Meta May Rent $10 Billion in Compute to Rival Anthropic
Welcome to the AI Compute Landlord Era

1What happened

The New York Times reported on July 16 that Meta is in talks to lease Anthropic up to $10 billion in compute over two years — a proposal Anthropic itself put forward in June. Discussions are early-stage, terms remain subject to change, and either side can exit. Meta shares closed down more than 2% the day the news broke (July 17), amid a broader tech selloff.

It's not Anthropic's first arrangement with a nominal competitor for capacity. In May, Anthropic struck a deal worth up to $40 billion with Elon Musk's SpaceX/xAI for the "Colossus 1" data center in Memphis — built for training Grok but running at only ~11% utilization, with 220,000+ GPUs and 300+ megawatts. Anthropic pays $1.25 billion a month through May 2029, with a 90-day exit clause. Reporting also points to a roughly $40 billion, five-year TPU deal with Google, and Amazon has committed up to $25 billion for up to 5 gigawatts of capacity.

2Why — this isn't about money, it's about chip delivery speed

The driver is that Anthropic's revenue is growing far faster than any capacity plan can account for. CEO Dario Amodei says the company planned for 10x growth and got 80x — annualized revenue went from $87 million in January 2024 to $30 billion in April 2026, then $47 billion in May, with full-year ARR now projected at $75-90 billion, while API volume is up nearly 70x year over year. Users feel it directly: reliability dips at peak hours, Pro-tier limits tighten, agentic workflows burn through allowances in hours.

What's actually scarce is never money — Anthropic just closed a $65 billion Series H, briefly touched a $965 billion valuation, and has confidentially filed for an IPO. What's scarce is how fast Nvidia's newest chips can physically arrive. So the strategy becomes: lease from whoever has idle compute, even a company building its own competing model. Musk's own comment sums up the logic: "the enemy of my enemy is a compute partner."

3Who wins, who loses — the second-order effect

Benefiting: Meta (turns AI infrastructure into a revenue line for the first time, if it closes), Anthropic (locks in scarce compute ahead of an IPO push), Nvidia (total chip demand doesn't shrink based on who owns the data center), and SpaceX/xAI (converted an 11%-utilized sunk asset into tens of billions in on-paper revenue). Exposed: Meta's own shareholders — the FT estimates a -29% return on Meta's AI infrastructure spend, the only negative figure among the big hyperscalers — plus confidence in the "AI capex always pays off" assumption, and independent cloud providers now facing uninvited new competitors.

4What the coverage is missing · a critical read

Most reporting compresses this into "Meta could make $10 billion leasing out compute" — the standard "revenue diversification" template — while skipping the one comparable, publicly disclosed data point available: SpaceX's compute-subletting segment lost money in its first reported quarter. Nearly every headline dollar figure circulating — Meta's $10B, SpaceX/xAI's $40B, the reported $40B Google deal — is a multi-year contract's nominal total, not confirmed current-period profit. That doesn't make the deals fake; it means this compute-subletting economy hasn't yet proven its business model covers its own costs.

One takeaway: next time a headline reads "Tech Giant Signs $X Billion AI Compute Deal," ask first whether that's booked revenue or a multi-year contract total, and whether the counterparty is a pure customer or also a competitor. If you hold Meta, Google, Amazon, or Nvidia, watch how tightly these mutual customer-supplier relationships are binding their income statements together — not any single headline in isolation.