The AI Data Center Power Crunch Is Now a Real Grid Crisis

On April 16, 2026, the Federal Energy Regulatory Commission told the public it would act by the end of June on a rulemaking that most Americans have never heard of, yet which will quietly shape the price they pay for electricity and the speed at which artificial intelligence gets built in their country. The docket, labeled RM26-4-000, would standardize how very large electricity users such as data centers connect to the interstate transmission grid. It exists because the AI data center power crunch has stopped being a forecast and become a present-tense operational problem that utilities, regulators, and tech companies are fighting over in public. The fight is no longer about whether AI needs enormous amounts of electricity. It is about who gets the power first, who pays for the grid upgrades, and whether the grid itself can absorb the load without breaking.

The headlines have framed this as a story about politics and electricity bills, and that framing is not wrong. Across the country, the prospect of rate hikes to support tech company expansion has prompted a significant public backlash. Energy Secretary Chris Wright invoked a rarely used authority under the Department of Energy Organization Act in October 2025 to direct FERC toward new federal rules, and the move set off a jurisdictional struggle between federal transmission authority and state regulators that is still unresolved. What that coverage tends to miss is the part that actually determines the outcome, which is technical. The reason this became a crisis, and the reason it might be solvable, both live inside the engineering of AI workloads themselves.

This is the dimension that a publication covering both energy and AI is positioned to explain. The demand curve that is straining the grid is not generic industrial load. It is the specific, measurable electricity consumption of training runs and inference serving, and those two activities behave very differently on the grid. Whether an AI data center can flex its power draw when the grid is stressed depends on whether it is training a model or serving one, on what kind of model it is, and on architectural choices that engineers make months before a facility ever requests an interconnection. Regulators are now writing rules that hinge on these distinctions, and most of the people debating the rules do not understand them.

This article explains the AI data center power crunch from both ends at once. It covers what the FERC rulemaking actually proposes and why the deadline matters. It explains the engineering of why AI compute consumes the power it does, and why the shift from training to inference is changing the shape of the demand. It examines the central technical dispute over whether data centers can genuinely curtail their power use, the dispute that one grid monitor bluntly called a regulatory fiction. It assesses what this means for the cost and geography of AI deployment, what businesses should understand before they commit to an AI strategy that assumes cheap and available compute, and where this collision between silicon and the power grid is heading next.

What FERC Is About to Decide and Why It Matters

The immediate event is a regulatory deadline. The Federal Energy Regulatory Commission committed on April 16, 2026, to take action by the end of June 2026 on Docket RM26-4-000, an Advance Notice of Proposed Rulemaking aimed at standardizing how large loads connect to the nation's transmission infrastructure. The proceeding traces back to October 2025, when Energy Secretary Chris Wright used Section 403 of the Department of Energy Organization Act, a rarely invoked authority, to direct FERC to consider new rules for connecting large electrical loads to the interstate transmission system. The threshold under discussion defines a large load as demand exceeding 20 megawatts, which captures essentially every serious AI data center being planned today.

The original request from the Department of Energy asked FERC to take final action by April 30, 2026. FERC did not meet that date and instead committed to acting by June, roughly two months later. That two month extension was not a delay born of indifference. It signaled a deliberate effort to build a rule that is fast enough to meet surging demand while being legally durable enough to survive the inevitable court challenges. The decisive design constraint, as legal analysts have noted, is the line between federal and state jurisdiction. FERC regulates the interstate transmission grid, but states have long held authority over retail electricity rates, and a federal rule that reaches into how large loads connect could be read as crossing that line.

The fourteen principles at the heart of the rule

The Department of Energy did not simply ask FERC to act. It supplied a set of fourteen principles to guide the rulemaking, and the substance of those principles reveals what the federal government is trying to achieve. The core idea is to accelerate the interconnection of large loads by treating them more like power generators, which already have standardized interconnection procedures. The proposal would allow customers to file joint, co-located load and generation interconnection requests, meaning a data center and the power plant built to serve it could be studied together rather than separately. The stated goal is to significantly reduce study times and grid upgrade costs.

