Key findings

Global electricity consumption from AI chip manufacturing rose from 218 GWh in 2023 to nearly 984 GWh in 2024, a year-on-year increase of more than 350%. Associated emissions increased over fourfold, from about 99,200 to 453,600 tonnes of CO₂ equivalent, driven largely by fossil-fuel-heavy electricity grids in East Asia. By 2030, electricity demand from AI chipmaking is projected to reach 11,550–37,238 GWh, up to 170 times 2023 levels and exceeding Ireland’s current electricity consumption

Introduction

The report examines an often overlooked component of AI’s environmental footprint: upstream chip manufacturing. While policy and corporate scrutiny has focused on data centres, advanced semiconductor production is highly energy intensive and geographically concentrated. The analysis focuses on electricity use and emissions from manufacturing leading AI chips, particularly GPUs and high-bandwidth memory, and assesses future impacts as AI demand accelerates.

The growing AI industry

AI investment has expanded rapidly, with global spending by major technology firms reaching hundreds of billions of US dollars annually. AI chips are central to this growth, enabling training and inference for advanced models. Production is dominated by East Asian manufacturers, including TSMC, SK hynix, Samsung and Micron. This concentration exposes regional power systems to rising industrial electricity demand and embeds AI supply chains within carbon-intensive grids.

The energy demand of AI

AI’s energy footprint spans chip manufacturing, model training and inference. Current debate has focused on downstream energy use, but the report shows upstream manufacturing is already material and growing quickly. Semiconductor fabrication is electricity intensive due to complex, repetitive processes, with wafer fabrication the largest contributor to energy use and emissions. As process complexity increases with advanced nodes, electricity intensity rises further.

Energy demand from data centers

Global data centres consumed an estimated 240–340 TWh of electricity in 2022, around 1–1.3% of global demand. AI workloads are expected to add approximately 200 TWh per year through to 2030, potentially doubling data centre emissions. While some large technology companies report matching operational electricity use with renewables, reliance on instruments such as unbundled renewable energy certificates raises concerns about real-world emissions reductions.

Energy demand from AI chipmaking

Using a bottom-up methodology, the report estimates wafer demand and electricity use for major AI models, including Nvidia A100, H100, H200 and B100/200, and AMD MI300X. Wafer demand more than tripled between 2023 and 2024. Taiwan saw electricity consumption from AI chipmaking increase from 83.4 GWh to 375.8 GWh, while South Korea rose from 134.6 GWh to 315.2 GWh. Emissions followed similar trends, reflecting fossil fuel shares of 83.1% in Taiwan, 68.6% in Japan and 58.5% in South Korea. Without grid decarbonisation, projected 2030 demand could result in 5.2–16.8 million tonnes of CO₂ equivalent annually.

Recommendations

The report calls for AI chipmakers and technology companies to target 100% renewable electricity across supply chains by 2030. High-impact sourcing, including power purchase agreements, direct investment and on-site generation, is prioritised over expanded fossil fuel or nuclear capacity. Companies are encouraged to engage suppliers and policymakers to accelerate renewable deployment, while governments in East Asia are urged to prioritise renewable supply for semiconductor manufacturing and avoid further fossil fuel lock-in.