The Geography of Machine Intelligence

The internet encouraged an assumption that digital economic activity was steadily becoming detached from geography. Software could be written in one country, hosted in another and consumed almost anywhere; cloud computing allowed businesses to purchase computing capacity without knowing precisely where the underlying machines were located; and improvements in telecommunications made physical distance appear progressively less important to large portions of the information economy.

Artificial intelligence is complicating that story because the systems required to produce machine cognition are becoming too large, too energy intensive and too physically demanding to locate without regard to place. Data centers need access to large blocks of electricity, transmission capacity, fiber networks, suitable land, cooling systems, construction labor and political permission, while the increasingly expensive computational hardware inside them creates strong incentives to locate where those resources can be obtained quickly enough to keep the machines productive. The intelligence economy may distribute its applications globally, but the physical production of that intelligence is becoming concentrated around places capable of supporting it.

The International Energy Agency estimates that nearly half of U.S. data-center capacity is already concentrated in five regional clusters, even though data centers remain a relatively modest portion of electricity consumption at the national level. That concentration helps explain why the consequences of AI-related power demand appear much larger in particular local markets than aggregate national statistics would suggest.

The emerging geography of machine intelligence will therefore be shaped less by where technology companies would ideally prefer to locate than by where the complete production system can actually be assembled. As the scale of computation rises, geography reasserts itself through megawatts, transmission corridors, fiber routes, permitting regimes, land availability, water systems, construction capacity and community acceptance, all of which determine whether proposed computational capacity becomes operational or remains an investment plan waiting for the physical economy to catch up.

Computing Is Becoming Geographically Selective

Traditional technology clusters formed around unusually dense combinations of skilled labor, research universities, venture capital, entrepreneurial networks and established companies. Silicon Valley became the obvious example because the value created by those relationships reinforced itself over time: engineers moved between firms, experienced founders created new companies, investors accumulated specialized knowledge and suppliers emerged around the expanding ecosystem.

Data-center geography follows a related but materially different logic. Human talent remains important, particularly for frontier model development, software engineering and research, but the large facilities performing the computation do not necessarily need to sit beside the people consuming their output. Their location is determined increasingly by the physical economics of power, connectivity and expansion capacity.

Power availability has consequently moved ahead of some factors that once dominated site selection. CBRE's 2026 outlook reports that securing large electrical deliveries has become sufficiently difficult that power cost and delivery speed now outweigh pure connectivity considerations in many site-selection decisions, with developers increasingly seeking locations capable of providing more than 300 MW within schedules that match the economics of the proposed facility.

JLL reaches a similar conclusion from the global market perspective, describing “speed to power” as the primary site-selection criterion in 2026, followed by community support, latency and proximity to customers. Its analysis estimates that the world could add roughly 100 GW of data-center capacity between 2026 and 2030, requiring investment approaching $3 trillion once real estate, infrastructure and tenant computing equipment are considered.

A metropolitan area may possess excellent universities, advanced telecommunications and a large technology workforce, yet still lose a major project if another 500 MW cannot be delivered for eight years. Another region with less historical technology prestige but abundant electrical capacity, large industrial sites, fiber connectivity and a predictable permitting environment can suddenly become more attractive. Machine intelligence is producing a technology map that does not neatly resemble the geography of the previous software economy.

Power Is Pulling Compute Toward New Places

The transformation is already visible in North American development patterns. Northern Virginia remains the continent's largest established data-center market, with CBRE reporting approximately 4,496 MW of inventory during the first half of 2026, but available capacity had fallen to only 10.8 MW and vacancy to 0.2 percent. Atlanta, meanwhile, overtook Northern Virginia for the first time in total capacity under construction, with roughly 2,882 MW underway, while Dallas-Fort Worth, Phoenix, Chicago and other markets continued substantial expansion.

