Market Clusters & the Site-Scoring Playbook

The earlier chapters of Part 3 each isolated one siting variable: 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 practice those variables arrive bundled, pre-correlated and named. A market cluster is that bundle. Pick Northern Virginia and you have also picked a permitting regime, a grid operator, a water table, an incentive package and a community-opposition climate. The market is the one decision that pre-loads a dozen constraints, as the workload archetype did in Chapter 1.1, but for the site rather than the machine.

The chapter compares 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), APAC (Japan, Singapore, Malaysia, India) and Brazil. For each it names the constraint that binds and the workload the cluster serves. It then builds the site-scoring playbook: weighted matrices versus pass/fail gates, the stage-gated diligence funnel, backup-site and portfolio discipline, and the memo, board package and assumptions register that a board and a lender will fund.

A market deserves to be the top-of-funnel decision because clusters are self-correlating. A market becomes a cluster by doing one thing well: cheap stranded power, dense fiber, a generous tax regime or a cold climate. That attracts operators, then the supply chain, the labor pool and the substations, which make the next build cheaper than a greenfield elsewhere. The same gravity then pushes the cluster into its own binding constraint, because 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 has ballooned. The constraint that bites is predictable from the cluster's maturity, so a disciplined search starts by naming it.

So do not score markets on a flat list of pros and cons. Ask first: what is the binding constraint here, and can my workload tolerate it? Only markets that pass that screen earn a weighted score. Training that checkpoints and can stop on request suits flexible interconnection, provided the job's completion date can absorb the interruptions. Inference can live on a curtailable site if failover and replication elsewhere keep it inside its availability and latency SLO. For cost-led batch work, set the metro premium against the capacity you can actually get, the network you need and how much sooner the site starts earning. Match workload and market before you score parcels, because a mismatch can strand capital.

This tour covers the markets that matter, organized by the constraint that binds each one. Read each as a worked example of the bundle: one binding constraint, and the workload it therefore suits. Figures are dated where they appear, and the key-number block below carries the sources.

Northern Virginia / PJM remains the largest data center market on earth. CBRE counted about 4,040 MW of inventory there at the end of 2025, 37% more than a year earlier. On 1 March 2026 Loudoun County alone counted approximately 233 data center buildings, built or under construction, covering 56.5 million ft². A large share of internet traffic is often said to transit the Loudoun–Prince William–Fairfax corridor, though that share is unverifiable. Fiber density, federal-market proximity and low latency make the region a natural home for latency-sensitive enterprise and edge serving.

Its binding constraint is now twofold. Firm power and entitled land are scarce: vacancy is near zero, and CBRE's 18 March 2026 release put about 96% of 2026 scheduled supply as already committed, an end-2025 position. The social license has hardened too. Loudoun ended by-right approval in March 2025, so uses once allowed by right now need special-exception approval, though qualifying applications accepted before 12 February 2025 proceed under a grandfathering resolution. On 15 September 2026 the Board voted 7-1-1 to pause final votes on legislative data center and substation applications for up to 12 months, with the formal resolution set for 20 October. Administrative and grandfathered applications sit outside the pause (→ 3.11). NoVA is a latency-first market that has run out of power-first inputs. A training campus, or a centralized frontier-reasoning campus whose SLO does not need the metro, pays a premium for proximity it does not use.

The serving utility's own numbers show why a contract is not load. Dominion Energy told PJM on 29 September 2026 that it holds about 54 GW of data-center capacity under contract (electric service agreements plus construction and engineering letters of authorization), against a data-center billing peak of 4.8 GW in August 2026. Its forecast leaves out 19 GW of the engineering-stage letters, 59% of them, and assumes a typical site ramps for five years before drawing 63% of its contracted capacity. In Dominion's forecast, firm load means an executed service agreement or a construction letter; a project that holds only an engineering letter may not appear in it at all. Base a NoVA power date on a signed service agreement with energization milestones, nothing weaker.

The water gate hardened as well. Virginia DEQ's July 2026 Potomac Aquifer study found limited capacity for significant new withdrawals east of I-95, and its modeled 3-MGD evaporative-cooling case failed the permit criteria. Groundwater-cooled designs in eastern Virginia carry a steep permitting burden (→ 3.7).

