The picks and shovels play is renting Meta's AI training GPUs to outside labs, with roughly 10 gigawatts of data center commitments anchoring a recurring revenue business worth about $10 billion a year.
Meta signed more than five gigawatts of AI training capacity in the first six months of 2026, the largest such build-out the company has ever attempted. The company is now preparing to rent portions of that fleet to outside customers, its frontier-model rivals included, per Bloomberg reporting surfaced by QbitAI.
A GPU-leasing consideration Bloomberg calls Meta Compute would package the company's AI accelerators into multi-year leases modeled on the neocloud contracts SemiAnalysis has spent the last quarter cataloging. For a company whose in-house models have not closed the gap with OpenAI, Anthropic, or Google, the playbook is straightforward: if Meta cannot win the model race on its own, it can rent the road to anyone who can.
The revenue model borrows from SpaceX-style economics. SemiAnalysis pegs the long-term leasing rate at roughly $50 billion in annual revenue per gigawatt. At that benchmark rate, a 200 megawatt slice routed to external customers could plausibly generate about $10 billion a year at very high margins, even before Meta layers in Bedrock-style third-party model hosting or premium pricing for ad-recommendation compute. The contract structure matters as much as the rate. SpaceX pioneered a roughly three-year term with a 90-day mutual cancellation window, effectively a quarterly auto-renew, so Meta can pull capacity back to its own superintelligence lab (MSL) the moment an in-house model needs it.
Meta's infrastructure scale backs the bet. Roughly 10 gigawatts of data-center and compute commitments have been signed since early 2024, per SemiAnalysis, with the two largest campuses under construction representing about 2.5 gigawatts on their own. Ten gigawatts is not a single facility. At SemiAnalysis's rough equivalence, that is closer to ten hyperscale campuses, large enough to host both Meta's training pipeline and a meaningful external customer base.
The monetization urgency is the capex. Meta's 2026 capital expenditure guidance is $125–$145 billion, and Q1 alone burned $19.84 billion, CNBC has reported. Selling internal training capacity on multi-year contracts turns a depreciating build-out into a recurring-revenue asset, the re-rating investors have rewarded across the rest of the AI infrastructure stack. Mark Zuckerberg told the same outlet in May that a cloud business was "definitely on the table."
The market read was immediate. Meta shares rose roughly 9% on Monday as the Bloomberg report circulated, while pure-play GPU lessors sold off, CoreWeave and Nebius among the names hit hardest, CNBC noted. The trade mirrors the one investors made on the earlier SpaceX-Anthropic compute deal, covered by Marketwise and reported by TechCrunch and others. When hyperscalers pivot to leasing, neoclouds take the loss. (In this context, "neocloud" refers to GPU-only rental providers like CoreWeave and Nebius, distinct from general-purpose clouds such as AWS or Azure.)
Anthropic is reportedly in late-stage talks for a private Claude instance running on Meta infrastructure, alongside a Bedrock-style third-party model marketplace that SemiAnalysis describes. The talks — which SemiAnalysis characterizes as approaching anchor-tenant status — have not been confirmed by either company. "Bedrock" here refers to AWS's existing service for hosting outside AI models, a category also occupied by Microsoft's Foundry and Google's Vertex. In the hypothetical marketplace SemiAnalysis lays out, Anthropic as a marquee counterparty would put Meta Compute on the same shelf as those hyperscaler marketplaces, assuming the deal closes and assuming those competitors complete their own counter-moves.
The pivot reads as a hedge against Meta's own stalled model progress rather than a victory lap. Internal morale at MSL is described as a 20-year low, Gemini access was recently restricted, and the next-generation model Watermelon remains in training with compute budgets far above its predecessor. The picks-and-shovels framing, the same one QbitAI's headline uses to describe the move, only requires that AI training demand stays roughly where it is and that someone has to rent the GPUs.
Either way, roughly $20 billion a quarter in AI infrastructure spend now carries a forward lease contract attached. The picks-and-shovels thesis does not depend on Llama closing the gap with GPT-5; that decoupling from model success is what moved the stock.