The important signal is not that one metropolitan area has replaced another, because mature clusters retain powerful advantages in fiber, customers, experienced labor and network effects. What is changing is the radius within which developers are willing to search for additional capacity. JLL reported in August 2026 that frontier markets accounted for 77 percent of North American data-center capacity under construction, while identifying energy-rich and development-friendly areas such as West Texas, Ohio, Louisiana and the Carolinas among places receiving substantial new investment. Earlier in the year, its research found that 64 percent of the North American construction pipeline was already located outside traditional mature hubs.

When the resource limiting expansion becomes scarce in established markets, investment begins searching geographically for places where that resource remains abundant. In the data-center industry, the scarce resource is increasingly not land by itself and not fiber by itself, but deliverable power combined with development certainty. A site becomes attractive when sufficient electricity can be delivered within an acceptable timeline, when the electrical system can support later expansion, when telecommunications networks can connect the facility to major markets and when local institutions can actually permit and build the supporting systems required to bring the project into operation.

This marks a meaningful reversal from an earlier period in which compute tended to follow connectivity and population concentrations. Increasingly, connectivity itself is being extended toward power-rich locations because developers can often build fiber more readily than they can manufacture an unconstrained electrical grid where none exists. CBRE reported more than 20,000 route miles of new U.S. fiber projects announced during 2026 as network providers expanded national and regional backbones toward emerging AI markets. Once sufficiently valuable computational loads begin appearing in new places, telecommunications investment can follow, reinforcing the formation of additional clusters.

The Intelligence Map Will Not Look Like the Software Map

This shift has consequences for how regions think about participation in the technology economy. During the software era, places outside major technology centers often attempted to replicate the visible ingredients of Silicon Valley by cultivating startup ecosystems, incubators, coding programs and venture networks. Those efforts sometimes produced successful companies, but the strongest technology agglomerations proved difficult to reproduce because their advantages depended upon decades of accumulated relationships and human capital.

The economy beneath AI creates a different entry point. A region does not need to become a world center for frontier-model research in order to become important to the intelligence economy, just as a state does not need to host an automobile company's headquarters to participate substantially in automobile manufacturing. Regions can specialize at different layers of the intelligence stack. One metropolitan area may concentrate researchers and model developers, another may host hyperscale computational capacity, another may manufacture semiconductors, another may become important for electrical equipment, and another may support power generation or industrial facilities that use machine intelligence intensively. Those activities remain economically connected even when they are geographically separated. The geography of machine intelligence may therefore resemble a network of specialized industrial regions more than a single technology cluster.

That possibility is particularly important for places with substantial inherited industrial assets. Older manufacturing regions often possess characteristics that appeared less valuable during the era of lightweight software development: large industrial parcels, rail corridors, transmission capacity, substations, natural-gas pipelines, utility expertise, engineering talent and municipalities familiar with accommodating large physical projects. Some of those assets can acquire new economic value when technology becomes industrial again.

A former manufacturing site with high-capacity electrical service and fiber access may be more relevant to the intelligence economy than an empty field closer to a fashionable technology center. Brownfield locations can offer existing power and connectivity that would be expensive and time-consuming to reproduce, and CBRE notes that constrained European markets are already seeing increased interest in brownfield redevelopment partly because existing power and fiber can shorten project-delivery timelines. The broader lesson is that technological opportunity changes when the production function changes. AI does not erase the advantages accumulated by existing technology centers, but it creates new advantages around physical capacity that previous generations of software did not value nearly as highly.

Training and Inference May Produce Different Maps

The geography becomes more complicated once the different forms of AI computation are separated. Training large models tends to reward enormous concentrations of computational hardware because processors must exchange vast quantities of information while working on the same problem. That encourages very large clusters where energy availability, network architecture and economies of scale dominate the site-selection decision. Training workloads can also tolerate greater physical distance from end users because the customer is not waiting for the model to complete its initial training process in real time.

