We didn't expect Elon Musk to admit defeat. Not publicly. Not like this.
But there it was: a post on X, praising his biggest rival. "Anthropic is clearly the leader in AI right now," he wrote. The crypto-native part of me flinched. Because when a billionaire who built his entire brand on technological dominance suddenly bows down to a competitor, the market usually misses something deeper.
And the market is missing it. I'm talking about the fact that behind Musk's humble admission lies a $12.5-billion-a-month GPU lease contract between his own company, xAI, and Anthropic. A contract that locks 220,000 Nvidia GPUs into a single facility for the next six years. This isn't just a tech story. This is a story about centralization of power, about the death of the trustless ideal, and about the uncomfortable marriage between the world's most capital-intensive technology and the world's most capital-inefficient industry: artificial intelligence.
Let me break this down the way I break down any complex protocol: by stripping away the narratives and looking at the code — in this case, the capital flows.
Context: The Unholy Alliance
For those who haven't followed the AI Cold War closely, the landscape is brutal. Anthropic — the startup built by ex-OpenAI employees who wanted to make AI safe — has been squeezing out state-of-the-art models. Their Fable 5 model now tops the Artificial Analysis Intelligence Index with a score of 54, ahead of OpenAI's GPT-5.5 at 52, and Anthropic's own Opus 4.8 at 48. xAI's Grok 4.5 ranks fourth. Musk himself admitted that Grok 4.5 is only competitive with "the previous generation of Claude."
But the real story is about what happens behind the models. At the end of 2025, Anthropic — which was burning cash faster than any unprofitable tech unicorn I've audited — signed a deal with xAI to lease compute from the Colossus 1 supercomputer facility. The terms: $12.5 billion per month, for 220,000 Nvidia GPUs, contracted through 2029.
Yes, you read that right. $12.5 billion per month. That's an annualized cost of $150 billion.
To put that in perspective: the entire global revenue of the cloud computing industry in 2025 was about $2 trillion. Anthropic alone is spending 7.5% of that just on compute rental. This is the kind of concentration that makes central banks nervous and makes crypto maximalists weep.
I learned to stop preaching and start listening after my 2022 burnout. I stepped back from the daily charts and the endless yield farming analysis. I realized that the human cost of technological acceleration was invisible to most analysts. Now, when I see a company spend $150 billion a year on compute, I don't just see a competitive advantage — I see a systemic risk.
Core: The Infrastructure Crisis of the Soul
Let me share three data points that changed my perspective on this deal.
First, the power. 220,000 H100 GPUs, each consuming up to 1,400 watts at peak, means total power draw of about 308 megawatts. That's the equivalent of a small nuclear reactor's output. Colossus 1 will consume about 27 billion kilowatt-hours per year. To put that in crypto terms: Bitcoin's entire network consumes around 100 TWh annually. This one facility will use about 27% of that. The carbon footprint alone is staggering. And yet, there is no 'proof-of-work' consensus mechanism here — just a single company's proprietary training run.
Second, the chip dependency. Every single one of those GPUs is made by Nvidia. The deal locks Nvidia into the AI narrative for the next six years. If Nvidia's roadmap falters, or if export controls tighten (as they almost certainly will for next-gen chips), Anthropic's entire competitive advantage disappears. The data doesn't lie: the GPU supply chain is the new Saudi Arabia, and Nvidia is the new OPEC. But unlike oil, these chips are not fungible. They are proprietary, programmable, and tied to a single ecosystem (CUDA). This is the antithesis of what we built in crypto — an open, permissionless, trustless infrastructure.
Third, the relationship. Musk publicly called Anthropic a "hypocritical company" just a year ago. Now he's their largest compute provider. This is the ultimate example of 'co-opetition': Musk controls the compute pipeline, yet he praises their AI while failing to beat them. Why? Because the contract is a $150 billion annual revenue stream for xAI. By locking in Anthropic, Musk turns his competitor into a customer. He gets cash flow, technical insights into their training runs, and the option to pull the plug — though he promised he wouldn't. This is the kind of power dynamic that decentralized systems were supposed to eliminate.
I've been in this industry long enough to know that when you see a single entity controlling the resources for a frontier technology, you see a extractive rent. In 2017, I co-hosted "Chain of Thought," a podcast about the ethical implications of smart contracts. Back then, we debated whether ICOs were democratizing access to capital. Now, I'm watching a single contract between two corporations determine the trajectory of artificial general intelligence.