Texas / ERCOT is the power-first counter-pole and the fastest-growing US cluster. ERCOT's June 2025 forecast put non-crypto data center load at 16.8 GW in 2028 and 25.0 GW in 2030. Both are forecasts of load; neither is connected capacity, and neither promises power to any particular site. Texas has an independent, build-friendly grid and abundant gas and wind for behind-the-meter and hybrid supply. It studies loads of 75 MW and above through a large-load interconnection process, but that threshold is a study category, not a fast track.

The 765 kV backbone is advancing in stages. ERCOT's board approved the Eastern Backbone (about $9.4B, 1,109 miles) in December 2025, but its eastern segments still need PUCT route certificates. Oncor reports PUCT approval on 28 August 2026 of two Permian import lines, Dinosaur–Longshore and Longshore–Drill Hole, with completion planned for 2029. Tie your power date to the specific serving line, its certificate and its energization commitment (→ 3.2).

The binding constraint is the curtailment bargain. SB6 requires serving utilities to set firm-load-shed protocols and install the curtailment equipment before eligible transmission-voltage loads connect after 31 December 2025, with exclusions for specified critical industrial and natural-gas loads. Large loads of 75 MW or more can also be paid, through a separate reliability service, to cut demand in emergencies. Underwrite your site's actual curtailment protocol, service participation and recovery terms, because expected curtailment hours become a line item (→ 3.2). A checkpoint-tolerant training or batch workload absorbs the bargain cheaply. Inference fits only when its availability and latency budgets tolerate the curtailment and the network path.

In August 2026 the large-load gate turned political: the governor froze the process pending a project-by-project audit. On 3 September 2026 ERCOT issued its Batch Zero classifications, all of them conditional. A classification does not mean the study is finished, and it does not give permission to interconnect. New interconnections still wait on the audit, and ERCOT expects its report no later than 10 December 2026. Already-connected sites are unaffected, but any 2027–28 Texas energization date carries a state-audit contingency. State permits are held too: on 21 September 2026 the governor told TCEQ to stop issuing permits to data centers until ERCOT finishes the audit (→ 3.9). Chapter 3.2 has the governing process and the queue figures.

US secondary markets (Columbus, Indianapolis, Salt Lake City, Reno, San Antonio, the Permian Basin, the Upper Midwest) are taking the demand that NoVA and the coastal primaries can no longer serve. They offer available power and developable land at a discount, often with eager economic-development incentives. Their binding constraints are fiber depth, water, labor and, increasingly, the mid-project moratorium risk that follows the buildout as opposition organizes (→ 3.11). They are structural winners for training, batch and centralized frontier inference, but they trade primary-market certainty for a thinner ecosystem and a less-tested permitting regime.

Two states show how fast that regime can change. New York is closed for now. Executive Order 62 (14 July 2026) freezes DEC discretionary applications, pending and future, for covered data centers able to draw at least 50 MW, unless DEC had already deemed the application complete. The hold ends when the Department of Public Service submits the final statewide environmental review (GEIS) report and findings statement, so it has no fixed end date. The PSC followed with a positive declaration and draft scope under case 26-E-0045 on 17 September 2026, and comments close on 20 October. Local approvals sit outside the hold (→ 3.9, 3.11). Pennsylvania's EO 2026-05 (18 August 2026) removed data center projects from PA Permit Fast Track. A project above 25 MW peak demand can sign a project-specific GRID Consent Order and Agreement, which lets DEP review its new applications on a rolling basis. Qualifying permits still issue only once the required local approvals are in hand. Without that agreement, DEP review waits for local approvals and water-withdrawal and discharge authorizations (→ 3.9).

Ireland (Dublin) is the cautionary tale of a cluster that hit its constraint hard. Data centers used about 23% of national electricity in 2025, up from about 22% in 2024 and about 5% in 2015, and CRU imposed a de facto Dublin-region connection moratorium in 2021. In December 2025 CRU opened a conditional route for new data centers to apply, and on 20 May 2026 EirGrid published the documents that put it into practice. Applicants with a maximum import capacity of 10 MVA or more must provide dispatchable on-site or nearby generation or storage that covers that capacity on a derated basis and offer it into the wholesale market. From 1 MVA up, additional Irish renewables must supply at least 80% of annual demand on a six-year glide path, and every applicant must clear a location-specific technical assessment.

