AI, Data Centres and Electricity Demand on the Grid

AI cuts both ways on the grid. On the demand side, data centres consumed about 415 TWh in 2024 (roughly 1.5% of world electricity) and are projected to more than double to around 945 TWh by 2030, driven mainly by AI ([IEA, 2025](https://www.iea.org/reports/energy-and-ai/executive-summary)). On the supply side, the same AI techniques can unlock up to 175 GW of extra transmission capacity on existing lines and cut outage durations by 30-50% ([IEA, 2025](https://www.iea.org/reports/energy-and-ai/executive-summary)). The practical question for utilities is not whether AI raises demand, but whether you deploy it fast enough on the operational side to absorb that demand.

The two-sided story you cannot ignore

There are two AI-and-electricity stories running at the same time, and most coverage tells only one of them. The first is alarming: AI workloads, concentrated in data centres, are pushing power demand up faster than grids were built to handle. The second is quieter but just as real: AI is one of the most effective tools available for squeezing more capacity, reliability and flexibility out of the assets you already own.

Both are true. If you run a utility, a distribution or transmission system operator, or sit on an energy leadership team in the Netherlands or wider Europe, holding both ideas at once is the whole job. Treating AI purely as a demand threat misses the levers that make it manageable; treating it purely as a productivity miracle ignores the physical strain already showing up on connection waiting lists.

This piece walks through the demand side with the numbers, then the productivity side with the numbers, then the synthesis - and what it means for a grid like the Dutch one, where netcongestie is no longer a forecast but a daily operating constraint. For the operational detail, we link out to our spokes on grid optimisation, congestion and flexibility markets and demand forecasting throughout.

Side one: AI is driving electricity demand up, sharply

Start with the baseline. Data centres accounted for about 1.5% of global electricity consumption in 2024, around 415 TWh (IEA, 2025). That sounds modest until you look at the slope: data-centre electricity demand has grown roughly 12% per year since 2017, more than four times faster than total electricity consumption (IEA, 2025). Slow-moving averages hide fast-moving local pressure.

The forward projection is where the strain shows. In the IEA Base Case, global data-centre electricity consumption is set to more than double to around 945 TWh by 2030, just under 3% of total global electricity (IEA, 2025), and continues rising to around 1,200 TWh by 2035 (IEA, 2025). The driver inside that growth is specifically AI: electricity use in AI-focused data centres roughly triples by 2030 (IEA, 2025), and accelerated (AI) servers are projected to grow about 30% per year versus roughly 9% per year for conventional servers (IEA, 2025).

The 2025 numbers confirmed the trajectory rather than softening it. Data-centre electricity demand rose 17% in 2025 to about 485 TWh, with AI-focused use surging about 50% (IEA, 2025). This is not a single-source story, either. Goldman Sachs Research independently projects global data-centre power demand to rise around 50% by 2027 and as much as 165% by 2030 versus 2023, with capacity climbing from roughly 55 GW today to about 122 GW by end-2030 (Goldman Sachs, 2025). McKinsey expects global data-centre capacity demand to nearly triple, from about 82 GW in 2025 to around 219 GW by 2030, with AI making up roughly 70% of that total (McKinsey, 2025).

Where the load lands: not evenly

Aggregate growth is one problem; concentration is another. Data-centre load does not spread itself evenly across a continent. In 2024 the geographic split was United States 45%, China 25% and Europe 15% of global data-centre electricity consumption (IEA, 2025). The growth concentrates further: China and the US together account for nearly 80% of global data-centre demand growth to 2030, with US demand rising about 240 TWh (+130%) and China about 175 TWh (+170%) versus 2024 (IEA, 2025).

The US case is the sharpest illustration of how fast a local share can move. Lawrence Berkeley National Laboratory put US data-centre use at 176 TWh in 2023, up from 58 TWh in 2014 - about 4.4% of total US electricity (LBNL, 2024), and projects 325 to 580 TWh by 2028, roughly 6.7% to 12% of national electricity (LBNL, 2024). EPRI's scenarios land in a similar band, with data centres reaching up to about 9% of US electricity generation by 2030 (EPRI, 2024).

Europe is smaller in share but rising on the same curve. McKinsey expects European data-centre demand to grow from about 10 GW today to roughly 35 GW by 2030, lifting electricity demand by around 85 TWh and reaching about 5% of total European power consumption, up from roughly 2% today (McKinsey, 2024). It is worth keeping perspective alongside the alarm: DNV projects that AI reaches about 3% of global electricity by 2040 (DNV, 2025) - a serious load, but not, system-wide, the whole story. The acute pain is local and near-term, which is exactly why grid planning, not just generation planning, is the binding constraint.

The Dutch reality: netcongestie is already here

Nowhere is the gap between demand growth and grid capacity more visible than in the Netherlands. This is not a 2030 projection - it is the queue today. TenneT's high-voltage waiting list holds 212 offtake requests totalling 38 GW, on top of 14,044 requests totalling 9 GW on regional operators' lists (TenneT via NL Times, 2025). Peak offtake demand on the Dutch grid is about 19 GW now and expected to reach roughly 27 GW by 2030 (TenneT, 2025).

