Hook: A Number Without a Megawatt
The most important number in the Anthropic data center report is 70–80. The least reliable part is everything readers may infer from it.
The report states that Anthropic has signed, or is pursuing, between 70 and 80 letters of intent for data center capacity. It does not disclose total megawatts, committed capital, contract duration, geographic distribution, power source, hardware allocation, or the conversion rate from preliminary agreement to binding lease. It also does not establish whether the capacity is intended for model training, inference, internal safety evaluation, or enterprise deployment.
That distinction matters. A letter of intent is not a cluster. It is not a power delivery date. It is not a purchase order for accelerators. In infrastructure markets, an LOI can reserve negotiating position while both parties test price, permitting, financing, and technical feasibility. The document may never become operational capacity.
Still, the number is a material signal. Anthropic is planning around a future workload that its existing cloud allocation may not comfortably absorb. The company is moving from model development as a software exercise toward model delivery as an industrial system. That transition creates opportunity. It also creates a liability profile that cannot be evaluated from model benchmarks alone.
Verify the proof, ignore the hype. The proof begins with power, contracts, and utilization.
Context: What an LOI Actually Says
Anthropic operates in a market where computation is the principal production input. A large language model requires accelerators for pretraining and post-training. Once deployed, every request consumes additional compute. Training produces a concentrated, scheduled load. Inference produces a distributed, persistent load shaped by geography, customer traffic, latency requirements, and model selection.
The infrastructure decisions are therefore different. Training clusters prioritize high-bandwidth interconnects, synchronized accelerator performance, storage throughput, and uninterrupted availability. Inference networks prioritize proximity to users, capacity elasticity, caching, request routing, and predictable tail latency. A single data center can be excellent for training and inefficient for serving global enterprise traffic.
Seventy to 80 LOIs may indicate a distributed strategy. Anthropic could be evaluating multiple operators, regions, and power markets to create an inference footprint with redundancy. It could also be reserving optionality because no single provider can deliver enough suitable capacity on schedule. These are plausible explanations, not confirmed facts.
The reported source provides little technical detail and appears to rely on limited information. That lowers confidence. The reader cannot responsibly convert the headline into a precise estimate of Anthropic’s compute inventory. A facility with ten megawatts of usable critical load is not equivalent to a facility with ten megawatts of contracted utility capacity. Transformer availability, cooling architecture, rack density, and commissioning dates determine what the customer can actually run.
The difference between reserved capacity and energized capacity is often measured in quarters or years. A power interconnection may be approved while substations, transmission upgrades, and backup systems remain unfinished. A building may be available while liquid cooling retrofits are incomplete. Hardware may be purchased while networking equipment is delayed. The headline compresses all of these dependencies into one integer.
That integer still has strategic meaning. Anthropic appears to be preparing for demand beyond a narrow API product. The likely customer base includes enterprises requiring service-level commitments, regional processing, compliance controls, and private deployment options. Those customers do not buy benchmark scores. They buy availability, auditability, predictable latency, and contractual recourse.
Core: From Model Company to Capacity Buyer
The technical question is not whether Anthropic wants more data centers. Every serious model provider does. The question is whether the proposed capacity will be productive enough to support its financing and operating costs.
A useful starting model separates three workloads: training, inference, and safety operations. Training capacity is expensive but episodic. Inference capacity is expensive and recurring. Safety operations require isolated environments for red teaming, evaluation, monitoring, and incident response. If Anthropic expands inference aggressively, it must maintain enough idle capacity to handle spikes and failures. That idle margin is operationally prudent but financially inefficient.
Suppose, purely as an analytical range, that each LOI represents 10 to 20 megawatts of potential capacity. The aggregate would be 700 to 1,600 megawatts. This is not a reported Anthropic figure. It is a sensitivity band showing why contract details matter. At the lower end, the program is already large. At the upper end, it approaches a regional industrial planning problem involving power procurement, transmission access, cooling water, construction labor, and accelerator logistics.
The estimate also exposes a common analytical error. Megawatts do not equal useful model output. Two sites with identical power envelopes can produce materially different token throughput because of accelerator generation, precision, utilization, networking topology, and software scheduling. A poorly balanced cluster can consume substantial electricity while delivering disappointing effective capacity.
Accelerator supply introduces another constraint. A facility cannot generate revenue until servers arrive, racks are installed, networking is configured, and models are optimized for the target hardware. If Anthropic relies heavily on one supplier, delivery schedules and pricing become strategic risks. If it diversifies across GPUs and custom silicon, software portability and kernel optimization become new risks. The capital decision is therefore coupled to the compiler stack, inference runtime, and model architecture.
This is where the reported number may hide an important shift. The company may be purchasing not only compute, but optionality. Multiple LOIs allow Anthropic to compare energy prices, delivery schedules, tax incentives, and operational terms. They may provide leverage in negotiations with cloud providers. They may also support a financing narrative by demonstrating that the company has access to future infrastructure.