Several of the principles point in the same direction, toward flexibility. At least six of the fourteen suggest that FERC and the transmission providers it regulates should facilitate, and even encourage, the co-location of new generation or storage with large loads so that those loads present a flexible or curtailable net demand to the grid. One principle states that load and hybrid facilities should be studied together with generating facilities to the extent practicable. Another contemplates expedited studies specifically for curtailable loads. The logic is straightforward. A data center that can reduce its power draw during grid stress, or that brings its own generation, is far easier to integrate than one that demands constant uninterrupted power around the clock.

The jurisdictional fight nobody can avoid

The National Association of Regulatory Utility Commissioners, which represents state utility regulators, filed comments arguing that FERC asserting jurisdiction over load interconnection would interfere with state authority over retail rate cases. This is not a technicality. It is the question that will determine whether the rule survives. If FERC draws the line in a way that overreaches, the rule gets challenged and potentially struck down, and the regulatory uncertainty that is already slowing data center development continues. If FERC draws it too cautiously, the rule fails to deliver the acceleration the AI industry is demanding. How FERC resolves this in June will either clarify or further complicate the relationship between federal and state energy authority for years.

The reason any of this commands urgency is that the underlying demand is real and growing fast. Analysts project that nearly 100 gigawatts of new data center capacity will be added globally between 2026 and 2030, effectively doubling global capacity, with the sector expanding at roughly a 14 percent compound annual growth rate through 2030. The same analysis describes an infrastructure investment supercycle that could require up to 3 trillion dollars by 2030. These are not abstract figures. They represent physical power plants, transmission lines, and substations that have to be built and paid for, and the question of who pays is exactly what has turned this into a public controversy.

The Backlash That Turned Compute Into a Political Problem

The political heat behind the FERC rulemaking comes from electricity bills. At a time when the cost of living is already straining household budgets, the prospect of rate hikes to support data center expansion has prompted a backlash across the country. The grievance is intuitive even to people who know nothing about AI. If a tech company builds a facility that consumes as much power as a small city, and the utility has to upgrade the grid to serve it, and those upgrade costs get spread across everyone's monthly bill, then ordinary ratepayers are subsidizing the AI buildout. Whether that is precisely what happens is more complicated, but the perception alone has been politically potent.

The data on rates is genuinely murky, and honest analysis has to acknowledge it. Research cited by energy analysts in 2026 found no clean relationship between higher electricity demand and higher retail rates, noting that bills are rising for many reasons at once, including power plant retirements, inflation, market design, wildfire prevention, and grid modernization. One analysis cautioned that attributing rising bills to a single visible source such as data centers oversimplifies a more nuanced reality. At the same time, it did not hold data centers blameless, and it acknowledged real research gaps on the connection between data centers and rates. The honest summary is that the link is regional, complicated, and contested.

How regulators are already responding

States have not waited for FERC. According to the Edison Electric Institute, as of May 2026, twenty three states had approved at least one large load tariff, with another seven states having tariffs pending. These tariffs are legally binding rate and service rules specifically for large customers such as data centers, designed to determine how much they pay and on what conditions they connect. The central purpose is to require new data centers to bear the incremental costs of the capital they force the utility to spend, rather than socializing those costs across existing ratepayers.

Some jurisdictions have gone further than tariffs. Oregon passed the Protecting Oregonians With Energy Responsibility Act in June 2025, becoming the first state to create a dedicated electricity rate class for data centers. The law applies to facilities consuming more than 20 megawatts and requires ten year minimum contracts with a baseline demand payment regardless of actual consumption, separate cost allocation so that data center infrastructure does not subsidize residential rate increases, and transmission cost pass-through requiring large users to pay for the grid upgrades they trigger. New York lawmakers went in a blunter direction, with legislators proposing a three year pause on new data centers in February 2026.

The pledge with no teeth

The federal executive branch has also tried to manage the backlash. At a White House gathering in March 2026, several prominent tech companies including Microsoft, Meta, OpenAI, and Amazon signed onto a Ratepayer Protection Pledge. The pledge is a voluntary agreement to secure their own power for data centers, to pay for any power lines or other infrastructure utilities need to build to move that power, and to hire locally in the communities where they build. As reporting has noted, the pledge itself has few specifics or enforcement teeth, and the White House did not set up oversight mechanisms to ensure the commitments are honored. It functions more as a political signal than a binding constraint.

