AI energy consumption: the IEA numbers and the bill that's coming

Ask a CTO what next year's AI project will cost and you'll get an answer in licenses, tokens and person-days. Almost nobody answers in megawatt-hours. Yet energy is becoming the constraint sitting behind all the others: the International Energy Agency estimates that a typical AI-focused data center consumes as much electricity as one hundred thousand households, and the largest ones under construction today consume twenty times as much. The figure comes from the most comprehensive report ever published on the topic, built on a new global model of demand.

The IEA numbers

In 2024 data centers consumed 415 terawatt-hours, around 1.5% of world electricity, growing 12% a year since 2017: four times faster than total electricity consumption. The IEA base case takes demand to roughly 945 terawatt-hours by 2030, more than Japan's entire electricity consumption today, and 1,200 TWh by 2035. The most conservative scenario stops at 700, already well above 2024 consumption; the most expansive exceeds 1,700.

Two details make it concrete. First: in the United States data centers already account for nearly half of electricity demand growth through 2030, and by the end of the decade they will consume more power than producing aluminium, steel, cement and chemicals combined. Second: Europe holds 15% of global data center consumption today, behind the US and China, but pays among the highest electricity prices of any large market.

Why the profile changed: from sprints to baseload

The ChatGPT era had a burst-shaped energy profile: giant clusters switched on for months to train a model, then silence. The inference era made consumption continuous but distributed: every conversation, every search, every suggested line of code. The era that has just begun belongs to persistent agents: processes that live permanently inside a data center, monitoring, deciding and calling tools while nobody watches. An agent running a supply chain is not a request-response transaction: it is a baseload load, like a furnace that never switches off.

The grid cannot handle it (yet)

The IEA estimates that, without intervention, around 20% of planned data center projects are at risk of delays for grid reasons: interconnection queues are long and complex, waiting times for transformers and critical cables have doubled in three years, and deliveries of turbines for new gas plants now measure in years. On emissions, electricity used by data centers goes from 180 million tonnes today to 300 million in the base case by 2035, and up to 500 in the most expansive scenario: still below 1.5% of the energy sector's total, but among the fastest-growing sources. Building energy capacity takes years, building data centers takes months: that temporal mismatch is the structural constraint of the decade.

Europe pays differently

According to Eurostat, in the second half of 2025 the average EU electricity price for businesses was €0.1837 per kilowatt-hour, with peaks such as Ireland at €0.2552 and Germany at €0.2264, against Finland at €0.0748. The same inference load served in different countries can cost two to three times as much in energy. For a market that wants to compete on AI, this means two things at once: a structural incentive to make every single workload efficient, and a legitimate question about where workloads should run, yours and your vendors'.

From macro to micro: the company AI bill

The energy constraint also arrives in invoices. The most instructive example is GitHub Copilot moving from a flat plan to consumption billing: since June 2026 it runs on AI Credits, with monthly allowances and budgets configurable per user, team and cost center. OpenAI and Anthropic publish equally transparent API price lists, but their APIs were born metered: what is new is a subscription product converting itself into a measured one, and it is where the whole market is heading. AI consumption thus becomes a resource to govern like bandwidth or compute: measured, routed, capped.

Governing means three concrete operations. First, picking the right model for the task: sending every request to the flagship model is the equivalent of sending the senior architect to change a lightbulb. Second, using prompt caching wherever context repeats, because cached tokens cost a fraction of the rest. Third, measuring what the AI actually produces, so consumption links to outcomes rather than impressions of speed. All three are ordinary engineering practice, and at enterprise volume the difference shows up as tens of percentage points of spend.

Tokenmaxxing: when consumption becomes a sport

There is also a case study in how NOT to govern this consumption. It is called tokenmaxxing: treating the number of tokens burned as a proxy for productivity. In April 2026 the tech press reported internal leaderboards at a major platform that rewarded employees for usage, with a top contender reaching 281 billion tokens in thirty days (TechCrunch, April 17, 2026). The data collected by analytics platforms tells you the result: declared acceptance rates for generated code at 80-90% drop to 10-30% once later revisions count, and teams with the largest token budgets get twice the throughput paying ten times as much. It is Goodhart's law applied to inference, plus the degradation that excess context stuffing brings. Every token burned to climb a leaderboard is baseload consumed for nothing.

Efficiency does not close the account

The standard objection says efficiency solves everything: better chips, smaller models, cheaper inference. It is the right argument, read badly. Since 1865, when William Stanley Jevons observed it on coal, it has been known that when a resource gets cheaper demand does not shrink: it finds uses that were not economical before, and grows beyond the freed-up space. Even the most interesting research direction confirms the picture: generative thermodynamic computing proposes letting the physics of a system generate structure directly from noise with minimal heat emission, demonstrated so far only in simulation. Real solutions, for specific workloads, on long timelines. Meanwhile efficiency makes uses economical that do not exist yet: which is exactly the mechanism behind the IEA curve pointing upward even in its highest-efficiency scenarios.

What it means for a small business

Three operational consequences, none of which requires turning yourself into a utility. First: measure. If AI consumption is not a visible, attributed line item, there is no governance to be had: tokens per team, per workflow, per model. Second: govern, applying the three levers above (right model, cache, evaluation) to your real workflows. Third: reconsider where stable workloads run. When the load is continuous, predictable and data-sensitive, the energy-and-cost equation of on-premise starts smiling again, and it is the same logic documented in cloud repatriation: not a return to the past, a placement decision made on numbers. And where installation efficiency enters the equation, Italian incentives exist and deserve reading before signing any supply contract: from CAM ICT to deductions for hyper-amortisable capital goods.

Whoever decides today where their AI workloads will run is also choosing which share of AI's energy bill becomes theirs and which stays with somebody else: an architecture decision before it is a procurement one.

In short

Energy is the real constraint of AI's next cycle, with the IEA's official numbers behind it: 415 terawatt-hours today, nearly double by 2030, and Europe pays more for the same thing. Three actions are useful: measure your own consumption, govern it with routing, caching and evaluation, and consciously decide where each workload runs.