Every time a data center in Virginia hums a little louder to render a ten-second video of a cat playing piano, somewhere on the same transmission line a grid operator is recalculating a margin that used to be comfortable and no longer is. That juxtaposition a trivial, disposable piece of content sitting on top of an increasingly fragile physical system is the least discussed part of the AI energy story. Most public debate treats "AI's energy use" as a single, slowly rising line on a chart. The reality captured in the International Energy Agency's Energy and AI Observatory is messier and more interesting: efficiency per unit of AI work is improving quickly, even as the total electricity bill for AI is exploding, because the mix of what people ask AI to do is shifting toward far more energy-intensive tasks. Text prompts are being crowded out by video generation, and that shift is arriving at precisely the moment grids have the least slack to absorb it.

What the Observatory Actually Shows

The IEA built its Energy and AI Observatory to solve a specific problem: there was no reliable, continuously updated public dataset on how much electricity AI actually consumes, even as trillion-dollar investment decisions were being made on the assumption that the answer is "a lot, and growing." The latest data behind the Observatory shows that global electricity demand from data centers grew 17% in 2025, but electricity consumption specifically from AI-focused data centers surged 50% in the same year three times faster than the broader data-center category. The capital being thrown at this buildout is staggering: the largest technology companies spent more than $400 billion on data-center capital expenditure in 2025, a figure the IEA expects to jump another 75% in 2026, meaning the combined capex of just five technology companies now exceeds global investment in oil and natural gas production combined. Using satellite-based tracking, the IEA also found that purpose-built "AI factories" data centers designed specifically for AI workloads rather than general cloud computing have more than tripled in capacity in eighteen months.

IEA Executive Director Fatih Birol has framed the stakes bluntly, noting that "AI is quickly emerging as one of the most important technologies of our time", and the Observatory's own framing is even more direct: "there is no AI without energy." That second phrase is worth sitting with, because it captures why this is no longer a niche sustainability footnote. AI's growth trajectory is now a variable that national grid operators have to model explicitly, the same way they model industrial electrification or EV adoption except AI demand is spatially concentrated, growing faster, and far less predictable.

The Efficiency Paradox Hiding in Plain Sight

Here is the part that gets lost in most coverage: the underlying AI models are, in fact, becoming dramatically more energy-efficient per unit of output. Inference costs per token have fallen sharply as chip architectures, quantization, model distillation, and data-center cooling have all improved. Google has claimed a typical Gemini text prompt now consumes roughly 0.24 watt-hours, while other independent benchmarks for comparable chatbot exchanges land closer to 2.9 watt-hours either way, a small fraction of the energy of, say, brewing a cup of coffee. If efficiency gains were the whole story, AI's grid footprint should be flattening even as usage grows, the same way computing's energy intensity per calculation has fallen for decades even as total compute has exploded.But the IEA's own three-factor framing of AI energy demand improvements in efficiency, surging uptake, and changing model capabilities makes clear that efficiency is only one lever, and it is being overwhelmed by the other two. Major model providers have reported roughly a threefold increase in active users, and every generation of models unlocks qualitatively new, far more computationally demanding use cases. This is a textbook rebound effect: as each unit of AI output gets cheaper and greener, people don't hold total consumption steady, they use dramatically more of it, and they use it for heavier tasks. The efficiency curve is real, but it is being run over by the demand curve, and video generation is the clearest example of a heavier task arriving at scale.

