AI is often presented as weightless software: a chatbot in a browser, an image generator in an app, an assistant embedded in a workplace tool. But the rapid expansion of artificial intelligence is producing a decidedly physical race. The decisive choices are increasingly being made not at the interface, but on parcels of land near substations, transmission corridors, fiber routes and water systems.
AI data centers are moving toward places where large amounts of reliable electricity can be delivered, cooling can be managed and permits can be secured. Proximity to users still matters for services that require very low latency, such as some financial systems, gaming and real-time communications. Yet for many of the biggest AI workloads, access to infrastructure can matter more than being close to a major city.
That shift is redrawing the cloud computing geography. It is also making data centers a more visible local political issue, as communities weigh jobs and tax revenue against pressure on power grids, water supplies, land and public infrastructure.
AI computing has a physical center of gravity
Modern AI systems depend on large groups of specialized processors, networking equipment, storage systems and backup infrastructure. Training a frontier-scale model can involve sustained computing across thousands of accelerators. Serving that model to millions of users creates a different but potentially enormous demand: inference, the repeated process of generating answers, images, recommendations or other outputs.
Both tasks need more than chips. A facility needs dependable electricity at every hour, high-capacity fiber connections, equipment that can be replaced or expanded, and systems that keep servers within safe operating temperatures. The more densely computing hardware is packed into a building, the harder data center cooling becomes.
This is why AI infrastructure cannot be understood only as a competition among model developers. It is also a competition for land, electrical interconnection, transformers, switchgear, construction capacity and cooling equipment. In many regions, those inputs are slower to obtain than servers.
Traditional cloud computing favored locations near major internet exchanges and population centers, alongside places with favorable taxes and available real estate. Those factors remain relevant. But a proposed facility may now be constrained first by whether a utility can supply its requested load, whether the transmission network can support it, and how long an interconnection study and upgrade process will take.
Electricity is becoming the first constraint
The defining feature of an AI-focused data center is not simply that it uses a lot of energy. It is that it may require a very large, highly reliable and relatively constant power supply. Computing equipment can be adjusted, paused or shifted in limited cases, but data-center operators are generally designed around continuous service and redundancy.
This creates a difficult planning problem for utilities. They must accommodate new industrial-scale customers while maintaining reliability for existing households and businesses. At the same time, many grids are retiring older generating plants, adding variable wind and solar generation, and facing growing demand from electrified vehicles, heating and manufacturing.
Forecasts of electricity demand from AI vary substantially because analysts must make assumptions about chip shipments, model use, efficiency improvements, facility utilization and the speed of construction. It is also difficult to separate AI workloads from conventional cloud services inside a mixed data center. The broad direction, however, is clear: data-center energy use is becoming material enough to shape utility investment plans in a growing number of markets.
For developers, a signed power agreement is not the same as physical capacity. A local distribution network may need upgrades. A high-voltage substation may need to be expanded. New transmission can take years because it crosses jurisdictions, requires land rights and can face its own opposition. As a result, a site with apparently cheap electricity can become less attractive if the connection timeline is uncertain.
Clean-energy claims depend on the clock
Power sourcing adds another layer of complexity. A company can buy enough renewable-energy certificates or contract for renewable generation to match its annual electricity consumption. That may support investment in clean generation, but it does not necessarily mean the data center is running on carbon-free electricity in every hour.
Hourly matching is a stricter standard. It asks whether clean power is available on the same grid, or through credible delivery arrangements, at the time the facility consumes electricity. Achieving that requires a mix of generation, storage, transmission and demand flexibility. It is harder than annual accounting, especially in regions where clean power is abundant at some times of day but scarce at others.
The distinction matters because AI facilities add demand to real grids with real operating constraints. A sustainability claim should therefore be read alongside information about location, timing, power contracts and the utility system serving the site.
Water is essential, but the accounting is complicated
Data center water consumption is often discussed as though it were one simple number. It is not. Facilities can use water directly on site for cooling, and they can also be associated with water use at the power plants that generate their electricity. These are separate pathways and should not be casually combined without explaining the method.
Direct water use depends on cooling design, weather, server density and operational practice. Some facilities use evaporative cooling, which can be energy efficient in appropriate climates but consumes water as it evaporates. Others rely more heavily on air cooling, mechanical chillers or closed-loop systems. Newer high-density AI deployments are also accelerating interest in liquid cooling, where liquid carries heat away from chips or servers more effectively than air alone.
Water metrics require care. Withdrawal describes water taken from a source, some of which may be returned. Consumption generally refers to water that is not returned to the same watershed in an available form, often because it evaporates. A facility can have a large withdrawal figure but lower consumption, or modest direct water use while relying on electricity with significant upstream water impacts. Local conditions determine which burden is most consequential.
A gallon used in a water-rich region is not equivalent, in social or ecological terms, to a gallon consumed during drought in a stressed watershed. That is why broad corporate water totals can obscure the question residents actually need answered: where does this facility get its water, how much will it use in peak conditions, and what happens when supplies are constrained?