That route does not reopen Greater Dublin. EirGrid rates the region fully constrained. Adding conventional generation next to a data center would push short-circuit levels further past their limits, and the network still cannot move large blocks of power. Meeting the generation requirement therefore does not establish deliverable capacity (→ 3.2). 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 changes the energy-strategy math (→ 3.4, 3.5).

Ireland also adds a compliance gate on 27 October 2026. CRU approved Grid Code modification MPID 345 on 22 September 2026, and it applies to existing and new transmission-connected demand facilities. Each facility has to stay connected through specified voltage disturbances and a rate of change of frequency (RoCoF) of ±1 Hz/s. Once the fault clears and voltage is back to 0.9 per unit, it has 500 ms to return to 90% of its pre-fault active power. Compliance is measured at the point of connection.

The rule exists because many data centers cut their grid draw almost instantly during a fault, and EirGrid treats 900 MW as the most demand the system can lose at once before stability is at risk. An existing site that cannot meet the standard has to seek a derogation, which can cap its non-compliant demand at a fixed monthly Demand Utilisation Threshold; EirGrid proposed setting it from the site's highest monthly consumption in 2025. The group derogation routes are open only to facilities connected or connecting before the end of 2026, so a site energizing in 2027 or later should be designed to comply from the first day. Put the site's compliance or derogation terms into its power-capacity gate (→ 3.2).

The Nordics (Norway, Sweden, Finland) are the power-first, sustainability-first cluster. In Finland alone, the grid operator Fingrid says data center projects holding signed connection agreements added up to more than 3 GW of planned capacity by the end of June 2026. That is contracted load, not capacity in operation. Norway's grid is about 98% renewable, the cold climate offers a cooling advantage, and district-heating offtake can turn waste heat into a revenue line and a social-license asset (→ 3.7). Calculate economizer hours and annual PUE from your heat-rejection design and hourly site weather, not from regional averages.

The binding constraint is latency to the major European demand centers, plus grid headroom outside a few nodes. That is why the Nordics win training, batch and centralized reasoning workloads; place latency-sensitive serving there only when the measured end-to-end SLO and replication plan clear the network path. Queue rules are also moving. Under a proposal Finland published in September 2026, demand and storage connection requests would be sorted into four priority groups from 1 January 2027. Earlier requests would be sorted too unless they already hold a connection agreement or a binding conditional one. Where the grid is congested, a project could move up by bringing weather-independent flexible generation or new local production. Record the proposed treatment and your executed-agreement status in the power schedule (→ 3.2).

Iberia (Spain, Portugal) is the rising EU power-first alternative: abundant solar and wind, available land and grid capacity compared with the saturated FLAP-D core, Atlantic subsea-cable landings, and aggressive regional incentives (Aragón, Madrid's periphery, Lisbon). Its binding constraints are water in an increasingly drought-stressed climate (→ 3.7) and the grid-build pace needed to firm an intermittent renewable base. Iberia is selling itself to EU operators as a way to get power-first economics while staying inside the Single Market and its data-residency comfort (→ 3.12).

The Gulf (UAE and Saudi Arabia) is the sovereign-AI cluster, and its binding constraint is geopolitics, not power: cheap gas and solar are abundant. The UAE–US AI campus hosting Stargate Abu Dhabi 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 now targets roughly 3 GW by 2030 (October 2026). The decisive variable is US export control, and chip access now differs by country. For the UAE, a BIS rule effective 10 July 2026 lifted the license requirement for listed recipients, including G42, Core42 and eight US AI companies with their UAE subsidiaries; the G42 and Core42 listing lapses on 6 April 2027 unless BIS extends it. HUMAIN still buys under its November 2025 approval of up to 35,000 GB300-equivalents, because BIS has published no equivalent Saudi rule. A Gulf site can have firm power energized faster than almost anywhere and still be throttled by a license denial or a lapsed listing. Geopolitics is a first-order siting gate here (→ 3.12).

Physical risk is now part of that gate. Drone strikes in March 2026 damaged AWS facilities in the UAE and Bahrain, and in a September 2026 status update AWS said its Bahrain region cannot be restored (→ 11.2). AWS said the damage ran across several availability zones, beyond what its multi-AZ services are built to withstand, so price diversity across regions, not just zones, into any Gulf deployment.

