Site Selection Strategy & the Reordered Criteria Hierarchy

Site selection used to balance a familiar set of criteria: latency to population centers and peering hubs, land that was cheap, flat and zoned, fiber routes and carrier diversity, fiscal incentives such as sales-tax exemptions and property-tax abatements, and power. Northern Virginia, the largest data center cluster on earth, grew on dense fiber, proximity to the federal market, generous Virginia incentives and an established utility base. Even there, power was holding back new projects before the generative-AI boom began (CBRE, H2 2022).

At AI scale, megawatts energized by a date turned into a binding input in any market where grid delivery would trail the workload's need date. Chips, land and fiber still have their own pass/fail conditions. US generator interconnection queues hold more than 2,000 GW of requested capacity (LBNL Queued Up: over 2,060 GW active at end-2025, down from about 2,290 GW at end-2024). On the load side, ERCOT's large-load request queue has multiplied since end-2024; Chapter 3.2 tracks it. In PJM, RMI finds that the average time from interconnection application to commercial operation for new generation rose from under two years in 2008 to more than eight years in 2025. Where firm power is the scarcest, slowest input, investigate it early so you do not spend on sites that cannot energize. Workload connectivity, water and lawful land use remain mandatory gates, and only the sites that clear them reach the weighted comparison.

This chapter sets out the constrained search. It covers early screening of the longest unresolved dependency, how the workload archetype from Chapter 1.1 drives the search, the single-campus versus distributed-fabric fork, the four build modalities and what each buys, and the site-selection funnel of mandatory pass, fail and unknown tests, with diligence sequenced by cost and the binding need date. A slab cannot be moved once it is poured, so an ordering mistake here compounds for the life of the asset.

The reordering changes what gates the deal, not just the weights. Power was always mandatory. What AI scale changed is how early you must prove it will arrive by the need date. A site must now clear its power, network, cooling, land and legal requirements before any scored advantage counts. Speed-to-power is a mandatory gate, and when grid delivery controls the need date its evidence should be gathered early. A site that cannot show a credible path to firm capacity by your in-service date is eliminated before its fiber, land or tax score is computed.

The break is sharp because AI campus load is large, dense and growing faster than transmission can be built. A single hyperscale AI building now draws 100+ MW, and a gigawatt campus draws as much as a mid-sized city. The grid cannot absorb that on the old timeline, so firm capacity by your need date now comes through a multi-year grid process that many projects drop out of. That process, not construction, is the long pole. Mastt's April 2026 practitioner estimate is 18–24 months for greenfield hyperscale-campus construction, while the grid path behind it takes 3 to 7+ years, and longer in the densest hubs (→ Chapter 3.2). That gap defined 2024–2026 siting. It is why 'bring-your-own-power' (behind-the-meter gas, batteries, fuel cells, eventually SMRs) entered first-power comparisons where fuel, permits and availability can narrow the gap. By mid-2026 Cleanview's tracker showed roughly 90 GW of US behind-the-meter generation announced and about 2 GW running.

Speed-to-power is the gate because speed is worth more than almost anything else the site decision controls. At SemiAnalysis's contested June 2026 estimate of $12–13B a year per GW of contracted critical IT, a fully monetized 200 MW phase that enters service six months early brings about $1.2–1.3B of gross revenue forward (Chapter 2.2 works the example and its limits). Meanwhile the GPU fleet starts depreciating against a contested 2–3-year bear-case economic clock the moment it is racked. So rank sites on contracted margin, ramp, bridge costs and remaining obligations. A site with cheaper power and a later service date still wins if its discounted cash flows come out ahead.

Workload then shapes how you meet the gate. Training and batch inference can accept bounded interruptions in exchange for an earlier or larger service tranche, if their checkpoint, restart and deadline budgets fit the curtailment terms on offer. A synchronous training job already checkpoints and resumes, so it can tolerate non-firm or curtailable interconnections, bridge power and phased energization. Online inference is less forgiving. Unserved load breaks its serving SLO and costs revenue, so it can take a curtailable site only if backup supply or surviving replicas keep the SLO intact. That is why it pays the proximity-and-firmness premium. Some utilities and ISOs will offer an earlier or larger connection to a load that agrees to curtail during system stress. The Duke Nicholas Institute (February 2025) estimates about 98 GW of headroom across 22 US balancing authorities for new load willing to be curtailed for 0.5% of its maximum potential annual energy use, before any local network constraint. Test the project's actual utility terms and operational response against the mechanics in Chapter 3.2, because a workload label establishes neither flexibility nor an earlier connection. The energy-supply strategy that backs the gate is in Chapter 3.4 and Chapter 3.5.

