Year: 2026
Type: article
Abstract: Europe is now the fastest-warming continent, with 2024 its hottest year on record.1,2 Recent summers have been devastating, with heat contributing to an estimated 61 700 deaths in 2022 and 47 000 in 2023, with roughly half of all heat- and pollution-related mortality reflecting cardiorespiratory causes.3,4 Public-health agencies note that heat already accounts for >95% of climate- and weather-related deaths in Europe, a risk amplified by ageing populations and intensifying urban heat.5 Dry conditions are also fuelling more frequent and severe landscape fires with >440 000 hectares having burnt in Europe this year, over twice the area in 2024 (189 000 hectares).6 Collectively, these events are driving surges in air pollution and heat-triggered spikes in ischaemic events, heart-failure decompensation, dehydration-related acute kidney injury, and heat stroke, overwhelming emergency departments. A new and fast-escalating factor in global-warming projections is the enormous energy demand of artificial intelligence (AI). Fossil fuels are expected to provide more than half of data centre power through 2030, with the International Energy Agency (IEA) projecting data centre electricity use to nearly double to ∼945 TWh, about 3% of global demand.7 This surge was absent from climate scenarios even 5 years ago. From Dubai to Dublin, AI training workloads are prompting new investments in fossil-fuel generation, as these cannot be met by renewable sources. More than 85 large gas-fired plants are now in development worldwide to feed data centre expansion, with China and India also meeting rising demand with coal. In 2024, 44 GW of new coal capacity came online, while only 25.2 GW retired, expanding the global coal fleet by 18.8 GW.8 China proposed 94 GW of new coal capacity in 2024, its highest since 2015 and India added 38 GW; together accounting for 92% of all proposals globally. In contrast, Europe continues its phase-out: EU27 countries retired 11 GW of coal in 2024, led by Germany (6.7 GW), and the UK closed its last coal plant that same year. A typical 1 GW coal unit emits roughly 5–6 million tons of CO2 per year, equivalent to several million passenger cars, and produces substantial PM2.5 and toxic metals.9 Gas generates fewer pollutants but methane leakage substantially erodes its climate advantage. Training large-scale AI models, such as the latest language and foundation models, is an intensely energy-hungry process. Artificial intelligence workloads can run for days or weeks, drawing on thousands of Graphics Processing Units (GPUs) or Tensor Processing Units (TPUs) operating at full capacity. The resulting energy use is both enormous and irregular. A single training run, for instance, can consume as much electricity as thousands of homes use in a year. Running inference, which is the process of deploying trained models for real-time use, may require less energy, but at a global scale with billions of requests, it can rival or exceed that of training. Both training and inference create sharp, sometimes unpredictable spikes in demand that are poorly aligned with the daily and seasonal production patterns of wind and solar power. In contrast, the energy demands of data storage are both steady and relentless. Storing AI model weights, massive training datasets and user content in hyperscale data centres, requires a continuous, high-reliability power supply. The mismatch between AI-driven load requirements and renewable energy generation is creating unique structural and operational challenges. Wind and solar output vary hour to hour, day to day, and season to season. While short-duration battery storage can smooth some fluctuations, covering multi-day or seasonal lulls remains economically and technically challenging. The problem is magnified during unplanned AI training runs or inference surges, during periods of low renewable output, forcing grids to rely on gas turbines or oil-fired boilers. Without significant changes in the way AI workloads are scheduled and powered, this unfettered growth is becoming a structural driver of fossil-fuel dependence at the very moment when emissions need to be falling sharply. Meeting this challenge will require a combination of zero-carbon strategies such as advanced nuclear, geothermal, or hydropower, combined with pumped storage. The latter are long-duration energy storage options that extend far beyond the 4–8 h range of current lithium–ion systems. Additionally, demand-management practices that shift non-urgent AI workloads to periods of abundant clean power are a must. Given this large mismatch between power needs for AI throughout the world, how are global technology companies, including Amazon, Meta, and Google greenwashing themselves as sustainable and/or ‘100% matched’ with renewable energy? Energy credits, where companies can purchase credits or can make investments directly through power purchase agreements in clean energy years out in the future, are a big reason. For example, Amazon may pay for wind or solar energy to be produced for a decade, earning the right to claim credits representing the additional clean energy. These arrangements, being purely financial in nature, do not change the fossil-fuel power requirements of the company through the electricity grid. Additionally, the credits do not have to match the location where energy is being used nor the time of day when it is needed. Thus, a company could purchase carbon credits in Brazil for AI energy requirements in Ireland or Germany. This arrangement clearly dissociates shifts in clean energy, both spatially and temporally, to a nebulous future state, which may or may not transpire. Managing the dual challenge of accelerating AI-driven electricity demand and heat-related cardiovascular health impacts requires integrated action. First, heat adaptation should be embedded into core health system planning, with funding for heatwave early warning systems, cooling access programmes, and targeted communication directed at patients at high risk for heat-related cardiovascular events.10 Second, governments must link permits for large-scale AI and data centre expansion to verifiable clean power procurement. Thus, buying and/or investing in green carbon credits alone should not be allowed to sideline and ignore green commitments. Finally, policymakers should formally account for the healthcare costs of extreme-heat- and air-pollution-induced cardiovascular morbidity and mortality in energy and infrastructure planning processes, ensuring that climate, energy, and health policy operate in a unified framework rather than in silos. This integrated approach would not only slow emissions growth but also protect public health from the direct and indirect effects of a warming planet. Nothing to declare.
Source: European Heart Journal
Language: English
Subfield: Health, Toxicology and Mutagenesis
Field: Environmental Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W7155189347