In the last 72 hours, a figure has been circulating across crypto and tech media, reportedly pulled from within Anthropic's real estate procurement department: seventy to eighty letters of intent for data center capacity. Crypto Briefing reports it, but as with most supply-side leaks, the primary source is obscure. Whether it appears in official SEC filings remains unknown. However, numbers like this never materialize out of a vacuum; they are either highly coordinated leaks or, less likely, the result of an agency clipping from a private placement memo. As an on-chain data strategist, I don't have access to that internal memo. I do have something better: the industry physics of power and compute. Ignore the AI narrative for a moment. These 70-80 LOIs paint a picture of an entity moving beyond model-centric games, entering a phase of infrastructure and capital expenditure primacy—a phase that everyone in this sector attempted to avoid in 2022 with self-sabotaging narratives.
The Ledger of Power For the past several years, AI training infrastructure has been governed by a simple rule: throw more compute at the problem. The parameters expanded, the GPUs multiplied, but what remained constant was the dependency on cloud partners (AWS, Azure, GCP). For Anthropic, this is about to change, fundamentally. The definitive confirmation of this is the protocol of Letters of Intent (LOIs) to power suppliers, representing more than just grid connection requests.
Let's audit a basic metric used by ME for years: the arbitrary baseline of a hyperscale data center. The average medium-sized deployment accepts around 50 to 100 MW per site, primarily for BF3 and HGX pods. If we take a highly conservative figure, say, 15 MW per letter, we are looking at a cumulative total capacity of at least 1,050 MW. If these are larger clusters, for instance, 100 MW facilities, we're talking about 7 GW to 8 GW. Let's put that number into perspective: 7 GW translates to about 5.6% of estimated global data center energy consumption of borderline industrial. A major firm like AWS typically completed about 20,000 projects per region per year. This order of magnitude isn't a tactical expansion; it's the creation of a separate compute grid, its own dedicated network, completely independent of third-party cloud dependencies.
This type of forecast reveals a tension that I've observed in on-chain analyses of institutional positions. Often, transactions are stored with too much complexity—the ratio of 70-80 different domestic providers or LOIs rather than two or three. The structure creates a piecemeal torture for the analysis: If this were exclusively for training infrastructure, we'd see two large clusters (Texas or Virginia), perhaps 3 to 5 LOIs total. We'd see low-signature capitalization events. The fact that we're looking at 70-80 fits with a philosophy of very high distributed infrastructure for inference or edge locations, replicating the average NFT wash-trading cluster map: up to 20 primary wallets, the rest are collectible. The core wallet is the one never exposed publicly.
"G"-Scale: Overextension or Strategic Depth? The financial implication is straightforward: capital commitment at the gigawatt scale. Given a current market cap for new buildouts (including Opex on top of leases) of at least 50 million per MW, the total price tag for this ambition comes to a radical estimate. Compounding, we're looking at a $0 investment between $50 billion and $85 billion over the next five years. That is slightly beyond valuations the industry has currently assigned to AI companies. The latest round valued them around $180B. That would represent an enormous percentage of the pie to convert to physical assets. In finance, this is disclosed as CapEx pressure.
Here, we see an immediate disconnect in previous market narrative: In Q1 2025, twice a week a trend was presented to the board: 'Anthropic uses Azure/AWS as a service provider; this is the thin layer that allows them to have a zero-capital model.' The buzzer stops. A critical audit is required, and if I had access to their balance sheet ledger, I would look at EbitDA versus expected interest rates deployment schedules.
But, from a register of an energy margin, the rumored letters of intent reveal something else. The politically divisive component is bigger. In these large, renewable purchasing zones in the Sun Belt, global operators are fighting for capacity slots. If those 70+ letters get signed with a specific power source commitment, we can forecast a concentration city... The next phase is for utilities.
Data from Power Jecutive Fourth quarter shows an aggregate of electric utilities reporting massive load expansion, transmission backlogs into 2027, and not enough sites for dispatchable generation. The series effectively bills us to a state that is no longer predicated on GPU brinkmanship.
A Contrarian View on The 'Break Glass' Slope Call me skeptical as a data analyst, an experienced audience is nothing but likely. I admit to reading this news with elevated suspicion. A key & major blind spot made me doubt: the arena of speculative volume. The infrastructure is not the night's model. And signup protocols are not the new 'Token Profit' or 'Instruction Cascade.' They’re commitments, legal commitments on financial liability to cover certain time spans, generally. The probability of 'breakage' for home-built cluster of 70-80 deal is high.
Walker six prior standards of confidentiality: In 2021, another well-funded startup signed the same amount of structures under demands to comply with high chip allocation. The campaign revealed later than 40% of the speculative margins were never used to house computation on the grounds that routing applied. What actually happened is that the cap was used to secure DDL (Deed of Delay) rights against bad power supply. That flows into this insight: Without downstream binding equipment orders (GPU Title Support) or an immediate commitment to distribution safety certifications, the announcements are mostly a data privacy, cost-control mandate, not a 'shovel-ready' power project.
The real problem with the Excess capacity is the stability of the network. In a bearish institutional setting, an AI company will hoard LOIs to signal strength. Yet, Silicon Valley analysis tags this as an 'all-in' bullish signal. I disagree. Precisely distributed skill and valuation: In this traditional world of high layers, the killers try to push lawsuit: 70-80 letters mean they have capacity options. But it's also a clear beta of the option to overburden the tree line. The equipment blow. Solid identities delayed price c.
At base, The Correlation vs. Causation issue appears. Does signing a specialist portfolio to data center improve an artificial network? Causes-deeply: No. Correlation exists: a model R&D arm that over-expands capacity, losing sight of the actual product. An Aggressive, most shadow-infamous connection in the sector was false: capacity = moat. User benefit and model alignment remain the only endpoint. Wipe the floor controlled by efficient volume. It's the halo of securing power, like a 'sip living', rather than distillation above-board. Those genuinely displaced say If training, inference density.
Assessing this phenomenon from a on-chain perspective: The current network should act as compliance warehouses, not as positions that modify based on trending services. The verifiable transactions (Q: 'completed L') are subscription, regular signs launch. additional user expansion Actually concrete commitment terminates each entity site signed.
What I would look for next (Q2 2026)
A dist