Version 3

Meta is building a cloud business to sell excess AI compute capacity to outside developers and enterprises, structured like AWS Bedrock with per-token model access and raw GPU rental. [1][2] Bloomberg's July 1, 2026 report sent META shares up approximately 10% while CoreWeave fell roughly 12.7% and Nebius fell roughly 12.4%. [6] SemiAnalysis reported Meta contracted over 5 GW of external cloud and colo capacity in H1 2026 alone and is in final talks with Anthropic for private Claude instances at potentially $10B+ scale. [7] Anthropic, however, has independently assembled a 10 GW multi-provider compute portfolio spanning Amazon, Google, Microsoft, Fluidstack, and SpaceX — making it a sophisticated multi-supplier buyer rather than a captive customer of any single provider. [8]

Meta would be the first major US social media company in enterprise cloud, converting $125–145 billion in planned capex into a direct B2B revenue line. The split between Wells Fargo's 85% margin surplus model and SemiAnalysis's premium-contract model determines whether the financial case holds. Anthropic's multi-GW compute strategy shows enterprise AI demand for Meta's compute is real, but also that buyers have options — which limits Meta's pricing power.

On July 1, 2026, Bloomberg reported that Meta is building a cloud business to sell excess AI compute capacity to outside developers and enterprises. The planned offering resembles AWS Bedrock — per-token access to Meta's Llama and Muse Spark models alongside raw GPU compute rental — with Meta charging developers directly for infrastructure built primarily for its own AI systems. [1][2] Meta shares rose approximately 10% on the day. Mark Zuckerberg had signaled the direction in May 2026, noting that outside companies ask Meta for compute almost every week, and a January 2026 infrastructure initiative had established the groundwork. [3][4] The neocloud market reaction treated the announcement as a competitive threat: CoreWeave fell approximately 12.7% and Nebius fell approximately 12.4%, despite Meta having committed more than $35 billion in take-or-pay contracts to CoreWeave. [5][6]

SemiAnalysis, reporting on July 2, challenged the 'excess capacity' framing. Meta contracted over 5 GW of external cloud and colo capacity in the first six months of 2026 — not counting its own self-build activity — and is in final talks with Anthropic to access private instances of Claude, potentially at $10 billion or more in value. [7] Meta also plans to scale its advertising recommendation systems by more than 10x in complexity, with Meta Superintelligence Labs consuming the bulk of incremental compute. SemiAnalysis argues Meta will pursue SpaceX-style premium on-demand contracts rather than commodity bare-metal IaaS and will not operate at 30% gross margins: 'all its options are high value.' If internal demand grows as fast as capacity, the surplus available for resale may be structurally narrower than financial models assume.

The Anthropic dimension is more complex than it first appeared. In May 2026, Anthropic announced it had signed an agreement with SpaceX to access the full Colossus 1 data center — over 300 megawatts and 220,000 NVIDIA GPUs — and disclosed a multi-provider compute portfolio totaling over 10 GW of future capacity across Amazon, Google/Broadcom, Microsoft/NVIDIA, Fluidstack, and SpaceX. [8] Anthropic is simultaneously pursuing orbital compute development with SpaceX and has expressed a preference for partnering with providers in democratic countries with secure supply chains. This context matters for evaluating the Meta-Anthropic talks: Anthropic is a sophisticated compute buyer with multiple active supply relationships, not a captive customer, which limits how much pricing leverage Meta can extract from any single deal.

The financial debate remains unresolved. Wells Fargo models $20 billion in revenue per gigawatt at 85% operating margins, projecting meaningful EPS accretion per GW resold. [2] Milk Road AI argues that Meta's implied $20B/GW pricing validates CoreWeave and Nebius at roughly twice their current market valuations, framing the July 1 sell-off as a mispricing rather than a correct competitive read. [5] Bloomberg analyst Mandeep Singh takes the opposite view: hardware alone does not make a cloud business, and Meta lacks the billing systems, uptime commitments, security infrastructure, and developer tooling that enterprise customers require — making the move look more like fallback revenue than a credible AWS alternative. [1] Neither a launch date nor pricing for Meta's cloud service has been publicly disclosed.