Inference introduces a different set of location requirements because users interact with the trained model during everyday operations. Once a model has been trained, millions of users and machines may interact with it continuously. Applications involving autonomous systems, industrial control, financial transactions, medical diagnostics, communications or interactive software can become increasingly sensitive to latency, which creates an incentive to position at least some computational capacity closer to the locations where intelligence is being consumed.

JLL's 2026 analysis anticipates this transition occurring in stages. It expects the present phase of AI inference to remain highly centralized around major cloud regions, but projects a later shift toward regional inference hubs and eventually much more distributed edge intelligence as latency-sensitive applications become common. Its outlook suggests that healthcare, finance, manufacturing and other sectors could create specialized regional clusters where rapid response becomes more important than simply maximizing the scale of centralized computation.

If that progression occurs, there will not be one geography of AI but several overlapping ones. Frontier model training may concentrate near enormous blocks of inexpensive power, while large-scale cloud inference remains concentrated in major regional data centers where economies of scale dominate. Latency-sensitive inference may move closer to population centers and industrial users, while embedded intelligence increasingly runs directly inside factories, vehicles, medical equipment, phones, robots and other devices. The intelligence economy could therefore develop a hierarchy extending from a relatively small number of immense computational centers through regional facilities and ultimately outward toward billions of intelligent endpoints.

That structure would resemble other networks in which enormous centralized facilities coexist with increasingly distributed local systems. Electricity generation itself has operated this way for more than a century, combining major power stations, regional transmission networks, local distribution systems and end-use equipment throughout the economy. Artificial intelligence may develop a comparable geography of central production and distributed consumption.

Fiber Creates the Nervous System Between Energy and Markets

The ability to separate compute from its users depends upon telecommunications capacity, which is why fiber remains essential even as power moves to the center of site-selection decisions. A data center supplied with abundant electricity but poorly connected to the wider network cannot serve as an effective computational hub. Large AI clusters require massive internal networking to coordinate accelerators and equally capable external networks to exchange data with cloud regions, enterprises and users around the world.

Power and fiber therefore shape location together, with electricity determining where computational capacity can be physically supported and fiber determining how efficiently that capacity can be connected to markets. CBRE's 2026 North American research describes network companies extending long-haul and regional fiber toward emerging hyperscale markets, with financing flowing into national backbones and regional networks designed specifically to connect new computational clusters.

Once several large data centers arrive, the process can become self-reinforcing. Fiber providers gain greater incentive to build additional routes; stronger connectivity makes the region attractive to additional compute; more facilities create specialized construction and technical workforces; utilities respond to larger loads; and supporting businesses begin locating nearby. This is how a cluster becomes an ecosystem, except that the initial anchor may be physical rather than purely intellectual.

Land Is Valuable Only When the Surrounding Systems Work

The extraordinary size of emerging AI campuses adds another geographic constraint. CBRE reports that site selectors increasingly seek properties capable of supporting 250 MW or more on parcels of roughly 125 acres or greater, while some proposed developments are considerably larger. In power-constrained markets, procurement timelines can extend beyond ten years, encouraging developers to examine colocated generation, behind-the-meter natural gas and future advanced nuclear configurations as alternatives to waiting for conventional grid expansion.

A rural county may possess thousands of inexpensive acres, but those acres have limited computational value if the nearest high-voltage line lacks capacity, fiber is distant, road access cannot handle construction traffic or the utility has no credible route for supplying the required load. Conversely, land beside existing electrical and telecommunications capacity can command strategic importance far beyond what acreage alone would suggest.

The intelligence economy consequently changes the meaning of industrial real estate. The most valuable site may not be the property with the lowest purchase price but the site where multiple systems already converge. That helps explain the growing interest in locations near existing power plants, major substations and retired industrial facilities. Those places often contain grid connections whose replacement would otherwise require years of planning and construction, while surrounding land may already be zoned or culturally accustomed to heavy development. Energy capacity effectively becomes embedded value in the land.

