半年狂揽5GW：SemiAnalysis拆解Meta算力棋局，市场抛售系误读

TL;DR

• In the first half of 2026, Meta signed over 5GW of capacity in cloud and colocation, excluding its simultaneously accelerated self-built data centers.
• The new computing power could be directed towards MSL training, ad recommendations, private instances for Claude, and short-term, high-priced external transactions.
• CoreWeave and Nebius' RPO may benefit, but MSL's catch-up efforts and contract flexibility remain risks.

The sell-off in Neocloud stocks triggered by Meta may have been misguided. A report published by SemiAnalysis on July 2nd states that Meta has secured over 5GW of IT capacity in cloud services and colocation in the first half of 2026, a figure that does not include its simultaneously accelerated self-built data centers.

This contradicts market concerns from the past few days. According to a Bloomberg report on July 1st, Meta is developing a cloud business to sell its excess AI computing power. The plans are still in development and subject to change. Following this news, shares of Neocloud companies like CoreWeave and Nebius experienced a sell-off, as investors feared Meta transitioning from a major customer to a potential competitor, leading to a rapid oversupply of AI data centers.

SemiAnalysis offers an alternative narrative: Meta is not reducing external procurement; instead, it is leveraging third-party Neocloud providers to obtain capacity faster. Since the beginning of 2024, Meta has signed contracts totaling nearly 10GW, with most of the new capacity still acquired through third parties. For suppliers like CoreWeave and Nebius, Meta's orders could actually continue to drive up their Remaining Performance Obligations (RPO).

Quarterly breakdown of Meta's compute transactions: Cloud and colocation contracts signed in H1 2026 cumulatively exceed 5GW, distinguishing between new data center leases and Neocloud GPU deals.

Market Fears Meta Becoming a Seller, Report Sees an Even Larger Buyer

The crux of this debate isn't whether Meta will dabble in cloud compute resale, but rather who builds, consumes, and bears the revenue risk for the massive influx of new computing power.

If Meta merely subleases its GPUs, becoming a bare-metal IaaS provider with roughly 30% gross margins, then the market's concern over Neocloud valuations would be justified. A major customer entering the supply side would weaken the bargaining power of existing suppliers and could potentially lead the industry into a price war.

However, in SemiAnalysis's framework, Meta's new capacity resembles more of a "flexible compute pool." It can be allocated between internal frontier models, ad recommendations, enterprise-grade model services, and short-term high-priced external transactions, rather than being limited to low-margin GPU subleasing.

This is also key to the report's refutation of the claim that "only 5GW of data centers are under construction in the US." Meta's two largest data center campuses under construction alone account for approximately 2.5GW of capacity. Adding third-party cloud and colocation contracts, the actual construction intensity is higher than some pessimistic estimates.

In simpler terms, the market's current question isn't whether Meta buys computing power, but whether this vast capacity can be absorbed by high-value use cases.

Four Avenues to Absorb New Compute, MSL is Not the Only Outlet

The top priority for Meta's new computing power remains Meta Superintelligence Labs (MSL) for frontier model training. This is the most direct narrative for capital expenditure: Meta needs sufficiently large training clusters, talent, and room for experimentation to catch up with OpenAI and Anthropic.

But even if MSL's progress doesn't fully meet expectations, Meta isn't forced to merely lease out its GPUs at low prices.

The second avenue is the ad recommendation system. Meta's official earnings report shows a 19% year-over-year increase in ad impressions and a 12% year-over-year increase in average price per ad in Q1 2026. Meta Engineering previously noted that the GEM-related training stack achieves a 23x improvement in effective training FLOPs, approximately 1.43x MFU improvement, and a 16x expansion in GPU scale. Doubling the GPUs for GEM training led to a 5% and 3% improvement in Instagram and Facebook Feed ad conversion rates, respectively.

This path is easier for investors to understand: if more compute can enhance ad conversion rates, it's not just "burning cash on GPUs" but a component of advertising revenue and pricing power. As for the specific ranking metric improvements mentioned in the report, independent public sources are limited, making them more suitable as assumptions in SemiAnalysis's model rather than fully confirmed facts by Meta.

The third avenue is a model service platform, akin to AWS Bedrock or Google Vertex. SemiAnalysis states that Meta is in negotiations with Anthropic for private instances of Claude and is attempting to build a "Token-as-a-Service" platform. This capacity could be used internally, sold to SaaS providers, or distributed externally. However, such transactions should still be viewed as "potential developments" rather than realized revenue.

The fourth avenue is large-scale, short-term, high-premium on-demand compute transactions, similar to SpaceX's approach. This is perhaps the most striking set of numbers in the report.

