Study · AI Infrastructure · July 2026

Data Center Alley

By Aydin Ali · July 18, 2026 · ~7 min read

Loudoun County, Virginia — twenty-some minutes from where I live — is the densest concentration of data centers on Earth. This isn't an abstract theme I read about; it's the physical corridor several of my actual holdings are built to serve. This is a map of that corridor, connected to the public companies that profit across its stack, and an honest look at what that proximity does and doesn't tell me about my own portfolio's concentration risk.

The corridor

Loudoun County hosts the highest concentration of data-center capacity in the world, a position that traces back to Ashburn's role as an early internet interconnection point and has compounded for two decades since. Estimates of the share of global internet traffic that passes through the county vary widely by source and methodology — from roughly 35% to as high as 70% — and I'd treat any single precise figure with real skepticism; what's consistent across sources is that it's the largest concentration by a wide margin, not a close contest. As of 2024, Loudoun had 71 operating data centers, the most of any locality in Virginia, and data-center revenue funds roughly 40% of the county's budget.

How Northern Virginia compares

"The largest concentration in the world" is a claim worth putting a number next to. As of early 2026, Northern Virginia carries roughly 4,040 MW of built data-center inventory — nearly three and a half times more capacity than every secondary U.S. market combined — and 2025 net absorption (new space actually leased) rose 144% year-over-year while vacancy fell to just 0.5%, the tightest of any primary U.S. market. It is not a close contest.

The gap is real but not necessarily permanent: several forecasts expect the Southeastern U.S. (Georgia, the Carolinas, Texas) to become the largest regional market by power capacity by 2030, precisely because Northern Virginia's own grid constraint — covered next — is pushing new development toward markets with more available power. AEA holds no direct Atlanta- or Texas-specific data-center real estate exposure; this is a corridor study, not a claim that Northern Virginia's lead is durable forever.

The binding constraint isn't chips — it's power

The framing that's taken hold across the neocloud and power names I track directly is that the AI buildout is "starting to look less like a scramble for GPUs and more like a fight over electricity." Northern Virginia is the clearest real-world evidence of that: JLL's 2026 Global Data Center Outlook cites an average wait of roughly seven years for a new 100 MW grid connection in the region, and Dominion Energy's own interconnection queue for large commercial load stretches beyond 36 months for new substation service. That constraint is why new development is increasingly sited 20–40 miles outside the traditional Ashburn cluster — not because land in Ashburn ran out, but because the grid did.

Who captures the economics

The corridor's value chain runs from raw land and power, through the physical building and its cooling, to the compute inside it, to the company that actually sells that compute to an AI workload. I hold real positions at three of those four layers.

TLN and VST aren't sited in Northern Virginia specifically — Talen's Susquehanna nuclear plant serves AWS datacenters under a long-term PPA, and Vistra's generation fleet spans multiple regions — but they're the same structural bet: merchant power producers converting AI-driven demand growth into long-duration, contracted cash flow, the exact dynamic playing out in Loudoun's own interconnection queue. VRT builds the power and liquid-cooling infrastructure that shows up inside the buildings themselves, in this corridor and others.

Layer 5 — hyperscale demand — is the layer actually paying for all four layers below it. The AI Capex Reality Check is the demand-side companion to this piece: a look at whether GOOGL and META's own cash flow is keeping pace with the capex funding this exact buildout.

The other side of the ledger: water, noise, and local pushback

A corridor study that only covers capacity and tickers would be incomplete, and possibly dishonest by omission. Loudoun's data centers are genuinely controversial with the people who live next to them, for reasons that show up in local government meetings, not just op-eds. Loudoun data centers used an estimated 1.6 billion gallons of water in 2023 alone — nearly 10% of the county's total water use — and the county faced mandatory water-use restrictions during 2024 drought conditions. A new Virginia law taking effect in 2027 will require aggregated public reporting of data-center water usage for the first time; individual-facility figures remain protected as trade secrets today, which is itself part of what residents are pushing back on.

Beyond water, the recurring complaints at county meetings cluster around five things: cooling-tower and backup-generator noise audible at adjacent property lines, rising electricity rates tied to grid expansion built to serve data-center load, construction traffic, and visual/lighting impact on rural and historic land. None of this shows up in a value-chain diagram of tickers, and all of it is a real input into how much friction — permitting delays, local moratoria, rate-case fights — the next decade of buildout actually faces. A corridor that's this economically important to Loudoun County (data-center revenue funds roughly 40% of the county budget) has real countervailing pressure to keep permitting it moving, but "the money wins eventually" is an assumption, not a fact I've verified.

The honest risk

This is the section a purely promotional version of this piece would skip. My own book carries 60.93% thematic concentration in AI infrastructure, compute, and power — a figure computed live from Holdings weights on the Thematic Exposure page, not asserted once and forgotten. Living twenty minutes from the physical center of this buildout makes the theme easier to research in depth; it does not make the theme less concentrated. If hyperscaler AI capex decelerates — the same risk argued in the Physical Limits of Compute whitepaper — multiple layers of this value chain would likely reprice together, not independently, because they're the same bet expressed five different ways. A local research edge is a real asset for underwriting the thesis. It is not a hedge against the thesis being wrong.

Methodology & sources