AI Now Consumes 210 TWh Annually as Inference Overtakes Training for the First Time

The energy consumed by AI data centres reached approximately 210 terawatt-hours per year in 2026 — up 38% from the prior year — with a structural shift that has significant implications for how the industry manages its environmental footprint: inference now accounts for 63% of the AI lifecycle's total electricity use, surpassing training for the first time.

Why the Shift Matters

From roughly 2022 through 2024, training dominated AI's energy profile. Building a frontier model required enormous computational resources concentrated over weeks or months, while inference — running the model for users — was comparatively modest per query. That relationship has now inverted.

The crossover point came in 2025, driven not by a sudden increase in per-query energy but by the exponential growth in deployment volume. Hundreds of millions of users now interact with large language models daily. Enterprise applications route financial analysis, code generation, document review, and customer interactions through AI systems continuously. A single advanced query against a reasoning-mode model can consume between 6 and 14 watt-hours; at scale, even modest per-query figures aggregate into substantial grid demand.

Current measurements put individual query energy consumption at 0.84 watt-hours for a standard GPT-5.5 chat request and 0.61 watt-hours for Gemini 3. Extended context requests and reasoning tasks are far more expensive: a Claude Opus 4.7 query using an 800,000-token context consumes 14.1 watt-hours, nearly seventeen times the standard figure.

Efficiency Improvements Have Not Offset Growth

The per-query efficiency picture is not uniformly bleak. Architectural advances — mixture-of-experts designs, KV-cache optimisation, and sparse attention mechanisms — have decoupled raw capability improvement from energy growth. One industry estimate suggests that model capability roughly doubled in the past 18 months while per-query energy fell approximately 12%. That decoupling is real, but it is not keeping pace with the 38% annual growth in total consumption driven by deployment volume.

Water consumption adds a parallel dimension to the resource picture. The global average for data-centre cooling is approximately 1.8 litres of water per kilowatt-hour consumed, but this varies dramatically by geography: data centres in Phoenix or Singapore require around 4.3 litres per kilowatt-hour, while those in Iceland or the Pacific Northwest use as little as 0.4 litres. Locating new AI infrastructure in warm, water-stressed regions — which often offer cheap land and good connectivity — carries costs not captured in electricity-price comparisons.

The Sustainability Gap

Absolute AI energy use growing at 38% annually poses a direct challenge to corporate decarbonisation targets and to national grid plans premised on gradual electrification. The International Energy Agency's projections for data-centre demand underestimated actual growth through 2025; revised 2026 estimates suggest the sector may account for between 1% and 1.5% of global electricity consumption by 2027.

Researchers have proposed several approaches to containing the trajectory. The most aggressive involves combining neural networks with symbolic reasoning systems, an architecture claimed to reduce per-inference energy by up to 100 times for certain categories of task by offloading structured computation to deterministic algorithms rather than probabilistic inference. Whether this scales to the full range of AI applications remains unproven.

Regulatory approaches are also entering the picture. The EU's Green Deal obligations require large data-centre operators to report energy use and, from 2026, to demonstrate progress toward efficiency targets. The European Commission's environment directorate published findings in early 2026 supporting research into repurposing data-centre waste heat for industrial water purification and carbon capture, an initiative framed explicitly as a mechanism for aligning AI infrastructure growth with circular-economy commitments.

What to Watch

The inference-dominance finding changes how AI companies should prioritise efficiency investments. Training-time improvements — more efficient hardware, better distributed-training algorithms — were the primary lever when training dominated. Now that inference accounts for nearly two-thirds of total consumption, deployment-time optimisation becomes at least as important: model quantisation, speculative decoding, smaller distilled models for common queries, and smarter routing between model tiers are the efficiency tools with the largest marginal impact.

Whether the industry moves fast enough on those fronts to stabilise its energy trajectory before grid and regulatory constraints bite will be one of the defining operational questions of the next two years.