APAC splits along the same power-versus-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). Since March 2025 METI and MIC have run the Watt-Bit public-private council, which plans power and telecom infrastructure for new data centers together. SoftBank's Tomakomai campus in Hokkaido, sized above 300 MW and fed by local renewables, is due to open in fiscal 2026; SoftBank presents it as a counterweight to the clustering of data centers around Tokyo and Osaka (SoftBank, April 2025). Singapore is the textbook saturated metro. It imposed a moratorium in 2019 and partly relaxed it through the 2024 Green DC Roadmap (at least 300 MW of new capacity, plus a potential ~200 MW more for green-energy deployments). The December 2025 DC-CFA2 call then provisionally allocated 200 MW to Digital Realty, Equinix, Keppel DC and STT GDC (about 50 MW each, on Jurong Island) in August 2026, conditional on at least 50% green energy. New Singapore capacity is therefore allocated competitively, not available on request. Rent reflects the scarcity: in CBRE's Q1 2026 survey, wholesale space for 250–500 kW requirements averaged about $403/kW-month in asking rent, the highest of any market CBRE tracks.

Malaysia (Johor) takes Singapore's overflow and is now Southeast Asia's fastest-growing hub. Johor's pipeline runs to multiple GW, and Malaysia's trade ministry (MITI) projected in June 2025 that the state will host about 60% of Malaysian capacity by 2030. Its constraints are water and grid pace, and water is now an explicit approval gate. On 5 January 2026 Johor's state infrastructure chairman said every project goes before a Special Technical Committee, which looks at its water and electricity supply, utility issues and whether the location suits it. The state sets a water usage effectiveness (WUE) threshold of 1.8. Meeting it is not enough if local water is short, and a project that needs cooling water must look at alternatives such as recycled sewage water. Size the design's water draw against the source and volume the state will approve; switching to dry heat rejection or reclaimed water can alter that draw (→ 3.7). India is the large-demand, latency-first growth market with a residency tailwind. Its binding constraints are power reliability, land and a grid that is immature but improving. Visakhapatnam (Vizag) on the east coast is becoming its AI cluster: Google plans about $15B over 2026–2030 for a gigawatt-scale campus there with AdaniConneX and Airtel, with new subsea cables landing beside it (Google, 14 October 2025). Andhra Pradesh's answer to the power constraint is a license. Under G.O. Ms. No. 32 (22 April 2026), an investor with 300 MW or more of data-center load in the state may hold a Deemed Distribution Licence and buy directly from generators; critics argue the order goes beyond the Electricity Act. APAC is where the saturated latency-first metro and its power-first overflow neighbor repeat most cleanly.

Brazil is Latin America's emerging power-first cluster, built on a grid that runs mostly on hydro, wind and solar; its binding constraint is transmission access, not energy. Scala Data Centers plans its AI City campus in Eldorado do Sul (Rio Grande do Sul) at up to 4.75 GW, starting with a 54 MW first phase (Scala, September 2024). In May 2025 the energy ministry accepted the campus's 525 kV connection route in Portaria SNTEP/MME 2.942, but that grants no capacity. The grid operator ONS must still confirm spare system capacity, ANEEL must authorize access, and the recognition lapses if no transmission-use contract is signed by the end of 2033. In the Northeast, TikTok confirmed a campus at the Pecém port in Ceará on 3 December 2025, with Omnia and Casa dos Ventos as partners. Its first phase has 200 MW of IT, new wind farms for supply and operation expected in 2027, and local Indigenous leaders and the consumer group Idec contest its state environmental license.

Brazil's federal Redata regime (Law 15,504, 15 September 2026) suspends federal taxes on data-center equipment. To qualify, a site must source all its electricity from renewable or low-emission supply and commit at least 10% of its capacity to the Brazilian market (8% in the North, Northeast and Center-West), or direct that share to research institutions or the public sector, or invest an extra 10% of the benefited purchases in research and development instead. The government still has to issue the implementing rules (→ 3.10). Underwrite Brazil for training and batch work, and treat a ministry connection recognition as a route to the grid, not a power date.

