SemiAnalysis refutes the "overcapacity argument": Meta's computing power expansion far exceeds imagination, next year's capital expenditures will be "astonishingly high."

A piece of news about Meta "selling computing power" severely hit the AI hardware sector, with CoreWeave dropping 13% in a single day and Nebius falling by 15%. The narrative of "excess computing power" quickly spread. However, the well-known semiconductor research institute SemiAnalysis believes the market's interpretation that Meta "leasing out computing power = cutting expenses" is wrong. On July 3, the institution released an analysis report stating: "We believe both interpretations are incorrect. Meta's data center and computing power procurement will accelerate, not slow down. Capital expenditure in 2027 will be astonishingly high." To support this judgment, SemiAnalysis provided specific numbers: In just the first half of 2026, Meta has already signed contracts for over 5GW of data center capacity, covering cloud leasing and managed rooms, not including the full progress of self-built projects. Satellite/aerial photos of Meta's two largest ongoing data center campuses—the combined capacity under construction in these two campuses alone reaches 2.5GW. The report also refuted another widely spread market narrative—"Half of US data center projects are delayed, and only 5GW nationwide is under construction." SemiAnalysis pointed out: Meta's two campuses alone already equal half of that figure. "These headlines are completely wrong." Four Paths Forward: Why Meta Can Continue to Bet on Computing Power SemiAnalysis believes the market misjudged because it only saw the "selling computing power" action, but not why Meta has the confidence to keep expanding. The report outlines four high-value monetization directions, each fundamentally different from the bare-metal IaaS model of ordinary Neoclouds. First: Advanced AI Models (MSL) remain core. SemiAnalysis clearly states that Meta has not abandoned training advanced models. Meta Superintelligence Labs (MSL) remains the largest destination for incremental computing power. The report says the team is "excited" about its progress and will soon release an in-depth report to assess MSL's chances of catching up to Anthropic and OpenAI. Second: Advertising Recommendation Systems (RecSys): 10x expansion potential. SemiAnalysis believes Meta is confident it can increase the complexity of its advertising recommendation systems by more than 10 times to accelerate revenue growth. This requires investments in both inference and training computing power. Larger, more expensive RecSys models are already driving advertisers to pay higher prices while maintaining strong return on ad spend (ROAS), and keeping users engaged longer on Meta apps, expanding monetizable ad inventory. Third: Bedrock-like Model API Services. SemiAnalysis exclusively reveals that Meta is in the final stages of negotiating an agreement with Anthropic, to obtain private deployment rights for Claude, similar to how Amazon accesses Claude through Bedrock, except it will run in Meta's own data centers. This means Meta will not only sell its own models but can package Claude in its own computing power and platform for external services. SemiAnalysis lists three monetization paths: - Internal use: Meta itself needs Claude tokens, while Anthropic can't meet demand; private instances also offer enhanced security and privacy. - Selling Claude services externally: Like the Bedrock model, Meta controls the full tech stack from CPUs to GPUs to networks with high security; yet building enterprise client relationships is a challenge for a newcomer. - Vertical integration to create application layers: As one of the world's largest ad platforms, Meta can develop AI Agent products in sales and marketing areas, deeply integrating frontier models. SemiAnalysis also points out that distributing models to free social media users and Meta's hardware ecosystem (such as smart glasses) is a potential option and is strategically valuable for OpenAI and Anthropic: "They are likely willing to make concessions to get access to these distribution channels." Fourth: "SpaceX-style" bulk computing power leasing. Musk created a new market—Meta wants a share. This is the most impactful insight in the report. The computing power leasing agreement between SpaceX and Google shocked the whole AI infrastructure circle—its pricing is four times industry equivalents and three times that of Anthropic's agreements. SemiAnalysis's AI Cloud TCO team tracks hundreds of GPU cloud deals annually and holds the world's most complete GPU pricing database. Their conclusion: "We've never seen an agreement of this scale and duration. It's nominally a three-year contract, but