I’ve spent the past decade watching decentralized networks promise to flatten power structures. Yet here we are in 2026, and the most dangerous concentration of power isn’t in a bank—it’s in a 220,000-GPU cluster owned by one man. Elon Musk just admitted that his own AI company, xAI, is no longer competing with Anthropic. He called them the clear leader. But the real story isn’t the ego—it’s the $1.25 billion monthly rent check Anthropic is paying to Musk’s other company to keep those GPUs humming.
Let me back up. I launched my first DAO in 2017 from a co-working space in Cape Town. We raised $120,000 in ETH, built a community of 500 artists, and then watched it die because I hadn’t accounted for gas fees. That failure taught me something no whitepaper ever could: decentralization without robust infrastructure is just a prayer. Today, the same lesson applies to AI. The “leader” in artificial intelligence isn’t the one with the best research—it’s the one who can write a $90 billion contract for compute over six years.
The numbers are staggering. Anthropic, the company Musk once called “hypocritical,” is now renting over 220,000 Nvidia GPUs from xAI’s Colossus 1 facility. The price: $1.25 billion per month, locked in until 2029. That’s $15 billion annually—more than the entire revenue of most Fortune 500 companies. Meanwhile, Grok 4.5, xAI’s flagship, ranks fourth on the Artificial Analysis Intelligence Index, behind three Anthropic models (Fable 5, Opus 4.8) and one from OpenAI (GPT-5.5). Musk himself admitted that Grok 4.5 only competes with “previous-generation Claude.” That’s not a battle—it’s a resignation.
What fascinates me is the architecture of this deal. It’s not a vendor relationship; it’s a symbiotic stranglehold. Musk controls the compute that powers his rival. By leasing the GPUs instead of selling them, he turns Anthropic’s success into his own cash flow. At $1.25 billion per month, xAI generates more revenue from this one contract than most AI startups will ever see. But there’s a catch: if Musk ever decides to flip the switch, Anthropic’s entire operation grinds to a halt. He promised not to cut off supply—publicly. But promises are only as strong as the balance sheet behind them.
I’ve seen this pattern before. During DeFi Summer in 2020, I chased yields across three protocols simultaneously, made $15,000, and nearly burned out. The lesson? When you don’t own the infrastructure, you’re renting your upside. Anthropic is renting its future. Yes, they have the best models today—Fable 5 is the king of the benchmark hill—but their cost structure means they must win continuously. Every month that passes without a breakthrough, the $1.25 billion bleeds. And while Musk’s xAI may be a step behind in model quality, they’re sitting on a cash machine that doesn’t require innovation—just uptime.
The contrarian take: Musk’s public humility isn’t weakness—it’s a hedge. By praising Anthropic, he inflates the value of the compute he’s selling. And by locking them into a six-year contract, he gains insider knowledge of their roadmap. He predicts “Mythos 2 coming soon”—a detail he likely knows from the partnership. That’s more valuable than any benchmark score. In Web3, we say “code is law, but people are truth.” Here, the law is the contract, and the truth is that Musk now owns a piece of every Anthropic breakthrough.
This is the shape of the AI arms race: not a battle of algorithms, but a war of capitals. The market is already responding. Nvidia’s backlog just became a multi-year annuity. Small AI labs that can’t afford $1.25 billion a month are being squeezed out. And the real winner might not be any single model, but the infrastructure layer itself. Embrace the volatility, find the signal. The signal here is that compute is the new land—and Musk is the biggest landlord.
What happens when that landlord decides to raise the rent? Or when the next Scaled Law breakthrough makes 220,000 GPUs obsolete? I don’t have the answers, but I know this: the communities that will thrive are the ones building their own infrastructure, not renting someone else’s. In Cape Town, we learned that the hard way. Maybe it’s time for the AI industry to learn it too.