PJM Interconnection, the operator powering thirty U.S. states, projected demand growth from data centers of around 30 GW between 2025 and 2030. AI is producing unprecedented infrastructure demand, and the institutions tasked with meeting it have not moved at the same pace.
The essentials
- PJM projected data center load growth of around 30 GW between 2025 and 2030, without identifying 48.5 GW as blocked power in its interconnection queue (PJM Interconnection, 2026).
- In April 2026, the federal administration invoked the Defense Production Act to accelerate supply chains and electric infrastructure capacity.
- The bottleneck stems from both institutional constraints, insufficient capacity and supply delays: environmental permits, jurisdictional conflicts, local opposition, and interconnection rules designed for a twentieth-century grid.
- The European Union has common mechanisms, notably EuroHPC and its AI Factories, to finance and pool computing infrastructure.
- Institutional barriers affect the capacity of the three major powers to deploy the infrastructure necessary for AI.
A grid designed for another century
The American electrical grid was conceived to power factories and homes, not computing warehouses each consuming the equivalent of a mid-sized city. PJM Interconnection manages a territory that produces roughly 180 GW at peak. Seeing an additional 30 GW of demand appear from data centers within a few years asks the grid to absorb in one decade what it took a century to build.
The central problem is procedural. Connecting a new site to the American grid requires successive environmental impact studies, agreements between local operators, approvals from state authorities that don’t always communicate with each other, and negotiations with landowners no one is obligated to pressure into rapid sales. This procedure was calibrated for conventional power plants. PJM launched a complete reform of its interconnection process in 2023, moving from a serial queue to a cyclical process, even if this reform alone cannot absorb the new digital demand.
Data center operators—Microsoft, Google, Amazon, Meta—have the capital. They have the projects. What they lack is grid access that fits their timelines. Interconnection delays of one to three years are documented for hyperscale facilities according to available federal reports. Some projects are abandoned before even reaching that timeframe.
Invoking the Defense Production Act, a signal of impasse
In April 2026, the federal administration invoked the Defense Production Act to accelerate supply chains and electric infrastructure capacity. This wartime tool, designed to mobilize industry in a national emergency, had also been used after the pandemic for energy and electrical capacity and supply chains.
Its invocation signals something specific: neither the market nor the federated states resolved the bottleneck. The market provided demand and capital. The states managed their competencies without coordinating them. The federal government concluded that neither level alone could unblock the situation and chose an exceptional tool.
But the Defense Production Act does not eliminate environmental permits. It does not resolve jurisdictional conflicts between state regulators. It accelerates certain federal procurement procedures and can mobilize prioritized funds, but leaves intact most of the barriers blocking interconnections. By June 2026, a significant volume of projects still awaited interconnection. The emergency tool signaled the problem’s scale without solving it.
Carl Benedikt Frey observes, in How Progress Ends, that decentralized democracies explore technology effectively by multiplying laboratories, startups, and experiments, but struggle to scale when that scaling demands institutional coordination that no one has been mandated to ensure. The PJM-States conflict illustrates this mechanism: technological exploration is flourishing, while scaling infrastructure remains paralyzed.
New York concentrates the bottleneck
The New York case merits separate examination. The State of New York has its own decarbonization goals, its own regulators, and a tradition of mistrust toward large-scale industrial energy projects. Several data center projects situated near the metropolitan region were slowed by local environmental requirements incompatible with federal timelines.
This federal-state conflict is a structural feature of the American system. States retain significant authority over electricity production and land use, but the federal government also shares authority over interstate transmission and wholesale markets. The federal state can incentivize, partially finance, and invoke emergency powers, but cannot impose every local interconnection. It does possess limited powers that can prevail in certain projects of national interest transmission.
Dani Rodrik, who has long analyzed the tension between industrial policy and institutional architecture, would pose a direct question here: effective industrial policy requires institutions capable of coordinating actors at multiple levels. The United States has built a de facto industrial policy for AI—subsidies, federal priority, diplomatic pressure—without constructing the institutional coordination architecture that would make implementation possible. The result is visible in the PJM queue.
Perspective with the challenges of climate adaptation is instructive: in both cases, money exists or can be mobilized, but institutions responsible for allocating it and coordinating actors remain the principal bottleneck.
Europe regulates, without building
The European Union adopted the AI Act. It is an ambitious regulatory framework, the first in the world to attempt to classify risks associated with AI by level and impose obligations on developers and users. It has been in force since 2024, with progressive implementation through 2026.
The AI Act does not provide for financing of a shared infrastructure to operate the models that the regulation oversees. The European Union combines AI regulation and deployment of shared computing infrastructure via EuroHPC and its AI Factories, without constituting an electrical equivalent comparable to PJM. Each member state advances with its own champions—Orange in France, Deutsche Telekom in Germany, Ferrovial in Spain. The Union has a continental mechanism for pooling computing via EuroHPC, even if it is not functionally comparable to an electrical grid like PJM.
