Meta is negotiating a two-year deal to lease computing power to Anthropic for as much as $10 billion, according to a New York Times report that pushed Meta shares off their Friday lows. If it closes, a company best known for selling ads against your attention will start selling the silicon that trains its rivals’ models — and Anthropic, locked in a frontier race with OpenAI and xAI, will pay a social-media giant for the privilege of keeping up.
🔍 THE BOTTOM LINE
Compute is now the scarce input, and the companies that hoarded it are turning into landlords. Meta leasing capacity to Anthropic isn’t partnership — it’s arbitrage. Meta monetises infrastructure it already paid for; Anthropic gets GPUs it can’t build fast enough. The real signal: the “AI cloud” market is no longer a niche of neoclouds. It’s now big tech selling to big tech.
What the Deal Actually Looks Like
The reported terms are a two-year lease worth up to $10 billion, with monthly payments from Anthropic and exit clauses on both sides, per Whalesbook’s summary of the NYT and Bloomberg reporting. Meta would lease a “significant portion” of its computing capacity — the same GPUs it bought for its own Llama training runs and its recommendation models — to Anthropic, which needs them to train and serve Claude.
No final agreement has been announced. Both sides have included clauses allowing an early exit, which is a polite way of saying neither wants to be trapped if the compute market turns. But the shape of the deal is clear: Meta becomes an infrastructure-as-a-service provider, and Anthropic becomes a tenant.
Why Anthropic Is Paying Meta Instead of Building Its Own
Anthropic has been on a compute shopping spree. Weeks before the Meta talks surfaced, Anthropic signed a deal with Elon Musk’s SpaceX to lease capacity at the Colossus 1 data center in Memphis — 300 megawatts, 220,000 GPUs, live within a month. That followed earlier capacity arrangements tied to its Amazon and Google partnerships. The pattern: Anthropic would rather rent than build, because renting converts a capital problem into an operating one and lets it scale faster than construction permits allow.
The Meta talks are the same logic on a different counterparty. Anthropic is signing wherever GPUs are sitting idle — and the uncomfortable truth for the rest of the industry is that Meta, after years of overbuilding for its own AI roadmap, has idle capacity. That’s not a sign of weakness. It’s a sign that Meta bought enough silicon to survive a demand miss.
Why Meta Is Selling
Meta has spent tens of billions on data centers and NVIDIA GPUs to train Llama and run its advertising models. For years that infrastructure was a cost center — a giant line item that Wall Street scrutinised every quarter. Leasing it to Anthropic converts that cost into revenue, and potentially high-margin revenue at that: the marginal cost of running already-paid-for GPUs is power and cooling, not the chips themselves.
The strategic upside is bigger than the revenue line. If Meta becomes a credible AI infrastructure provider, it competes directly with the neoclouds — CoreWeave, Nebius, and the rest — that have been eating the rent-GPU market. Nebius signed a $27 billion, five-year deal with Meta in March, with Meta as the buyer. Six months later Meta is reportedly on the other side of the table, selling capacity to Anthropic. The same company is both a neocloud customer and a neocloud competitor. That’s not a contradiction — that’s a company with enough scale to play both sides of the market.
The Bigger Pattern: Big Tech Selling to Big Tech
The Meta-Anthropic talks fit a broader reshuffling. The NYT documented last December how Microsoft, Meta, and Google were offloading AI buildout risk onto upstarts — Microsoft leasing tens of billions in compute, Meta financing a Louisiana data center without taking on the debt, Google renting capacity and reselling some of it to OpenAI. The pattern then was big tech pushing risk down. The Meta-Anthropic pattern now is big tech selling capacity across to each other.
The implication is that the AI compute market is bifurcating. At the top, a handful of labs (OpenAI, Anthropic, xAI, DeepMind, Meta itself) need more compute than any single infrastructure owner can provide. At the bottom, neoclouds and specialised providers compete to rent out capacity. The Meta-Anthropic deal collapses the layers: a lab becomes its own neocloud’s landlord. Expect more of these. OpenAI has its own infrastructure deals with Microsoft; xAI has Colossus; Google has TPU capacity it can sell. The labs that don’t have an internal compute reservoir — Anthropic, Mistral, the Chinese labs — are the natural tenants.
NZ Angle
None of this directly moves the needle for New Zealand’s sovereign AI position, but it sharpens the existing dilemma. If compute is becoming a market where only a handful of counterparties can supply at scale — Meta, Microsoft, Google, AWS, xAI — then any NZ organisation training frontier models is renting from one of those five. The Crown-owned AI institute and the universities’ HPC clusters aren’t in the running for frontier-scale training; they’re running inference and fine-tuning. The Meta-Anthropic deal is a reminder that the “sovereign compute” question isn’t going away — it’s getting more acute as the suppliers consolidate.
❓ FAQ
Is the deal finalised? No. Multiple outlets report it as “in talks” with exit clauses on both sides. Anthropic and Meta have not publicly confirmed, and Anthropic has been simultaneously negotiating with other capacity providers.
Why is Meta willing to help a competitor? Anthropic’s Claude competes with Meta’s Llama in the model market, but Meta’s strategic interest in monetising infrastructure likely outweighs the competitive concern. Meta also doesn’t sell Claude — it sells GPUs. The revenue and the infrastructure-as-a-service market position are worth more than the marginal competitive harm from Anthropic having more compute.
How does this compare to the SpaceX deal? The SpaceX/Colossus deal was 300MW and 220,000 GPUs live within a month — operational, near-term capacity. The Meta deal is reported as a two-year lease structure worth up to $10B, suggesting a longer-term, larger-scale commitment. Anthropic appears to be layering short-term leases (SpaceX) and long-term leases (Meta) to cover different planning horizons.
What’s the risk for Meta? Internal capacity contention. Meta has its own expanding AI roadmap — Llama training, recommendation models, the Ray-Ban Meta AI features. Leasing capacity to Anthropic means less capacity for its own workloads unless it has overprovisioned. The exit clauses in the reported deal terms are the hedge.
🔍 THE BOTTOM LINE
The AI arms race has produced a strange new category: the compute landlord. Meta selling capacity to Anthropic isn’t an alliance and isn’t a capitulation — it’s a market forming in real time, where the labs that overbought silicon become suppliers to the labs that didn’t. Expect Microsoft, Google, and AWS to play the same game within quarters. The question isn’t whether big tech will sell compute to its rivals. It’s which rivals, and at what price.
📰 Sources
- Quartz — Anthropic in talks to lease Meta computing power in $10 billion deal
- TipRanks — Meta Platforms to Lease Computing Power to AI Startup Anthropic
- Whalesbook — Meta In Talks For $10 Billion AI Compute Deal With Anthropic
- Reuters — Nebius signs AI infrastructure deals with Meta worth up to $27 billion over 5 years
- New York Times — How Tech’s Biggest Companies Are Offloading the Risks (Dec 2025)
- Bloomberg — Meta in Talks to Sell Computing Power to Anthropic
