Hook
Anthropic reportedly approached 70 to 80 data center operators with letters of intent for future capacity. That number is the headline. It is also the least useful number in the report.
A letter of intent does not deliver a megawatt. It does not reserve a transformer, secure a grid connection, allocate a single accelerator, or create a paying customer. It records interest during negotiation. The distinction matters because the artificial intelligence industry now converts preliminary paperwork into evidence of inevitable expansion.
The report provides no aggregate capacity figure. It identifies no locations, delivery dates, operators, chip suppliers, power contracts, or financial commitments. It also does not establish whether the documents came from Anthropic, its cloud partners, or intermediaries marketing future capacity. The source, Crypto Briefing, is better known for cryptocurrency coverage than for primary infrastructure reporting. That does not make the claim false. It makes verification mandatory.
The ledger does not lie, but the narrative does. In this case, the ledger is still missing.
Context
Anthropic operates in a market where model quality is only one part of the product. Customers purchase predictable access, controlled latency, regional availability, privacy assurances, and contractual service levels. Every one of those promises eventually reaches physical infrastructure.
Training requires concentrated clusters with high bandwidth between accelerators. Inference requires something different. It requires capacity distributed across regions, enough reserve for demand spikes, and network paths that keep response times within a commercial threshold. A company serving global enterprises may therefore negotiate many facilities without intending to build 70 or 80 giant campuses.
The same documents could represent several strategies. Anthropic might be seeking inference capacity for its Claude application programming interface. It might be preparing dedicated environments for large customers in regulated industries. It might be comparing locations before selecting a smaller group. It might simply be using competing offers to improve pricing from cloud and colocation providers.
These interpretations have different consequences. A distributed inference network could improve availability while increasing the operational burden of moving models, synchronizing weights, managing security boundaries, and maintaining hardware across jurisdictions. A training program would imply concentrated capital expenditure and long procurement cycles. Private enterprise deployments would require stronger isolation and support commitments, but could attach revenue to each installation.
The report answers none of these questions. It supplies a quantity of intentions and invites the reader to supply the infrastructure.
Core Analysis
The first audit question is capacity. Suppose each letter concerned 10 megawatts. The aggregate would be 700 to 800 megawatts. Suppose the average was 20 megawatts. The range would rise to 1.4 to 1.6 gigawatts. Those calculations demonstrate scale, not reality. Without the actual term sheets, the average could be far lower, and some documents could concern options rather than firm reservations.
A power number is also incomplete. Data center operators distinguish between utility capacity, critical load, and usable information technology load. A facility may advertise a large grid connection while delivering materially less power to servers after cooling, backup systems, and electrical losses. The relevant measure is not the maximum connection. It is the accelerator load available on a specified date.
The next question is hardware. A megawatt of general data center capacity cannot automatically run an artificial intelligence workload. High-density accelerator deployments require specialized cooling, networking, power distribution, and floor design. Liquid cooling may be necessary for newer systems. High-speed interconnects determine whether a cluster can train efficiently. A facility suitable for conventional cloud workloads may be unsuitable for a tightly coupled training run.
This creates a hidden conversion problem. The reported LOIs may indicate available buildings, but not artificial intelligence-ready capacity. They may indicate power, but not NVIDIA, AMD, or custom accelerator allocation. They may indicate land and permitting, but not the networking fabric required to turn separate rooms into a functioning computational system. Source code is the only truth that compiles. Infrastructure claims require an equivalent operational record.
The commercial test is more severe. Anthropic must pay for reserved capacity before every megawatt produces revenue. Long-term commitments can lower unit costs and improve supply certainty, but they also create fixed obligations. If customer demand grows more slowly than expected, the company carries idle capacity, cancellation penalties, or expensive cloud substitutions.
This is where an LOI count can mislead investors. A large negotiation pipeline may be rational procurement. It may also be a signaling device ahead of financing. Infrastructure ambition strengthens a growth narrative, and growth narratives support higher private valuations. But valuation is not cash flow. If the agreements are nonbinding, they cannot be treated as contracted revenue or debt collateral without additional documentation.