This is the environment into which the AI data center power crunch has thrown the country. The demand is real, the costs are real, the backlash is real, and the policy responses so far have been a patchwork of state tariffs, legislative pauses, and voluntary pledges. The FERC rulemaking is the first serious attempt at a federal framework, which is why its June deadline matters well beyond the energy sector. But to understand why this is so hard to resolve, you have to understand the thing generating the demand, and that requires looking inside the AI workloads themselves.

The AI Dimension That Energy Coverage Keeps Missing

The AI data center power crunch is not a generic story about big computers using electricity. It is a story about two distinct kinds of AI computation, training and inference, that have different power profiles and different implications for the grid. Almost all energy coverage treats data center demand as a single monolithic load. That framing is wrong in a way that matters, because the entire regulatory fight over flexibility and curtailment turns on the difference between these two activities.

Training is the process of building an AI model by running enormous volumes of data through a neural network and adjusting billions of parameters over many iterations. A frontier model training run can occupy tens of thousands of specialized chips operating in tight synchronization for weeks or months. The compute used to train frontier AI models has been doubling roughly every five to six months, a pace far faster than the historical improvement of computer chips. Critically, this growth is not coming from making transistors smaller. It comes from deploying ever larger clusters of chips, which means the power demand scales almost linearly with the ambition of the model. A bigger model means a bigger cluster, which means more megawatts.

Inference is different. It is the act of running an already trained model to produce an answer, the moment when you send a prompt and the model responds. Inference for a single query is far cheaper than training, but inference runs continuously, at whatever volume users demand, every hour of every day. As AI moves from research labs into products used by hundreds of millions of people, inference is becoming the dominant driver of demand. In 2025, AI represented roughly a quarter of all data center workloads, with training driving most of that. The projection is that inference will overtake training as the dominant AI requirement around 2027, and that AI could represent half of all data center workloads by 2030. This shift from a training-dominated world to an inference-dominated one is the single most important and least understood fact in the entire power debate.

Why the training-to-inference shift changes everything

The reason this distinction is decisive is that training and inference have very different relationships to time. A training run, in principle, has scheduling flexibility. If the grid is stressed at 6 p.m. on a hot day, a training job can sometimes be paused or slowed and resumed later, because what matters is that the model finishes eventually, not that it computes during any particular hour. This is the technical basis for the entire concept of a flexible or curtailable AI load, and it is what the DOE principles are counting on when they propose to fast-track curtailable data centers.

Inference does not work that way. When a user sends a query to a deployed AI product, they expect an answer in seconds. You cannot tell a customer to wait three hours because the grid is stressed. Inference serving for a live product is, by its nature, far less flexible than training. As inference becomes the dominant workload, the fraction of AI compute that can genuinely curtail its power draw shrinks. The grid is being asked to accommodate a load that is not only enormous and growing but increasingly inflexible. This is the engineering reality sitting underneath the regulatory fight, and it is why the curtailment debate has become so contentious.

The hardware reality behind the megawatts

The physical source of the demand is the accelerator chip, the specialized processor that performs the matrix multiplications at the heart of neural networks. Modern AI accelerators draw far more power per chip than previous generations of processors, and they are deployed in dense racks that concentrate enormous power and heat in small physical footprints. A single large AI training cluster can require the kind of dedicated power infrastructure that historically would have served an industrial plant. This concentration is why a data center campus can show up on a utility's planning horizon as a sudden, gigawatt-scale request that the local grid was never designed to handle.

The density also creates a cooling problem that compounds the power problem. Removing heat from racks packed with high-power accelerators requires substantial additional energy, often delivered through liquid cooling systems that themselves consume power and water. The total facility power draw therefore exceeds the raw chip consumption by a meaningful margin. When a developer requests an interconnection for a large AI campus, the number on the application reflects not just the compute but everything required to keep that compute running and cool. Understanding this is essential for any organization at KriraAI's level that is reasoning about the true cost and footprint of the AI systems it deploys, because the electricity bill and the carbon footprint of an AI product trace directly back to these hardware realities.

The Curtailment Fight That Will Define the Rule

The most revealing dispute in the entire proceeding is about whether AI data centers can actually curtail their power use when the grid needs them to. This is not a side issue. It is the hinge on which the whole federal strategy turns, because the DOE principles propose to give faster grid access to loads that can flex. If flexibility is real, the rule makes sense. If it is a fiction, the rule fast-tracks loads that will strain the grid at exactly the wrong moments.