Europe combines regulation and deployment of shared infrastructure via EuroHPC, but continues to depend largely on computing capacity situated elsewhere. The United States produces more notable AI models, but leading-edge models also include Chinese actors. Applying the AI Act to these models from Brussels amounts to regulating a factory whose key you don’t hold. This tension is visible in the uneven deployment of generative AI in European public administrations, which encounter governance difficulties, data protection, organizational readiness, and technological sovereignty.
The approach of ThaiLLM in Asia, which chooses to build its own capacity rather than rent foreign infrastructure, illustrates the same tension from a different angle: computational sovereignty requires infrastructure, and infrastructure requires institutions capable of planning and deploying it.
China coordinates, but at what cost
China presents an opposite institutional profile. The central state has the capacity to decide, finance, and build data centers without passing through the procedures that block federal democracies. Reports suggest rapid expansion of AI-dedicated infrastructure in China, with indications of coordination between major national operators and provincial authorities.
But China encounters its own institutional barriers, of a different nature. Access to advanced semiconductors remains limited by American export restrictions. Models developed on less performant chips must find alternative architectures, which costs time and engineers. Centralized coordination accelerates physical construction but does not resolve the technological gap in components.
Frey would observe that centralized regimes solve the coordination problem for scaling, but at the price of reduced exploration: fewer independent laboratories, fewer competing bets, and thus fewer chances of finding the next technological breakthrough. China builds infrastructure quickly for current models. It may be less well positioned to invent the next generation.
The three major powers encounter difficulties combining decentralized exploration and institutional scaling. This is the heart of the problem.
What is at stake by 2030
The question posed by the PJM bottleneck goes beyond connecting a few hundred data centers. It concerns the window of time available for AI to produce the growth cycle its proponents announce.
Productivity gains from AI do not materialize solely in training models. They require massive deployment in businesses, hospitals, administrations, and factories. This deployment requires computing infrastructure that is available, reliable, and accessible at reasonable cost. If interconnection capacity remained blocked over an extended period, infrastructure deployment would be delayed, and with it, productivity gains.
Frey identifies in How Progress Ends moments when promising technology collides with unsuitable institutions and slows without ever delivering on its initial promise. American electricity took forty years to spread through factories after its invention, for lack of institutions capable of organizing grid access. AI could reproduce this pattern on a compressed timescale.
Two trajectories are plausible for 2030-2035, without either being certain.
In the first, the United States reforms its interconnection procedures sufficiently, via a combination of the Defense Production Act, new federal permitting rules, and negotiations between states and grid operators, to unblock a significant portion of waiting capacity by 2027-2028. Productivity gains begin appearing in macro data from 2028-2029 onward. Europe, noting its infrastructure lag, launches a common financing mechanism, likely backed by the European Investment Bank, to build sovereign data centers in the best-connected energy hubs: Scandinavia, the Iberian Peninsula, and eastern Germany. The window remains open.
In the second, institutional conflicts persist. The most populous American states maintain regulatory requirements incompatible with federal timelines. The Defense Production Act produces isolated effects without systemic reform. Data centers partially relocate to less constrained states, but this redistribution would not suffice to meet needs. Europe combines regulation and infrastructure investment, even as the Commission acknowledges that available capacity remains insufficient against demand.
China remains confronted with certain technological dependencies, but the performance gap between AI models and those of the United States has narrowed. The three powers enter the 2030s with unresolved institutional problems, affecting the scaling up of computing capacity.
The signals that would distinguish these two trajectories are readable now. The first: the number of GW actually interconnected in PJM by end of 2027. The second: the existence or nonexistence of a dedicated European financing mechanism for sovereign computing infrastructure. The third: China’s capacity to develop competitive model architectures on domestic chips, something MERICS publications track quarter by quarter.
None of these signals are yet settled. AI is the most explored technology of recent history and one that tests scaling institutions substantially reformed since the twentieth century, even if their current procedures remain contested. As the analysis on robotics in territories showed, what drives disparities comes down to the capacity to deploy technology where it is needed.
Institutional reform is not glamorous. It does not make headlines at technology conferences. But the significant queue of interconnection projects at PJM reminds us that institutional reform, ultimately, constitutes an essential factor for enabling technology to produce its effects.
Sources
- Zone Armée, “Political Conflicts Slow Data Center Development and Energy Infrastructure in the United States,” September 2026, https://www.zonearmee.com/conflits-politiques-freinent-le-developpement-des-centres-de-donnees-et-linfrastructure-energetique-aux-etats-unis/
- Carl Benedikt Frey, How Progress Ends: Technology, Innovation, and the Fate of Nations, Princeton University Press, https://press.princeton.edu/books/hardcover/9780691233079/how-progress-ends
- PJM Interconnection, interconnection queue data 2026 (pjm.com)
- MERICS, reports on AI infrastructure in China, 2026 (merics.org)
- European Commission, AI Act, Regulation (EU) 2024/1689 (eur-lex.europa.eu)