The congestion has spread from industry to households. Liander recently placed around 7,300 households on a waiting list for the first time, with waits of up to three years for new or upgraded connections, while Stedin has effectively closed parts of Utrecht's network to new capacity (Liander / Stedin via NL Times, 2026). For a country building data centres and electrifying industry at the same time, that is a hard ceiling. About 90% of Dutch businesses now experience direct or indirect consequences of grid congestion, affecting expansion, renewable projects and fleet electrification (Strategic Energy Europe, 2025).

The capital response is enormous but slow. The Dutch government estimates about EUR 200 billion in grid investment is needed through 2040, funded partly by tariff increases averaging 4.3% to 4.7% annually until 2034 (Dutch government / TenneT via PPC Land, 2025). Across the EU the picture rhymes: the European Commission's Grids Action Plan put EUR 584 billion of electricity-grid investment as needed by 2030, with 40% of distribution grids over 40 years old (European Commission, 2023), and ENTSO-E has raised its cross-border investment estimate from EUR 2 billion per year to EUR 5 billion per year up to 2030 (ENTSO-E via Bruegel, 2025). Copper and steel cannot arrive fast enough on their own. That is the opening for the second half of the story.

Side two: AI is also how you make the grid absorb that demand

If new lines take years and billions, the fastest available capacity is the capacity you already have but cannot fully use. This is where AI stops being a problem and becomes a tool. The headline from the IEA's own analysis is striking: applying AI tools - remote sensors, dynamic line rating, AI-based grid management - to existing transmission lines could unlock up to 175 GW of additional transmission capacity without building new lines, which the IEA notes is more than the increase in data-centre power load to 2030 in the Base Case (IEA, 2025).

Read that comparison carefully, because it reframes the debate. The same report that projects the demand surge also finds that intelligent operation of existing assets could roughly offset the transmission impact of that surge. The constraint is not physics - it is deployment speed and operating practice. On reliability, AI-based fault detection can cut outage durations by 30-50% by identifying and pinpointing faults faster (IEA, 2025), which directly improves how much firm load a congested network can responsibly carry. We cover that operational pattern in our anomaly and fault-detection spoke and our grid congestion and flexibility-markets piece.

The cost case is large too. The IEA estimates that widespread AI adoption in power-plant operations and maintenance could deliver up to USD 110 billion in annual savings by 2035 from avoided fuels and lower costs (IEA, 2025). DNV's modelling points the same way over the long run, projecting that AI and deep digitalisation could cut total power-system costs by 6 to 13%, including a USD 1.3 trillion decrease in clean-energy generation costs and a USD 188 billion reduction in grid-equipment costs by 2050 (DNV, 2024).

Forecasting: the highest-leverage place to start

Most of AI's grid value flows through one capability: better forecasting. If you can predict load, wind, solar and price more accurately, you can commit generation more efficiently, curtail less, and run lines closer to their true limit with less risk. The IEA finds that AI improves the forecasting and integration of variable renewables, helping operators minimise wind and solar curtailment alongside demand shifting and storage (IEA, 2025). IRENA's G7 analysis quantifies the gain, reporting forecasts up to 45% more accurate than traditional methods, with examples from Australia, India and the UK (IRENA, 2025).

The operational payoff is well documented. Google DeepMind applied a neural network to weather forecasts and historical turbine data across 700 MW of wind capacity, predicting output 36 hours ahead, and boosted the value of that wind energy by roughly 20% through day-ahead delivery commitments (Google DeepMind, 2019). At system scale, an NREL study found that integrating better short-term forecasts into unit commitment could save up to about USD 5 billion per year across the Western US grid (NREL, 2015).

The modelling literature backs the accuracy claims. Deep-learning load forecasters now reach very low error - one LSTM-Attention model reported a MAPE of 0.52% on an Australian dataset and 0.53% on a US dataset (LSTM-Attention study, 2022), and a multi-country PV framework cut day-ahead RMSE by about 10% for Bulgaria and 9% for Greece (ScienceDirect, 2025). If you want the practical build pattern rather than the research, our demand-forecasting utility guide and price-forecasting piece for traders and utilities walk through it.

Flexibility and storage: where forecasts turn into capacity

Better forecasts only pay off if the system can act on them. That is what flexibility - storage, demand response and smart dispatch - provides, and it is scaling fast. BloombergNEF reports a record 112 GW / 307 GWh of batteries added worldwide in 2025, up 48% from 2024 (BloombergNEF, 2025), and projects battery deployment to jump 17-fold, from 223 GW in 2025 to 3.8 TW by 2035 (BloombergNEF, 2026). Dispatching and valuing those assets profitably is itself a forecasting-and-optimisation problem, which is exactly where AI earns its place.