But optionality has a carrying cost. Reserving capacity can create deposits, minimum commitments, cancellation penalties, or take-or-pay obligations. If the agreements are nonbinding, the cost may be limited. If they become binding before customer demand is contracted, Anthropic could acquire a fixed cost base ahead of revenue. Utilization then becomes the central variable.
The same principle applies to inference economics. A model provider earns revenue per request, token, seat, or enterprise contract. The provider pays for compute, power, cooling, network transit, storage, personnel, and cloud overhead. Gross margin improves when accelerators remain busy and software extracts high throughput from each watt. It deteriorates when customers demand low latency and the provider must maintain excess capacity.
Model releases can make this calculation worse. A more capable model may increase demand, but it may also require more computation per response. Long context windows, tool use, multimodal inputs, and agentic workflows increase average request cost. Enterprise customers may value these features while negotiating lower unit prices. Revenue growth alone does not prove that infrastructure expansion is economically sound.
My experience auditing smart contracts in 2017 established a durable rule: stated intent is not executed behavior. I found integer overflow risks in rate calculation functions that automated scanners missed because the surrounding assumptions looked reasonable. Infrastructure announcements have the same failure mode. Management intent is not energized capacity. A pipeline is not production.
A serious assessment would require several measurements. It would track contracted megawatts against delivered megawatts. It would compare accelerator utilization with reserved capacity. It would separate training hours from inference hours. It would measure cost per million output tokens by model and hardware class. It would disclose customer commitments, cancellation terms, and the share of capacity funded directly by Anthropic versus cloud partners.
The new insight is that the 70–80 figure may be less a capacity measurement than a bargaining map. It could reveal how fragmented the supply problem has become. Anthropic may be searching across sites because the constraint is not simply money. It is the joint availability of power, suitable buildings, networking, cooling, and accelerators. In that environment, the company with the largest LOI count is not necessarily the company with the most compute. It may be the company facing the most severe procurement friction.
This also changes the competitive interpretation. OpenAI benefits from Microsoft’s infrastructure relationship. Google controls substantial TPU capacity and the surrounding software ecosystem. Anthropic’s partnerships with major cloud providers provide resources, but dependence can limit scheduling control and margin visibility. A distributed capacity strategy could improve resilience and enterprise reach. It could also increase systems complexity, complicate observability, and create inconsistent performance across regions.
Code is law, but bugs are reality. In AI infrastructure, the equivalent bug is an untested assumption about utilization. A cluster that is technically available but commercially underused remains a balance-sheet liability.
Contrarian: The Security Blind Spot Is Contractual
The standard interpretation is that dozens of LOIs demonstrate confidence in Anthropic’s growth. The contrarian interpretation is that they may demonstrate uncertainty about where growth will materialize.
A company with clear demand, defined deployment regions, and committed customers can often negotiate a smaller number of precise agreements. A large count of preliminary documents may instead reflect broad market testing. Anthropic may be asking many operators to preserve options while customer demand, model economics, and hardware availability remain unsettled.
There is also a security blind spot. Distributed infrastructure expands the attack surface. More sites mean more identity systems, network boundaries, vendors, maintenance personnel, physical access points, telemetry pipelines, and compliance jurisdictions. A security program designed around model behavior cannot compensate for weak infrastructure access controls or poorly segmented management networks.
Data sovereignty adds another layer. Enterprise prompts may contain regulated financial, medical, or proprietary information. Routing requests across jurisdictions can trigger retention, transfer, and audit obligations. Geographic redundancy improves availability, but it can complicate legal residency. Security claims must therefore include the facility, the cloud control plane, the hardware supply chain, and the data deletion process.
Environmental commitments face the same evidentiary test. A company may procure renewable energy certificates while drawing power from a constrained local grid. That accounting may satisfy a reporting framework without resolving transmission congestion or community opposition. “Green” capacity is not a substitute for transparent hourly energy data.
Investors should also resist treating LOIs as direct evidence of imminent financing or valuation expansion. They may support a capital raise, but they may equally increase dilution pressure and debt requirements. Large infrastructure commitments can strengthen a platform when utilization rises. They can accelerate cash burn when demand misses plan. The balance sheet is part of the protocol.
Takeaway: Watch Conversion, Not Headlines
The next reliable signal is not another announcement. It is conversion. Track binding contracts, power delivery milestones, accelerator deployments, inference latency, utilization, and gross margin. Track whether enterprise revenue grows faster than fixed infrastructure expense.
If most LOIs expire, the headline was an option inventory. If they convert into energized, well-utilized capacity, Anthropic will have built a meaningful distribution advantage. The vulnerability forecast is straightforward: the greatest risk is not a shortage of ambition, but a mismatch between contracted infrastructure and paid demand. When will the first audited operating metrics show whether 70–80 was foresight or overcommitment?