Research Dossier · AI Infrastructure · September 2026
The Capacity Test
An AI-infrastructure claim is not proved by a model launch, a construction announcement, or a large backlog alone. It becomes investable only when four separate systems line up: customers want the capacity, the physical system can deliver it, the financing can sustain it, and the resulting economics convert into cash. This dossier is a way to test each link without confusing one for another.
The four-link chain
| Link | What it establishes | Primary evidence | What would weaken it |
|---|---|---|---|
| 1. Demand Someone needs capacity. | Customers have a reason and a budget to consume compute, storage, networking, or power. | Customer capex guidance, contracted demand, utilization disclosures, and renewal behavior. | Reduced budget plans, delayed deployments, weaker consumption, or a lower willingness to pay. |
| 2. Delivery Capacity can arrive. | Land, interconnection, generation, equipment, and construction can become usable capacity on time. | Interconnection queues, energization milestones, permitting, equipment lead times, and utility filings. | Power delays, equipment shortages, cost overruns, or capacity that is announced but not energized. |
| 3. Capital The build can be financed. | The sponsor can fund the build without turning a growth plan into an unmanageable balance-sheet claim. | 10-Ks, 10-Qs, debt maturities, lease obligations, project finance terms, and free cash flow. | Funding needs that outrun operating cash, more expensive refinancing, or recurring equity dilution. |
| 4. Cash conversion Capacity earns an economic return. | Revenue, gross profit, working capital, and maintenance needs leave a credible path to cash. | Unit economics, margins, backlog conversion, receivables, maintenance capex, and cash flow statements. | Revenue that grows without margins, poor collection, weak utilization, or capital intensity that absorbs the return. |
The important distinction: each link is necessary, and none is sufficient. A company can have spectacular customer interest with a delayed power connection. It can have an energized facility with a fragile capital structure. It can report strong revenue from capacity built in an earlier cycle while the next cycle is already becoming less attractive. The work is to locate the weak link before the headline does it for you.
Start with the physical constraint, not the market narrative
AI infrastructure is a chain of real assets and contracts. The relevant question is usually more specific than “Is AI demand strong?”: which buyer is committing funds; which facility can be energized; which vendor has lead-time exposure; which financing structure absorbs the construction period; and which revenue stream remains after the facility is live? The answer may be different for a cloud platform, a generator owner, a cooling supplier, a memory manufacturer, and a highly levered capacity developer.
That is why a theme should be researched as a set of constraints, not as a collection of tickers. A single broad demand story may support several businesses while leaving their economics sharply different. A supplier with paid orders can have a stronger near-term record than a developer with a large stated pipeline. A power asset with contracted demand can still face regulatory, interconnection, or renewal risk. The causal chain makes those differences legible.
What good evidence looks like
A credible research file has a source hierarchy. The first read is the company’s own filing and investor presentation, not a chart reposted without context. The next read is the system around the company: a grid operator’s auction report, a regulator’s docket, a customer’s capex disclosure, or a supplier’s earnings release. Only then does commentary become useful—as a way to identify questions, not as a substitute for the record.
Read the buyer
Trace the claimed demand to a customer budget, contracted obligation, or observed use. A supplier’s backlog is evidence about the supplier; it is not automatically evidence about end-market permanence.
Read the system
Check the power, interconnection, equipment, and construction path. A data center is not productive capacity until the system around it is ready too.
Read the statements
Follow the claim through revenue recognition, margins, working capital, interest expense, maintenance needs, and free cash flow. This is where a narrative becomes an underwriting exercise.
A practical monitoring board
The output should be a short list of observable tests, not an endless stream of updates. For demand, watch customer capital plans, capacity utilization, and pricing. For delivery, watch energized megawatts, equipment lead times, and project milestones. For capital, watch net leverage, maturity walls, and equity needs. For cash conversion, watch backlog conversion, gross margin, receivables, and the relationship between earnings and operating cash flow.
Each test should answer one pre-written question. “Did the company beat estimates?” is usually too broad. “Did management maintain the customer-funded portion of the build plan?” or “Did energized capacity rise at the stated pace without an increase in financing needs?” gives the evidence somewhere concrete to land. A good monitoring board makes it harder to move the goalposts after a result arrives.
Source trail
The following records are useful starting points for the physical and financial portions of this work:
- SEC EDGAR for company filings, including 10-Ks and 10-Qs.
- PJM’s 2026/2027 Base Residual Auction Report for capacity-market conditions and clearing results.
- FERC’s reliability technical-conference record for the regulatory questions around large loads and system reliability.
- NERC’s Large Loads Task Force white paper for the operating risks created by emerging large loads.
The conclusion is a conditional one
Infrastructure themes reward precision because the story can be directionally right while an individual investment is wrong. A demand forecast can be real but overpaid for. A power constraint can be durable but show up in a different asset than expected. A backlog can be valuable but not cash-generative. The purpose of the Capacity Test is not to make the conclusion neat. It is to make the uncertainty explicit, the evidence proportional, and the next reason to change the view visible in advance.