How to Read AI’s Energy Impact Before Trusting the Numbers

The same technology now blamed for a 50% jump in electricity use at AI-focused data centres in a single year is also being sold as the best available tool for cutting energy waste across the entire global economy.

Both claims come from serious sources. Both are, in their own way, true. That is what makes AI’s energy footprint one of the harder questions in energy investing right now.

The stakes are not abstract. According to the International Energy Agency (IEA), AI-specific data centre demand is on track to roughly triple between 2025 and 2030, grid operators are already reporting localised stress from AI cluster buildouts, and infrastructure vendors are simultaneously pitching “AI efficiency” as a bullish thesis to the same investors watching those grids strain.

You need a way to hold both ideas at once without picking a team.

What follows here separates the verified efficiency gains from the speculative ones, so you can judge both sides of the story against the same evidence. By the end, you will know which claims survive independent scrutiny, which do not, and the three questions to ask before you accept any AI efficiency figure at face value.

How much electricity AI actually uses, and where it is heading

Start with what is already on the meter. Global data centre electricity demand grew 17% in 2025, reaching roughly 485 TWh, or about 1.5% of worldwide electricity consumption, according to the IEA’s April 2026 update on energy and AI.

That headline number hides a sharper story underneath it. Over the same period, electricity use in AI-focused facilities specifically grew 50%, more than sixteen times the roughly 3% growth in total global power demand.

AI data centre demand is being shaped by a convergence of inference growth, cooling density increases, and geographic concentration, each of which compounds the pressure on transmission infrastructure in ways that aggregate TWh figures alone do not fully capture.

The trajectory is where the weight lands. The IEA projects total data centre consumption to roughly double to around 950 TWh by 2030, lifting data centres from 1.5% to approximately 3% of global electricity, with AI-specific demand tripling over that window.

The IEA’s energy and AI projections form the statistical backbone of the demand case, covering data centre consumption trajectories, regional breakdowns, and sensitivity scenarios that independent analysts use to stress-test the base forecasts.

Goldman Sachs Global Investment Research frames the same buildout in capacity terms: global data centre power near 55 GW in early 2025, forecast to rise toward 84-92 GW by 2027.

Why is this structural rather than a passing cycle? Five drivers compound on one another:

• Usage growth: The IEA notes a threefold rise in active AI users and a fivefold jump in provider revenue over a single year, with longer prompts and multimodal outputs raising the compute cost per interaction.
• Model scale: Frontier models demand vast GPU clusters running for weeks, and efficiency gains per model tend to enable larger, more frequent training runs rather than fewer.
• Inference dominance: The continuous serving of models to millions of users has overtaken episodic training as the dominant load.
• Cooling density: Goldman Sachs projects power density climbing from approximately 162 kW to approximately 176 kW per square foot by 2027 (IT load only).
• Geographic concentration: AI facilities cluster in favourable jurisdictions, triggering transmission bottlenecks and long interconnection queues.

The inference point is the one that should change how you read this curve. Allianz research documents a greater than 280-fold decline in inference costs between late 2022 and late 2024, and that collapse in price has stimulated so much extra usage that per-task efficiency gains are being swamped by sheer volume.

That is the mechanism to keep front of mind. Cheaper AI does not mean less AI energy. It means more AI, and more energy, unless something intervenes on the demand side. For anyone assessing the generation, transmission, and storage buildout through 2030, these figures set the floor, and it is already showing up in real interconnection queues.

What “net-positive AI energy” actually means, and why the concept is contested

Here is the optimistic case, presented on its own terms. The World Economic Forum (WEF) framework, “From Paradox to Progress,” defines a net-positive AI energy balance as a state where the energy AI saves across grids, buildings, and industry exceeds the lifecycle energy the AI systems themselves consume.

What makes this different from swapping an old boiler for an efficient one? Conventional efficiency substitutes lower-consumption equipment. AI instead uses real-time data to optimise the timing and operation of whole systems, extracting savings that hardware substitution alone cannot reach.

The WEF sets out a six-step methodology to get there:

1. Measure baseline consumption.
2. Integrate real-time data sources.
3. Identify inefficiencies.
4. Forecast demand and variability.
5. Take automated action.
6. Remeasure results.

The IEA lends support to the optimistic reading, noting that the per-task energy efficiency of AI has improved at a historically fast rate. In its High-Efficiency Case, the agency suggests stronger efficiency progress could cut global data centre demand by more than 15% by 2035, though this scenario figure carries more uncertainty than the agency’s base projections and should be read as conditional.

Why the optimistic case is harder than it looks

Then the counterargument arrives, and it is a serious one.

The Jevons paradox describes what happens when a technology makes something cheaper: total consumption of that thing tends to rise, not fall. Applied to AI, cheaper inference means more inference, which means more total energy, regardless of how much each individual query improves.

Academic reviews, including work published in Frontiers in Energy Research and by UN University, estimate economy-wide rebound effects commonly sit at 30-60% or higher, eroding a large share of the savings AI achieves before they ever reach the system level.

This is where the sceptics cluster. BloombergNEF, Allianz, and academic reviewers argue that infrastructure bottlenecks and the elasticity of usage will keep AI a net energy consumer unless strict governance is applied.

For readers wanting to situate the Jevons paradox argument within a broader economic framework, our dedicated guide to the AI productivity paradox examines why efficiency gains at the task level do not automatically translate into aggregate economic or energy savings across an economy.

The read you should take from this is that net-positive AI is not primarily an engineering question. Whether it works in a lab is almost beside the point. Whether governance frameworks constrain the predictable rebound is a policy question, and that is the variable most likely to decide the outcome. When you hear an efficiency claim from a vendor or utility, that mental model, technology performance versus system-level result, is the distinction that protects you.