The same logic applies to expansion. A project that can operate at 200 MW today but has no realistic pathway toward 500 MW may be less attractive than a site capable of growing over several phases, because hyperscale facilities are increasingly designed as campuses whose ultimate buildout unfolds over years. Developers are therefore purchasing not simply current electricity but future electrical optionality.

Water and Cooling Add Another Geographic Variable

Electricity dominates the current conversation, but cooling requirements introduce another consideration because computational hardware ultimately converts much of the electricity it consumes into heat. Cooling technology is evolving rapidly as higher-density GPU racks increasingly push facilities away from conventional air cooling and toward liquid systems capable of removing much greater quantities of heat from a smaller physical area. CBRE notes that server densities have increased from the roughly 3-to-10-kW-per-rack range common in older CPU facilities toward GPU racks around 100 kW, making closed-loop liquid cooling increasingly important for new high-density deployments.

Water use therefore varies considerably according to cooling architecture, climate, facility design and the source of electricity itself. Some modern closed-loop systems can sharply reduce direct water consumption compared with conventional evaporative cooling, while facilities relying heavily on evaporative systems may place greater demands on local water resources. There is no universal water profile for an AI facility, since a development in a water-constrained region using one cooling system can have very different local implications from a facility built in another climate with a closed-loop architecture.

Climate can also affect operating economics. Cooler environments may offer greater opportunities for economization, while hotter locations increase cooling requirements; however, those differences must be weighed against power prices, land, taxes, fiber, natural-disaster exposure and other site-specific characteristics. Competitive locations emerge from combinations rather than from a single variable.

Community Acceptance Has Become a Development Constraint

Perhaps the most important nontechnical change in data-center geography is that political and community acceptance has become a practical site-selection factor. For years, major projects have encountered opposition over land use, electricity demand, transmission lines, water consumption, noise, backup generation, aesthetics and the distribution of economic benefits. The enormous scale of proposed AI campuses is bringing many of those issues together in communities that may previously have had little experience with data-center development.

CBRE's latest North American research argues that local opposition and zoning delays have become sufficiently consequential that community approval now belongs alongside power and fiber availability in site-selection decisions. Its H1 2026 assessment describes development risk as having changed fundamentally because developers can secure electricity and land yet still lose years navigating local resistance or entitlement uncertainty. JLL similarly ranks community support immediately behind speed to power among current data-center site-selection factors. Its midyear North American analysis places community acceptance at the center of the industry's next expansion phase as development moves beyond established markets into regions encountering projects of unprecedented scale.

A community does not have to approve every proposed data center in order to remain competitive, nor should it. Large computational projects create legitimate questions about grid costs, water, land use, tax incentives, employment, noise, transmission construction and whether scarce capacity could produce greater local value if allocated elsewhere.

Predictability matters to investors because it allows the costs and conditions of development to be assessed before capital is committed. A region capable of establishing clear rules governing infrastructure costs, zoning, water, utility rates and community benefits allows companies to price those conditions into the project before committing capital. A region where expectations change after hundreds of millions of dollars have already been committed introduces a different category of risk.

The ability to say either “yes, under these conditions” or “no, this site is unsuitable” within a credible process can therefore be more economically valuable than nominally favorable policy surrounded by years of uncertainty. In that sense, governance becomes part of the site itself.

The Host Community Needs to Understand the Exchange

The growing economic significance of data centers does not mean every region should compete to host them indiscriminately. A hyperscale facility can involve billions of dollars of investment while employing relatively modest permanent staffs compared with many conventional manufacturing operations. At the same time, it can generate substantial property-tax revenue, construction employment, utility investment and demand for supporting services, with the balance varying considerably according to local tax structures and project characteristics.

For municipalities, the appropriate question is therefore not whether data centers are generically “good” or “bad,” but what a particular project contributes relative to the resources it consumes and the alternatives available. If a community possesses abundant electrical capacity that would otherwise remain unused, attracting a computational campus can monetize an existing advantage while supporting broader grid investment. If the same project absorbs virtually all remaining substation capacity and requires substantial public subsidies while producing little local economic spillover, the calculation changes.

