AI data centers threatened with outages in the United States

The power grid, artificial intelligence's new hardware frontier

The global race for artificial intelligence is often portrayed as a battle of models, graphics chips and capital. Announcements of new accelerators from Nvidia, AMD or Google, the massive investments of Microsoft, Meta, Amazon and OpenAI, as well as the construction of gigantic computing campuses, have placed GPUs at the center of attention. But another physical factor is now asserting itself with growing force: access to electricity.

In the United States, this constraint is taking on a particularly concrete dimension in the area covered by PJM Interconnection, the operator of the country's largest power grid. According to information published by TechCrunch under the headline “Data centers may face temporary power cuts to prevent blackouts on largest US grid”, PJM is considering a mechanism that would allow temporary outages or consumption reductions for certain customers, notably data centers, in order to preserve the stability of the power system during critical periods.

The issue goes far beyond the purely technical question of load shedding. It illustrates a transformation in the hierarchy of resources needed to develop AI. Having land, fiber optics, servers and processors is no longer enough. It is also necessary to obtain a sufficiently powerful power connection, within a timeframe compatible with the pace at which digital infrastructure is deployed, and to ensure that this supply will remain available during demand peaks or periods of strain on the grid.

The grid operated by PJM covers all or part of 13 U.S. states as well as the District of Columbia. It coordinates the balance between electricity production and consumption in real time across a vast area of the eastern and Midwestern United States. Its role is therefore fundamental: a local failure can spread, and the operator must prevent generation or transmission constraints from escalating into a broader outage.

In this context, the prospect of asking data centers to temporarily reduce their consumption represents a significant shift in tone. Data centers have long been regarded as particularly attractive industrial customers for regions: they promise investment, tax revenues, construction projects and a lasting presence by major technology companies. Their growth now requires grids to treat those same facilities as electrical loads that can potentially be adjusted in an emergency.

This development is directly linked to the rise of generative AI. Training large language models, generating images or videos, inference at scale and the growing use of automated assistants require considerable computing capacity. Unlike a traditional digital service, modern AI infrastructure often concentrates thousands, or even more, accelerators in buildings designed to operate very dense equipment without interruption. Energy powers not only servers, but also cooling systems, networking equipment, storage devices and safety infrastructure.

The finding is not that every use of AI immediately causes an outage risk. It is that the accumulation of projects in already constrained areas can abruptly alter the assumptions on which grids were planned. For decades, growth in U.S. electricity demand was relatively moderate in several regions. The simultaneous arrival of large cloud computing campuses, AI sites, electrified industrial facilities and new loads associated with the energy transition is restoring a degree of importance to power planning that the technology industry had sometimes underestimated.

PJM facing demand that is growing faster than infrastructure

The mechanism mentioned by TechCrunch follows a classic power-security logic: when available supply or transmission capacity is insufficient to meet all demand, it is better to target reductions in the consumption of certain very large customers than to let the system approach an imbalance liable to cause uncontrolled outages. In a power grid, the balance between production and consumption is not an operational preference; it is a physical necessity.

Load reductions, sometimes called load shedding or demand response depending on their framework and degree of voluntariness, have existed for a long time in power systems. Industrial sites can agree to temporarily reduce their consumption in exchange for compensation or under specific contracts. What is drawing attention today is the possible application of this logic to the data center sector, which has become one of the symbols of new electricity demand linked to AI.

TechCrunch reports that PJM is considering temporary outages in order to prevent failures on its grid. The essential point is the preventive nature of the approach: the aim is not to punish data center operators or call their grid connection into question, but to have an instrument of last resort to protect the entire system when it is under exceptional pressure.

For technology companies, the distinction is important, but it does not remove the challenge. A data center is generally designed to provide high availability. Online AI services, cloud platforms, enterprise applications, payment systems, digital public services and communications infrastructure all rely on architectures that seek to limit interruptions. A power reduction may therefore involve shifting workloads to other regions, slowing certain non-urgent calculations, activating backup generators or reducing particularly energy-intensive computing operations.

