Set as preferred source
Insights from Key Questions on Energy and AI (IEA, 2026).
AI is often discussed as if its main constraint is intelligence: better models, better chips, better software, better products. The IEA report points to a less comfortable reality. The next phase of AI will be shaped as much by electricity systems, capital markets, industrial supply chains, and skilled operators as by model performance.
That does not mean AI will overwhelm the global energy system. The numbers are more specific than that. The IEA’s base case sees global data-centre electricity use rising from about 485 TWh in 2025 to around 950 TWh in 2030, or roughly 3% of global electricity demand. At global scale, that is material but not destabilising. The problem is concentration. Data centres do not arrive as evenly distributed demand. They arrive in clusters, near fibre, land, incentives, customers, and power. In those local markets, AI load can become large enough to reshape grid planning, electricity prices, permitting politics, and infrastructure finance.
That is the structural shift. AI is moving from a software scaling problem to an infrastructure execution problem.
The older cloud assumption was relatively simple: build capacity where demand, connectivity, and economics made sense. AI changes the equation because the facilities are larger, denser, more variable, and far more capital-intensive. AI-focused data centres are projected to more than triple their electricity use by 2030, reaching roughly 465 TWh. That growth is tied not just to user adoption, but to accelerator shipments, high-bandwidth memory, cooling architecture, rack density, power electronics, and grid interconnection. In practical terms, the AI value chain is becoming coupled to the energy value chain.
This is why the capital signal matters. The largest technology companies are still financially strong, but the scale of the build-out is moving beyond normal corporate expansion. Capital expenditure by major technology companies exceeded USD 400 billion in 2025 and is expected to rise by another 75% in 2026. Leading hyperscalers and neo-clouds are now pointing toward roughly USD 715 billion in 2026 capex, a figure larger than annual investment in the entire US energy sector. When projects require tens of billions of dollars, the question is no longer only whether the technology works. It is whether the revenue model can support the capital intensity, whether debt markets remain open, whether power can be secured, and whether the facility can actually be energised on time.
Markets are already starting to make that distinction. In the first part of 2025, the AI value chain still looked like a broad capital-market winner: hyperscalers added roughly USD 2 trillion in market capitalisation between March and October, chip designers added about USD 2.5 trillion, and neo-clouds gained around USD 0.5 trillion. Then the signal changed. Neo-clouds gave up almost all of those gains as concerns over monetisation and capital intensity became harder to ignore. Chip manufacturers continued to benefit because they sit closer to the physical bottleneck. The market was no longer rewarding AI exposure equally. It was rewarding control over scarce capacity.
That changes how capacity announcements should be read. A gigawatt-scale data-centre project is not the same as deliverable capacity. In several regions, announced AI “gigafactory” pipelines are far larger than today’s installed base. Some will be built. Some will be delayed. Some will never move beyond strategic signalling. The gap between announced capacity and executable capacity is now one of the most important filters in the sector.
The reason is straightforward: AI infrastructure is running into physical ceilings. Grid connections can take years. Transformer lead times are averaging two to three years. Gas turbine deliveries can stretch to five years or more. Advanced chips depend on packaging capacity and high-bandwidth memory, where shortages may persist until late 2027. Helium adds another layer of fragility. It is essential in advanced chip manufacturing, and supply disruptions linked to Qatar and the Strait of Hormuz show how upstream geopolitics can affect AI infrastructure far from the data-centre site itself.
That is the less visible risk. A company may secure land, capital, local incentives, and even a power strategy, yet still be constrained by semiconductor inputs. The AI infrastructure race is not one bottleneck. It is a stack of bottlenecks, and the binding constraint can move.
The physical design of the data centre is changing just as quickly. The report’s most vivid comparison is a single advanced AI server rack: a structure roughly the size of a large refrigerator that could draw as much peak power as 65 households by 2027. Between 2020 and 2025, AI server power density increased 11-fold; by 2027, it is expected to rise another fourfold, with next-generation racks moving toward 1 MW of peak demand. That is not just an engineering detail. It explains why old assumptions about cooling, backup power, power conversion, and grid integration are breaking down.
Efficiency will help, but it will not remove the infrastructure problem. The report is clear that energy use per simple AI task has fallen sharply. A basic text prompt is not the issue. The more important question is what AI is becoming. Reasoning models, video generation, multimodal systems, and agentic workflows consume far more compute per completed task than simple text generation. As tools become cheaper and more capable, usage expands. This is the compute version of Jevons paradox: efficiency lowers the unit cost, and lower unit cost expands demand.
So the old assumption was that efficiency gains would offset load growth. The new reality is more complicated. Efficiency is necessary, but it also enables broader deployment. AI demand will likely scale into whatever infrastructure becomes economically and physically available.
That is why data centres are starting to look less like passive electricity customers and more like grid participants. The next generation of AI facilities creates large and rapid swings in power demand. Managing those swings requires batteries, advanced power electronics, cooling systems, and more sophisticated connection arrangements. By 2030, the IEA estimates that around 20–25 GW of battery storage could be installed in data centres globally. Under the right market rules, those batteries can do more than protect the facility. They can provide flexibility to the grid, support demand response, smooth load, and potentially accelerate interconnection through non-firm connection models.