The PJM Interconnection Independent Market Monitor, which oversees the largest grid region in the United States serving 67 million people, released a report attacking the premise of voluntary data center curtailment. The monitor labeled it a regulatory fiction. The argument is that data centers, especially those running revenue-generating inference, have strong commercial incentives never to curtail, and that voluntary promises to reduce load during grid stress will evaporate the moment curtailing means turning away paying customers. The monitor further argued that data centers should not be allowed to skip the interconnection queue unless they bring their own generation to the table.

Where the AI companies stand

The AI companies themselves are divided, and their positions map directly onto their workload profiles. OpenAI offered strong support for the Wright proposal, arguing that flexible loads should receive interconnection priority over inflexible loads because they enhance overall system efficiency. This is a sensible position for a company that runs large training operations with genuine scheduling flexibility. Meta took the opposite view, arguing that utilities should not limit the expedited process based on curtailability. Meta wrote that such a limitation would be incompatible with the operational requirements of many data centers, which often allow little downtime over the course of a year. That is the voice of a company running always-on services where curtailment means degraded products.

This split is not a disagreement about values. It is a reflection of the engineering truth described earlier. Companies whose load is training-heavy can credibly offer flexibility, while companies whose load is inference-heavy cannot, at least not without hurting the products their users depend on. The regulatory question of whether to reward curtailability is, underneath the surface, a question about which kind of AI workload the grid should preferentially accommodate. Few participants in the public debate frame it this way, but it is the core of the matter.

The genuine case for AI as a grid solution

The flexibility debate has a counterintuitive flip side that deserves serious attention rather than hype. The same AI data centers that strain the grid could, under the right design, become assets that help stabilize it. Industry analysts have argued that in 2026, data centers can play a more active role in stabilizing the grid and mitigating cost increases by promoting load flexibility through curtailment, and by securing strategic investment in new generation and storage. A large data center that can shed load on command, or that pairs itself with on-site batteries and generation, looks less like a threat to grid reliability and more like a controllable resource the grid operator can lean on during peak stress.

There is also a distinct and more mature use of AI in the energy system that should not be confused with the data center question. AI models are increasingly used to forecast electricity demand, predict renewable generation from wind and solar, optimize the dispatch of resources, and detect faults on the grid before they cascade into outages. These applications are real and deployed today, and they represent genuine value at the intersection of AI and energy. The honest framing is that AI is simultaneously the largest new source of grid stress and one of the more promising tools for managing a more complex grid. Both things are true at once, and serious analysis has to hold them together rather than picking the more convenient one.

What the Data and Geography Actually Show

The clearest evidence that the AI data center power crunch is binding right now is that grid limits are already reshaping where data centers get built. The geographic distribution of new AI projects is increasingly dictated by power availability rather than the traditional factors of network latency and proximity to users. This is a profound shift. For decades, data centers clustered near population centers and internet exchange points because milliseconds mattered. Now access to megawatts matters more than access to low latency, and development is migrating toward regions with energy surpluses.

The market evidence is concrete. In May 2026, Texas was reported to have overtaken Northern Virginia as the world's top primary data center market, a position Northern Virginia had held for years on the strength of its network connectivity. Texas won not because of better connectivity but because of power availability and a regulatory environment friendlier to large loads building their own generation. Major projects illustrate the scale involved. A single flagship campus broke ground in Texas at 1.2 gigawatts of capacity, and a chipmaker announced a partnership to deploy up to 5 gigawatts of AI infrastructure globally. These are power figures that would have been unthinkable for a single computing facility a decade ago.

The signals of a binding constraint

When supply cannot meet demand, you see specific symptoms, and the data center sector is showing all of them. In the first quarter of 2026, a record number of data centers were canceled, according to industry tracking data. Cancellations at record levels are the signature of a constraint, not of healthy growth. Some potential data center customers paused decision-making in the PJM region because of uncertainty around the grid operator's process for handling them. Internationally, the strain is visible too. In Denmark, the grid operator Energinet paused new grid connection agreements for large electricity consumers, citing the sheer volume of requests from data centers, batteries, and other large loads overwhelming its capacity to process them.