The ability to do so varies greatly from one operator to another. Certain tasks, such as model training or some internal batch processing, are more flexible than production services used continuously by millions of people. Cloud providers already distribute their infrastructure across multiple regions and availability zones. However, this redundancy comes at a cost, and it does not guarantee that capacity will always be available elsewhere at the exact moment when a region faces a power constraint.

The issue is also geographical. A national or continental grid does not operate like a perfectly interchangeable electricity reserve. Transmission lines have their own limits, interconnections between regions are not infinite, and power needs are sometimes concentrated in areas close to major digital hubs. Building a data center where electricity is theoretically abundant does not automatically solve the issue: the power must still be delivered to the site, the appropriate connection must be obtained and, if necessary, local substations and lines must be reinforced.

PJM is not an isolated case in its confrontation with the rise in loads linked to data centers. Grid operators and regulators are observing, in several regions, connection queues, grid reinforcement needs and disagreements over how the cost of new infrastructure should be allocated. But PJM's size makes the signal particularly visible. When an operator of this scale examines the possibility of adjusting the consumption of very large customers to prevent a blackout, the debate leaves the abstract realm of energy forecasts and enters that of the system's concrete operation.

The question is no longer only how many data centers will be built, but in which areas, at what pace, with what guaranteed capacity and with what obligations during periods of strain.

This situation also reveals a timing mismatch. AI companies can announce new computing clusters, order servers and sign real estate contracts in a few months. Core electrical equipment, meanwhile, often requires permitting procedures, studies, construction work and investments that extend over several years. The availability of accelerators is therefore tied to that of the grid: an unpowered GPU produces no token, no image, no prediction and no commercial value.

After GPUs, electric power becomes the bottleneck

Since 2023, the shortage of specialized AI accelerators has shaped the strategy of major technology companies. Nvidia has benefited from this demand with its chips for training and inference, while hyperscalers have developed or strengthened their own components, such as Google's TPUs, AWS's Trainium and Inferentia chips, and Microsoft's Maia accelerators. These initiatives respond to imperatives of cost, performance and technological sovereignty.

Yet multiplying chips only shifts part of the problem. A modern computing cluster is a complete system: electrical supply, power conversion, high-speed networking, cooling, buildings, orchestration software and operations staff. Efficiency gains per chip can reduce the energy needed for a given operation, but they do not necessarily prevent total consumption from increasing when computing volumes grow faster than unit gains.

This dynamic is well known in the digital economy: when a technology becomes more efficient and less costly to use, it can also be deployed in more products and services. With generative AI, uses are multiplying. Companies are no longer limited to training a base model; they also run inference for users, tailor models to specific needs, process new types of content and integrate AI into business processes that had previously seen little automation.

From this perspective, the proposal being considered by PJM marks an important step. Electricity no longer appears merely as a cost item in data centers' operating accounts. It is becoming a resource capable of determining the very ability to market AI services at scale. For a platform, uncertainty over the power supply can translate into uncertainty over service availability, the price of computing, the commissioning schedule for a campus and the location of future investment.

Major cloud groups have considerable advantages in responding to this challenge. They can negotiate electricity contracts, sign power purchase agreements, distribute their loads across multiple regions and finance dedicated infrastructure. They can also design their software to identify deferrable computations. But these capabilities do not remove reliance on the public grid. Even when an operator secures low-carbon generation by contract, the electricity consumed must be physically integrated into a transmission and distribution system whose stability depends on collective rules.

The U.S. debate also highlights a distinction often blurred in corporate communications: buying or producing enough energy on an annual basis does not necessarily mean having the necessary power at every moment and in every location. Grids must meet demand peaks, manage power plant outages, anticipate extreme weather events and maintain a reliability margin. A facility may have an ambitious energy strategy while still depending on a saturated local connection.

For data center developers, the consequence is potentially profound. The cheapest land or the site closest to a large market will not necessarily be the best location. Location criteria must include immediately accessible electrical capacity, required work, connection timelines, exposure to load-reduction mechanisms, the availability of water or other cooling means, and the possibility of connecting the site to multiple power sources.

