Much of the public debate about artificial intelligence begins with a dramatic question: what happens if AI becomes powerful enough to threaten the people who created it? The question has generated an enormous literature on alignment, regulation and machine control, but it may be overshadowing a less cinematic vulnerability that could arrive much earlier.
What happens if businesses, hospitals, governments and entire industries reorganize themselves around the assumption that machine intelligence will always be available, only to discover that the electricity and computing capacity behind it are no longer as abundant as their operating models require?
This is not an argument that a future energy crisis will suddenly switch off every AI system on Earth. A disruption large enough to do that would also cripple telecommunications, finance, transport, water systems, hospitals and much of modern industry; the loss of generative AI would hardly be the defining problem. The more plausible scenario lies between uninterrupted abundance and total collapse. Power remains available, but grid connections become harder to obtain, electricity becomes more expensive, large users are asked to reduce demand, and some forms of computing must be delayed, shifted, downgraded or rationed.
That possibility matters far more in a society that has already redesigned itself around permanently available AI than it does today. The deeper vulnerability is therefore not simply how much electricity AI consumes, but how much human and organizational capacity may disappear because institutions assume the machines consuming that electricity will remain continuously available.
AI looks like software, but its foundations are industrial
For most users, artificial intelligence appears almost weightless. A question goes into a browser and an answer comes back seconds later, giving little indication that the interaction ultimately depends on semiconductor factories, transmission lines, transformers, substations, cooling systems, backup generators, networks and warehouses full of specialized computers.
Those physical requirements are becoming harder to ignore. The International Energy Agency estimates that data centers consumed about 485 terawatt-hours of electricity worldwide in 2025 and projects roughly 950 TWh by 2030 in its central case, equivalent to around 3 percent of global electricity demand. Electricity consumption at AI-focused data centers grew by about 50 percent in 2025 and is projected to roughly triple between 2025 and 2030. The same IEA analysis also provides an important corrective to overheated claims that AI is about to consume the world’s power supply: data centers remain a relatively small share of global electricity consumption, while improvements in hardware and software efficiency continue at an extraordinary pace.
The global percentage, however, conceals the problem that utilities actually have to solve. Data centers are not distributed evenly across an abstract worldwide grid. They arrive as enormous concentrated loads connected to particular transmission systems, often in regions where generation capacity, transformers and grid connections were planned years before the current AI boom.
Berkeley Lab‘s 2026 update illustrates the difference. Its reference-case estimate puts U.S. data-center electricity consumption at 649 TWh in 2030, equivalent to about 11.8 percent of total U.S. electricity use, while its uncertainty scenarios range from roughly 9.5 to 15.3 percent. The width of that range matters because nobody knows precisely how quickly AI adoption, hardware efficiency and data-center construction will evolve, but utilities still have to make multibillion-dollar infrastructure decisions while those variables remain unsettled.
PJM Interconnection, which operates the electricity market across parts of 13 states and the District of Columbia, offers an unusually clear example of the timing problem. PJM has identified data centers as a primary source of recent demand growth and has warned that new large loads can arrive two or three times faster than many of the generating resources needed to serve them. In July 2026, its auction for the 2028–29 delivery year secured 138,318 MW of capacity but still finished 6,831 MW below the system’s reliability requirement. That result does not make blackouts inevitable, but it shows why rapidly expanding digital demand can collide with physical infrastructure whose planning and construction cycles are measured in years rather than model releases.
The mismatch is fundamentally one of time: AI models, servers and data-center projects can change within months or a few years, while major generation, transmission projects and specialized grid equipment generally require much longer planning, permitting and construction cycles.
Technology companies are already behaving as if power is a strategic constraint
If electricity were merely an abstract future concern, technology companies would have little reason to reorganize the way they obtain it. Their behavior increasingly suggests otherwise.
In February 2026, The Washington Post reported on a growing group of large U.S. data-center projects pursuing their own generation rather than relying entirely on conventional utility connections. Its investigation identified projects across Texas, New Mexico, Pennsylvania, Wyoming, Utah, Ohio and Tennessee, many relying heavily on natural gas because developers wanted large amounts of dependable electricity faster than existing grids could provide it. Meta, OpenAI, Oracle and other companies were among those associated with projects using or considering versions of this approach.