Video Generation: The Category That Breaks the Curve

Text-based efficiency gains simply do not transfer to video. Independent estimates of OpenAI's Sora 2 model suggest a single ten-second video clip can consume somewhere close to a full kilowatt-hour of electricity comparable to running a load of laundry while more detailed technical benchmarking from researchers modeling video diffusion models estimated that Sora 2 Pro can draw more than 1,300 watt-hours for a twelve-second clip rendered at 1080p resolution. Journalists have translated this into more visceral comparisons: CNET's reporting, cited widely across tech outlets, estimated that a generated video can carry roughly 2,000 times the energy cost of a single text response from the same company's chatbot. Put another way, the efficiency gains that took years of engineering effort to shave watt-hours off a text query can be erased by a single button press that asks for eight seconds of synthetic video instead of a paragraph of synthetic text.This matters because video generation is not a marginal feature. It is being positioned as a mainstream consumer product, embedded into social apps, marketing pipelines, and short-form content platforms, precisely the kind of high-frequency, casual, repeat-use context that historically drives exponential demand curves once a product finds product-market fit. Analysts modeling deployment at scale even a hypothetical four million users generating just two short clips a day have projected electricity consumption in the hundreds of gigawatt-hours over a six-month period, comparable to the annual electricity use of well over a hundred thousand average households, for content that is largely disposable and rarely rewatched. Model architecture choices compound the uncertainty: researchers benchmarking open video models found that a smaller model with fewer parameters can consume more energy per video than a larger one, meaning the industry does not yet have a reliable rule of thumb like "bigger model, more energy" for predicting or regulating this category's footprint. That absence of predictability is itself a policy problem, not just a technical curiosity.

Grids Were Not Built for This Shape of Demand

The efficiency-versus-surge tension would be an interesting academic debate if grids had comfortable headroom. They do not. The IEA's parallel Electricity 2026 analysis finds more than 2,500 gigawatts of renewable, large-load, and storage projects currently stalled in interconnection queues worldwide, and estimates that annual grid investment needs to rise roughly 50% above today's $400 billion by 2030 just to keep pace. The mismatch is structural: a data-center campus can be permitted and built in two to three years, while the transmission lines, substations, and large power transformers needed to actually deliver electricity to it can take seven to ten years, and in the most congested markets, more than a decade. Power and distribution transformers unglamorous but essential are reportedly facing supply shortfalls of 30% and 10% respectively, a bottleneck that no amount of chip-level efficiency gain can route around.The geographic concentration makes the strain worse. Data centers already account for something close to 80% of Dublin's electricity consumption and between roughly a third and over 40% of consumption in Amsterdam, London, and Frankfurt, prompting Irish regulators to publicly flag the trajectory as unsustainable without major reinforcement. In the United States, data centers now account for roughly half of the country's incremental electricity demand growth, and in Texas alone, grid operator ERCOT projects peak summer demand could approach 145 gigawatts by 2031, driven substantially by AI infrastructure. Reliability incidents are no longer hypothetical: in 2024, a protection-system fault in Virginia's "Data Center Alley" triggered a disconnection event in which roughly sixty of more than two hundred co-located data centers suddenly dropped off the grid and switched to backup generation simultaneously the kind of concentrated, correlated load behavior that traditional grid planning was never designed to anticipate, let alone the more exotic risk that coordinated shifts in AI workload scheduling could someday be used to intentionally destabilize local frequency regulation.

Why This Trade-off Stays Underexplored

Most public narratives about AI energy use resolve into one of two comfortable stories: either "efficiency will save us", pointing to falling per-query costs as evidence the problem is self-correcting, or "AI is an ecological emergency", pointing to gigawatt-scale buildouts as evidence of impending catastrophe. The IEA's own data suggests neither framing captures what is actually happening. Efficiency gains are real and are not being wasted without them, today's electricity bill for AI would be substantially higher than it already is. But those gains are being consumed almost entirely by a shift in what AI is used for, not just how much of it is used. A population that mostly typed short chat prompts a few years ago is migrating toward workloads video, agentic multi-step reasoning, image generation at higher resolutions that are inherently, unavoidably more electricity-intensive per interaction, regardless of how efficient the underlying silicon becomes.This is precisely the kind of trend the IEA warns is difficult to forecast with confidence, since it depends on consumer adoption patterns, model capability jumps, and corporate deployment decisions that shift faster than grid planners, utilities, or regulators can update their assumptions. The practical consequence is that grid operators are being asked to plan multi-decade, multi-billion-dollar infrastructure investments against a demand signal that could look completely different in eighteen months if, say, real-time AI video becomes as casually embedded in social apps as photo filters did a decade ago. Closing that gap will likely require the kind of systematic energy-consumption disclosure from AI developers that the IEA has explicitly called for, treating watt-hours per generated output the way food labels treat calories per serving not because any single video is significant, but because billions of them, generated daily, are what actually determines whether the grid holds.