Compute geographies

This project studies where AI compute is built, how those locations are selected, and what dependencies emerge around land, energy, water, tax incentives, network connectivity, chips, and operational labour. Rather than treating compute as an abstract cloud resource, we investigate it as situated infrastructure: built in specific places, governed by local constraints, and connected to wider industrial supply chains.

The geography of AI compute is becoming a central question for public institutions, research organisations, and policymakers. Training and serving advanced models depends on specialised facilities, high-capacity power connections, fibre routes, cooling systems, chip availability, and long-term procurement contracts. These requirements concentrate capability in some places while leaving others dependent on remote infrastructure and commercial access terms.

Research focus

The project maps the relationship between compute facilities and the conditions that make them possible. We examine where data centres and cloud regions are located, what kinds of energy systems they draw upon, how land and permitting decisions shape expansion, and how public-sector strategy interacts with private infrastructure investment.

Our work asks three connected questions. First, what locations are becoming important for AI compute, and why? Second, how do local environmental, regulatory, and economic conditions influence the infrastructure that can be built there? Third, how do these geographic patterns affect claims about AI sovereignty, research autonomy, and public-interest access?

The aim is not simply to produce a map of facilities. We are interested in the relationships behind the map: grid capacity, ownership structures, procurement agreements, regional development narratives, tax incentives, public resistance, and the technical standards that determine whether a facility can support contemporary AI workloads.

Why geography matters for AI

Cloud platforms are often described through a language of availability zones, regions, capacity, and elastic scaling. That vocabulary is useful for developers, but it can hide the political and environmental choices that make cloud services possible. A model endpoint may be accessed through an API, yet the request is served by a chain of physical dependencies: substations, backup generation, cooling systems, water rights, fibre landing points, logistics routes, security staff, maintenance teams, and hardware replacement cycles.

These dependencies are not distributed evenly. Some regions have a combination of cheap energy, permissive planning environments, existing industrial land, tax incentives, and network connectivity. Others may have strong research institutions and policy ambitions but lack grid capacity or suitable facilities. The resulting geography shapes who can train models, who can deploy them at scale, who can negotiate favourable access, and who must accept the conditions set by external providers.

For public institutions, this matters because compute location is not only a technical parameter. It affects procurement resilience, data governance, environmental accountability, regional economic policy, and democratic oversight. A public agency may procure AI services from a cloud region described as local, while important operational dependencies remain controlled elsewhere. Conversely, a smaller research compute cluster may provide less raw capacity but more meaningful institutional control.

Infrastructure layers

We study compute geography across several layers. The facility layer includes data centres, colocation sites, cloud regions, high-performance computing centres, and specialised AI clusters. The energy layer includes grid connections, power purchase agreements, backup generation, cooling requirements, and local debates about industrial electricity demand. The network layer includes fibre routes, exchange points, latency requirements, and interconnection markets.

The hardware layer focuses on accelerator supply, server design, procurement timelines, repair capacity, and the replacement cycles that determine whether a facility can support frontier or applied AI workloads. The governance layer includes planning permissions, tax arrangements, environmental reporting, procurement rules, security classifications, and contractual terms. The labour layer includes the engineers, technicians, construction workers, facility managers, and policy teams required to keep infrastructure operational.

Seeing these layers together helps avoid two common simplifications. The first is the idea that compute is simply wherever a cloud provider says a region exists. The second is the idea that local ownership alone guarantees local control. In practice, compute capability is assembled through layered dependencies, and different actors control different parts of the stack.

Method

The project combines infrastructure mapping, policy analysis, and interviews. We collect public information about data centres, cloud regions, energy constraints, chip clusters, research computing facilities, and national compute initiatives. We then connect those sources to planning documents, grid connection debates, company reports, procurement frameworks, and public investment programmes.

This mixed approach is necessary because compute geography is only partly visible. Some infrastructure is publicised through cloud-region announcements or national AI plans. Other elements are hidden in leasing agreements, colocation contracts, energy infrastructure applications, or vendor-specific capacity planning. We therefore treat the geography of compute as an evidence problem as much as a mapping problem.

Our mapping work follows facilities through multiple sources rather than relying on a single dataset. We compare company announcements, real-estate filings, planning applications, energy regulator documents, procurement records, industry reports, and local news coverage. Where a facility is publicly visible but its AI relevance is uncertain, we mark it as such instead of overclaiming. Where a national compute initiative depends on commercial cloud access, we distinguish between public funding, public ownership, and public control.