A New York Times investigation published the following month documented the same pressure from another angle, describing technology companies and data-center developers turning to gas engines, mobile generators, batteries and other alternatives while confronting long waits for conventional grid connections. The significance is not that the American electricity system is collapsing; it plainly is not. The more revealing point is that some of the world’s best-capitalized companies are willing to build pieces of the energy system themselves because access to power has become important enough to affect how quickly computing capacity can be deployed.
There are legitimate objections to interpreting every private-generation project as evidence of crisis. Developers argue that onsite power can reduce pressure on utility customers and allow projects to proceed without forcing the grid to finance every new connection, while critics argue that data-center developers may still compete with utilities for turbines and other scarce equipment and that isolated generating systems create their own reliability and environmental risks. The disagreement is useful because it prevents a simplistic conclusion: AI does not face a universal shortage of electricity, but reliable power at the right location and on the right schedule is becoming a material competitive constraint.
That constraint also reaches household economics. The Washington Post documented disputes in several U.S. states over whether ordinary electricity customers were carrying part of the infrastructure costs created by very large data-center loads. Its reporting also noted the necessary caveat: fuel prices, aging plants, extreme weather, supply-chain problems and interest rates can all raise electricity prices independently of AI. Data centers may compound those pressures without being the sole cause of them.
An energy crisis would probably ration AI before it eliminated it
The most plausible failure mode is therefore not a global blackout of machine intelligence. Modern data centers use redundant feeds, batteries and backup generation; large cloud providers distribute infrastructure geographically; some workloads can be moved across regions or shifted to different times of day; and smaller models can replace larger ones when capacity becomes expensive or constrained.
There is now direct evidence that at least some AI workloads can behave this way. A field demonstration published in Nature Energy tested a software-based power-management system on a 256-GPU cluster running representative AI workloads in a hyperscale cloud facility in Phoenix. During periods of peak grid demand, the system reduced cluster power use by 25 percent for three hours while maintaining its specified quality-of-service guarantees, without adding batteries or modifying the underlying hardware. The experiment was limited in scale and should not be extrapolated to the entire AI industry, but it establishes an important mechanism for this discussion: computing capacity can become a flexible load before it becomes an unavailable one.
Electricity systems also have mechanisms for reducing very large loads before the grid itself fails. Texas provides a useful example. Senate Bill 6 created requirements and procedures affecting certain very large transmission-level customers and directed ERCOT to develop ways for large loads to reduce demand during anticipated emergencies. During Winter Storm Fern in January 2026, the U.S. Department of Energy also authorized ERCOT, under specified emergency conditions, to direct backup generators at data centers and other large commercial and industrial facilities to operate. ERCOT later reported that it never entered emergency conditions severe enough to use that authority, an important distinction because the event demonstrates emergency planning rather than an actual forced data-center shutdown.
If electricity were seriously constrained in the future, computing would therefore be more likely to acquire priorities than simply disappear. Training runs could be postponed, batch jobs shifted to off-peak periods, larger models replaced by smaller ones and low-value consumer applications throttled before critical medical, cybersecurity or grid-management systems were allowed to fail.
At that point, a question that abundant computing allows us to avoid would become unavoidable: which forms of machine intelligence justify scarce electricity? A radiology triage system and an AI-generated advertising video are both computing workloads, but few societies would treat them as equally important during a severe shortage. Scarcity would turn what is now largely an economic allocation of compute into an operational and eventually political one.
The more consequential problem, however, would have been created before the shortage began.
The vulnerability is created during the years when AI works well
Imagine a company that can produce the same output with 6,000 employees and extensive AI assistance that previously required 10,000 people. There is nothing irrational about taking advantage of that productivity improvement. Competitors will do the same, investors will reward lower costs, and the remaining workers may genuinely spend less time on repetitive work.
Over time, however, the organization changes in ways that cannot be reversed simply by switching software off. Manual procedures are used less often and stop being updated. Entry-level employees no longer learn every stage of a process because the machine handles some stages from the beginning. Experienced workers retire and take tacit knowledge with them. Internal systems are redesigned around automated classification, summarization, coding or decision support, while staffing levels gradually come to reflect the productivity of humans and machines operating together.
Research on AI-related deskilling has begun to examine this structural problem. Avigail Ferdman, writing in AI & Society, describes “capacity-hostile environments” in which people have fewer opportunities to acquire, exercise and transmit capabilities because automated systems increasingly mediate the activities through which those capabilities were once developed. The value of this argument is that deskilling does not have to arise because individuals become lazy or careless; it can emerge gradually from an environment that no longer requires people to practice certain skills in the first place.