This is an important operating-model shift. The data centre is no longer just buying electricity. It may become part of the reliability architecture of the power system. But that only works if tariffs, grid rules, and commercial incentives reward flexibility rather than treating every facility as a fixed load. The value moves from access to power to useful behaviour inside the power system.
Onsite gas generation is part of the same story, but it is not the simple workaround it sometimes appears to be. The United States is where this model is most advanced, with the IEA projecting that roughly 15–27 GW of onsite natural gas could power data centres by 2030, mostly in the US. Yet the economics carry a hidden redundancy tax. To achieve grid-level reliability using onsite gas alone, a data centre may need to overbuild generation capacity by 30% to 70% relative to actual demand. A 1 GW facility may need 1.3 GW to 1.7 GW or more of onsite capacity to manage outages and maintenance. That turns “self-supply” from a speed solution into a capital-allocation problem.
The United States is the most exposed market because it combines the largest data-centre base, the deepest AI capital pools, strong state-level incentives, large AI-factory projects, and rising interest in onsite gas generation. That creates an advantage, but also a concentration of risk. Northern Virginia has the connectivity and installed base. Texas has land, energy resources, and market flexibility. Both models face pressure from interconnection queues, generation adequacy, local opposition, turbine availability, and the politics of who pays for grid upgrades.
Europe has a different problem. It wants more digital sovereignty and has a policy ambition to expand data-centre capacity, but energy constraints remain central. The opportunity is less about winning a pure scale race and more about building credible low-emissions, grid-integrated infrastructure. That may be slower, but it could prove more durable if electricity affordability and public acceptance become harder constraints.
Asia-Pacific is more complex. Southeast Asia is attracting project pipelines through power availability, policy incentives, and regional demand growth. China is pursuing something strategically distinct. The report points to an application-first approach: more emphasis on industrial productivity, open-source enterprise models, robotics, manufacturing adoption, and cost efficiency. That matters because the competitive threat may not come only from frontier AI models. It may come from factories, logistics networks, battery supply chains, and energy-intensive industries that use AI to improve yields and lower operating costs.
This is where the energy story broadens beyond data centres. AI can reduce energy costs in energy-intensive industries by an estimated 3–10 percentage points through process optimisation. In sectors such as steel, aluminium, cement, chemicals, and refining, that is not a marginal gain. Energy is a core production input, and margins are often thin. A domain-specific AI model that improves furnace control, maintenance timing, process stability, or energy use can become a direct cost advantage.
The deeper transition is from predictive systems to adaptive systems. Predictive tools help forecast failures, demand, prices, or equipment behaviour. Adaptive systems go further: they adjust operations in real time, learn from changing conditions, and correct performance without waiting for manual intervention. That is where AI becomes more than analytics. It becomes part of the operating system of industrial production.
Yet the report also shows why adoption will be uneven. The largest barrier to AI uptake in the energy sector is not hardware. It is digital skills. From 2018 to 2024, the share of AI professionals in utilities, oil, gas, and mining was on average 40% lower than in sectors such as education, financial services, technology, information, and media. Fragmented data, legacy equipment, cybersecurity concerns, and limited digital capability all slow implementation. A company can buy software. It cannot instantly buy institutional learning.
This talent constraint may become one of the most underestimated cost drivers in the AI-energy transition. The market is focused on chips, power, and data centres because those bottlenecks are visible. But the ability to apply AI inside energy systems, industrial plants, grids, and physical assets depends on people who understand both models and operations. That combination is scarce.
The energy-sector upside from AI is therefore selective. The IEA does not support a broad “AI lifts all energy assets” thesis. The stronger signals are concentrated in bottleneck segments: transformers, switchgear, gas turbines, power electronics, batteries, nuclear offtake, geothermal, grid services, and specialised infrastructure. Utilities in data-centre-heavy regions may benefit, but only if regulation allows cost recovery without triggering affordability backlash.
This is where market structure changes. Broad exposure to electricity demand is less powerful than exposure to the constraints that make electricity usable for AI. The value is in interconnection, flexibility, dispatchability, cooling, equipment availability, and credible delivery. The same applies inside the AI sector. Model quality still matters, but infrastructure execution is becoming equally important. A better model has limited value if the capacity behind it cannot be powered, cooled, financed, staffed, or supplied.
The risk map is also becoming more physical. A capital-market reversal could slow the build-out if monetisation fails to justify infrastructure spending. Grid-cost backlash could emerge if households and businesses are seen to subsidise data-centre expansion. Equipment bottlenecks could delay projects even when demand is real. Semiconductor input shortages could cap server deployment before facilities are filled. Cyber-physical exposure will rise as energy systems become more digital and AI-enabled.
None of these risks imply that AI demand disappears. They imply that the next phase will be more disciplined. The market will separate announced capacity from executable capacity, generic AI access from domain-specific operating leverage, and passive electricity consumption from grid-aware infrastructure.
The most important conclusion from the IEA report is not that AI will consume too much energy. It is that AI is forcing a new test of industrial coordination. Capital, power, chips, data, skilled labour, regulation, and public acceptance now have to move together. When one part of that system lags, the whole growth story slows.
That is the real strategic shift. AI is no longer only scaling through code. It is scaling through substations, turbines, transformers, batteries, cooling systems, memory supply, permitting offices, and industrial teams capable of turning models into operating performance.
The companies and regions that understand this first will treat energy not as a back-office input, but as a strategic condition for AI deployment. The constraint is becoming the strategy.