The PJM case also shows that public policy and energy regulation are returning to the foreground in the strategy of digital players. AI companies depend not only on a global semiconductor supply chain. They also depend on local decisions: building permits, connection rules, line planning, tariffs, capacity markets and the operating conditions of grid operators. Building AI infrastructure is not merely an IT operation; it is becoming an industrial and territorial operation.

This reality may benefit companies able to plan far ahead and commit significant resources. Conversely, it may complicate the position of smaller players, specialized cloud providers or startups that had counted on rapid availability of computing capacity. If electricity becomes scarce in certain areas, access to compute could be increasingly determined by geography, long-term contracts and the financial capacity to absorb grid costs.

A warning for Europe and French projects

The PJM case cannot be mechanically transposed to France or the European Union. Electricity markets, generation mixes, the responsibilities of grid operators and regulatory frameworks differ. France notably has a power system heavily shaped by nuclear energy, while Europe relies on interconnections between states and on an institutional organization distinct from that of the United States. It would therefore be imprecise to present the U.S. situation as an automatic indication of equivalent outages in France.

Nevertheless, the signal is relevant. Europe is also seeing growth in cloud infrastructure and data center projects. Computing needs associated with AI, digital sovereignty requirements, data localization and the desire to develop European capacity are prompting companies and public authorities to take a closer interest in physical infrastructure. The availability of land and fiber remains essential, but it cannot be separated from the energy issue.

In France, the balance of the power system is managed by RTE on the transmission grid, while distribution is handled in particular by Enedis across most of the country. For a data center project, feasibility is not limited to a national average of electricity generation. It depends on the required voltage level, proximity to suitable infrastructure, the state of the local grid, necessary reinforcements and the connection schedule. Announcements of digital campuses can therefore run up against infrastructure delays that do not match the speed of technology cycles.

The issue is particularly sensitive around major areas of digital connectivity. Data centers have historically tended to concentrate where users, telecommunications networks, Internet exchange points and business customers are located. This concentration improves latency and simplifies interconnection, but it can increase pressure on areas that are already heavily used. As AI needs grow, trade-offs between market proximity, available power and local acceptability become more difficult.

The European Union has already begun to increase visibility into the sector's energy consumption. The revised European Energy Efficiency Directive notably provides reporting obligations for operators of data centers exceeding a certain threshold of installed IT power. The goal is to obtain comparable data on the energy and environmental performance of these facilities. This approach does not resolve the connection issue, but it shows that data centers are now considered energy infrastructure as much as digital infrastructure.

For the French-speaking AI market, the challenge is twofold. On the one hand, companies developing models, assistants or inference services need competitive access to computing. On the other hand, France and Europe are seeking to retain room for maneuver in an industry dominated by major U.S. groups and, increasingly, Asian players. If new electrical capacity or connections become difficult to obtain, this could slow the creation of local computing capacity and strengthen dependence on foreign infrastructure.

The development of sovereign AI is therefore not played out solely in model quality or access to data. It also plays out in transformers, high-voltage lines, substations, cooling systems and administrative procedures. This dimension is sometimes less visible than model launches, but it is decisive. A sovereign computing policy without sufficiently robust energy and grid planning risks running up against a physical limit.

Public officials will also have to arbitrate between different uses of electricity. Transport electrification, heating through heat pumps, reindustrialization, hydrogen production and new digital needs can accumulate. Presenting data centers as an isolated sector would be misleading: they are part of a more general increase in power demand and a transformation of the electricity system. But their profile, characterized by very high concentrated power requirements and an expectation of continuous availability, makes them a particularly demanding case.

For local authorities, the arrival of a large data center can represent both an economic opportunity and a planning challenge. Projects require construction work, consume land and may raise questions about water use, noise, waste heat or the actual contribution to local employment. The AI argument now adds another dimension: the infrastructure does not merely host websites or storage; it supports strategic computing capacity that may be sought after by companies and public administrations.