Once that process has gone far enough, an AI-dependent organization is no longer an ordinary organization that happens to use sophisticated software. Its normal productive capacity has become the combined result of human labor, machine intelligence, network access and electricity.
If machine capacity is suddenly cut in half, the former workforce does not reappear with it. Skills that have not been practiced for years cannot necessarily be restored during an emergency, and workflows eliminated for efficiency cannot always be reconstructed on demand. A technical loss of capacity can occur in hours; rebuilding institutional capability may require months or years.
This difference between technical recovery and organizational recovery is the central risk.
A hospital makes the distinction easier to see
Consider a future hospital in which physicians still diagnose patients and nurses still provide care, but AI has become routine in radiology triage, scheduling, coding, record summarization, medication screening, insurance documentation and bed management.
No single application needs to replace a physician for the institution’s staffing model to change. If many small tasks become faster, fewer administrative staff may be required, clinicians may handle larger caseloads, and the hospital may gradually organize itself around an amount of throughput that would have been impossible under its previous manual processes.
Now suppose a regional energy emergency or infrastructure failure sharply reduces access to external AI computing for several days. The hospital still has electricity. Its physicians have not forgotten medicine, and its nurses still know how to care for patients. Nevertheless, hundreds of small cognitive and administrative tasks return to human hands at the same time. Reports take longer, scheduling slows, documentation accumulates, clinicians spend more time retrieving and processing information, and patient throughput falls.
The resulting failure is not primarily a loss of medical knowledge but a loss of operational capacity. Modern societies depend not merely on somebody knowing how to perform an essential task, but on institutions performing enormous numbers of those tasks at the required speed. If AI allows organizations to increase that speed and then to remove some of the people and processes that once supplied equivalent capacity, reduced access to AI becomes fundamentally different from losing an optional productivity tool.
Electricity scarcity would then become, indirectly, a constraint on usable cognitive capacity.
AI dependency turns an energy constraint into a capacity problem
Most discussion of AI and electricity asks how much power data centers will require, where the electricity will come from, whether household bills will rise and how much carbon will be emitted. Those questions are important, but they stop one step before the problem considered here.
The deeper issue is what happens when electricity increasingly supports machines to which economically valuable cognitive work has been delegated.
A small but growing body of research has begun to describe AI as a form of “cognitive infrastructure,” meaning systems that increasingly mediate how people search, interpret information, plan and make decisions. The terminology is still emerging rather than settled, and it should not be treated as an established scientific consensus, but the concept is useful for one narrower reason: it highlights the possibility that AI can move from being an optional tool to becoming something institutions quietly build normal operations around.
The physical side of that transition deserves equal attention. Expertise distributed across people and institutions also depends on infrastructure, education, communications systems and organizational routines, so it would be misleading to portray human cognition as infrastructure-free. The relevant difference is concentration. Machine-assisted cognition can depend heavily on a comparatively small number of data centers, cloud platforms, model providers, semiconductor supply chains and power systems. When many organizations share those dependencies, a failure or constraint at one layer can propagate much farther than an isolated tool failure.
This is where an energy problem becomes something more than a utility problem. If society converts part of its working cognitive capacity from people and local procedures into machine services faster than it builds ways to function when those services become scarce, then energy or compute constraints can produce a much larger loss of effective capacity than the electricity shortage alone would suggest.
That is the central argument of this article, and it does not require AI to malfunction.
Efficiency may reduce the energy problem without eliminating the dependency problem
Any serious discussion of AI energy demand has to account for the speed at which the technology is becoming more efficient. The IEA estimates that energy use per AI task has fallen by at least an order of magnitude annually in recent years for some applications. Simple text generation now requires far less electricity than many widely circulated comparisons imply, while reasoning models, video generation and agentic systems can consume hundreds or thousands of times more energy per task than basic text generation.
This makes long-range electricity forecasts unusually uncertain. Better chips, model compression, routing, quantization and smaller specialized models could reduce the amount of power required for a given amount of useful work, and some of today’s most alarming projections may turn out to be too high.
Efficiency, however, addresses consumption rather than dependency. If an AI task becomes ten times cheaper and society responds by using it one hundred times more often, total demand can still rise. More importantly, a highly efficient service can still become indispensable. GPS requires little energy at the point of use, yet modern aviation, shipping, telecommunications and financial systems depend heavily on the timing and positioning services built around it. The vulnerability comes from what society builds around a service, not merely from the amount of electricity the service consumes.
