This is a guest blog authored by Humane Intelligence volunteers, exploring topics related to AI evaluations and sociotechnical topics in AI. Co-authors are Maria Di Fonzo, Humam Seddik Akkad, Nkechika Ibe, Theodora Skeadas.
Examining AI’s environmental footprint
The expansion of AI has brought renewed attention to its environmental footprint, particularly the energy, water, and material demands of data centers that power modern AI systems. While AI evaluations and practices like red teaming contribute to compute use in the short term, they also offer a potential path toward greater efficiency over time by helping organizations identify failures, inefficiencies, and risks earlier in the lifecycle of AI systems. At Humane Intelligence, our work sits within this tension: contributing to the use of AI infrastructure today while aiming to reduce wasteful deployments, unnecessary retraining, and poorly performing systems in the long run. As AI becomes increasingly bound to physical infrastructure, understanding how data centers shape geopolitics, governance, and evaluation practices is essential to grappling with the true costs and consequences of AI at scale.
The rapid rise of AI has tethered its potential not just to algorithms and data, but to the physical infrastructure of data centers. Since compute power underpins AI capabilities, the geography of where data centers are located, who owns them, and how they are governed now has profound geopolitical implications. As governments and companies both vie for technological supremacy, the concentration of data centers in a few jurisdictions is reshaping AI evaluations, geopolitics, and global governance.
Opportunities from data centers for evaluators
Data centers offer opportunities for evaluators. Red teaming as a process may save power and compute in the longer run if it can help organizations identify problems with their AI systems early on, though it contributes to the use of AI, which produces environmental waste, in the shorter run. Data centers also create new failure modes that red teaming must surface, such as carbon-aware scheduling that deprioritizes certain regions, model performance shifts under load or energy constraints, and supply chain and geopolitical risks.
Growing consolidating around data centers and compute capacity
Data centers, once simply back‑end infrastructure for cloud services, have become strategic assets. For example, the World Economic Forum notes that the U.S. alone hosts roughly 51 % of the world’s data centers, making other countries heavily reliant on U.S.‑based cloud infrastructure. This concentration raises risks: Compute access becomes a chokepoint, data flows favor dominant jurisdictions, and national strategies for “digital sovereignty” emerge in response. Recent research found that U.S. companies operate 48 % of non‑U.S. data center projects by investment value, meaning local build‑out does not guarantee sovereignty.
At the same time, governance frameworks struggle to keep pace with how rapidly compute power is concentrated in a handful of nations. As the Federal Reserve reports, the U.S. now controls an estimated 74% of global high-end AI compute capacity, far ahead of China (14%) and the EU (4.8%). This concentration has outpaced regulatory responses and raises urgent questions about global oversight, equitable access, and sovereignty. These patterns underscore how the location and ownership of compute infrastructure are as central to AI governance as algorithms or data policy.
This growing gap in compute capacity is a feature, not a bug, of modern ‘chokepoint’ statecraft. AI geopolitics now reaches deep into the semiconductor supply chain, with the U.S. using export controls to starve rival nations of high-end GPUs, critical for the development and expansion of cutting edge AI. By locking down the most advanced hardware, IP-dominant nations effectively impose a ceiling on other countries’ research potential, adversaries and allies alike. As a result, data center strategies cannot be viewed in a vacuum, they represent the end of a highly politicized supply line dependent on Dutch machinery, Taiwanese fabrication, and American innovation and geopolitical dominance.
The physical clustering of AI‑capable data centers also triggers infrastructure, environmental, and social consequences. In the U.S., some states, like Virginia, now see data‑center electricity demands consuming 26 % of local supply and other states reaching double‑digit shares of grid use. Since many data centers locate near existing fiber, power and cooling infrastructure, development often concentrates in a handful of regions—accelerating local strain on utilities and amplifying environmental burdens. Studies show that AI‑capable data centres clustered in rural or less‑resourced regions are associated with increased local strain on water resources and infrastructure, and some locations face air‑pollution risks due to backup power systems. These conditions prompt critical questions for governance around how and where data‑centre build‑outs happen, including community participation in siting decisions and whether benefits are equitably shared.
The discourse, however, extends beyond simple grid capacity to the more urgent imperative of energy security, and by extension, national security. In a bid to decouple from instability and the potential complex dynamics surrounding public grids, hyperscalers are moving to directly capture nuclear assets and renewable fleets. This trend signals a privatization of critical energy infrastructure that effectively blurs the lines between corporate sustainability mandates and national energy strategy. Second-order effects of this shift would see energy-rich regions transform into geopolitical heavyweights, leading to a potential strategic realignment that the national security establishment, currently outpaced by the corporate imperative for rapid scaling, has yet to fully appreciate.
Addressing dual challenges around national sovereignty and transnational interdependence
As AI infrastructure centralizes, global governance frameworks must contend with dual challenges: national sovereignty and transnational interdependence. A comparative governance study outlines how the U.S. leans market‑driven, the EU a precautionary risk‑based model, and Asia a state‑guided deployment strategy, each shaped by their compute geography and existing development styles. Initiatives like Gaia‑X in Europe aim to build federated cloud infrastructure for digital sovereignty by design, reflecting a response to heavy reliance on U.S. or Chinese‑dominated compute infrastructure. Concurrently, infrastructure investment deals underscore the scale of consolidation: a $40 billion deal involving Nvidia, Microsoft and BlackRock to acquire major data‑center operator Aligned Data Centers reflects compute concentration and raises questions about infrastructural gate‑keeping. Increased funding by top technology companies such as Google, Meta and Microsoft under the framework of AI for Good in Africa highlights new spheres of influence. Growing critiques of these investments and funding underscores rising concerns around digital dependence through AI infrastructural dominance by the West, especially the United States.
Ultimately, the geopolitics of AI infrastructure is rapidly becoming a central front in global competition. With compute power and data‑center capacity concentrated in a few geographic and corporate zones, questions of sovereignty, equity, and sustainability come to the fore. Effective AI governance today must extend beyond data and algorithms—it must grapple with where and how the machines of intelligence are physically built, powered and ruled. For governments, companies, and communities alike, the compute infrastructure map is quickly becoming a blueprint for power.