Over the past 90 days, while the crypto market has been grinding sideways in a chop that’s testing even the most patient LPs, a different kind of infrastructure war has been quietly brewing. Anthropic, the AI safety darling that emerged from the OpenAI split, has reportedly signed 70-80 letters of intent for data center capacity. At first glance, this is just another tech giant’s CapEx spree—a company that hasn’t broken even yet, betting billions on a future where everyone talks to a Claude. But for those of us who have spent years auditing the trust assumptions of centralized systems, that number screams something else entirely. It’s not just a demand signal; it’s a confession. A confession that the current model of AI infrastructure—rented cloud, opaque supply chains, single points of failure—is fundamentally broken. And I’ve seen this movie before.
Context: The Infrastructure Gap and the Illusion of Scale
Let’s strip away the hype. A letter of intent (LOI) in the data center world is a non-binding, exploratory document. It says, "We’re interested in leasing 10-20 MW of capacity in your facility, contingent on due diligence, pricing, and availability." The number 70-80 is staggering—not because it guarantees 70-80 data centers, but because it signals a level of desperation that only a company with a massive, unfilled pipeline would exhibit. Anthropic’s current cloud infrastructure, likely hosted on AWS and GCP, can’t keep up. Their enterprise clients—banks, hospitals, governments—are demanding private, low-latency deployments that the hyperscalers can’t provide without premium pricing and long lead times. So Anthropic is going direct.
I’ve been in these negotiations. In 2022, during the bear market, I spent six months at ZKSync deep-diving into scalability solutions. I learned that the difference between a successful protocol and a failed one often comes down to infrastructure procurement. The best teams don’t just buy compute; they design for it. Anthropic’s LOIs are a bet that they can outrun the hyperscalers by building their own network. But the math raises more questions than answers. If each LOI averages 15 MW, total capacity is 1,050-1,200 MW—roughly 1.5x the size of a typical hyperscale campus. That’s enough to power a small city. And it’s all going to one company, running one family of models.
Core: The Technical and Ethical Dimensions of Compute Concentration
This is where my blockchain background kicks in. When I audited the first 50 ERC-20 tokens for the Ethereum Foundation in 2017, I discovered that 60% of them had logic flaws that would have allowed founders to drain liquidity. The common thread was a lack of distributed trust—the code assumed a benevolent administrator. Anthropic’s data center strategy is no different. By centralizing compute in a handful of owned or long-leased facilities, they are creating a single point of failure for the entire AI ecosystem. If Anthropic’s leadership decides to prioritize a specific client, or if a disgruntled employee inserts a backdoor, there’s no on-chain audit trail. The models become black boxes hosted in black boxes.
The technical analysis is revealing. Let’s assume each data center is equipped with NVIDIA H100 or B200 GPUs. A 15 MW facility can support roughly 2,000-3,000 H100s (assuming 7-8 kW per GPU with overhead). That’s 140,000-240,000 GPUs across all LOIs. At $30,000 per H100, the hardware alone is $4.2-7.2 billion. Add in construction, power, cooling, and networking, and the total bill could exceed $15 billion. That’s more than the entire market cap of most altcoins. And yet, Anthropic’s revenue in 2024 was estimated at less than $1 billion. The gap between CapEx and revenue is the definition of a speculative bubble.
But here’s the hidden insight: this is not just about training. The LOIs are likely for inference. Training runs are batch jobs; they can be queued and prioritized. Inference is real-time. If Claude becomes the default assistant for millions of users, the latency requirements force geographic distribution. A single data center in Iowa can’t serve a user in Tokyo. Anthropic needs a global network. That’s why 70-80 LOIs—they need presence in every major region. This is a land grab, not a compute grab.
Contrarian: The Case for Optimism—and Why It’s Still Dangerous
The contrarian view is that these LOIs are a net positive for the decentralized compute ecosystem. Here’s the logic: if Anthropic is struggling to secure capacity, it validates the need for distributed, verifiable compute networks. I’ve been working on this since 2022 when I dove into ZK-proofs. The idea is simple: use blockchain to create a marketplace for compute, where AI models can be run on untrusted hardware and still be verified. Protocols like Akash, Render, and Golem are trying to do this, but they’ve lacked the anchor tenant—a major AI company that needs to offload capacity.
Anthropic’s LOIs could be the catalyst. If they oversubscribe and can’t build fast enough, they might turn to decentralized compute to fill the gaps. Imagine a scenario where Anthropic buys 10 LOIs and then leases the remaining 60 from a decentralized network of GPU providers. The economic incentive is clear: decentralized compute is cheaper, more flexible, and doesn’t require long-term commitments. For a company that’s burning cash, that’s a lifeline.
But here’s where the contrarian breaks down. The very nature of AI inference—especially for safety-critical applications like medical diagnosis or autonomous driving—requires SLAs that decentralized networks can’t yet guarantee. If a node goes offline during a surgery, the patient dies. Blockchain’s current scalability and latency constraints make it unsuitable for real-time inference at Anthropic’s scale. The LOIs are a bet on centralized reliability, not decentralized resilience.
Takeaway: The Fork in the Road
We are at a fork. Anthropic’s 70-80 LOIs represent the last gasp of the centralized infrastructure model for AI—the same model that gave us FTX, the same model that gave us the 2017 ICO frauds. The alternative is a trustless, verifiable compute layer that uses blockchain to ensure that every model inference is auditable, every parameter is immutable, and every user’s data is sovereign. I’ve been building toward that vision for a decade, from the Ethereum Foundation audit to the DeFi for Humans workshops to the Agents of Truth campaign.
The question is not whether Anthropic will build these data centers; they will. The question is whether the AI models they run will be accountable to the users they serve. Blockchain provides the only viable framework for that accountability. The next bull market won’t be about DeFi or NFTs; it will be about decentralized AI infrastructure. And the winners will be the ones who understand that compute is the new trust asset.
It’s not obvious to the casual observer, but the numbers we see today—70-80 LOIs, thousands of GPUs, billions of dollars—are the last chapter of the old story. The new story is being written in code, on-chain, one ZK-proof at a time. Based on my experience at the Ethereum Foundation, I know that the first mover advantage is real, but only if you build on decentralized foundations. The market is chopping now, but the signal is clear: position for the decentralized compute narrative. The signal is in the LOIs, but the noise is in the centralized execution. The math doesn’t lie, but the narrative does. I’ve seen this movie before. The ending is always the same: centralized systems fail, and decentralized ones rise. The only question is timing.