AI may eventually create the same distinction between low operating cost and high systemic importance.
Resilience does not require preserving every old job
There is an obvious way to take this argument too far. If technological dependency creates risk, one could conclude that society should retain large numbers of employees and obsolete procedures merely in case automation fails.
That would be economically wasteful and, in many cases, counterproductive. Humans make errors, become tired and often perform routine information processing badly. AI can improve safety, reduce costs and allow scarce expertise to be used where it is most valuable. A hospital that refuses useful automation merely to preserve every old administrative workflow could become less efficient without becoming meaningfully more resilient.
The better engineering principle is graceful degradation: a system should be able to lose capability without losing its essential functions.
An AI-dependent hospital should know which operations must continue if external model capacity falls sharply. A bank should know which decisions require a local or manual fallback. A government agency should know which essential services depend on a single model or cloud provider. Some organizations may need smaller local models for emergency use, while others may need to preserve human expertise only in a narrow set of genuinely critical tasks rather than maintain duplicate staffing for every automated function.
The difficult part is economic because redundancy always looks inefficient during years in which nothing fails. Spare electrical capacity costs money, backup systems require maintenance, manual procedures require training, and employees retained partly for resilience appear less productive than a leaner competitor’s workforce. Competitive markets are very good at identifying unused capacity and removing it, which is usually a source of efficiency but can also reduce the ability of a tightly optimized system to absorb an unexpected loss of capacity.
AI may push that familiar efficiency-versus-resilience tradeoff into a domain where societies have less experience: cognition itself.
AI competition is increasingly becoming infrastructure competition
The global AI race is normally described in terms of algorithms, chips, models, data and talent, but electricity is steadily joining that list. The IEA now describes a scramble not only for chips and capital but also for electricity, grid connections and energy equipment, while technology companies increasingly treat generation capacity and grid access as part of the AI supply chain.
This matters because the performance of an AI system is no longer determined only by the quality of an individual processor or model. At scale, computing depends on the performance of entire systems combining processors, networking, memory, cooling, software, buildings and power infrastructure. A country can possess excellent researchers and advanced chips yet still struggle to deploy large quantities of computing if it cannot deliver sufficient reliable electricity to the places where those machines operate.
A company whose models perform the same useful work with half the electricity therefore gains more than a lower utility bill. Under constrained conditions, it can obtain more computational output from the same physical infrastructure.
One day we may care as much about useful machine intelligence per megawatt as we now care about model performance per dollar. There is no accepted metric for “useful intelligence,” and the phrase should not be mistaken for an established technical measure, but the economic direction is clear: once computation becomes an industrial-scale consumer of power, energy efficiency becomes part of strategic AI capacity.
The risk may become serious only after AI stops feeling remarkable
New technologies attract the most attention while they are unfamiliar. Dependency usually develops later, after people stop noticing them.
Modern societies do not wake each morning astonished that the electrical grid works, that GPS satellites remain overhead or that cloud servers are available. Those systems became critical because businesses and institutions quietly redesigned themselves around their reliability.
AI may follow the same pattern. If it does, the important threshold will not necessarily be the day a model surpasses another benchmark or displays a dramatic new capability. A more consequential transition may occur gradually, when organizations conclude that maintaining people and procedures capable of performing certain tasks without AI no longer makes economic sense.
An energy shortage before that transition would mainly reduce productivity. The same shortage after it could remove part of an organization’s effective operating capacity because the old human and procedural substitutes would no longer exist at sufficient scale.
The useful question is therefore not whether an energy crisis could switch off every AI system. Except during a much larger catastrophe, it probably could not. The more important question is how much of society’s ability to analyze, coordinate and process information we are willing to make dependent on continuously available computation, and how much reduced capacity our institutions could tolerate before ordinary inconvenience became systemic disruption.
AI does not need to take control of society for that problem to matter. It only needs to become useful enough that organizations gradually stop maintaining the alternatives on which recovery would depend.
AI Dependency & Resilience
This essay is part of a three-part series on what happens when machine intelligence becomes ordinary infrastructure: how energy constrains it, what organizations lose when human fallback disappears, and how scarce compute is allocated.
01 — Energy & Availability
When Society Depends on AI, Can It Withstand an Energy Crisis?
02 — Human Fallback
What Happens When Companies Forget How to Work Without AI?
03 — Compute Allocation
When AI Compute Becomes Scarce, Who Gets It First?