The utility industry itself is split on whether to welcome this demand, which tells you how genuinely difficult the tradeoff is. The CEO of one major utility, Eversource, flatly declared he is not interested in data centers because they are only going to drive up the price of energy. Other utilities are pursuing data center pipelines of many gigawatts while taking pains to insist they are protecting ratepayers from cost shifts. The grid operator's independent market monitor cited data centers in reporting a large wholesale cost jump in the first quarter of 2026, even as many of the biggest proposed projects have not yet arrived on the grid. The strain is being felt before most of the announced capacity has even been built, which is the most alarming data point of all.

The aging grid underneath it all

None of this is happening on a modern, resilient grid with room to spare. Much of the U.S. grid was built between the 1950s and 1970s, and by industry estimates roughly 70 percent of it is approaching the end of its operational life. The unprecedented load growth from AI is landing on infrastructure that was already overdue for replacement. This is why the FERC rulemaking and the state tariff actions feel so urgent to the people inside the system. They are not managing growth on a healthy grid. They are trying to absorb the largest new electrical load in a generation on a grid that needs hundreds of billions of dollars in modernization regardless of whether a single new data center gets built.

What Business and Technology Leaders Should Understand

For any organization building or buying AI capabilities, the power crunch is not a distant energy-sector concern. It is a direct input into the cost, availability, and strategic risk of an AI program. The era in which compute could be assumed to be cheap, abundant, and instantly scalable is ending, and the leaders who internalize that first will make better decisions. KriraAI works with organizations that are moving AI from experiment to production, and the power constraint changes the calculus of that transition in ways that deserve to be made explicit.

The first thing leaders should understand is that the cost of inference is becoming a structural line item, not a rounding error. As inference overtakes training as the dominant AI workload around 2027, the recurring electricity cost of serving a model in production will increasingly shape the unit economics of AI products. An AI feature that looks cheap in a pilot can become expensive at scale, because every query consumes compute that consumes power that is getting more contested and, in many regions, more costly. Reasoning about total cost of ownership for an AI system now requires reasoning about its inference power profile, not just its development cost.

Practical implications worth acting on

There are several concrete things organizations should weigh as the AI data center power crunch tightens. Each of these follows directly from the engineering and market realities described in this analysis. None of them require deep energy expertise to act on, only a willingness to treat power as a real constraint.

1. Model and task fit matters more than chasing the largest model, because a smaller model that meets the requirement consumes less compute and therefore less power per query, which compounds across millions of inferences into real cost and footprint differences.

2. The physical location of the compute serving your AI now carries strategic risk, because regions with constrained grids face higher costs and the possibility of curtailment, while regions building generation alongside load offer more stable economics.

3. Latency-tolerant AI workloads have a hidden advantage, because batch and asynchronous processing can be scheduled into off-peak hours where flexible pricing and grid availability are better, turning the training-versus-inference flexibility distinction into a procurement strategy.

4. The regulatory environment is now a material variable in AI planning, because the outcome of the FERC rulemaking and the spread of state large load tariffs will affect the cost and timeline of any large-scale AI deployment that depends on new data center capacity.

5. Efficiency engineering is no longer optional polish, because techniques that reduce the compute required per inference translate directly into lower power consumption and better resilience against a constrained and more expensive energy supply.

The second thing leaders should understand is that the flexibility of their AI workloads is becoming a genuine asset. An organization whose AI processing can tolerate scheduling, that can run heavy jobs overnight or pause them during grid stress, is positioned to benefit from the very rules FERC is writing. This is where the engineering of a system meets the economics of the grid. KriraAI builds production AI systems with these real-world constraints in mind, designing for the actual environment of cost, latency, and infrastructure rather than the idealized one in which compute is free and infinite. The companies that treat power as a first-class design constraint will have a durable advantage over those that discover it only when the bill arrives.

The third thing to understand is that this is a fast-moving regulatory and infrastructure story, and standing still is itself a decision. The June FERC ruling will not be the end. It will be the start of a multi-year reshaping of how AI infrastructure gets sited, powered, and priced. Organizations that build an informed view now, and that design their AI strategies to be robust across a range of regulatory outcomes, will navigate the transition far better than those that assume today's conditions will persist.