either party can cancel within 90 days—essentially a self-renewing three-month agreement." Why has no one done this before? Because very few companies can. Ordinary Neoclouds need multi-year contracts to cover financing costs and can’t offer a 90-day cancellation. The three hyperscale cloud giants (Microsoft, Amazon, Google) have the technical capability but prefer longer-term value: Microsoft took OpenAI’s equity and IP, Amazon promotes Bedrock and Trainium, Google works on TPU and Vertex. Result: Only two companies can truly replicate the SpaceX model: Oracle and Meta. SemiAnalysis is blunt in its assessment of Oracle: "This is a significant blow. Oracle could have monetized its several GW of computing power better." The report compares Oracle's and SpaceX’s valuations to illustrate this divergence. Meta's advantages? Lots of computing power, builds fast, contracts can be canceled anytime. At a price of $5 billion per GW per year, allocating just 200MW to external customers would bring in more than $1 billion in annual revenue with extremely high margins. The 90-day cancellation clause means if Meta Superintelligence Labs needs more power, it can reclaim it at any time. The report also highlights Meta's "tent-style" ultra-fast data center construction strategy—SemiAnalysis was first to track this design last year, and these facilities are now being rapidly deployed across the US. Quick deployment, quick monetization, highly compatible with the SpaceX model. SemiAnalysis anticipates: Meta will soon announce a SpaceX-like bulk computing power leasing deal, most likely with Anthropic. "CFO's Dream": High Optionality Makes Meta More Comfortable with Buying This is one of the core logics in the entire SemiAnalysis report. The simultaneous existence of four monetization paths means every GW of Meta's computing power has multiple high-value outlets. This is not "buying too much means losses," but "buying more means more options." The report says: "This is basically a CFO's dream, making all-in computing power very easy. We bet Susan (Meta CFO Susan Li) saw the SpaceX agreement pricing and did a 180-degree turn!" The logic is obvious: If Meta Superintelligence Labs succeeds, all the computing power is used internally—highest ROI. If it hits a temporary setback, it can divert some power to SpaceX or Bedrock models and immediately generate high-margin revenue. If RecSys expansion underperforms, there are other outlets too. This high optionality brings another effect: Meta can continue outsourcing computing power from CoreWeave, Nebius, and other third-party Neoclouds, because even if it "subcontracts" this power out again, profit margins are still high enough to cover costs. SemiAnalysis reassures Neocloud investors—“Meta will not become a bare-metal IaaS supplier with only 30% gross margin. All its monetization options are high-value, giving enough room for profit. While providing computing power to external clients, Meta can keep buying capacity from Neoclouds to accelerate expansion.” In other words, Meta is more likely to be a key driver of growth in companies like CoreWeave’s RPO, not a competitor. RecSys: The Most Overlooked Computing Power Monetization Engine Another dimension is the advertising recommendation system. From late 2022 to early 2023, the market generally thought Meta was reaching maturity. But in the following years, Meta's revenue growth accelerated sharply. SemiAnalysis believes GPU investment was the key trigger. The logic: Larger, more expensive RecSys models → more precise advertising → advertisers willing to pay higher prices → strong ROAS (Return on Ad Spend) → virtuous cycle. At the same time, upgraded content recommendation systems increased user time spent on Meta apps, further expanding ad inventory. So, how far can RecSys’s AI expansion go? SemiAnalysis believes Meta is confident it can increase model complexity over tenfold. SemiAnalysis's core conclusion: Meta is leasing computing power not because it bought too much, but because it has enough power to support multiple high-value strategies simultaneously, and each option is highly profitable. Capital expenditure in 2027 will be much higher than market expectations. For investors, this report basically says: That wave of sell-offs may have been a misjudgment. But SemiAnalysis also leaves an important footnote: Whether MSL can truly catch up with Anthropic and OpenAI remains the largest uncertainty. Risk Tip & Disclaimer The market carries risks; investment requires caution. This article does not constitute individual investment advice and does not consider the special investment goals, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article suit their specific circumstances. Investment based on this is at your own risk.