Where AI has actually delivered measurable energy savings

Move from theory to the ledger. The evidence for AI-driven savings is strongest when you look sector by sector at documented cases with hard numbers attached.

Start in the data centre itself. Google’s deployment of DeepMind AI to optimise cooling cut cooling electricity by roughly 15%, saving over £1 million annually and avoiding approximately 760 tonnes of CO2 per documented site each year.

Google DeepMind cooling benchmark Approximately 15% reduction in cooling electricity, over £1 million in annual savings, and around 760 tonnes of CO2 avoided per site per year. This remains the most widely cited and independently referenced efficiency result in the dataset.

Buildings show the widest range. Systematic reviews from 2025-2026 report 20-40% savings from AI-IoT building systems, with office HVAC reaching up to 37% and cross-deployment averages landing at 22-28%.

Grids offer portfolio-scale evidence. A Chinese AIoT platform coordinating wind, solar, storage, EVs, and industrial loads achieved 100,000 MWh of electricity savings and a 5% cut in peak demand across multiple sites.

Industry rounds out the picture. The IEA cites AI deployments at Nvidia’s Guadalajara and Schneider Electric’s Wuxi factories that delivered energy-intensity improvements of 25-42%, and a 2025 estimate put realised industrial AI savings at roughly 8 TWh globally in 2024.

Now the aggregate. The IEA estimates AI could save more than 300 TWh in buildings globally by 2035, and roughly 3 EJ in industrial process optimisation over the same horizon.

Here is the caveat that keeps you honest. The localised cases are genuine, but reaching those aggregate figures requires the same efficiency gains to scale uniformly across millions of facilities with wildly different baseline conditions. That assumption is doing a lot of work, and it is worth interrogating before you treat the headline totals as banked. For investors tracking grid modernisation, building management, and industrial decarbonisation, these numbers mark where the thesis has empirical grounding today, and where it is still extrapolation.

Why the numbers are often overstated, and what rigorous evaluation actually shows

Now the cold water. A 2025-2026 impact evaluation by NYSERDA examined 654 sites running advanced energy-management solutions and found that independently verified electricity savings averaged only 48% of what vendors originally claimed.

Gas told the same story, with verified savings at 47% of claimed figures. In some cases, entire categories of AI-HVAC vendors showed near-zero verified savings despite confident marketing.

Why does the gap between claim and reality stay so wide? Three structural reasons:

• Metrics ignore lifecycle impacts: Standard measures like power usage effectiveness (PUE), the ratio of a facility’s total energy to the energy reaching its computing equipment, leave out embodied energy in hardware manufacturing, cooling water use, and land footprint.
• The attribution problem: Basic digitalisation gains get folded in and credited to AI specifically, inflating the AI-attributable figure.
• The rebound effect: As covered earlier, cost-lowering erodes gross savings before they reach the system level.

The scale of that first gap matters more than it looks. A 48% verification ratio means that for every two percentage points of efficiency a vendor advertises, the real-world outcome is closer to one. Apply that haircut across a portfolio and a net-positive AI energy thesis can flip to net-negative.

How to evaluate an AI efficiency claim

Turn the critique into a checklist. When you meet an efficiency claim in a vendor pitch, a regulatory filing, or a corporate sustainability report, ask three questions:

• What is the baseline methodology? A saving is only meaningful against a defined and honest starting point.
• Was the outcome independently verified? The WEF and IEA both recommend mandatory third-party verification, precisely because self-reported figures run roughly double reality.
• Does the figure account for lifecycle energy and rebound effects? Energy per useful output, lifecycle carbon, and water usage are the metrics that separate credible claims from inflated ones.

This article is for informational purposes only and should not be considered financial advice. Investors should conduct their own research and consult with financial professionals before making investment decisions.

What the evidence tells investors and analysts assessing AI’s energy trajectory

Put the two halves together and the honest verdict is asymmetric. AI’s demand growth is structural, already visible in grid data, and supported by consistent IEA and Goldman Sachs projections. AI’s efficiency potential is real but contingent, dependent on governance, measurement discipline, and clean power that is not yet built at scale.

Three variables will decide which way this tips through 2030:

• Clean energy contracting pace: Watch whether hyperscaler procurement keeps up with demand. Microsoft’s roughly $16 billion, 20-year PPA for approximately 835 MW from Three Mile Island Unit 1, signed September 2024, and Google’s carbon-free energy rising from 64% in 2023 to 66% in 2024-2025 reporting, are the reference points.
• Verification standards: Watch adoption of rigorous third-party verification, without which the NYSERDA 48% gap persists.
• Governance on rebound: Watch whether frameworks constrain the rebound effects that predictably follow any cost-lowering technology.

Clean power contracting has become the primary lever hyperscalers use to manage reputational and regulatory exposure from rising consumption, but the pace of PPA execution varies significantly across operators and geographies, and not all agreements deliver the additionality that underpins a credible net-zero claim.

That contracting pace is more than an ESG signal. It is a forward indicator of where new generation capacity will be contracted and built over the next decade, which is exactly where the infrastructure investment story meets the efficiency debate.

The WEF three-condition test Net-positive AI requires designing for efficiency, deploying for high-impact use cases, and shaping demand wisely.

The clear-eyed conclusion: the infrastructure demand story has stronger near-term evidence than the net-positive thesis, but the efficiency opportunity is real enough in grids, large-scale buildings, and industrial process control to warrant serious attention from investors with longer time horizons. You need both sides of the equation, the demand floor that makes new capacity essential and the efficiency ceiling that decides whether AI helps or hinders decarbonisation.

Past performance does not guarantee future results. Financial projections are subject to market conditions and various risk factors, and forward-looking scenarios are speculative and subject to change based on technology, policy, and market developments.