A report about Meta Platforms, Inc. (META) potentially selling compute capacity triggered a sell-off in AI hardware stocks, with CoreWeave dropping 13% and Nebius falling 15% in a single day, fueling a rapid spread of the "compute surplus" narrative.
However, the prominent semiconductor research firm SemiAnalysis argues that the market's interpretation of Meta Platforms, Inc.'s compute leasing as a sign of spending cuts is incorrect. In a report published on July 3rd, the firm stated:
"We believe both interpretations are wrong, and Meta's data center and compute procurement will accelerate, not slow down. The capital expenditure for 2027 will be staggeringly high."
To support this view, SemiAnalysis provided specific figures: in the first half of 2026 alone, Meta Platforms, Inc. has already contracted for over 5GW of data center capacity, covering cloud leases and colocation facilities, and this does not include the full progress of its self-built projects.
The report included satellite/aerial images of Meta Platforms, Inc.'s two largest data center campuses currently under construction, which together represent a capacity of 2.5GW.
The report also countered another widespread market claim that "half of U.S. data center projects are delayed, with only 5GW under construction nationwide." SemiAnalysis stated that Meta Platforms, Inc.'s two campuses alone equal half of that figure, calling such headlines "completely wrong."
Four Strategic Avenues: Why Meta Can Keep Betting on Compute
SemiAnalysis believes the market misjudged the situation because it only saw the action of "selling compute" without understanding why Meta Platforms, Inc. has the confidence to continue expanding.
The report outlines four high-value monetization paths, each fundamentally different from the standard "bare-metal IaaS" model of typical Neocloud providers.
First, frontier AI models (MSL) remain the core focus. SemiAnalysis explicitly stated that Meta Platforms, Inc. has not abandoned training frontier models. The Meta Superintelligence Labs (MSL) is still the primary destination for incremental compute. The report noted the team is "excited" about its progress and will release a dedicated deep-dive report to assess MSL's chances of catching up to Anthropic and OpenAI.
Second, the advertising recommendation system (RecSys) has potential for 10x expansion. SemiAnalysis believes Meta Platforms, Inc. is confident it can increase the complexity of its ad recommendation system by more than tenfold to accelerate revenue growth, requiring simultaneous investment in inference and training compute. Larger, more expensive RecSys models are already driving advertisers to pay higher prices while maintaining strong return on ad spend (ROAS), keeping users engaged on Meta's apps longer, and expanding the inventory of monetizable ad impressions.
Third, a model API service similar to Bedrock. SemiAnalysis exclusively disclosed that Meta Platforms, Inc. is in final negotiations with Anthropic to secure rights for a private deployment of Claude, similar to how Alphabet (GOOG) accesses Claude via Bedrock, but with the distinction of running it within Meta Platforms, Inc.'s own data centers. This means Meta Platforms, Inc. could in the future offer services bundling Claude with its own compute and platform, not just its own models.
SemiAnalysis listed three monetization paths for this: internal use to meet its own Claude token demand where Anthropic's supply lags; selling Claude services externally in a Bedrock-like model with high security; and vertical integration to build application-layer AI Agent products for sales and marketing, leveraging its position as one of the world's largest advertising platforms.
SemiAnalysis also noted that distributing models to free social media users and within the Meta hardware ecosystem (e.g., smart glasses) are potential options, which hold high strategic value for OpenAI and Anthropic, who "would likely be willing to make concessions to gain access to this distribution channel."
Fourth, "SpaceX-style" bulk compute leasing. This is one of the report's most impactful assertions. The compute leasing deal between SpaceX and Alphabet shocked the AI infrastructure sector with pricing four times the industry average; its deal with Anthropic was priced three times higher.
SemiAnalysis's AI Cloud TCO team, which tracks hundreds of GPU cloud deals annually, concluded they had "never seen a deal of this size with such a short term. The contract is nominally for three years, but either party can cancel with 90 days' notice—effectively a 3-month contract that auto-renews."
The report explains that few companies can replicate this model. Typical Neocloud providers need multi-year contracts to cover financing costs and cannot offer a 90-day cancellation option. While the three hyperscalers (Microsoft (MSFT), Amazon, Alphabet) technically could, they have higher-value long-term binding strategies. The only companies left capable of copying the SpaceX model, according to SemiAnalysis, are Oracle and Meta Platforms, Inc..
Meta Platforms, Inc.'s advantages are its vast compute resources, rapid build-out capability, and the ability to cancel contracts at short notice. At a pricing of $500 billion in revenue per GW per year, allocating just 200MW to external clients could generate over $10 billion in annual revenue with extremely high margins. The 90-day clause means Meta Platforms, Inc. can reclaim that compute if its Superintelligence Labs needs it.
The report also highlighted Meta Platforms, Inc.'s "tent-style" ultra-fast data center construction strategy, which SemiAnalysis first tracked last year and is now being rapidly deployed across the U.S. This rapid deployment and monetization aligns perfectly with the SpaceX model. SemiAnalysis expects Meta Platforms, Inc. to soon announce a similar bulk compute leasing deal with a major client, most likely Anthropic.
The "CFO's Dream": High Optionality Fuels Confidence
This is one of the core logics of the SemiAnalysis report. The coexistence of four monetization paths means every gigawatt of compute Meta Platforms, Inc. acquires has multiple high-value outlets. This isn't about buying too much and facing losses, but buying more to have options.
The report stated: "This is basically a CFO's dream, making it very easy to go all-in on compute. We'd bet that Susan [Meta CFO Susan Li] did a complete 180 after seeing the pricing on the SpaceX compute deal!"
The logic is straightforward: if the Meta Superintelligence Labs succeeds, all compute is used in-house for the highest ROI; if it faces setbacks, a portion can be leased via the SpaceX or Bedrock models for immediate high-margin revenue; if RecSys expansion underperforms, other outlets are available.
This high optionality has another effect: Meta Platforms, Inc. can continue procuring compute from third-party Neoclouds like CoreWeave and Nebius because even if it "sub-leases" that capacity, the profit margin is sufficient to cover costs.
SemiAnalysis offered reassurance to Neocloud investors: "Meta Platforms, Inc. will not become a bare-metal IaaS supplier with 30% gross margins; all its monetization options are high-value. This gives it enough profit margin to provide compute to external clients while continuing to procure capacity from Neoclouds to accelerate expansion." In other words, Meta Platforms, Inc. is more likely to be a significant source of RPO growth for companies like CoreWeave than a competitor.
RecSys: The Most Overlooked Compute Monetization Engine
Another key dimension is the advertising recommendation system. From late 2022 to early 2023, the market widely believed Meta Platforms, Inc. had entered a period of maturity. However, revenue growth accelerated significantly in the following years. SemiAnalysis attributes this to GPU investment as a key trigger.
The logic chain is: larger, more expensive RecSys models → more precise ad targeting → advertisers willing to pay higher prices → advertisers maintain strong ROAS → creating a virtuous cycle.
Simultaneously, upgrades to the content recommendation system increased user time spent on Meta's apps, further expanding ad inventory. SemiAnalysis believes Meta Platforms, Inc. itself is confident it can increase model complexity by more than 10x from current levels.
SemiAnalysis's core conclusion is clear: Meta Platforms, Inc. is leasing compute not because it bought too much, but because it has enough to simultaneously support multiple high-value strategies, each sufficiently profitable.
Capital expenditure in 2027 will be much higher than market expectations. For investors, this report suggests the recent sell-off may have been based on a misjudgment. However, SemiAnalysis added an important caveat: whether MSL can truly catch up to Anthropic and OpenAI remains the largest variable and uncertainty.
Comments