Demand-side flexibility matters as much as supply-side storage, and it directly addresses the data-centre load. AI-led optimisation of heating, cooling and flexibility in buildings could deliver around 300 TWh of potential global electricity savings (IEA, 2025), and widespread AI adoption in light industry could cut process energy use by 8% by 2035 (IEA, 2025). Controlled studies are concrete: reinforcement-learning HVAC control has delivered up to 26.3% energy savings versus conventional control in validated simulations (Springer, 2025) and around 25% on top of existing PID control in a factory setting (ACM e-Energy, 2020).

The institutional groundwork is already being laid in Europe. ENTSO-E's RDI Roadmap sets milestones for AI-based decision-support and analysis in transmission-system operations over 2024-2034, framing AI as part of a digital backbone for the energy transition (ENTSO-E, 2024). And the appetite is real on the operator side: DNV's survey found nearly half of around 1,300 senior energy professionals plan to integrate AI-driven applications into operations within the coming year (DNV, 2024).

Assets, maintenance and the value beyond the wires

Some of AI's grid value sits inside the physical assets rather than the control room. Predictive maintenance keeps generation and grid equipment running longer and failing less, which adds effective capacity without a single new connection. McKinsey's classic analytics work found that predictive, analytics-driven maintenance reduces machine downtime by 30-50% and increases machine life by 20-40% (McKinsey, 2017).

Offshore wind is a strong example, given how expensive a vessel call-out is. Peer-reviewed work shows autoencoder-based condition monitoring detecting component degradation with 99% classification accuracy, flagging anomalies up to 60 days before reported failures in transformers, gearboxes, generators and hydraulic groups (PMC, 2025), and a Bayesian deep-learning framework reaching 99.14% accuracy for gearbox bearing faults (PMC, 2022). Our offshore-wind predictive-maintenance spoke goes into the deployment detail.

Beyond physical assets, AI changes the economics of the energy business itself. McKinsey estimates that across agriculture, chemicals, energy and materials, generative AI could create an additional USD 390 billion to USD 550 billion of value (McKinsey, 2024), and BCG finds AI in renewable operations can lift worker productivity by 15-25% and energy yield by 1-3 percentage points, with 10-15 use cases capturing 60-70% of the value (BCG, 2026). On the customer side, generative AI could reduce human-serviced contacts by up to 50% and add productivity worth 30-45% of current function costs in sectors including utilities (McKinsey, 2023).

The honest synthesis: not a contradiction, a sequencing problem

Put the two sides together and the paradox dissolves into a sequencing problem. AI adds load - settled by the IEA, Goldman Sachs, McKinsey, LBNL and EPRI numbers above. AI also adds usable capacity, reliability and efficiency to the existing grid - settled by the same IEA report and a wide research base. The two do not cancel out automatically. The demand arrives on the schedule of a data-centre construction project; the AI-driven grid gains arrive on the schedule of your own deployment program. Let the first outrun the second and you get the Dutch waiting list.

Two honest caveats keep this credible. First, the productivity figures are potentials, mostly tagged to widespread-adoption scenarios - the 175 GW, the USD 110 billion, the 300 TWh are what is achievable, not what a tool guarantees. Second, AI for grid operations is itself compute that consumes power; the case rests on the operational gains comfortably exceeding that overhead, which the evidence supports but which still has to be engineered, not assumed.

The practical takeaway is unglamorous and correct: treat AI-driven forecasting, congestion management, fault detection and predictive maintenance as grid infrastructure, on the same footing as steel and copper, because they deliver capacity faster. The technologies on the data-centre demand bill are the same ones that can pay part of it back on the operations side.

What this means for your team - and where Crux fits

If you operate or plan a grid, the move is to stop treating AI demand and AI productivity as separate conversations owned by separate teams. The same EUR 200 billion of Dutch grid spend goes further if the network is also run more intelligently. Start where the leverage is highest and the data already exists: load and renewable forecasting, dynamic line rating, congestion forecasting, and predictive maintenance on your most failure-prone assets.

This is the work Crux Digits does for energy and grid clients - DSOs, TSOs and utilities. We are a boutique AI consultancy in Nieuwegein, in the Utrecht region, working in English and Dutch, and we run fixed-scope engagements rather than staff augmentation: a EUR 2,500 Audit to find the highest-value use case in your data, a EUR 20,000 Proof of Concept to prove it on your own operations, and a Production build from EUR 50,000. We are not a Power BI reseller; we build forecasting, anomaly-detection and optimisation systems that go into operation. The plumbing under all of it - clean, reliable data pipelines - is covered by our data-engineering practice, and where the use case is asset inspection, our computer-vision work applies.

If your near-term problem is a connection queue, a curtailment bill, or forecasts that are not accurate enough to commit against, that is a scoped, provable engagement rather than an open-ended program. A short conversation is usually enough to tell whether AI moves the needle on your constraint. You can reach the team here, or read how we work with grid operators in the Netherlands.

Frequently asked questions