Meta's total compute capacity forecast from 2023 to 2027, using a stacked bar chart to distinguish between MSL, Other AI, and Non-AI, showing significant capacity expansion from 2026 to 2027.

High-Price Short-Term Deals Change the Revenue Perception of "Selling Compute"

The key to SpaceX-style deals isn't just "renting GPUs," but the differences in pricing and contract structure.

SemiAnalysis estimates that the annualized revenue per GW from the SpaceX-Anthropic deal is approximately $3.1 billion, which is 2.6 times the average price of a typical five-year Neocloud IaaS contract. The SpaceX-Google deal is even higher, at about $4.8 billion/GW/year, or roughly 4 times the average. Independent public sources offer limited confirmation of these contract details, so these figures are best treated as scenarios within the report, illustrating potential high premiums for short-term scarce compute.

If Meta were to allocate just 200MW to similar external transactions, based on calculations in the report's public pages, the annualized revenue could exceed $10 billion. This scale could change the market's intuition about "Meta selling compute externally": it might not necessarily be low-margin subleasing, but could involve selling time windows to top-tier customers in urgent need of compute, leveraging its rapidly deployable data center capacity.

The report also mentions that Meta's fast-deploying data center design is suited for such transactions. Its value lies less in long-term, lowest-cost leasing and more in delivering large-scale compute capacity faster when model companies, AI applications, or major clients have immediate needs.

However, this remains an optional path, not a stable, realized revenue stream. Meta having the conditions to replicate some high-premium deal structures does not mean it has already become a SpaceX-style compute seller.

Comparison of SpaceX pricing premiums: A typical 5-year Neocloud IaaS deal is ~$1.2B/GW/year, SpaceX with Anthropic is ~$3.1B, and SpaceX with Google is ~$4.8B.

Pressure on CoreWeave, Nebius Doesn't Necessarily Stem from Disappearing Demand

For Neocloud companies like CoreWeave and Nebius, the market's prior concern was: if Meta builds its own or resells compute, its external procurement would decrease, potentially draining industry orders.

However, based on existing contracts, Meta is still accelerating its use of third-party Neocloud providers. Public information shows CoreWeave holds a $21 billion contract with Meta, and Nebius's contract with Meta could be worth up to $27 billion. In its Q1 2026 shareholder letter, Nebius mentioned signing a second major deal with Meta, with a contract capacity exceeding 3.5GW, and referenced commitments from Microsoft and Meta as customers.

Meta's willingness to pay a premium for speed is precisely why third-party suppliers retain value. As long as Meta believes the compute can be absorbed by MSL, ad systems, model services, or short-term high-price deals, it has reason to let Neocloud providers build clusters first, rather than waiting for its self-built projects to deliver slowly.

Assessing "capacity oversupply" also shouldn't focus purely on total GW figures. The truly scarce elements in AI data centers are often not paper power, but rather available GPUs, networking, data center delivery speed, customer migration costs, and contract flexibility. If Meta needs large blocks of capacity quickly, third-party Neocloud providers remain useful.

This doesn't mean Neocloud companies are without risk. Their valuations still depend on customer concentration, financing costs, GPU depreciation, long-term contract quality, and whether customers actually consume the future capacity. If the RPO growth driven by Meta corresponds to high capital expenditure and high customer concentration, the market will still apply a discount.

If MSL Fails to Catch Up, 5GW+ Becomes a Capital Expenditure Burden

The most crucial point of restraint in this report is not to portray every optional path for Meta as already successful.

Whether MSL can catch up with OpenAI and Anthropic remains highly uncertain. Frontier model competition isn't solvable by GPU count alone; data strategy, research teams, training stability, product distribution, and inference costs all influence outcomes.

Contract terms also affect risk levels. SemiAnalysis notes that SpaceX-like deals often include 90-day mutual cancellation clauses. This arrangement provides flexibility for both parties: if a team's progress falters, compute can be quickly reclaimed; if demand changes, it avoids being locked in long-term. Details of these clauses lack independent public confirmation and should be treated as report assumptions.

For Meta, flexibility itself holds value. It can allocate sufficient power and GPUs for MSL's frontier experiments while diverting some capacity to ad recommendations, private Claude instances, or high-price short-term deals.

Conversely, if Meta ends up signing a large volume of long-term compute deals lacking flexible exit options, the risk increases. If its frontier model pursuit stalls and advertising or model services cannot absorb the new capacity, the over 5GW of newly outsourced compute could more directly transform into a capital expenditure burden.