Where the AI Data Center Power Crunch Is Heading Next

The trajectory of this collision is becoming clear in outline even though the details remain contested. The first signal to watch is the FERC ruling itself by the end of June 2026. If FERC produces a rule that successfully fast-tracks flexible and co-located loads while holding a defensible jurisdictional line, it will accelerate a particular kind of AI infrastructure, the kind that brings its own generation and offers curtailment. If the rule overreaches and gets tied up in litigation, the uncertainty that is already canceling projects will deepen, and the geographic flight toward power-rich, regulation-light states will intensify.

The second signal is the continued migration of the inference workload. As inference overtakes training around 2027 and AI heads toward half of all data center workloads by 2030, the share of AI compute that can credibly offer flexibility will shrink. This puts a clock on the curtailment-based regulatory strategy. The rules being written now assume a meaningful fraction of AI load is flexible, but the workload mix is shifting toward the inflexible end. The policy and the engineering are moving in opposite directions, and that tension will force either better inference scheduling technology or a harder reckoning with the limits of voluntary curtailment.

The two paths from here

Broadly, two futures are plausible, and the choices made in the next year will tip the balance between them. In the first, the buildout outpaces the grid's ability to adapt. Costs rise, the ratepayer backlash hardens into more state-level pauses and stricter tariffs, projects continue to cancel at elevated rates, and AI deployment becomes constrained not by algorithms or chips but by the unglamorous physical reality of electricity. In this future, power is the binding constraint on the AI economy, and the companies that win are the ones that learned to do more with less compute.

In the second future, the combination of co-located generation, energy storage, smarter scheduling, AI-driven grid optimization, and a workable federal interconnection framework allows the buildout and the grid to expand together. Data centers become controllable resources that help stabilize the grid rather than threats to it, and the enormous capital flowing into AI infrastructure also funds the modernization of a grid that desperately needs it. This future is achievable, but it requires the engineering of AI workloads and the design of the grid to be treated as a single integrated problem rather than two separate fights. That integration is precisely the kind of real-world systems thinking that separates durable AI deployment from fragile experimentation.

The most likely outcome is some messy blend of the two, varying sharply by region depending on local grid conditions, state policy, and the specific mix of training and inference in each market. What is certain is that the question of where the electricity comes from has moved from the back office to the center of AI strategy. The AI data center power crunch is no longer a footnote to the AI story. It is becoming the main constraint that shapes how fast, how cheaply, and where the technology can actually be deployed.

Conclusion

Three insights stand out from this analysis of the AI data center power crunch. The first is that the demand straining the grid is not generic industrial load but the specific, measurable consumption of AI training and inference, and the difference between those two workloads is the hidden hinge on which the entire regulatory fight turns. The second is that the controversy framed in the news as politics and electricity bills is, underneath, a technical dispute about whether AI loads can genuinely flex their power use, a dispute that one grid monitor bluntly called a regulatory fiction. The third is that the constraint is already binding, visible in record data center cancellations in early 2026, in grid operators pausing new connections, and in the migration of the world's top data center market from Northern Virginia to power-rich Texas.

These insights point to a larger truth about the current moment in AI. The bottleneck on AI progress is shifting from algorithms and chips toward the physical infrastructure of electricity, and the resolution of the FERC rulemaking by the end of June 2026 will be an early and consequential marker of how the United States chooses to manage that shift. The decisions made now about interconnection, curtailment, cost allocation, and the federal-state jurisdictional line will shape the cost and geography of AI deployment for years. This is the kind of development where the surface news story and the deep technical reality diverge sharply, and where understanding both is the only way to make sound decisions.

This is the work KriraAI is built for. KriraAI helps organizations make sense of the AI developments that current events are accelerating and revealing, translating fast-moving intersections like the collision of AI compute and the power grid into clear strategic understanding. KriraAI builds production AI systems designed for the real world with all its complexity, where power is a finite resource, inference cost is a structural line item, and infrastructure constraints are first-class design considerations rather than afterthoughts. The organizations that treat the energy reality of AI as central to their strategy, not peripheral to it, will deploy more durable and more economical systems than those that discover the constraint only when the electricity bill arrives. If your organization is navigating the AI landscape that events like the FERC rulemaking are actively shaping.