The financing structure deserves equal attention. Anthropic could fund capacity through equity, strategic investment, cloud credits, equipment financing, or customer-backed commitments. Each route changes the risk profile. Equity extends the runway but dilutes existing holders. Debt preserves ownership but adds fixed payments. Cloud credits reduce early cash expense while preserving dependence on the provider. Customer-backed construction lowers demand risk only if the customer has signed a binding, enforceable contract.
My audit work on the proposed spot Bitcoin exchange-traded products taught me to examine the boring layer first. Custody architecture, key-management latency, withdrawal procedures, and responsibility boundaries mattered more than the product narrative. The same method applies here. The central question is not whether Anthropic wants more compute. Every serious model company does. The question is who bears the cost when planned demand fails to arrive.
Geography introduces another variable. A global network may reduce user latency and jurisdictional concentration, but it increases exposure to data sovereignty rules, export controls, local energy markets, water restrictions, and physical security threats. A model serving European customers from one region may face different requirements from a model serving financial institutions in the United States or healthcare providers elsewhere.
The company’s safety identity also creates an infrastructure obligation. Anthropic cannot describe safety as a model property while treating the surrounding operational system as an afterthought. Isolated evaluation clusters, access logging, incident response, model version control, and secure administrator pathways consume capacity and money. A safe system must be auditable at the machine boundary, not only persuasive in a policy document.
There is a further accounting issue. Inference capacity is not simply proportional to model size. Quantization, caching, batching, context length, response speed, and utilization determine the economics of each request. A more capable model can require more hardware, but software efficiency can reduce the required fleet. Conversely, enterprise users demanding long contexts and low latency can consume capacity far faster than consumer traffic.
That produces a new insight absent from the original report: the meaningful metric is not the number of facilities under discussion, but reserved compute multiplied by contracted utilization. A 100-megawatt fleet with no minimum usage may be less valuable than a 30-megawatt fleet backed by customers paying for guaranteed availability. Investors should seek accelerator-hours, utilization floors, gross margin per inference unit, and cancellation terms. Those metrics connect infrastructure to business performance.
The competitive implications remain substantial. OpenAI benefits from Microsoft’s infrastructure relationship. Google controls specialized tensor processing hardware and large internal facilities. Anthropic can reduce its disadvantage through strategic capacity reservations, but reservations do not create a proprietary supply chain. If the company remains dependent on external operators and cloud vendors, its bargaining power may improve without becoming independent.
Silence in the data is a confession, though not necessarily of misconduct. Here it signals incomplete reporting. Until Anthropic identifies signed partners, power quantities, deployment schedules, hardware commitments, and financing sources, the market is measuring communication intensity rather than installed capability.
Contrarian Angle
The bullish interpretation is not irrational. Anthropic may be preparing for a real surge in enterprise inference. Large customers often reject unreliable service, and capacity shortages can block sales even when a model performs well. Negotiating with many operators may be prudent risk management. Geographic distribution could improve resilience, lower latency, and reduce dependence on one cloud provider. Early reservations may also secure scarce power before competitors consume it.
The company may therefore be doing exactly what a growing infrastructure business should do. The problem is not ambition. The problem is measurement. Bulls are correct that compute availability has become a strategic asset. They are wrong only when a preliminary document is treated as proof that the asset exists.
My post-Merge infrastructure verification reached the same conclusion in a different system. A transition can function publicly while client-level delays and configuration mismatches remain hidden. Operational success is a measured condition, not a press release. Merges change the mechanics, not the incentives. The same incentives now reward Anthropic for presenting optionality as momentum.
Takeaway
Anthropic’s reported 70 to 80 letters of intent should be filed as a procurement signal, not a capacity balance. The next evidence must be contractual. Watch delivered megawatts, accelerator deployment, utilization, customer commitments, financing terms, and service performance.
Volatility is the tax on unverified consensus. The gap between promise and proof is fatal when fixed infrastructure costs outrun recurring revenue. Anthropic will not be judged by how many facilities it can discuss. It will be judged by whether those facilities become efficient, auditable machines that customers are willing to fund.
