IEA Sees Global Power Demand Growing Fastest in 15 Years as AI Load Strains Grids

IEA Sees Global Power Demand Growing Fastest in 15 Years as AI Load Strains Grids The agency says annual grid investment must rise 50% from $400 billion through 2030, while a separate study finds data-centre flexibility could cut peak load 45%. Global electricity demand is rising at its fastest pace in 15 years and will keep climbing at least through the end of the decade, the International Energy Agency said in a report carried this week (2026-05-19). Data centres, advanced manufacturing and electrification are driving the increase, with the agency putting annual average demand growth at 3.6% between 2026 and 2030.⁷ For power markets that have spent a decade pricing flat or falling consumption in mature economies, that is a regime change. The IEA estimates the world would need to lift annual grid investment by roughly 50% from $400 billion to keep pace through 2030. Higher industrial load, electric vehicles, air conditioning and AI compute all pull in the same direction.⁷ The question is whether the wires can be built fast enough. A study reported by Montel on Monday (2026-05-18) found that operating data centres flexibly could cut their contribution to peak power demand by up to 45% in 2035, avoiding 4 GW of fossil-fuel back-up. The premise is that AI training and inference loads, unlike homes and factories, can be shifted in time.¹ That flexibility matters for how utilities plan reserve margins. If a large share of data-centre demand can be curtailed or rescheduled around system peaks, grid operators avoid building peaking plants that run a few hundred hours a year. Cut the peak, and you cut the fossil back-up that would otherwise be locked in for decades.¹ But flexibility is a promise, not a delivered megawatt. Hyperscalers have shown little appetite for throttling expensive GPU clusters to suit a grid operator's duck curve, and the same compute that the study treats as movable is the compute that data-centre economics demand run flat-out. The 45% figure is a 2035 ceiling under favourable assumptions, not a base case.¹ The harder near-term constraint is interconnection. Battery storage developers in the United States are seeing surging interest from AI data centres but face years-long queues to connect to the grid and a supply chain still heavily dependent on China, Reuters reported on 18 May (2026-05-18). Storage is the obvious bridge between intermittent renewables and round-the-clock compute, yet the bridge itself is stuck in the queue.⁴ Without that storage, the load gets served by whatever is already dispatchable. AI data centres may keep fossil fuels in use despite rapid solar growth, one analysis argued on 19 May (2026-05-19), because round-the-clock demand cannot wait for the grid to decarbonise on its own schedule. The gap between demand growth and clean supply additions is where gas burn and coal life-extension live.³ The Southeast Asia case shows the scale. Wood Mackenzie expects data-centre power demand in the region to quadruple from 2.6 GW to 10.7 GW between 2025 and 2035, reaching 3-4% of peak demand from 1% now. A Bain and Standard Chartered report estimates data centres, EVs and green industrial parks will add more than 100 TWh over the next three to four years, requiring over $200 billion.^5,26 The same report flags an estimated $18 billion annual shortfall in grid investment by 2035, and asks whether power systems can expand quickly enough to support industrial growth without locking in higher emissions. That is the IEA's $400 billion question reframed at the regional level: the demand is contracted, the grid is not.² None of this has moved benchmark prices yet. Henry Hub front-month sat near $3.12 as of Friday's close (2026-06-12), and AI power demand remains a multi-year thesis rather than a balance-sheet event in the gas curve. Power demand growth at 3.6% a year compounds quietly; it does not gap a screen.⁷ For traders the read-through is in the build-out, not the headline. Watch interconnection queue data and storage deployment rates in US markets, because those govern how much new AI load gets served by gas peakers versus batteries. Watch whether hyperscalers actually sign demand-response contracts that turn the 45% flexibility study into committed capacity.^4,1 The structural story is the one to size correctly. A decade of stagnant demand made grid investment a regulatory afterthought; a 50% step-up in annual spending is now the price of keeping the lights on for compute. Whether utilities and regulators move at that pace, or let the queue ration growth, is the variable that decides which fuels clear the gap.^7,6