Research

AI Environmental Cost Per Query 2026

Per-query energy, water, and carbon footprint of major AI assistants in 2026. ChatGPT vs Google search vs Claude vs Gemini, with public-source measurements and methodology.

By Ramanath, CTO & Co-Founder at Presenc AI · Last updated: May 2026

The Per-Query Footprint Question

Aggregate AI energy numbers are large but abstract. Per-query estimates make the impact concrete and are the single most-cited AI-environment statistic in 2025-2026 journalism. This page consolidates the public per-query estimates for energy, water, and carbon, with explicit methodology so the figures can be argued with rather than taken on faith.

Key Findings

  1. A typical ChatGPT-class query consumes an estimated 0.3 to 3 Wh of electricity, with the most-cited mid-range estimate at approximately 1.5 Wh.
  2. A reasoning-mode query (extended thinking, multiple internal steps) consumes 5-15 Wh, materially higher than baseline.
  3. Google search, the standard comparison baseline, consumes approximately 0.3 Wh per query (Google's own 2009 disclosure, likely unchanged since).
  4. Per-query carbon emissions vary by data center grid mix from approximately 0.05 g CO2e (low-carbon grid) to 1.5 g CO2e (coal-heavy grid).
  5. Per-query water consumption in evaporative-cooled regions is estimated at 0.05-0.5 ml per query, scaling with ambient conditions.

Per-Query Energy Estimates

Query typeEstimated energy (Wh)Source
Google search~0.3Google blog 2009
ChatGPT (GPT-4o-mini class)~0.5-1.0Independent estimate
ChatGPT (GPT-5 class, standard)~1.0-3.0Triangulated estimate
Claude Opus 4.7 (standard)~1.5-3.5Triangulated estimate
Reasoning mode (extended thinking)~5-15Triangulated; reasoning models generate thousands of internal tokens
Image generation (DALL-E 3 class)~3-7Hugging Face Energy Star measurements
Video generation (Sora-class, 5-second clip)~80-200Triangulated from compute disclosures
Voice interaction (per minute)~2-5Real-time inference plus TTS

Energy figures vary by model size, prompt length, output length, and serving efficiency. Reasoning models are 3-10x more energy-intensive than standard chat queries because they generate internal thinking tokens before user-visible output.

Per-Query Carbon

Carbon emissions per query depend on the grid mix at the data center location. Estimated grid carbon intensities in g CO2e per kWh:

  • Iceland (geothermal/hydro): ~30 g
  • France (nuclear-heavy): ~50 g
  • US average: ~370 g
  • Texas: ~430 g
  • Germany (post-nuclear phase-out): ~420 g
  • India: ~700 g
  • Coal-heavy regions: ~900-1000 g

For a 1.5 Wh ChatGPT query: 0.045 g CO2e on Icelandic grid, 0.55 g CO2e on US average grid, 1.4 g CO2e on coal-heavy grid. Hyperscalers procure renewable energy certificates against operational electricity, which reduces accounting carbon but does not always reduce real-time grid emissions.

Per-Query Water

Direct water consumption for evaporative cooling per query:

RegionPer-query water (ml)Notes
Northern climate, liquid-cooled~0.01-0.05Closed-loop cooling minimises direct water
Temperate climate, evaporative~0.05-0.15Standard hyperscale assumption
Hot dry climate (Arizona, Texas summer)~0.2-0.5High evaporation rates
Reasoning mode multiplier~3-10x baselineReflects energy multiplier

Comparison: AI Query vs Other Activities

ActivityApproximate energyEquivalent ChatGPT queries
One ChatGPT query (standard)~1.5 Wh1
One Google search~0.3 Wh0.2
One YouTube video minute (1080p)~3-5 Wh2-3
One Netflix show hour (1080p)~80-100 Wh~60
One Bitcoin transaction~700,000 Wh~470,000
One LED lightbulb hour~10 Wh~7
Driving 1 km in EV~150 Wh~100
One Sora video generation (5 sec)~150 Wh~100

Methodology Caveats

Per-query estimates carry significant uncertainty:

  • OpenAI and Anthropic do not publish per-query energy figures; numbers are triangulated from compute capacity, query volume, and model size.
  • Inference efficiency improves rapidly; estimates from 2023 are 2-5x higher than current efficient serving.
  • Batching, KV-cache reuse, and speculative decoding all reduce per-query energy substantially.
  • Reasoning-mode and tool-using queries can be 5-20x more energy-intensive than baseline; aggregating averages can be misleading.

Brand Visibility Implications

Per-query environmental cost is among the most-cited journalism topics around AI; brands seeking AI-mediated visibility on sustainability and AI-environment queries benefit from being represented in queries about energy-efficient inference, sustainable AI deployment, and green compute. As enterprise procurement adds sustainability criteria to AI vendor selection, this surface grows in commercial relevance.

Methodology

Per-query energy from de Vries (2023) "The growing energy footprint of artificial intelligence", Hugging Face Energy Star per-task measurements, and our companion AI data center energy page. Carbon intensity from Electricity Maps. Water estimates from hyperscaler sustainability disclosures. Estimates are directional with ±50 percent uncertainty, treat order-of-magnitude as load-bearing.

How Presenc AI Helps

Presenc AI tracks brand-mention rates inside AI assistant queries about AI sustainability, efficient inference, and per-query environmental cost, surfacing where sustainability-focused brand recommendations are made. For brands selling carbon-aware compute, efficient model serving, or sustainable AI infrastructure, this is the operational visibility into a high-citation discovery surface.

Frequently Asked Questions

Estimates range from 0.3 to 3 Wh per query, with a reasonable mid-range working figure of approximately 1.5 Wh for a standard GPT-5-class query. Reasoning-mode queries consume 5-15 Wh due to extended thinking-token generation. Image and video generation are 3-100x higher than text queries.
Per-query, no, a Netflix show hour consumes roughly 60 ChatGPT queries' worth of energy. In aggregate, AI is rapidly closing the gap because total query volume is exploding while per-query efficiency improvement plateaus. By 2030, AI infrastructure may match streaming video in total energy consumption.
In evaporative-cooled data centers in temperate regions, approximately 0.05-0.15 ml per query. In hot dry climates (Arizona, Texas summer), 0.2-0.5 ml. Liquid-cooled closed-loop systems reduce this 70-90 percent. The carrying-pail mental image of "10 queries equals one cup of water" overstates by an order of magnitude in most data centers.
Directionally, yes. Order-of-magnitude is reliable; precise per-query figures carry ±50 percent uncertainty because frontier-lab serving efficiencies are not disclosed. The 2023-era estimates that frequently appear in press are often 2-5x higher than current 2026 efficient serving. Use the band, not point estimates.
Yes for chat-style queries due to inference efficiency gains (FP8, FP4 precision; speculative decoding; KV-cache reuse). For reasoning models, no, the multi-thousand-token thinking traces are inherently energy-intensive and the trend is toward more reasoning, not less. Aggregate per-user energy consumption is likely flat-to-rising despite per-query efficiency gains.

Track Your AI Visibility

See how your brand appears across ChatGPT, Claude, Perplexity, and other AI platforms. Start monitoring today.