU and Vertex.
The result is that there are only two companies left that can truly replicate the SpaceX model: Oracle and Meta.
SemiAnalysis's evaluation of Oracle was blunt: "This is a major blow to Oracle. They could have monetized the several GW of compute capacity in their hands better." The report used the comparison of valuation trends between Oracle and SpaceX to illustrate how significant this divergence is.
Where lies Meta's advantage? Abundant compute capacity, fast construction, and contracts can be cancelled at any time.
Priced at $50 billion annual revenue per GW, allocating just 200MW to external customers can bring over $10 billion in annual revenue, with extremely high profit margins. And the 90-day cancellation clause means that if Meta Superintelligence Labs needs more compute capacity, it can be taken back at any time.
The report also pointed to Meta's "tent-style" ultra-fast data center construction strategy—SemiAnalysis tracked this design first last year, and currently such facilities are landing rapidly across the United States. Fast launch, fast monetization, highly compatible with the SpaceX model.
SemiAnalysis expects: Meta will announce a large client compute capacity leasing agreement similar to SpaceX in the near future, with the most likely target being Anthropic.
"CFO's Dream": High Optionality Makes Meta More Confident the More It Buys
This is one of the core logics of the entire SemiAnalysis report.
The simultaneous existence of four monetization paths means that every GW of Meta's compute capacity has multiple high-value outlets. This is not "buying too much leads to loss," but "buying too much means having options."
The report's exact words are: "This is basically a CFO's dream, making going All-in on compute capacity very easy. We dare to bet that after Susan (Meta CFO Susan Li) saw the pricing of the SpaceX compute capacity agreement, she did a direct 180-degree turn!"
The logic is very direct: If Meta Superintelligence Labs succeeds, all compute capacity is used internally, ROI is highest; if Meta Superintelligence Labs encounters setbacks periodically, a portion of compute capacity can be used for the SpaceX model or Bedrock model, immediately generating high gross margin revenue; if RecSys expansion is below expectations, there are also other outlets to absorb it.
This high optionality also brings another effect: Meta can fully continue to procure compute capacity from third-party Neoclouds such as CoreWeave and Nebius, because even if this compute capacity is "subcontracted" out again, the profit margin is still sufficient to cover costs.
SemiAnalysis gave Neocloud investors a reassurance here—"Meta will not become a bare metal IaaS supplier with only 30% gross margin; all its monetization options are high-value. This gives it enough profit margin to continue procuring capacity from Neoclouds to accelerate expansion while providing compute capacity to external clients."
In other words, Meta is more likely to be an important source of RPO growth for companies like CoreWeave, rather than a competitor.
RecSys: The Most Easily Overlooked Compute Capacity Monetization Engine
There is another dimension, which is the advertising recommendation system.
From late 2022 to early 2023, the market generally believed Meta had entered a mature growth period. But in the following years, Meta's revenue growth accelerated significantly. SemiAnalysis's judgment is: GPU investment is the key triggering factor.
The logic chain is: Larger, more expensive RecSys models → More precise ad delivery → Advertisers willing to pay higher prices → Advertiser ROAS (Return on Ad Spend) remains strong → Forms a positive cycle.
At the same time, the upgrade of the content recommendation system drove the dwell time of users on Meta-family apps, further expanding ad inventory.
So, how far can RecSys's AI expansion go? SemiAnalysis believes Meta itself is confident it can increase model complexity by another 10 times or more.
SemiAnalysis's core conclusion is clear: Meta rents out compute capacity not because it bought too much, but because there is so much compute capacity that it can support multiple high-value strategies simultaneously, and every option is profitable enough.
Capital expenditure in 2027 will be much higher than market expectations.
For investors, this report is actually saying: That sell-off might be a misjudgment. But SemiAnalysis also left an important footnote: Whether MSL can truly catch up to Anthropic and OpenAI remains the biggest uncertain variable.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
