AI Data Center Site Selection and Power Availability

#Why Power Availability Now Decides Where AI Data Centers Get Built

For two decades, data center site selection followed a familiar checklist: cheap land, low taxes, a mild climate, and reliable fiber. Power was assumed — utilities built substations to meet whatever load a developer requested, usually within twelve to eighteen months, and site teams treated that timeline as background noise rather than a design constraint. That assumption has collapsed. AI training clusters now request 100 megawatts to well over a gigawatt at a single site, loads that used to represent an entire city's demand growth over a decade, and a growing share of proposed campuses are sized for multiple gigawatts as training and inference capacity scales further.

Grid operators can no longer keep pace with requests at this scale. PJM, ERCOT, and other regional grid operators report interconnection queues stretching years, not months, with large-load requests waiting behind tens of gigawatts of prior applications competing for the same substations and transmission corridors. As a result, the first question site selectors ask is no longer 'where is land cheap?' but 'where can we actually get power, and when?' Fiber, water, and incentives have become secondary filters, applied only after a site clears the power-availability test, because a location with perfect logistics and no deliverable megawatts simply cannot host an AI data center on any usable timeline.

#Comparing the 5 Factors in Modern Site Selection

Power availability now sits above these five traditional criteria, but each still shapes the final decision once a site clears the power test.

• Grid capacity and interconnection timeline: How much firm capacity exists at the nearest substation or transmission node, and how long the interconnection queue runs — often the single factor that eliminates the most candidate sites before any other evaluation begins.
• Power cost and price volatility: Wholesale and retail electricity rates, plus exposure to real-time price spikes during scarcity events, directly affect the total cost of running power-hungry GPU clusters around the clock.
• Water availability for cooling: Traditional cooling systems can consume large volumes of water daily, making local water rights and drought risk a material constraint in many candidate regions.
• Fiber connectivity and latency: Proximity to internet exchange points and dark fiber routes determines whether a site can support low-latency inference workloads or only latency-tolerant training.
• Incentive programs and permitting speed: State and local tax abatements, equipment sales tax exemptions, and how quickly a jurisdiction processes permits still shift decisions between two sites that both clear the power bar.

#The Critical Gap: Power Availability Now Outweighs Traditional Site Factors

The gap between power-driven and traditional site selection shows up clearly in the data. ERCOT has reported that large-load interconnection requests grew nearly threefold in a single year, with the large majority tied to data centers and many individual requests exceeding a gigawatt. Lawrence Berkeley National Laboratory's federally funded research estimates that U.S. data center electricity use roughly tripled over the past decade and could double or triple again by 2028. No amount of cheap land, low taxes, or a fast permitting office compensates for a site that cannot receive power for five to seven years — a delay that outlasts most financing and offtake agreements a developer can reasonably underwrite.

This has inverted the traditional site-selection sequence. Instead of picking a region and then negotiating power, developers now start by mapping where grid operators and utilities have deliverable capacity — often down to a specific substation or queue position — and only then evaluate traditional factors within that power-constrained shortlist. Sites that once looked attractive on cost or incentives are being passed over entirely because the interconnection timeline runs years longer than the project's investment horizon can tolerate. Regulators have begun responding to this exact gap: FERC has directed regional grid operators to justify or reform how they process large-load interconnection requests, evidence that the power constraint has become a policy problem, not just a developer inconvenience.

#An Honest Assessment of Site Selection Resources

Several real resources help with the power side of site selection, though none replace hands-on grid engineering. PJM Interconnection and ERCOT both publish large-load interconnection queue data and process documentation showing, region by region, how much capacity is requested and how far behind an application sits — genuinely useful for narrowing a shortlist, though the queues shift month to month and neither operator can promise a firm delivery date up front. The U.S. Department of Energy has run formal requests for information to match data center developers with federal sites that already have power infrastructure, a useful model for how power-first screening works in practice even outside federal land. Commercial advisory firms such as CBRE publish annual data center market reports that rank metro areas partly on power availability, a reasonable starting point for regional comparison but not a substitute for a site-specific interconnection study, since market-level rankings can mask substation-level capacity constraints. None of these resources address how a developer should operate within a site's actual grid constraints once construction is finished and the facility is drawing load — that is an operations problem, not a research problem, and it is where most public site-selection guidance simply stops.

#The Empromptu Approach: Power-Aware Operations After Site Selection

Site selection answers where to build; it does not answer how to run a facility once the grid it depends on proves more volatile than the interconnection study assumed. Every grid has its own personality — ERCOT's scarcity pricing spikes, PJM's capacity performance penalties, and region-specific curtailment and demand response obligations all behave differently, and differently again during heat waves, cold snaps, or planned maintenance. Empromptu's Grid Guard capability is built for exactly this gap: it continuously monitors the specific grid conditions at the site actually selected, rather than treating power as a fixed input decided once during site selection.

Grid Guard ingests real-time and forecasted signals from the relevant grid operator or utility — pricing volatility, curtailment notices, demand response calls, frequency and voltage conditions — and translates them into orchestration decisions for AI workloads: shifting flexible training jobs, throttling non-critical load, or shedding load ahead of a mandatory curtailment event, all within the constraints negotiated at that specific site's interconnection agreement.

The result is that a good site-selection decision keeps paying off after it's made. A facility sited for a favorable interconnection queue position still needs active load management to avoid penalties, outages, or curtailment risk during grid stress. Grid Guard turns a site's specific power constraints into an operating parameter the AI platform respects automatically, rather than a risk an operations team has to track manually.