Market Clusters & the Site-Scoring Playbook

The previous twelve chapters of Part 3 each isolated a single siting variable and taught it to depth: speed-to-power (3.2), power cost (3.3), energy supply (3.4, 3.5), fiber (3.6), water and climate (3.7), land and geotech (3.8), permitting (3.9), incentives (3.10), community license (3.11), and geopolitics (3.12). In the real world those variables do not arrive one at a time — they arrive bundled, pre-correlated, and named. A market cluster is the bundle: pick Northern Virginia and you have implicitly chosen a permitting regime, a grid operator, a water table, an incentive package, and a community-opposition climate, all at once. The market is the single decision that pre-loads a dozen subsystem constraints — exactly as the workload archetype did in Chapter 1.1, but for the site rather than the machine.

This chapter does two things. First, it runs comparative deep dives across the first-class clusters of 2026 — the US markets (Northern Virginia/PJM, Texas/ERCOT, the secondary tier), the EU (Ireland, the Nordics, Iberia), the Middle East (UAE, Saudi Arabia), and APAC (Japan, Singapore, Malaysia, India) — naming for each one the constraint that actually binds and the workload it is built to serve. Second, it gives you the site-scoring playbook: weighted matrices versus hard pass/fail gates, the stage-gated diligence funnel, the backup-site and portfolio discipline, and the decision artifacts (site-selection memo, board package, assumptions register) that turn the whole exercise into something a board and a lender will fund. A market chosen for the wrong reason strands the same billion-dollar campus on a slab you cannot move.

The reason a market deserves to be the top-of-funnel decision is that clusters are not random collections of attributes — they are self-correlating. A market becomes a cluster because it once optimized some criterion well (cheap stranded power, dense fiber, a generous tax regime, a cold climate), which attracted operators, which attracted the supply chain, the labor pool, the substations, and the second-order economies that make the next build cheaper than a greenfield elsewhere. That same gravity then drives the cluster into its own binding constraint: the thing that made it attractive saturates. Northern Virginia ran out of firm power and entitled land; Dublin and Singapore ran out of grid headroom and imposed moratoria; Texas is racing transmission against a large-load queue that grew ~300% year-over-year. The constraint that bites is therefore predictable from the cluster's maturity, and a disciplined site search starts by naming it.

The strategic consequence is that you do not score markets on a flat list of pros and cons. You ask one question first — what is the binding constraint here, and can my specific workload tolerate it? — and only markets that pass that screen earn a weighted score. A training cluster that checkpoints and curtails can tolerate ERCOT's flexible-interconnection bargain that an always-on inference SLA cannot; an inference fleet bound to a latency budget can tolerate Northern Virginia's power premium that a cost-led batch workload would never pay. The market and the archetype must be matched, not optimized independently — and the most common, most expensive error in the playbook is running one workload's market search for the other's facility.

What follows is a tour of the markets that matter, organized by the constraint that binds each one. The figures are 2026-current and sourced in the keynumbers block below; read each cluster as a worked example of "the bundle" — one named binding constraint, and the workload it is therefore built to serve.

Northern Virginia / PJM remains the largest data center market on earth — roughly 20 GW of operational-plus-pipeline capacity in 2026, with Loudoun County alone hosting 160+ facilities across ~31 million ft², and an estimated 35% of global internet traffic transiting the Loudoun–Prince William–Fairfax corridor. Its old advantages (fiber density, federal-market proximity, low latency) make it the natural home of inference. Its binding constraint is now twofold: firm power and entitled land have gone scarce — vacancy near zero, ~96% of 2026 scheduled supply already committed — and the social license has hardened, with Loudoun ending by-right approval in March 2025 and pushing every project into special-exception hearings (→ 3.11). NoVA is a latency-first market that has run out of the power-first inputs, so a training campus that lands here is paying a metro premium for proximity it does not need.

Texas / ERCOT is the power-first counter-pole and the fastest-growing US cluster, on a path past ~40 GW by 2028 — roughly a third of projected US demand. Its advantages are an independent, build-friendly grid; a 75 MW large-load fast track; a ~2,500-mile 765 kV transmission backbone under construction; and abundant gas and wind for behind-the-meter and hybrid supply. Its binding constraint is the curtailment bargain: SB6 (effective for loads interconnecting after 31 Dec 2025) mandates a remote-disconnect "kill switch" and curtailability, turning expected-curtailment-hours into a line item in the underwriting (→ 3.2). A checkpoint-tolerant training or batch workload absorbs that bargain cheaply; an always-on inference SLA cannot. Texas is also where the large-load queue ballooned ~300% year-over-year, so the queue itself, not the grid's physics, is now the gate.

