For most of the software era, geography seemed optional. Code lived in the cloud, which was a useful marketing term for “somebody else's extremely physical building.” AI is making the euphemism harder to maintain.
Training and serving large models require dense computing infrastructure. Dense computing infrastructure requires electricity. Electricity requires generators, transmission, substations, transformers, switchgear, cooling, permits, land, and time. The glamorous part of the stack is currently being throttled by objects that weigh several tons and arrive on trucks.
The distance from the city is increasing.
Reuters reported in August 2026 that European hyperscale data-center projects planned for 2026–2028 are, on average, being built much farther from urban hubs than projects in the preceding years. Developers are looking for cheaper energy, cheaper land, and faster connections to power.
That is a structural shift. Traditional data centers valued proximity to dense networks, customers, exchanges, and urban infrastructure. AI training workloads can tolerate more geographic distance if the site provides enormous amounts of dependable power and high-capacity fiber.
The constraint hierarchy is changing.
A model is not useful if the grid connection arrives in 2031.
The International Energy Agency estimated global data centers consumed roughly 415 TWh of electricity in 2024 and projected consumption to reach around 945 TWh by 2030 in its base case. Later IEA analysis in 2026 reported that data-center electricity use surged again in 2025 and that AI-focused facilities are growing even faster.
What makes the problem difficult is not just total energy. It is location and timing.
A national grid can have enough annual generation on paper while a particular region has no substation capacity, no spare transformer, no transmission headroom, or a multi-year interconnection queue. A data center is a concentrated load. The grid must serve it at a particular place, with high reliability, not merely produce an equivalent number of megawatt-hours somewhere else.
The unit of AI competition is becoming the powered site.
That suggests a better way to compare regions.
Instead of asking which city has the most AI startups, ask which region can deliver:
• hundreds of megawatts of firm power;
• high-voltage interconnection without a decade-long queue;
• redundant fiber;
• industrial water or viable alternative cooling;
• large contiguous land parcels;
• transformers and switchgear on schedule;
• permitting that fits a two-to-three-year technology cycle;
• political tolerance for large loads;
• stable energy pricing;
• enough construction labor and supply-chain access to finish the facility.
That list looks more like industrial policy than software entrepreneurship.
Power density changes the building.
The IEA's 2026 energy-and-AI update notes that the power density of AI servers has risen dramatically and is expected to continue increasing. That pushes cooling, electrical distribution, and rack design closer to physical limits.
A traditional data center can spread heat across rows of relatively moderate-density racks. Advanced AI clusters concentrate more electrical load and more heat in less space. Liquid cooling becomes more important. Power delivery becomes more complex. Rapid load swings matter to grid operators and storage design.
This is not a side effect. It is part of the product architecture.
Cheap electricity is not the same as available electricity.
A region can advertise low wholesale power prices and still be useless to a developer if interconnection takes five years. Conversely, a site with moderately higher prices may win because the power is available immediately and the transmission path already exists.
This is why powered land is becoming its own asset class. The useful commodity is not dirt. It is dirt with credible megawatts attached.
Nuclear power gets a second look for a boring reason: capacity factor.
AI companies have shown renewed interest in nuclear power, including life extensions, power-purchase agreements, and prospective small modular reactors. The attraction is not mystical. Large compute facilities like steady, dependable supply. Existing nuclear fleets can provide high-capacity-factor generation without the intermittency of wind and solar.
That does not make nuclear automatically cheap or fast. New reactors carry capital, regulatory, and construction risk. But countries with existing nuclear capacity may possess a strategic asset that looks increasingly relevant as data-center demand rises.
Renewables still matter, but transmission matters with them.
The IEA expects renewables to supply a large share of incremental electricity for data centers. That requires storage, transmission, demand flexibility, and geographic coordination. A hyperscale site built beside abundant wind or solar can still face reliability problems if the grid and storage architecture cannot bridge production gaps.
The practical question is not “renewables or gas or nuclear?” It is: what portfolio can deliver reliable power at the site, on the deployment timeline, without making the surrounding grid worse?
The next AI map may look strange.
If compute follows power, regions currently outside the fashionable technology map can become important. Rural areas near transmission, industrial zones with retired generation infrastructure, regions with hydroelectric surplus, nuclear-heavy grids, and places with fast permitting can attract infrastructure disproportionate to their existing software ecosystems.
That creates second-order effects: tax revenue, construction booms, grid investment, land conflict, water demand, local price pressure, and political battles over who gets priority when power is scarce.
The cloud eventually becomes zoning.
What this evidence cannot establish yet.
The current evidence does not establish a single inevitable geography for AI infrastructure, nor does it prove that electricity will dominate every siting decision. Model efficiency, chip architecture, networking, regulation, energy prices, water constraints, tax policy, and workload latency can all move the balance. The claim here is narrower: power availability and infrastructure lead time are becoming first-order constraints that deserve to be measured alongside talent and capital.
AI infrastructure is moving from a chip-centered story to a power-and-grid story. The competitive advantage will increasingly belong to regions that can deliver reliable megawatts, fast interconnection, cooling, fiber, land, and permits together. The best map of the next generation of AI hubs may look less like a map of universities and venture capital and more like a map of substations, transmission corridors, nuclear plants, hydro resources, gas pipelines, and interconnection queues.
The index we should build
Cyberdelia should eventually maintain an AI Infrastructure Readiness Index that scores regions by power availability, grid queue time, energy mix, powered-land cost, fiber redundancy, water stress, permitting time, transformer availability, and political risk.
That would let us test the thesis instead of admiring it. If AI geography really is becoming an electricity map, the future deployment pattern should begin appearing in the infrastructure data before it appears in corporate press releases.