Flexibility, resilience and transparency: possible responses

PJM's consideration puts data center operators before a precise operational question: which loads can be interrupted, slowed or moved without endangering essential services? The answer cannot be uniform. A hospital, critical telecommunications infrastructure or a public service cannot be treated in the same way as model training scheduled over several days. Even within a data center, not all workloads have the same level of priority.

This differentiation opens the way to more refined management of electricity demand. Companies can separate critical loads from deferrable processing, schedule certain computations at less strained times or distribute training and inference across different sites. This kind of organization already exists in various forms in the cloud, notably to optimize costs and capacity. Pressure on grids could give it new importance.

However, the theoretical flexibility of data centers should not be overstated. Moving a computing load entails constraints involving networks, data, server availability, latency, security and compliance. A model trained at one site cannot always be moved instantly to another. Data volumes can be immense, distributed architectures complex and contractual commitments strict. Flexibility is a technological resource that must be built, not a switch that can be flipped at no cost.

Backup systems are another component of resilience. Data centers usually have uninterruptible power supplies and generators intended to keep operations running during an outage. But this equipment is primarily designed to ensure service continuity, not to durably replace grid power on a regional scale. Its use also raises questions about fuel, emissions and operating duration. The idea of systematically relying on backup solutions does not replace the need to invest in collective electrical infrastructure.

Over the longer term, AI growth will make closer coordination necessary among grid operators, energy producers, real estate developers and technology companies. Data center projects will need to incorporate the question of their consumption profile at a very early stage. For their part, grid managers will need credible forecasts, because not all project announcements turn into actually built facilities. A poor estimate, whether too high or too low, can lead to poorly sized investments or unnecessary strain.

Transparency becomes a central issue here. Grids need to understand which projects are firm, how much power they request and on what date. Companies need to know the actual conditions of their connection, including possible limitations in a crisis. Regulators must be able to assess who finances the necessary reinforcements and how costs are distributed between new customers and existing consumers. Without clear rules, the risk is turning access to electricity into an opaque advantage reserved for the most powerful groups.

The temporary consumption reduction considered by PJM can also be interpreted as a form of social contract between very large consumers and the rest of the grid. If a data center benefits from a high-capacity connection in a constrained area, should it contribute more to the system's flexibility? The answer will depend on regulatory frameworks, contracts and political choices. But the debate will be difficult to avoid, especially when households and small businesses fear higher costs or declining reliability.

For AI players, the strategic message is clear: raw performance is no longer enough. A model may be more efficient, a chip faster and software better optimized; without electricity available in the right place, the advantage remains theoretical. In the years ahead, the sector's most important announcements may therefore concern energy partnerships, connection capacity and grid infrastructure as much as new foundation models.

Toward an energy-determined geography of AI

The U.S. episode revealed by TechCrunch does not mean that data center expansion will stop. Rather, it indicates that this expansion is entering a more mature phase, in which invisible infrastructure is becoming a first-order constraint. The AI industry has already learned that semiconductor manufacturing capacity cannot be increased instantly. It is now discovering that power grids, too, are subject to long timelines, physical rules and complex decision-making processes.

This development could alter the global map of computing. Regions that combine abundant electricity generation, robust grids, rapid planning, high-quality digital interconnections and stable regulation will have an attractive advantage for cloud operators and AI companies. Conversely, areas where connections are saturated or unpredictable could see certain projects postponed, resized or moved.

For France and Europe, the challenge is not only to attract data center buildings. It is to decide what place this infrastructure should occupy in a coherent industrial, energy and digital strategy. AI can support productivity gains, scientific research, healthcare, industry and public services. But its benefits do not remove the need to measure the resources required for its operation or to organize fair access to them.

The possibility being studied by PJM of temporarily reducing data center consumption during periods of strain recalls a truth often obscured by the vocabulary of the cloud: computing is never immaterial. Behind every request sent to a model are chips, cables, transformers, power plants and grids. As AI becomes general infrastructure, the quality of these electrical foundations will determine an increasing share of its deployment speed, cost and resilience.

The next competitive advantage may therefore not be determined solely by the number of installed GPUs. It could depend on the ability of companies and states to align computing power with electrical power, without weakening the grids on which all economic and social activities depend.