After the workload sets the funnel, the next fork is concentration versus distribution. Do you fight for one enormous interconnection at a single gigawatt campus, or spread the same megawatts across several smaller, faster-to-energize sites joined by long-haul fiber? The choice sets the network architecture, the redundancy posture and which workloads the facility can run at all.

A single campus is the natural home of frontier pre-training. A tightly coupled supercomputer wants its GPUs in one place because synchronous all-reduce collectives run on every training step. Test any cross-site design against step time, collective bytes, topology, available bandwidth and added RTT, and reject it when measured goodput misses the target. Concentration also wins on cost: one substation, one fiber build, one security perimeter, one shared cooling and power yard. The price is concentration risk. You bet the whole campus on one interconnection fight, one grid node's reliability and one permitting jurisdiction's goodwill, and if that node slips two years, the entire investment slips with it.

A distributed fabric spreads the bet: several 100–300 MW sites, each with its own faster interconnection, aggregated to gigawatt scale. Several medium interconnections in less-constrained markets can energize faster in aggregate than one mega-interconnection in a saturated hub, and the fabric diversifies jurisdiction, grid and weather risk. But the workload must tolerate the inter-site latency. Inference geo-distributes naturally because requests are independent, and training schemes that are asynchronous or hierarchical (multi-data-center) can tolerate bounded staleness between sites. Run a strictly synchronous pre-training job across the fabric and you pay a goodput tax on every step that can erase the speed advantage you bought. Distribution buys power and speed, and a workload can cash that in only if its inter-site communication and goodput budgets pass.

The third fork is how you take the site to operating capacity. It is the lever that most directly trades speed against control and capital. The procurement archetypes from Chapter 1.6 map onto siting as four delivery modalities, each at a different point on the speed-control-capital surface.

Greenfield self-develop means buying raw land and originating the whole stack yourself: interconnection, substation, shell, power, cooling and fit-out. You control density, voltage architecture and cooling modality outright, and you can spread owned fixed costs across a large build. Whether that adds up to a cost advantage is settled by the matched cash-flow comparison in Chapter 1.8. The price is the longest, queue-gated schedule to a live cluster, the deepest capital commitment, and ownership of every permitting and interconnection risk. It is the right call only for a durable, well-forecast workload you intend to operate at scale for years. Build-to-suit (BTS) hands development and construction risk to a developer who builds to your spec on a long (often 15-year) lease. You get a purpose-built liquid-ready hall without fronting the land-and-shell capital, but you inherit the developer's site, including its interconnection position, and you pay a credit-tenant lease premium. Powered shell is the speed-and-optionality middle: a developer delivers an energized, weather-tight shell with power and water at the boundary, and you complete the IT fit-out. You keep your fit-out and density decisions while someone else absorbs the slowest, most uncertain part, the interconnection and the shell. It is the modality that most cleanly turns an irreversible decision into a deferrable one. Colocation (wholesale or retail) rents someone else's already-powered hall. You get a live cluster on the operator's schedule, with facility rent plus tenant-funded IT, at the cost of a shared shell and the least control over power and cooling design.

The decision is rarely all-or-nothing. A hybrid can combine these modalities: anchor durable base load in a self-develop or BTS campus, take a powered shell to preserve a density ramp you have not committed to, and rent colocation or neocloud capacity for burst and bridge while the owned interconnection matures. The build-vs-buy-vs-rent NPV is scored in Chapter 1.8; the long-lead equipment that gates every modality is in Chapter 2.3.

A defensible site-selection program is a funnel of ordered gates, not one weighted spreadsheet. The order matters as much as the criteria, because a funnel should fail candidates cheaply on the binding constraint before you spend diligence money on the rest. In 2026 that constraint is almost always firm power by a date, so it gates first, with the other mandatory constraints checked in parallel. Run the old order, scoring everything and checking power last, and you will spend months of geotech, fiber and incentive diligence on sites that were never going to energize in time.

