AI Consumption in Context

Water is local: ~2/3 of new US data centres since 2022 sit in high water-stress areas (Bloomberg/S&P Global), and a June 2026 Guardian analysis found the same pattern in what's planned: 517 of 809 upcoming US facilities are on land that spent the past year in drought, drawn there by cheap land, tax breaks and less equipment corrosion. A large site can use up to 5 million gallons a day - the water use of a town of 50,000 people. The same facility can be benign in a wet region and a real problem in a dry one; where water is drawn matters more than how much. Siting swings the total: the Cornell study finds the best versus worst distribution of US AI servers across states moves the 2030 water footprint by -52% to +354%. Texas is the live case: data centres drew ~0.75% of the state's water withdrawals in 2025 (including power generation) and could reach 3-9% by 2030-2040 depending on buildout (UT Austin Bureau of Economic Geology). Those are withdrawals, not consumption.

Design choices matter, disclosure lags: OpenAI says its data centres have largely moved off evaporative cooling (Feb 2026), and Project Rainier - the 2.2 GW AWS campus in Indiana that trains Claude - is air-cooled for ~98% of the year (PUE ~1.15), though AWS won't quantify its operational water use. Microsoft's new designs since Aug 2024 use closed-loop chip-level cooling that evaporates no water (pilots in Phoenix and Wisconsin in 2026; fleet WUE 0.30 L/kWh in FY2024, down from 0.49 in 2021), and Google aims to replenish 120% of the freshwater its data centres consume by 2030. Facility-level water data remains largely undisclosed across the industry. And boundaries still rule the numbers: Mistral's 45 mL per reply versus Google's 0.26 mL is a 170x gap of methodology (Mistral counts electricity-generation water), not thirst.

Methodology: TV/streaming hours from eMarketer/Nielsen (2025). ChatGPT queries from OpenAI data: 2.5B+ daily prompts / 900M weekly active users ≈ 2.8 queries/day. Energy: streaming at 77 Wh/hour (~72% of which is your TV, not the data centre), ChatGPT at 0.34 Wh/query. Even heavy users (22 queries/day ≈ 7.5 Wh) use ~70x less energy than average daily TV/streaming; reasoning-heavy use runs ~25-30x higher per query.

Note: Discussions of AI energy consumption often conflate different use cases. One 10-second AI video is estimated at ~2,900x the energy of a text query - though no official figure exists for closed video models, and measured open video models span 0.14-109 Wh per clip (an 800x range).

The maths is moving again: reaching a fixed capability is getting cheap, fast. The same ~27% accuracy on FrontierMath took ~43 million output tokens with o4-mini in April 2025 but ~5 million with GPT-5.2 in December 2025 - about 9x fewer tokens (roughly 3x cheaper after per-token prices) in eight months, a trend Epoch AI puts at 5-10x per year. And every major lab now ships a thinking dial (effort, reasoning_effort, thinking_level) plus adaptive thinking, making the ~10x token blow-up a per-request choice rather than a default - Anthropic claims Opus 4.5 matched its predecessor's SWE-bench score with 76% fewer output tokens (vendor claim). The catch: at the frontier those savings get spent on harder problems, so per-query token use keeps climbing even as tokens-per-fixed-task collapses.

Why no frontier model comparison? Only Google (Gemini: 0.24 Wh per median prompt) and Mistral (Large 2: 1.14 g CO₂e per reply) publish audited per-model figures. OpenAI offers only an unaudited CEO figure with no methodology, and Anthropic, xAI and Meta disclose nothing per-query as of mid-2026 - so a GPT-5 vs Claude vs Gemini energy table would be guesswork. Anthropic declined environmental disclosure in Stanford's transparency index (Dec 2025) and has no sustainability report, though it joined the Frontier carbon-removal coalition in June 2026 and is hiring emissions-reporting staff. The measured data above comes from open models (Hugging Face AI Energy Score v2, ML.Energy v3).

Technical factors: Newer accelerators (NVIDIA markets Blackwell as 10-50x more efficient than Hopper; independent teardown puts the real per-watt silicon gain at ~47%), mixture-of-experts architectures, and serving optimisations drive the gains. Caveats: Google's 33x figure is self-reported, and 8 of 14 comparable models measured in Dec 2025 used equal or more energy than their early-2025 predecessors - newer isn't automatically greener.

Location factor: Carbon intensity varies significantly by grid location. Quebec (32 g CO₂e/kWh) vs Central US coal regions (1,000 g CO₂e/kWh) represents a 16x difference. Source: Harvard/UCLA Study.

Kept in proportion: Data centres are ~8% of projected global electricity demand growth to 2030 - less than EVs or air conditioning (Carbon Brief). And the buildout may not materialise as promised: Microsoft froze ~1.5 GW of self-build capacity, Satya Nadella conceded "there will be an overbuild", and the Bank for International Settlements lists an AI capex bust among its top systemic risks (2026).

Frontier training in 2025-2026: The largest credibly-estimated run is Grok 4 at ~310 GWh (Epoch AI, ~6x GPT-4) - yet GPT-5 used less training compute than GPT-4.5, as OpenAI scaled post-training on a smaller base model. Frontier training is not monotonically growing. Beware GW-scale headlines (Stargate, Colossus 2): those are build-out capacity targets, not measured consumption. Below the frontier, disclosure improves - and newer doesn't mean hungrier: Meta's GPU-hour disclosures put Llama 4 Maverick (2025) at ~1.7 GWh, an eighth of Llama 3.1 405B a year earlier; DeepSeek's technical report implies ~2 GWh for V3; and the only fully measured runs are small open models (BLOOM at 0.43 GWh; AI2's OLMo 3 series at ~1.9 GWh including facility overhead). Google has published no Gemini training figure and Anthropic none for Claude.

US adoption (Pew Research, Feb 2026, n=5,119): 49% of US adults have used AI chatbots (up from 33% in 2024 and 23% in 2023); 24% use them daily. Platform reach: ChatGPT 44%, Gemini 24%, Copilot 17%, Meta AI 14%, Grok 8%, Claude 6%. ChatGPT also passed ~1 billion monthly actives in June 2026 (a different metric to the weekly figures charted above).