The interview component focuses on people who make infrastructure decisions: public-sector technology teams, research infrastructure managers, cloud procurement specialists, energy planners, data-centre developers, civil-society observers, and researchers who depend on compute access. These conversations help explain why certain options are chosen even when public documents present them as obvious or inevitable.

Case comparison

The project compares different kinds of compute sites rather than treating all capacity as equivalent. Hyperscale cloud regions usually offer high availability, mature tooling, and substantial capacity, but access is mediated by platform contracts and pricing models. National research computing facilities may provide more direct public-interest alignment, but they face constraints around funding cycles, hardware refreshes, staffing, and allocation procedures.

Colocation facilities occupy a middle ground. They can host public, academic, or private compute clusters, but they also embed those clusters in commercial real-estate and energy markets. University and regional compute facilities can support local research communities, yet they may struggle to keep pace with specialised AI hardware requirements. Edge and inference facilities raise a different set of questions around latency, data proximity, energy efficiency, and distributed governance.

By comparing these models, the project asks what each site actually enables. A facility optimised for enterprise cloud workloads may not support large-scale model training. A research cluster may be excellent for experimentation but difficult to use for continuous public-service deployment. A national procurement agreement may provide access to impressive capacity while leaving switching costs and audit rights unresolved.

Environmental and regional stakes

AI compute expansion places new pressure on regional energy systems. Data centres may compete with housing, industry, transport electrification, and climate-transition infrastructure for grid capacity. In some locations, new facilities are framed as engines of regional development. In others, they become flashpoints for concerns about water use, land conversion, noise, emissions, and unequal distribution of benefits.

The project does not treat these concerns as externalities that appear after the technical design is complete. They are part of the design space. A compute strategy that cannot secure sustainable power, local legitimacy, or credible environmental reporting is not a stable infrastructure strategy. Likewise, a regional development plan that counts data-centre investment as economic success without examining ownership, employment, and public value may overstate its benefits.

We therefore document both the enabling conditions and the local controversies around compute sites. This includes the language used to justify facilities, the actors included in consultation processes, the public information available about energy use, and the mechanisms for accountability when infrastructure imposes costs on host communities.

Policy relevance

Compute access increasingly shapes what kinds of AI research can be conducted, which organisations can experiment at scale, and which policy goals are technically realistic. A university, public agency, or civic organisation may have strong ideas about AI governance, but those ideas depend on whether suitable infrastructure is available under acceptable conditions.

Geography also changes the politics of AI infrastructure. A cloud region can be described as a national capability, but it may be operated by a foreign firm, depend on imported chips, draw on constrained local energy systems, and serve multiple jurisdictions. Conversely, a smaller public compute initiative may be limited in scale but more accountable to research and public-interest goals.

By tracing these tensions, the project contributes to a clearer understanding of what “compute capacity” means in practice. It helps distinguish symbolic infrastructure announcements from durable public capability, and it identifies the local dependencies that are often missing from high-level AI strategies.

Working questions

The project is guided by a set of practical questions that can be used by policymakers, researchers, and public institutions. What part of the compute stack is controlled locally, and what part is leased or contracted? Which energy, water, land, and network constraints determine whether capacity can expand? How are environmental costs measured and reported? Who receives the economic benefits of infrastructure investment? What contractual terms govern access, portability, auditability, and exit?

We also ask how compute geography changes research agendas. If some research groups have privileged access to large-scale clusters while others rely on small grants or retail cloud pricing, then the geography of compute becomes a geography of knowledge production. The location and ownership of infrastructure affects what questions are affordable to ask.

Outputs

The project will produce a set of public-facing maps, short policy briefs, and research articles. These outputs will compare compute geographies across regions, describe common infrastructure bottlenecks, and document how public institutions negotiate access to facilities that are usually owned or operated by commercial actors.

We are also developing a vocabulary for describing AI infrastructure that avoids both cloud abstraction and simple territorial ownership claims. The goal is to support more precise public debate about what kinds of infrastructure are needed, who controls them, and what trade-offs come with different models of provision.

Later outputs will include a typology of compute sites, a checklist for public compute procurement, and short regional case studies that show how energy systems, planning processes, and platform strategies combine to shape the practical availability of AI infrastructure.