US secondary markets — Columbus, Indianapolis, Salt Lake City, Reno, San Antonio, the Permian Basin, the Upper Midwest — are absorbing the overflow that NoVA and the coastal primaries can no longer serve. Their advantage is available power and developable land at a discount, often with hungry economic-development incentives; their binding constraint is fiber depth, water, labor, and — increasingly — the mid-project moratorium risk that follows the buildout as opposition organizes (→ 3.11). Secondary markets are the structural winners of the power-first era for training and batch, but they trade primary-market certainty for a thinner ecosystem and a less-tested permitting regime.

Ireland (Dublin) is the cautionary tale of a cluster that hit its constraint hard. Data center electricity demand reached ~22% of national consumption by 2024 (from ~5% in 2015), and CRU imposed a de-facto Dublin-region connection moratorium in 2021. The 2025 reset (CRU decision, with system operators publishing the new process by end-March 2026) reopens the grid — but conditionally: new large energy users must provide on-site generation or storage able to meet full demand, feed energy back via the wholesale market, and source ~80% of annual demand from renewables on a six-year glide path. Dublin is a latency-first FLAP-D metro whose binding constraint is grid-and-policy, and the price of entry is now a self-supply mandate that fundamentally changes the energy-strategy math (→ 3.4, 3.5).

The Nordics (Norway, Sweden, Finland) are the power-first, sustainability-first cluster: ~1.32 GW in 2025 heading to ~1.98 GW by 2031, with average annual temperatures below 10 °C delivering up to ~8,000 hours of free-air cooling, facility PUE as low as ~1.09, Norway's grid ~98% renewable, and district-heating offtake that turns waste heat into a revenue line and a social-license asset (→ 3.7). The binding constraint is latency to the major European demand centers and the depth of grid headroom outside a few nodes — which is why the Nordics win training and async-tolerant workloads and lose real-time inference for Continental users. It is the cleanest example in the world of a power-and-climate-first market.

Iberia (Spain, Portugal) is the rising EU power-first alternative: abundant solar and wind, available land and grid capacity relative to the saturated FLAP-D core, Atlantic subsea-cable landings, and aggressive regional incentives (Aragón, Madrid's periphery, Lisbon). Its binding constraint is water in an increasingly drought-stressed climate (→ 3.7) and the grid-build pace needed to firm an intermittent renewable base. Iberia is positioning as the place EU operators go when they want power-first economics without leaving the Single Market and its data-residency comfort (→ 3.12).

The Gulf — UAE and Saudi Arabia — is the sovereign-AI cluster, where the binding constraint is not power (cheap gas and solar are abundant) but geopolitics. The UAE's Stargate Abu Dhabi campus is planned at up to 5 GW, with a 1 GW first cluster and an initial ~200 MW phase targeted for 2026; Saudi Arabia's HUMAIN is building toward ~1.9 GW by the end of the decade. The decisive variable is US export controls: chip shipments to G42 and HUMAIN are gated by security assurances and government-to-government oversight (the late-2025 license approvals capped each "AI champion" at defined GB300-class volumes). A Gulf site can have firm power energized faster than almost anywhere — and still be throttled by a license denial. Geopolitics is a first-order siting gate here, not a footnote (→ 3.12).

APAC splits along the same power-vs-latency seam. Japan (Tokyo, Osaka) is a latency-first sovereign market with deep fiber and demand but constrained land, high power cost, and seismic diligence overhead (→ 3.8). Singapore is the textbook saturated metro: a 2019 moratorium, partially relaxed via the Green DC Roadmap and the DC-CFA2 call (≥300 MW unlocked, a further ~200 MW reserved for green-energy operators, applications closing March 2026), with the highest wholesale colocation pricing in the world (~$450/kW-month). Malaysia (Johor) is Singapore's overflow valve and now Southeast Asia's fastest-growing hub — pipelines on the order of multiple GW, expected to hold ~60% of Malaysian capacity by 2030 — with binding constraints in water (a live political issue in Johor) and grid pace. India is the large-demand, latency-first growth market with a residency tailwind, where the binding constraints are power reliability, land, and an immature-but-improving grid. APAC is where the single seam — saturated latency-first metro versus power-first overflow neighbor — repeats most cleanly.