• Power: required capacity by its need date (pass/fail/unknown). Is there a credible path to your required firm megawatts (or curtailable ones, where your workload can accept them) by your in-service date? That takes a real read on the local interconnection queue, substation headroom, transmission proximity and voltage class, and any BYOP bridge option. Sites without a path die here, whatever their other virtues.
• Delivered bill: affordability and contracted structure. Among sites that can energize, what is the all-in cost? Look at nodal/LMP pricing, congestion exposure, capacity and demand charges, curtailment terms, and who pays for network upgrades. In the cited Epoch case energy is the largest operating cost line, though chips and facility dominate total cost, so this is the main economic screen among survivors. → Chapter 3.3.
• Network: workload and route-failure budgets, including training and batch. Does the site meet the workload's documented percentile, TTFT, TPOT/inter-token latency, output-length, concurrency, replica-routing and network-RTT requirements? Those requirements, not a generic radius, decide which power-rich basins remain viable. → Chapter 3.6.
• Water & climate: lawful, sufficient cooling. At the design ambient, can your plant reject the heat it must while staying within the site's water rights and limits? No site skips this gate: cooling must be both sufficient and lawful. Model site WUE and annual and peak makeup water, then choose dry/non-evaporative or hybrid/adiabatic heat rejection to fit the water budget. → Chapter 3.7.
• Land, geotech, flood, zoning: a buildable, lawful layout. Look for enough contiguous, developable acreage with the right zoning to fit the phased layout, cooling yards, setbacks, expansion reserve, and the grid or on-site plant. The soil must bear the load, and the site must sit outside the floodplain. A gigawatt headline tells you neither the minimum acreage nor the building coverage; work out the gross-to-net land need before you count acreage. Rarely the binding constraint, but a hard stop when it bites. → Chapter 3.8.
• Fiber, fiscal, social license: service and approval gates before incentives. Carrier-diverse service and the operative community and permitting conditions must pass. Incentives enter the survivor comparison through dated eligibility, repeal and recapture cash scenarios. → Chapter 3.9, Chapter 3.10, Chapter 3.11.

A parcel with power ready can still miss its date because the permits are not. Take Pennsylvania as of October 12, 2026. Its GRID consent-agreement route covers projects with peak demand above 25 MW. Developers that sign enforceable commitments get a rolling DEP review, though no DEP permit is issued while any required local approval is still outstanding (Executive Order 2026-05 of August 18; Pennsylvania DCED). Developers who decline the agreement get a slower review. New York's separate, scoped DEC application moratorium is covered in Chapter 3.11, and the permit sequence in Chapter 3.9.

A failed must-have removes a site; an unknown condition holds the irreversible release until its evidence arrives. The full weighted scoring matrix, hard pass/fail templates and the stage-gated desktop→field→binding diligence sequence are built out in Chapter 3.13. This chapter identifies the constraints and sends their evidence to that single decision method.

Regional markets package these constraints differently, and the power-first screen is redrawing the map. Expansion is spreading into more markets as power and entitlement limits hit particular sites. Northern Virginia faces those limits (a Dominion/PJM shortfall, and zoning reform that ends by-right development), and so does California. Even so, Northern Virginia topped the primary markets for net absorption in H1 2026 and kept adding to its construction pipeline (CBRE, September 8, 2026). Texas, the Nordics, the Gulf and US secondary markets are now on the search list. Before you call any of their capacity deliverable, check the utility territory, parcel, network and approval path that actually apply.

PJM's January 2026 load forecast has its summer peak near 222 GW by 2036, about 66 GW higher over ten years. Yet it took 4.4 GW (2.6%) off its 2028 peak compared with the 2025 forecast, reflecting new electric-vehicle and economic assumptions and tougher screening of large-load requests. In the near-term years PJM now includes a large load only when it is backed by an electric service obligation, a construction commitment or a similar firm commitment; it classes projects further out as non-firm and derates them for their uncertainty. A project without those commitments therefore weighs less in PJM's planning for the next few years than its developer may expect.

Texas/ERCOT is still worth searching. As of October 12, 2026, though, a data-center project there moves forward only under the governor's August 3 directive on verification and audit, and its exceptions. ERCOT's conditional Batch Zero classifications of September 3 do not mean the studies are done, and they set no service date. Responses to ERCOT's State and Community Impact RFI were due October 12, and ERCOT expects its reports by December 10.

ERCOT's June 2025 forecast puts non-crypto data-center load at ~16.8 GW in 2028 and ~25.0 GW in 2030. Its preliminary 2026 long-term load forecast (April 15, 2026) counts total system demand, not data centers alone. It reaches about 368 GW by 2032 against an all-time peak of 85.5 GW, because it adds the medium and large loads that transmission and distribution companies report. ERCOT says that figure is higher than the growth it expects and is not a prediction of what will be built, so read it as request pressure, not as data-center load.

SB6's large-load framework makes curtailment a condition of service, not a faster queue. PUCT's April 2025 approval of the first 765 kV import paths for the Permian Basin covers a plan, not built lines. A state permit gate sits on top: on September 21 the governor directed TCEQ to halt permits sought by data centers until ERCOT finishes its audit (→ Chapter 3.9). None of these reserves capacity for your project, so test its classification, studies and authorized service tranche (→ Chapter 3.2).

