The biggest trade of 2025 isn't a model launch. It's a rental agreement.
Word on the street—and by street I mean the NYT wire and three anonymous sources in the know—is that Meta and Anthropic are negotiating a $10 billion compute lease over two years. That's $4.17 billion a month of raw GPU time. For context, that's about the same as what the US spends on ice cream annually. But in crypto terms? It's the kind of liquidity that makes a chain fork.
Context: Why Now?
Meta's been building data centers like they're buying distressed real estate in a bull run. CapEx this year alone hit $145 billion. That's double the previous year. And Zuckerberg himself admitted the spend hasn't yet borne fruit. Meanwhile, Anthropic is sitting on the world's second-most-wanted AI model—Claude—but they're gasping for compute. They already have a $45 billion deal with SpaceX (yes, Elon's rocket company) over three years. Yet they're still hungry. That tells you something about the scale of their training and inference needs.
Competitors becoming customers. That's the narrative. But it's deeper than that.
Core: The Deal
Here's what we know: The deal is two years, $10 billion, with monthly payments and exit clauses. Meta gets to turn idle GPUs into cash flow. Anthropic gets guaranteed compute at a time when every startup is begging for H100s like they're the last water in the desert. This isn't about training the next frontier model—it's about feeding the beast that is Claude Code and the exploding inference demand.
Speed is the only metric that survived the crash. And this deal is moving fast. Meta's been renting from CoreWeave and Nebius themselves. Now they're flipping the table and becoming the landlord. It's the ultimate pivot: from capital expenditure black hole to revenue-generating asset.
From my trading desk in Prague, I've watched compute contracts move like ETF flows. The signals here are clear. Meta's $145 billion spend was scaring investors. This two-year $10 billion lease is a lifeline—it proves their infrastructure can produce income. Not just sit there depreciating. For Anthropic, the message to IPO investors is: we have locked-in supply of the most scarce resource in AI. That's a strong narrative.
The Contrarian Angle
Everyone's talking about compute shortage. But the real story is compute assetization. Meta is using its infrastructure debt to generate yield. It's like turning a stranded mining rig into a staking validator. Social capital outpaced code in the ape arcade—and now the same applies to data centers.
But there's a dark side. Anthropic is putting its core IP—model weights, user queries, training data—on competitor metal. Even with hardware security modules and isolated networks, the risk of data leakage or IP theft is real. I've seen this in the ETF world: when you share the same broker, the order flow gets complicated. Here, the order book is the model itself.
And the exit clauses? They cut both ways. If Anthropic's efficiency improves dramatically and they need less compute, they can walk away. But that leaves Meta with stranded capacity. If Anthropic's growth explodes and they need more, they're locked into a contract that might become expensive. It's a game of timing, and timing in AI right now is measured in seconds, not years.
Reading the room while the order book burns—that's what Meta is doing. They saw the narrative around their massive spend was toxic. They turned a weakness into a strength. But this also signals that Meta's own models (Llama) might not be the priority. Why invest billions in catching up to Claude when you can just rent compute to Anthropic and profit from their success? That's a dangerous strategic bet.
Takeaway: What to Watch Next
The sprint doesn't end when the block confirms. This deal, if it closes, will be a template for the next wave of AI infrastructure play. Other overspenders—think Microsoft, Google, even Tesla—will start leasing out excess compute. We'll see a secondary market for GPU time that trades like futures.
But the real alpha is in monitoring the delivery. Can Meta actually provision that many GPUs quickly? Their data centers are optimized for internal workloads, not multi-tenant AI SaaS. If they stumble, Anthropic's IPO story cracks. And if regulators step in—calling this anti-competitive coordination—the whole house of cards wobbles.
Liquidity flows like adrenaline, not like water. This deal is a shot of adrenaline straight into the heart of the AI compute market. Watch the order books. Watch the wallets. The fastest runners will be the ones who read the signals before the block confirms.