The allocation of electrical capacity carries an opportunity cost, since a region capable of serving another 500 MW could direct that capacity toward one enormous data center, several manufacturers, a semiconductor facility, population growth or future businesses not yet known. Economic-development strategy should therefore treat electrical headroom as a productive asset whose allocation influences what else the region can build.

Infrastructure abundance is valuable partly because it reduces the severity of those tradeoffs. A region with enough generation and transmission to accommodate several forms of growth can accept new investment without treating every major project as a zero-sum contest for fixed capacity. Abundance expands the set of possible futures.

Industrial Regions May Receive a Second Geography

One of the more interesting consequences of AI development is the possibility that places shaped by earlier industrial eras could acquire new strategic roles. Regions around the Great Lakes, Appalachia, the Ohio Valley, Texas, portions of the Southeast and other historically industrial areas possess combinations of generation, transmission, pipelines, water, transportation networks, large parcels and engineering talent that can become relevant to computational development even when those regions were not dominant in the previous technology cycle.

The emergence of Ohio as one of the frontier markets receiving significant data-center investment illustrates the broader pattern identified by JLL, while federal redevelopment projects at former Department of Energy industrial sites demonstrate how legacy assets can be repositioned for AI-related investment. In 2026, DOE announced major projects at its Portsmouth site in Ohio and Paducah site in Kentucky involving large data-center campuses accompanied by new power generation and other development.

Assets built for one generation of economic activity can become valuable to another when the underlying requirements overlap, although the suitability of any former industrial property for a data center still depends upon its particular capabilities. A transmission corridor originally constructed around heavy industry can supply a GPU cluster as well as an electric furnace, while a large industrial parcel retains value because future technologies also require land. Rail and highway systems remain useful for moving construction materials and equipment, and pipelines and generation assets can support new electrical development.

This creates a potential bridge between old industrial geography and new computational geography. Regions that understand their inherited assets as a portfolio of productive capabilities rather than as remnants of a lost economy may discover opportunities that are invisible when economic development is reduced to attracting the fashionable industry of the moment.

International Geography Will Become More Unequal

The same forces operating within countries will influence the international distribution of machine intelligence. The United States, China and Europe currently dominate global data-center electricity consumption, with the IEA estimating their respective shares in 2024 at approximately 45 percent, 25 percent and 15 percent. Emerging and developing economies excluding China contain more than two-thirds of the world's population but account for less than one-third of global electricity generation and less than 10 percent of data-center capacity. Those disparities reveal why the AI divide cannot be understood entirely through access to software.

A country may be able to purchase access to frontier AI systems through the cloud without hosting substantial computational capacity itself, which makes machine intelligence more internationally accessible than industries requiring every country to reproduce the complete production system domestically. Nevertheless, economies with weak electrical grids, expensive connectivity and limited digital capacity remain disadvantaged in how deeply they can integrate AI into institutions and industry. The IEA notes that many emerging economies continue to face limited internet reliability, high data costs and much lower levels of electrical generation relative to population, conditions that constrain the systems upon which advanced AI applications depend.

At the same time, new regional hubs are developing outside the traditional centers. Southeast Asian data-center electricity demand is expected to more than double by 2030, helped by the growing Singapore–southern Malaysia cluster, while Middle Eastern markets are investing heavily in computational capacity as part of broader strategies to diversify their economies and leverage abundant energy resources.

The resulting map will therefore not remain static, but neither will it flatten into uniformity. Countries capable of combining abundant electricity, strong telecommunications, capital, political stability and market access can attract disproportionate shares of computational investment. Countries missing several of those ingredients may consume AI services produced elsewhere while capturing less of the capital formation associated with manufacturing the intelligence itself. That distinction is economically significant because the physical layer can bring engineering capability, network investment and related industrial activity even when the ultimate AI application is consumed far away.