Once you have chosen a market, you score candidate sites within it, and teams routinely conflate two instruments. A weighted scoring matrix gives each criterion a weight, scores each site 1–10 and sums to a ranked total. It suits continuous, fungible attributes (fiber diversity, incentive value, land cost, labor depth) among sites that have already cleared the binding constraints. A mandatory gate gives one of three results: PASS, FAIL or UNKNOWN. A FAIL rules the site out for the design you have now. UNKNOWN holds it until the specified evidence arrives, and it cannot be ranked as eligible. A site enters the weighted ranking only when every mandatory gate reads PASS. Strength on other criteria never buys back a gate. Speed-to-power by a date, lawful water for the chosen cooling design, 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 total while sitting eight years from firm power, and the sum hides the flaw. The defense is sequence: run the gates first, reject the failures, hold the unknowns, and only then score the eligible sites. That is the 2026 reordering from Chapter 3.1: speed-to-power moved from a matrix line item to the first gate. The matrix still exists, but only for sites that have 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 with very 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 its assumption failing. A queue position without an executed service agreement and a dated milestone schedule is not worth its face value, and generator-queue attrition rates do not substitute for the host utility's own large-load record (→ 3.2). In practice, score the attribute and also your confidence in it, and carry the contingency into the assumptions register described below.

Diligence costs money and calendar time, so you cannot run full diligence on every candidate. A stage-gated funnel spends more on fewer sites and lets a site advance only after it has earned the next stage 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 and climate normals. Running the hard gates on desk data culls a long list to a short one. There is no site visit and no commitment 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 hundreds of thousands of dollars per site. It converts desk assumptions into verified facts before any irreversible land purchase or construction commitment. If diligence needs a conditional land option, an access agreement or a utility study agreement sooner, sign it early and log its fee, expiry date and exit terms.
• 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. Here the project becomes real and the money stops being refundable, so release capital only after the power, network, cooling, land and approval gates have passed. Reach this stage 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 supply that cannot meet its cooling design) is killed before binding spend. The calendar it consumed is the cheap price of not stranding binding capital.

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 answers "why here, and why not the others" in a form a non-engineer can interrogate.
• Board package. The capital ask, set against what speed is worth. Show the in-service date and what a six-month slip costs in project cash flow, worked from the contracted capacity ramp with the Chapter 1.8 valuation method. For scale, renting a fully contracted 200 MW phase as AI cloud brings in over a billion dollars in gross revenue every six months (Chapter 2.2). Then list the irreversible commitments the board is authorizing, the contingency and backup, and the go/no-go gates still ahead. The board is not approving a site so much as a sequence of irreversible bets and the hedges against them.
• Assumptions register. The most under-valued artifact. List every load-bearing assumption (queue position, transformer delivery date, water-rights confirmation, zoning outcome, incentive durability, export-license eligibility) with its source, its confidence, its owner and the trigger that would invalidate it. Name the operative document behind each gate, such as ERCOT’s 3 September 2026 market notice on conditional Batch Zero classifications, the CRU's data-center connection decision or the serving utility's application form; a link is not a capacity allocation. The register makes risk-adjusted scoring auditable and turns a missed assumption from a surprise into a tracked, owned risk. The program manages against it for the life of the build.

Advance C, reject A and hold B under the assumed evidence. A fails on water, whatever its diagnostic score. B's carrier maps do not show ducts, entrances, rights, shared-risk segments or the service that survives a failure, so its route diversity is UNKNOWN and its score cannot enter the ranking. Only C is eligible, at 10 × (0.30×7 + 0.20×8 + 0.15×7 + 0.15×7 + 0.10×8 + 0.10×7) = 73 out of 100. A score-first ranking would have put A on top at 90, a parcel that cannot get its water.

Assign the power lead the executed service boundary, the phased capacity and date, and the curtailment conditions. Assign the network lead the service schedule, route rights, shared risks and failure test; the cooling lead the water quantity, restriction priority and discharge terms; and counsel the land and approval conditions. Conditional rights can buy access for diligence. Irreversible purchase waits for the gates.

The decision flips to B only if its route evidence passes the same survivor-capacity and RTT gate, with other assumptions unchanged; B's 80 then beats C's 73. If B fails a shared-duct test, it becomes FAIL instead, and raising the connectivity weight cannot cure that. An intake form such as the ESB Networks pre-application form starts the utility conversation; it is not service approval.