In 2020, I wrote "Why DeFi is a Protest Movement." It was a thread that went viral because I framed liquidity pools not as financial instruments but as social fabrics that rebuild trust after the 2008 crisis. Today, I look at the Anthropic-xAI deal and see the opposite: trust is being centralized, not rebuilt. The code that governs this arrangement is not open-source smart contracts — it's a PDF signed by lawyers. The execution is not automatic — it's subject to personal relationships and geopolitical whims.
The Numbers Don't Lie (But They Distort)
Let's run a back-of-the-napkin valuation. If xAI's Colossus 1 facility generates $150 billion annual revenue from this single contract, what is the facility worth? At a conservative 10x revenue multiple (standard for data center REITs), that's $1.5 trillion in enterprise value. For context, the entire crypto market cap in 2026 is around $2.5 trillion. A single compute contract is worth more than all of DeFi, all of Layer 2, and most of the altcoins combined.
This is not a bullish signal for decentralized technologies. This is a reminder that the real money — the capital that moves global markets — is still pouring into centralized, opaque, single-provider infrastructure. The crypto ecosystem talks about "the great wealth transfer" from boomers to millennials via Bitcoin, but the actual wealth is being created and locked inside colocation facilities owned by billionaires.
And here's the kicker: Anthropic is losing money. Fast. With $150 billion in annual compute costs, they would need to generate at least $200 billion in revenue to break even, assuming 75% gross margins on software. Their actual revenue in 2025 was estimated at around $10 billion. The gap is staggering. They are relying on constant infusions of capital from investors like Google, Amazon, and now effectively xAI (through the lease). This is a Ponzi-like dynamic on a massive scale — growth funded by ever-increasing debt and dilution.
In crypto, we call that "unsustainable tokenomics." In AI, we call it "market leadership."
Contrarian: The Case for Centralization (And Why It's Wrong)
I want to play devil's advocate here, because this is the part where honest analysis often gets uncomfortable.
One could argue that the massive compute concentration is necessary for safety. Smaller players can't afford to safety-test frontier models. Anthropic has a safety-first culture; they pioneered Constitutional AI. Giving them the most compute might lead to the most aligned AGI. Maybe the "trustlessness" ideal is impractical for AGI development — you need trusted entities to oversee training runs and prevent rogue behavior.
But this argument collapses under its own weight.
Trust is no longer a promise; it's a protocol. The very notion that safety requires centralization is a failure of imagination. We already have cryptographic proofs (SNARKs, STARKs) that can verify computations without revealing the data. We have on-chain governance mechanisms that could democratize compute allocation. We have decentralized physical infrastructure networks (DePIN) that tokenize compute resources. None of these are being used by Anthropic or xAI. They chose the traditional, closed, lawyer-driven model because it's faster and easier — not because it's safer.
Code is law, but empathy is the interface. The human impact of this centralization goes beyond market share. When a single facility burns 308 MW, it crowds out green energy projects. When a single contract ties up 220k GPUs, it stifles innovation by independent researchers. When a single relationship (Musk <-> Anthropic) controls access to frontier compute, it creates a bottleneck that can be exploited.
In my 2024 "Ethical Investor" webinar series, I spoke with traditional finance professionals who were excited about AI. They saw it as the next massive asset class. But I warned them: the most valuable asset in AI is not the model — it's the compute. And compute is becoming the new oil: finite, geographically concentrated, and controlled by a few players. The same people who told us to diversify portfolios and avoid single-stock risk are now piling into Nvidia and AI compute funds. The cognitive dissonance is deafening.
Takeaway: The Revolution Must Be Verifiable
So what do we do about it?
We build the pipes that make compute transparent. We demand that every GPU hour used by frontier AI companies be recorded on a public ledger. We create tokenized GPU resources that can be leased by anyone, not just billionaires. We deploy decentralized inference networks that let us verify that a model is actually being run on claimed hardware.
This is my next mission. In 2026, I founded the "Human-Centric Blockchain" initiative to ensure technology serves humanity. We organized a summit in Stockholm that brought together 500 developers and ethicists to discuss preserving human agency in an AI-driven economy. The manifesto we wrote, "The Soul of the Code," argued that blockchain's true value lies in verifying human intent, not just transactions.
Today, I'm extending that mission to AI compute. We need a protocol for compute attestation — a way to prove that a model trained on a given dataset without censorship, without central control, and without a single point of failure.
Trustless systems require trusting relationships. But trusting relationships in a system this big require verifiable, auditable, on-chain proofs.
Musk's confession may have been a moment of humility, but the system he's building is the opposite of humble. It's a cathedral of concentrated power. The only way to counter it is to build the decentralized alternative — not by fighting the cathedral, but by making it obsolete.
We didn't decentralize finance because it was efficient. We decentralized it because it was necessary. The same necessity now stares us in the face for intelligence.
Let's not ignore it.