Sovereign and Regulatory Requirements Will Shape the Map

Geography will also be influenced by rules governing where information can be processed. Governments and regulated industries increasingly impose requirements concerning data residency, privacy, cybersecurity and national security, which can limit the degree to which computation can simply migrate toward whichever country offers the lowest power cost. Financial institutions, healthcare systems, defense organizations and public agencies may require data or certain workloads to remain inside particular jurisdictions.

This creates a tension between physical efficiency and institutional sovereignty. From a purely economic perspective, centralizing computation into a handful of enormous energy-rich regions can create economies of scale. From a regulatory and resilience perspective, governments and businesses may prefer greater geographic diversification even when it carries higher costs.

JLL's 2026 global outlook expects demand for sovereign AI clouds to contribute to European and other regional capacity growth, demonstrating how legal jurisdiction itself can become a site-selection variable alongside power and fiber. The intelligence economy will therefore be shaped simultaneously by the physics of computation and the institutions governing information.

Geography Becomes Strategy for Municipalities and Regions

The emerging map of machine intelligence should change how economic-development organizations evaluate themselves. A city does not need to ask whether it can become “the next Silicon Valley,” a phrase that has generated more slogans than useful economic strategy. It needs to determine where it possesses genuine comparative advantages within the broader intelligence economy.

A region with inexpensive, abundant electricity and large industrial sites may be well suited to hyperscale computation, while a metropolitan area with major medical institutions could develop specialized healthcare inference capacity. A manufacturing center may benefit more from deploying AI throughout existing industry than from hosting frontier training clusters, just as a logistics hub could combine automation, robotics and transportation intelligence with its existing distribution economy. Different regions can participate at different layers.

The most useful analysis therefore begins with an inventory of assets rather than a predetermined industry label. Utilities should identify actual deliverable capacity and expansion timelines; telecommunications providers should map redundant fiber routes; local governments should understand industrial land, water capacity and permitting processes; universities should identify genuine research strengths; and economic-development agencies should examine how these systems interact rather than marketing each one separately. The question becomes less about whether a city possesses an “AI strategy” and more about whether its physical and institutional assets form a coherent economic proposition.

This approach also guards against chasing investments that do not fit local conditions. A water-constrained community with limited electrical headroom may have no reason to pursue an enormous training campus simply because AI is fashionable, while the same region may have outstanding opportunities in AI software, robotics, research or specialized inference applications that consume far fewer local resources. Economic development improves when geography is treated as a source of comparative advantage rather than an inconvenience to be overcome with subsidies.

Compute Will Move Toward Power, but Intelligence Will Move Toward Everything

The longer-term geography of AI will likely develop through two movements occurring simultaneously. At the largest scale, computation will increasingly seek regions capable of supplying enormous quantities of power. Training clusters, hyperscale cloud facilities and other concentrated workloads will gravitate toward locations where electricity, land and connectivity can be assembled economically.

Alongside the concentration of large computational facilities, intelligence itself will spread toward the places where it is used. As inference becomes less expensive and specialized hardware improves, AI will increasingly operate within regional data centers, factories, hospitals, vehicles, warehouses, homes and personal devices. The most computationally intensive workloads may remain centralized, while portions of machine cognition move progressively closer to the activities they are assisting.

The geography of AI can therefore become more concentrated at some layers of production while becoming more distributed at others. The factories producing enormous quantities of computation can concentrate in suitable locations while the intelligence derived from them becomes ubiquitous through networks and increasingly capable edge devices.

That pattern resembles earlier infrastructure revolutions. Electricity generation became concentrated into enormous plants even as electricity spread into virtually every building and machine, while telecommunications developed major switching centers and backbone networks as phones became commonplace. Cloud computing followed a similar pattern, concentrating enormous quantities of servers into hyperscale facilities while making digital services accessible from almost any location. AI may follow the same basic logic, except the commodity moving through the network is increasingly capable machine cognition.