back to top
Sunday, November 22, 2026

Top 5 This Week

The AI Power Paradox: Cheaper Answers, Doubling Electricity Use

Artificial intelligence is getting cheaper by the month. The cost of generating a single answer — a summary, a translation, a line of code — has been falling dramatically, and each new generation of models and chips squeezes more work out of every watt. By the ordinary logic of efficiency, that should mean less electricity, not more. Yet the opposite is happening: data centres are consuming more power than ever, and the International Energy Agency expects their total electricity use to double by 2030.

This is the AI power paradox, and it sits at the centre of one of the biggest questions in technology and energy today. If AI keeps getting more efficient, why does it keep using more power? And what happens to electricity grids, prices and climate targets if the trend continues?

Cheaper per query, more power overall

The numbers behind the paradox are striking. According to reporting highlighted this week, the energy needed for a single AI task has been falling by roughly ten times a year — a pace of improvement that would be remarkable in any industry. At the same time, data-centre power consumption rose by 17 percent in 2025, and by 50 percent at AI-focused sites. The cost per answer is collapsing; the total bill is soaring.

The explanation is an old idea in economics, named after the nineteenth-century economist William Stanley Jevons. Jevons noticed that when steam engines became more efficient, coal consumption did not fall — it rose, because cheaper steam power made it worth using steam for far more things. Efficiency lowered the cost of each use, and that lower cost expanded the total number of uses far beyond what the efficiency had saved.

AI is following the same pattern. When an answer costs a tenth of what it did a year ago, people do not simply pocket the saving and ask the same number of questions. They ask more questions, in more places, for more purposes. A tool that was once reserved for a difficult task becomes the default for everyday ones. A feature that was once too expensive to offer to every user becomes cheap enough to switch on for everyone. The efficiency gain is real — but it is swallowed, and then some, by the growth in use.

More users, bigger models, inference at scale

Several forces are pushing demand upward at once. The first is simply the number of people using AI. Chatbots and AI assistants have moved from a niche of early adopters to hundreds of millions of regular users, and AI features are now built into search engines, office software, phones and customer service systems. Every one of those interactions runs on a server somewhere, and each one draws power.

The second is the size of the models. Training a frontier AI model is an enormous computational task, requiring thousands of specialised chips running for weeks or months. But training, for all its scale, is a one-off event per model. The larger and more continuous load is inference — the everyday business of answering questions — and inference grows with every user and every query. As models become more capable, they also tend to become more computationally demanding per answer, using longer chains of reasoning to produce better results.

The third is competition. The largest technology companies — the so-called hyperscalers — are racing to build the biggest, most capable AI systems, and that race is measured in data centres. According to the reporting, hyperscalers have sold 284 billion dollars in bonds this year, much of it to finance the build-out of AI infrastructure. That is borrowed money, raised on the expectation that the investment will pay off, and it is being spent on land, chips, cooling systems and, above all, electricity supply.

What it means for grids and prices

For electricity grids, the AI boom arrives at an awkward moment. In many countries, grids were built for a world of steady, predictable demand, and they are already under strain from the electrification of transport and heating, and from the shift to variable renewable sources. A data centre is a very different kind of customer: it draws a huge, constant load, day and night, and it needs that power to be reliable, because even a brief interruption can damage equipment and disrupt services.

In some regions, utilities are now fielding connection requests from data centres that dwarf anything they have handled before, and the queue for grid connections stretches years into the future. That has consequences for everyone else on the grid. If data centres absorb the available capacity, other users — factories, homes, new housing developments — may wait longer for connections, or pay more for power.

Electricity prices are the other pressure point. In markets where prices are set by supply and demand, a large new source of demand pushes prices up unless supply grows to match. Some technology companies have responded by signing long-term deals to buy power directly from generators, including nuclear plants, or by building their own generation. But those deals do not add power to the grid for everyone; they secure it for one buyer, and the cost of new generation still has to be paid by someone.

Global energy markets remain volatile as demand, supply and geopolitics collide, and AI’s appetite for power is now one more source of that volatility. Energy planners who a few years ago worried mainly about electric vehicles and heat pumps must now factor in a new, fast-growing and geographically concentrated source of demand that did not exist at anything like its current scale a decade ago.

The climate question

The climate implications are the most contested part of the story. On one side, AI companies point out that data centres can run on clean electricity, and that AI itself can help the energy transition — by optimising grids, improving weather forecasting for renewables, or accelerating research into new materials. On the other, the sheer scale of new demand means that, in the short term, much of it is being met by whatever generation is available, including fossil fuels.

The International Energy Agency’s expectation that total data-centre electricity use will double by 2030 frames the scale of the challenge. Doubling in five years would make data centres one of the fastest-growing sources of electricity demand in the world, comparable in size to the entire consumption of some large countries. Whether that demand is met cleanly will depend on how fast renewable and nuclear capacity can be built, and on whether efficiency gains in chips and cooling can slow the growth.

There is also a subtler climate effect. AI’s falling cost per query means it is being embedded in more products and processes, some of which genuinely save energy — a building managed by AI uses less power, a logistics network planned by AI burns less fuel. Measuring the net effect, the energy AI consumes against the energy it saves elsewhere, is genuinely difficult, and honest assessments admit that the answer is not yet clear.

The market connection

This energy story is also the backdrop to the week’s sharp sell-off in AI-related stocks. Investors have begun asking whether the enormous spending on data centres will ever earn an adequate return. The bond figures tell part of the story: 284 billion dollars raised by hyperscalers this year is a bet on future demand, and bets of that size invite scrutiny when the payoff is uncertain.

The paradox sharpens the question. If each AI answer keeps getting cheaper, the revenue earned per answer falls too, even as the infrastructure needed to deliver those answers keeps growing. The companies building the infrastructure argue that volume will more than compensate — that cheaper answers mean vastly more answers, and vastly more paying customers. The sceptics argue that the volume will not arrive fast enough, or that competition will force prices down faster than costs fall.

Neither side can yet prove its case, because the technology and the market are both moving too fast. What is certain is that the energy dimension of the debate is no longer a footnote. A few years ago, AI was discussed mainly as a software story. Today it is an energy story too — about power plants, transmission lines, water for cooling and the price of electricity. The companies that win the AI race may be the ones that solve the power problem first.

For the rest of us, the paradox will show up in ordinary ways: in the electricity bill, in the debate over where new power plants and transmission lines are built, and in the quiet, steady growth of the infrastructure that answers our questions. AI is getting cheaper to use. It is not getting cheaper to run. That gap, and how it is closed, will shape the next decade of both technology and energy.

Figures in this article are as reported in the source material, attributed to the reporting and to the International Energy Agency. Source: KenRadio

Shayan Hussain
Shayan Hussain
Shayan Hussain is the founder and owner of Pakistan Morning Post. He is also a YouTuber and Freelancer. The sole purpose of Pakistan Morning Post is to share updates independently about Pakistan and rest of the world.

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Most Read Stories