Location decisions now follow infrastructure, not just customers
The ideal site for a large AI data center increasingly sits at the intersection of several systems: abundant land, high-voltage electricity, fiber connectivity, cooling options, a workable permitting process and a local government willing to host industrial development. Tax incentives and property-tax arrangements can influence decisions, but they cannot solve a missing transmission connection or an inadequate water supply.
Cooler climates can reduce the energy needed to remove heat. Regions with substantial renewable generation may offer lower-carbon electricity, particularly when transmission and contracts allow that generation to serve the facility. Industrial areas may already have useful electrical infrastructure. Locations near generation projects can reduce some grid constraints, although they may introduce others, including the need for new lines and the risk of concentrating development in a single region.
The result is a more fragmented cloud computing geography. Computing may be placed where power is available rather than where users live, while edge facilities and regional sites handle workloads that cannot tolerate delay. Training, batch processing and some inference can be more geographically flexible than interactive applications, but moving data and workloads still has costs, including network capacity, data-governance requirements and operational complexity.
Why communities are pushing back
Not every data center creates the same local impact. Its consequences depend on its size, cooling method, power source, existing grid capacity, tax arrangement and the condition of local water and road systems. Still, the scale of proposed projects has made residents and local officials more likely to ask questions that were once treated as technical details.
Common concerns include whether a utility will build expensive upgrades primarily for a new large customer, whether other ratepayers could bear part of those costs, and whether new demand could extend the life of fossil-fuel generation. In water-constrained areas, residents may question whether drinking-water supplies should support industrial cooling. Nearby communities may also raise issues around noise from generators and cooling equipment, land conversion, diesel backup systems and the transparency of negotiated incentives.
Developers, meanwhile, argue that data centers can broaden a tax base, support construction employment and attract related investment. Those benefits can be real, but they are not automatic, and permanent on-site employment is often more limited than the visual scale of a facility might imply. The central public-policy question is not whether data centers are good or bad in the abstract. It is whether a specific project fits the capacity and priorities of a specific place.
The most important local question is increasingly not whether a data center is “in the cloud,” but which public systems it will depend on when demand is highest.
Engineering can reduce impacts, but it cannot erase demand
The industry has several routes to lower the environmental and political costs of growth. None is a universal solution, because a design that minimizes water use may require more electricity, while a system that lowers energy use may be unsuitable for a hot or water-stressed location.
- Liquid cooling can handle dense AI hardware more efficiently than conventional air cooling in many deployments. Its water implications depend on whether the system is closed-loop, how heat is ultimately rejected and what cooling plant is used.
- Closed-loop cooling systems can recirculate fluid and reduce direct water consumption relative to evaporative approaches, though they may require more mechanical energy in some conditions.
- Reclaimed or non-potable water can reduce pressure on drinking-water supplies where municipal systems and treatment standards make it practical. It is not available everywhere, and it still has local infrastructure requirements.
- More efficient chips, servers and software can lower energy per calculation or per generated response. But efficiency gains do not guarantee lower total use if cheaper computing leads to far more AI activity.
- Workload scheduling can shift flexible tasks toward times or locations with lower-carbon, more abundant electricity. This is easier for training and batch jobs than for services that users expect to answer immediately.
- Heat reuse can make practical use of waste heat in certain climates and near compatible buildings or district-heating networks. It is promising but geographically limited: heat is difficult and costly to transport over long distances.
These measures should be evaluated in operational terms, not only in marketing language. A strong disclosure describes the facility’s expected power demand, direct water source, cooling approach, annual and peak use, backup generation, planned grid upgrades and the assumptions behind its clean-energy claims.
The trade-off is moving burdens, not making them disappear
Building near renewable resources or in cooler climates can lower some impacts. It can also require new transmission, increase competition for land, or place major industrial loads in smaller communities with limited administrative capacity. Building closer to users can reduce latency and possibly some network traffic, but may intensify pressure on already congested urban grids and expensive water systems.
There is no frictionless geography for AI. The challenge is to make the trade-offs visible early enough for utilities, regulators and communities to shape them. That means utility planning that identifies large-load requests before they become emergency infrastructure projects; permitting processes that disclose material resource assumptions; and contracts that clarify who pays for grid upgrades and what happens if promised development does not arrive.
What to watch next
Readers should look beyond announcements about new campuses and model capabilities. The more revealing signals will be the ordinary documents and decisions around them:
- Utility forecasts showing how large new loads are changing generation, transmission and distribution plans.
- Local approval hearings that reveal projected electricity demand, water sources, cooling designs, noise mitigation and tax terms.
- Rules for water-stressed regions, including restrictions during drought and requirements to use reclaimed water where available.
- More consistent reporting on direct water consumption, electricity use, hourly clean-power matching and the distinction between company-wide targets and site-specific performance.
- Evidence that hardware and software efficiency improvements are reducing total system impacts, rather than merely enabling faster growth in demand.
The future of AI will not be settled solely by algorithms. It will be negotiated through substations, reservoirs, planning commissions, transmission routes and utility rate cases. The cloud has always depended on physical infrastructure. As AI expands, that infrastructure is becoming impossible to ignore—and the places that host it will have a larger say in what the next era of computing looks like.