US secondary markets (Columbus, Salt Lake City, San Antonio, Reno, Indianapolis, the Permian) offer headroom some primary-market sites lack. The Nordics offer firm renewables, cold climates and heat reuse into district heating. Google reports a Hamina PUE of 1.09 for Q2 2026, but calculate your own site's free-cooling hours with the climate-bin analysis in Chapter 5.8. Finland has also proposed a new connection order. The draft, circulated September 18, 2026, would put demand and storage requests into groups by connection capacity and purpose. When capacity is scarce, a project's readiness and any dispatchable or new local supply would affect its priority. The target start is January 1, 2027, and earlier requests would fall under it only if no connection agreement or conditional connection agreement had yet been signed. As of October 12 it is a proposal, so test the project's agreement status and prospective priority (→ Chapter 3.2). The Gulf brings sovereign capital and plentiful gas. FTI Consulting's August 2025 forecast has GCC capacity reaching 3.3 GW by 2030, up from just over 1 GW.

Power availability is local, and it changes in both directions. A market that was full last year may have opened through new transmission or a bring-your-own-power regime, and a market that looked open may have closed. Dublin's moratorium on new connections ran from CRU's November 2021 direction to its December 2025 policy. Under it, any new data center asking for 10 MVA or more must bring generation, storage or both, on site or nearby, sized to that maximum import demand, and any project from 1 MVA needs a credible plan to meet the additional Irish renewable-electricity requirement (80% of annual demand within six years) (CRU decision, December 12, 2025; Chapter 3.2 and Chapter 3.4 cover the detail). EirGrid still lists Greater Dublin as fully constrained (May 2026). Singapore halted new data centers from 2019 to 2022; it now hands out capacity through capped, competitive calls. Its May 2024 Green Data Centre Roadmap aims for at least 300 MW in the near term. The second call (DC-CFA2) made its provisional awards on August 21, 2026: 50 MW each, 200 MW in total, to Equinix, Digital Realty, ST Telemedia Global Data Centres and Keppel Data Centres. All four will build on Jurong Island, where JTC is developing a low-carbon data-center park. An allocation still gives no date for energized delivery. The survivor set moves faster than a five-year siting study can track. Use Chapter 3.13 for the cluster deep-dives and to choose which authority and service territory to investigate; use Chapters 3.2–3.12 to test the actual parcel and contracts. A regional pipeline figure reserves no power for a particular project.

Site selection holds the most irreversible decisions in the lifecycle, so the discipline from Chapter 1.1 applies in full: over-build or hedge the irreversible decisions now; defer the reversible ones and keep them cheap to change.

Irreversible (decide once, at siting): the site itself, since you cannot move a slab; the interconnection capacity and voltage class, since the queue slot is the scarcest asset in the project and you do not get it back; the structural and water provisioning, meaning floor loading, facility water headroom and pipe-rack space for a density ramp you have not committed to; and the climate-and-water basin you have planted in. Floor loading follows the rack generation you will install. The guide's wet-rack screening band for GB200/GB300 NVL72 is roughly 3,000–3,300 lb, and HPE's GB300 QuickSpecs give 3,300 lb fully loaded with PGW. NVIDIA puts Vera Rubin NVL72 at roughly 4,000 lb (March 16, 2026) but does not say whether that weight includes coolant. Set design and handling loads from your OEM configuration (→ Chapter 6.2). These are where you spend the option premium. Provision headroom you may never use, because retrofitting it under a running hall is slow, disruptive and far costlier.

Potentially reversible (defer, re-decide cheaply): the accelerator generation and workload mix, inside the power and cooling limits you have accepted; future fit-out phases; how you supply energy on top of a secured interconnection (grid-only, grid plus bridge, or hybrid), which you can revisit as the BYOP and PPA markets move; and procurement or supply changes that your signed contracts allow. A powered-shell or colo lease lets you out only through break or termination rights you can enforce, or a permitted transfer that frees you of the tenant's obligations. Expansion rights preserve a future capacity choice, not an exit. Before you call a decision cheap to reverse, price what walking away would cost: the minimum bills and termination charges, the committed term, the fit-out already sunk, and any PPA or generation obligations still running (Chapter 1.6 and Chapter 1.8). The strategic move is to convert irreversible decisions into reversible ones wherever the premium is cheap: a powered shell instead of a full self-build preserves IT-fit-out optionality, and reserving floor loading and water turns an irreversible density ceiling into a deferrable choice.