Once a market is chosen, you are scoring candidate sites within it — and there are two distinct instruments, which inexperienced teams conflate at their peril. A weighted scoring matrix assigns each criterion a weight, scores each site 1–10, and sums to a ranked total. It is the right tool for trading continuous, fungible attributes — fiber diversity, incentive value, land cost, labor depth — among sites that have already cleared the binding constraints. A hard pass/fail gate is a binary: the site either clears it or is eliminated, and no amount of strength elsewhere buys it back. Speed-to-power-by-a-date, water availability in a stressed basin, export-license eligibility in the Gulf, and zoning that permits the use at all are gates, not scores.

The error that strands campuses is letting a weighted matrix average away a failed gate. A site with magnificent fiber, a generous abatement, and cheap land can post a dazzling matrix total while quietly being eight years from firm power — and the matrix, by summing, hides the fatal flaw behind a strong average. The discipline is sequence: run the gates first, eliminate ruthlessly, and only then score the survivors. The 2026 reordering from Chapter 3.1 is precisely this: speed-to-power moved from a matrix line item to the first gate. The matrix still exists — but only for sites that have already proven they can be energized on your in-service date.

Risk-adjusted scoring is the third refinement. Two sites can post the same weighted total while carrying wildly different variance: one has a firm utility commitment letter and an executed land option; the other has a verbal queue-position assurance and a parcel facing a rezoning fight. Risk-adjustment discounts each score by the probability and consequence of the underlying assumption failing — a queue position that 65–80% of a PJM cohort withdrew from before reaching an interconnection agreement is not worth its face value (→ 3.2). The practical mechanism is to score not just the attribute but the confidence in the attribute, and to carry the contingency explicitly into the assumptions register described below.

Diligence costs money and burns calendar, and you cannot afford to run full diligence on every candidate. The discipline is a stage-gated funnel that spends progressively more on progressively fewer sites, and refuses to advance a site to the next, more expensive stage until it has earned the right at the current one. Three stages structure the spend.

• Desktop screen (cheap, many sites). Public and purchased data only: ISO queue maps and utility load studies, FEMA flood layers, fiber-route databases, zoning maps, incentive statutes, water-stress indices, climate normals. The output is a long list culled to a short list by running the hard gates against desk data. No site visit, no commitment, no spend beyond analyst time and data subscriptions. Most candidates die here — correctly.
• Field diligence (moderate, few sites). Boots on the ground for the short list: geotechnical borings and seismic assessment (→ 3.8), a utility feasibility study and a real queue-position read (→ 3.2), water-rights and discharge-permit confirmation (→ 3.7), an environmental Phase I, a fiber field survey, and early, deniable community soundings (→ 3.11). This stage costs real money — hundreds of thousands per site — and exists to convert desk assumptions into verified facts before any binding commitment.
• Binding diligence (expensive, one or two sites). Executed land options or purchase, a signed utility commitment or large-load study agreement, permit pre-applications filed, the air permit started if BTM generation is in scope (→ 3.9), and the incentive package negotiated to term sheet. This is where the project becomes real and the capital becomes non-refundable. You reach it for the primary and — critically — for at least one backup.

The gate between stages is a kill-or-advance decision, not a formality. A site that surfaces a fatal flaw in field diligence (a seismic fault, a denied queue position, a water basin without rights) is killed before binding spend, and the calendar it consumed is the cheap price of not stranding the binding capital. The funnel's whole purpose is to make the expensive mistakes impossible by making them happen early and cheaply.

A site decision that survives board scrutiny and lender diligence is not a ranked spreadsheet — it is a set of artifacts that pin the decision and its assumptions to the page, signed before binding capital moves. Three artifacts carry the weight.

• Site-selection memo. The narrative of the decision: the market chosen and its binding constraint, the workload-to-market match, the gate results that eliminated the alternatives, the weighted scores of the survivors, the risk-adjusted ranking, and the recommendation with its primary and backup. It is the document that answers "why here, and why not the others" in a form a non-engineer can interrogate.
• Board package. The capital ask framed against the time-value of speed: the in-service date and what it is worth (at ~$10–12B of revenue per GW-year, the cost of six months' delay), the irreversible commitments being authorized, the contingency and the backup, and the go/no-go gates ahead. The board is not approving a site so much as approving a sequence of irreversible bets and the hedges against them.
• Assumptions register. The most under-valued artifact. Every load-bearing assumption — queue position, transformer delivery date, water-rights confirmation, zoning outcome, incentive durability, export-license eligibility — listed with its source, its confidence, its owner, and the trigger that would invalidate it. The register is what makes risk-adjusted scoring auditable and what turns a missed assumption from a surprise into a tracked, owned risk. It is the living document the program manages against for the life of the build.