AI Model Sustainability 2026: Energy Use Data

Frontier-model energy use stopped scaling linearly with capability in mid-2025 — the architectural compression of mixture-of-experts and aggressive KV-cache optimization broke the previous trend line. But absolute consumption is still climbing 38% year-over- year because deployment volume is outpacing per-query efficiency by a wider margin than ever.

This 2026 quarterly compiles per-query energy across the four leading frontier models, the training-vs-inference split (which crossed over in 2025 — inference now dominates), water-use data from datacenter operators, and the trend analysis from the Stanford AI Index 2026 environmental section.

The numbers are best-available estimates derived from public architecture data, ML.Energy per-FLOP measurements, and reasonable assumptions about data-center PUE and grid mix. They're not first-party figures from the model providers (none publish per-query energy yet) but they are the best quantitative view the field has compiled.

Three numbers anchor the 2026 sustainability picture: total data-center electricity attributable to AI, year-over-year growth in that number, and per-query energy at the frontier. All three are larger than the 2024 baselines, but the per-query trend has broken the linear-with-capability scaling that earlier eras assumed.

Per-query energy varies by model architecture, query type, and context length — by 20× or more across the realistic spread. The numbers below are estimates for typical call profiles, normalized to data-center kWh including PUE overhead. They represent the operational cost of a single user-facing call.

The reasonable read: standard chat queries land in the 0.6-0.9 Wh band — essentially equivalent to running a 60W laptop for 45-60 seconds. Reasoning queries scale 6-10× higher because of the internal reasoning traces. Long-context queries scale 15-25× higher because of attention's superlinear cost. Pick the mode that fits the task; reasoning and long-context are genuinely energy-intensive and shouldn't be the default.

The water-use story is more regional than the energy story. Average data-center water consumption is 1.8L per kWh of compute energy, dominated by evaporative cooling for heat rejection. In hot, water-stressed regions (Phoenix, Singapore, parts of Texas) it climbs to 3-5L per kWh. In cool, water- abundant regions (Iceland, Pacific Northwest, parts of Scandinavia) it drops below 0.5L. Siting matters.

The 2023-2024 narrative was that training dominated AI's energy footprint. That was true at the time — large frontier models took 50-200 GWh to train, and inference deployment volumes were orders of magnitude lower than they are now. By 2025, the math flipped: deployment volume scaled 3-5× per year, training spend scaled ~1.5× per year, and inference passed training as a share of lifecycle energy.

Two trend lines tell the 2024-2026 story. Per-query energy is slightly down (-15% over two years, despite ~2× capability improvement) thanks to MoE and KV optimization. Absolute energy is up significantly (+38% YoY) because deployment volume is growing faster than per-query efficiency. Both numbers are simultaneously true; the framing depends on which one matters for the question being asked.

• Use reasoning mode selectively. Reasoning traces (Deep Think, GPT-5.5 Pro Thinking) cost 5-12× the energy of standard inference. Default to non-reasoning; opt into reasoning for tasks that genuinely require it (math, complex agent planning, ambiguous reasoning over data).
• Cache aggressively at long context. An 800K cached call costs ~10% of an uncached one. For long-context workloads, cache topology is also the biggest sustainability lever the deployment owns. The cost gain and the energy gain point the same direction.
• Trim output amplification. A 200K input that elicits 10K of output uses dramatically more energy than the same input eliciting 500 output tokens. Set explicitmax_tokens budgets and output-shape system prompts to reduce per-call output by 60-80% on long-context workloads.
• Pick smaller models when capability allows.A 70B-class open-weight model running at FP8 uses dramatically less energy per query than a 1.6T MoE model. For workloads where capability isn't the bottleneck — summarization, classification, routing — use smaller models. The energy savings compound at deployment volume.
• Prefer providers with low-carbon grid mix and closed-loop cooling. Pacific Northwest, Iceland, Quebec, Scandinavia. The grid-mix difference between coal- heavy and renewable-heavy regions is 5-10× on emissions per query. Where data-residency allows, pick low-carbon regions.