In the quiet of a Dublin autumn, a whisper of 70 letters of intent echoed across the Atlantic. Each one a promise of silicon and steel, of compute that could reshape the digital landscape. But beneath the excitement, a question lingers: who holds the keys to these new cathedrals of code?
Code is law, but conscience is the compiler.
I first encountered this dissonance in 2017, during the ICO boom. I was a 22-year-old data science student, auditing a protocol called EtherSwap. The code was elegant, but the governance was a trap—whale wallets could bypass consensus with a single transaction. I refused to buy the tokens, instead writing a 4,000-word blog post titled "Code is Not Law if Power is Centralized." That post earned 50,000 views and a reputation I didn't ask for. Now, eight years later, I see the same pattern unfolding, not in a DeFi protocol, but in the infrastructure that powers the AI models we increasingly rely on.
Anthropic—the AI safety company behind Claude—has reportedly signed 70 to 80 letters of intent (LOIs) for data center capacity. The news, first reported by Crypto Briefing, comes from unnamed sources. No details on total megawatts, no confirmation from Anthropic. Just a number: 70-80. In the world of infrastructure, that number is a siren. It signals a shift from renting cloud capacity to building a private empire of compute. For a company that has publicly championed AI alignment and safety, the move raises a question that cuts to the core of the decentralization movement: can you build a safe AI on a centralized foundation?
Context: The Infrastructure Arms Race
Anthropic is not alone. OpenAI, Google, and Microsoft are all racing to secure data center capacity. The demand for GPU compute, especially for training large language models, has outstripped supply. NVIDIA's H100 chips are booked months in advance. Data center construction is booming, with global capacity expected to double by 2027. But the scale of Anthropic's ambition—if the 70-80 LOIs are genuine—is staggering. Assuming each LOI represents a commitment of 10 to 20 megawatts, the total capacity could reach 700 to 1,600 megawatts. That is the equivalent of one to two of the world's largest data centers, or enough to power a small city.
For context, the entire Bitcoin network consumes roughly 150 terawatt-hours per year, or about 17 gigawatts of average power. A single 1,600 MW data center is a significant fraction of that. Anthropic is effectively building a new digital city, designed to run the algorithms that will shape our interactions, our work, and our decisions.
But why letters of intent? LOIs are non-binding. They signal interest, not commitment. They are a way to negotiate favorable terms, to test the market, to create FOMO among data center operators. In the crypto world, we see the same pattern with token purchase agreements or node sale LOIs. The final conversion rate is often around 30-50%. So 70-80 LOIs might translate to 20-40 actual leases. Still, that is a massive bet on the future of AI.
Core: The Architecture of Dependence
Silence in the bear market is where truth compiles.
From my years auditing DAO governance, I've learned that the architecture of power is often invisible. The most dangerous centralization is not in the code, but in the physical infrastructure that runs the code. When we talk about decentralization in blockchain, we focus on consensus mechanisms, validator sets, and token distribution. But we rarely ask: who owns the servers? Who controls the power grid? Who supplies the chips?
Anthropic's LOIs are a stark reminder that the AI industry is repeating the same mistakes as the early internet—building on centralized infrastructure owned by a few corporations. The difference is that this time, the stakes are higher. AI models are not just storing data; they are making decisions, generating content, and increasingly, acting as agents. If the infrastructure is controlled by a single entity, that entity can censor, influence, or even shut down the AI. This is not a theoretical risk. It is the reality of the 2025 landscape.
Let me break down the technical implications. Assume each data center is equipped with NVIDIA H100 GPUs. An H100 has a peak compute of 60 TFLOPS for FP8. To train a model like Claude 3, you need thousands of GPUs running for weeks. The cost is in the tens of millions. But inference—the actual use of the model—is even more demanding. As Anthropic's user base grows, they need low-latency inference across the globe. That requires distributed data centers, close to users. The 70-80 LOIs likely include both training and inference clusters, spread across regions to reduce latency.
But here is the catch: the more dependent we become on Anthropic's infrastructure, the more we trust them with our digital lives. This is not unlike the oracle problem in DeFi. Chainlink, for example, relies on a network of nodes to bring off-chain data on-chain. But those nodes are often centralized in practice, run by a few operators. The same is true for Anthropic's AI. The model may be open-source, but the inference is run on Anthropic's servers. The code is transparent, but the execution is opaque.
We do not build walls, we weave nets of trust.
During DeFi Summer in 2020, I worked with LendFlow, a lending protocol that prided itself on technical efficiency. But when a liquidity scare hit, the community didn't flee because of the code. They stayed because of the trust we built through human connection. I spent hours in AMAs, translating yield farming mechanics into stories of financial sovereignty. That trust became the ultimate security layer. In the same way, Anthropic's LOIs are not just about compute. They are about building a net of trust—or a cage of dependence.
Contrarian: The Pragmatist's Argument
But let me play the devil's advocate. Perhaps the centralized infrastructure is necessary for safety. AI alignment research requires massive compute to run red-teaming experiments, to test for harmful behaviors, to ensure the model follows ethical guidelines. Decentralized compute networks, like Golem or Akash, are still too slow and unreliable for such tasks. The centralized model allows for rapid iteration and control. If Anthropic can build a safe AI that benefits all of humanity, does it matter if the servers are concentrated in a few locations?
Moreover, the LOIs could be a signal of a new business model. By owning the infrastructure, Anthropic can offer better SLAs, lower latency, and more secure private deployments for enterprise clients. This could generate the revenue needed to sustain their research. In the crypto world, we see similar moves with protocols like Ethereum, which is moving toward sharding and rollups to scale. But those are decentralized solutions. Anthropic's approach is the opposite: centralize the hardware, centralize the control.
Governance is not a vote, it is a vigil.
In 2024, I designed a quadratic voting system for CivicChain, a project that merged institutional finance with decentralized identity. The goal was to give smallholders meaningful influence. The system worked, but only because we designed it with vigilant oversight. Governance is not a one-time event; it is a continuous process of watching, questioning, and adjusting. Anthropic's LOIs are a governance decision—a decision to centralize the infrastructure. But who will watch the watchers? Who will ensure that this power is not abused?
Takeaway: The Vigil Begins
In the chaos of summer, we found our winter soul.
The AI winter of 2022 taught me that the market is not the only measure of value. The true value lies in the principles we uphold even when the market is down. The same is true for AI infrastructure. The LOIs may be a sign of growth, but they are also a sign of centralization. We must remain vigilant. We must build alternatives—decentralized compute networks, open-source models, and community-owned data centers. Otherwise, we will be tenants in someone else's digital empire, paying rent for our own thoughts.
I have seen this before. The ICO boom promised democratization, but delivered concentration. The DeFi summer promised trustless finance, but delivered whale domination. Now, the AI infrastructure race promises intelligence, but at the cost of independence. The question is not whether Anthropic will build these data centers. The question is whether we will let them become the sole gatekeepers of the machine.
Code is law, but conscience is the compiler. We must write the code that governs our own infrastructure, or we will be overwritten by others.