Quick take: Training and running large AI models consumes substantial electricity and water. AI data centers are growing faster than renewable energy capacity in many regions, and water used for cooling is becoming a local resource issue near major data centers. These costs are real and significant, though their scale relative to other industries and the potential for efficiency improvements makes the picture more nuanced than either dismissal or alarm.
Every prompt you send to ChatGPT, every image you generate with Midjourney, every AI feature you use in your apps runs on data centers consuming enormous amounts of electricity and water. The environmental cost of AI has been mentioned in passing by tech coverage but rarely examined in proportion to scale. Understanding what these costs actually are — and how they compare to alternatives and historical precedents — produces a more accurate picture than either ignoring them or catastrophizing.
The Energy Cost of Training
Training a large language model requires running thousands of high-end GPUs continuously for weeks or months. A 2023 estimate put the training compute for GPT-4 at roughly 2 x 10^25 FLOPs, consuming an estimated 50 GWh of electricity — roughly equivalent to the annual energy consumption of 5,000 average US households. This is for training once; subsequent version training runs have similar or larger costs.
The training cost is a one-time expense per model version — the training happens once, and the resulting model handles millions of queries. Inference — running the trained model to respond to individual queries — is a smaller per-query cost but adds up at massive scale. Microsoft estimated that adding AI to search queries would require additional data center capacity, specifically because AI inference is significantly more compute-intensive than traditional search. Each AI search query uses roughly 10x the energy of a traditional search query.
Data center electricity consumption is estimated at 1-2% of global electricity use currently, with AI contributing a growing share. Goldman Sachs projected AI could drive a 160% increase in data center power demand by 2030. In the US, data centers already consume roughly 4% of national electricity, and AI-driven growth is straining grid capacity in specific regions. Several AI data center projects have faced opposition from local utilities unable to supply the requested power.
Water: The Less-Discussed Cost
Data centers use water for cooling — evaporative cooling systems remove heat by evaporating water, which is more efficient than purely air-based cooling. A 2023 study estimated that training GPT-3 consumed approximately 700,000 liters of water. Microsoft reported that its global data center water withdrawal increased 34% between 2021 and 2022, attributable largely to AI workloads. Google reported similar trends.
The water impact is geographically concentrated — it affects communities near major data centers, often in arid regions where water is already stressed. Arizona and Nevada host major data centers specifically because land and power costs are low, but these states face water scarcity. The local impact — drawing on the same aquifers as communities and agriculture — is more concrete than global energy numbers suggest. Data center water consumption is becoming a planning and community relations issue in ways that were not anticipated a decade ago.
The energy cost of AI is not uniformly distributed — frontier models run on data centers that are increasingly powered by negotiated renewable energy agreements. Microsoft, Google, and Amazon have all signed large renewable energy purchase agreements partly in response to data center demand. The question is whether renewable capacity additions are keeping pace with AI-driven demand or whether AI growth is increasing fossil fuel use on the margin. The answer varies by region and time period.
Efficiency Improvements: The Countervailing Trend
AI hardware has become dramatically more energy-efficient over time. NVIDIA’s H100 GPU performs roughly 30x more AI computations per watt than its predecessors from five years ago. Model architecture improvements — smaller, more efficient models achieving similar performance through better training — also reduce per-query energy. Quantization (reducing numerical precision) and distillation (training smaller models to match larger model outputs) reduce inference costs substantially.
These efficiency improvements follow a pattern seen in previous computing waves: absolute consumption grows because scale grows faster than efficiency improves. The efficiency improvements are real and substantial — they mean AI is significantly less energy-intensive than it would otherwise be. Whether they’re sufficient to offset demand growth is an empirical question that changes yearly as both factors evolve.
Putting It in Context
Contextualizing AI’s energy use helps calibrate the concern. Bitcoin mining consumes roughly 150 TWh per year, comparable to the annual electricity consumption of Poland — several times the current estimate for all AI inference. Global data centers (pre-AI) already consumed roughly 200-250 TWh per year. Video streaming globally consumes more energy than AI inference. This context doesn’t mean AI’s energy use is unimportant — it suggests the concern should be proportional and that AI isn’t uniquely problematic relative to other digital services.
The harder question is growth rate. Bitcoin’s energy use grew rapidly then plateaued as adoption leveled off. AI’s energy use is growing rapidly and has not plateaued. If AI adoption follows the trajectory of the most optimistic capability predictions, energy demand from AI could become significant relative to global electricity generation within a decade. Whether that demand is met by fossil fuels or renewables — a policy and investment question — determines the climate impact.
Claims that AI will be environmentally benign because of efficiency improvements and renewables commitments should be evaluated against actual data. The efficiency trend is real; so is the growth trend. Renewable commitments involve purchase agreements that don’t necessarily mean the marginal data center power comes from renewables — they often involve accounting mechanisms that allow renewable credit purchases to offset consumption that is physically served by grid power with a mixed generation mix. The environmental case for AI depends on what the actual marginal energy source is, which varies by location and time.
- Training GPT-4 consumed an estimated 50 GWh — equivalent to 5,000 US households annually; inference adds per-query cost at massive scale.
- AI data center water use for cooling is significant and geographically concentrated near arid-region data centers.
- GPU efficiency has improved ~30x per watt in five years, but scale growth outpaces efficiency gains in absolute terms.
- AI inference energy is comparable to Bitcoin mining but growing faster; streaming and pre-AI data centers use more currently.
- Renewable energy commitments exist but don’t guarantee the marginal power is renewable — location and time matter.
- The climate impact depends on whether AI demand growth is met by renewables or fossil fuels — a policy question as much as a technology one.
Frequently Asked Questions
How much energy does ChatGPT use per query?
Estimates vary and OpenAI doesn’t disclose energy use per query. IEA estimated roughly 10x the energy of a Google search per ChatGPT query, which would be approximately 0.001-0.003 kWh per query. At millions of queries per day, this aggregates to meaningful consumption. Image generation is substantially more compute-intensive than text generation; multimodal reasoning falls between.
Is AI bad for the environment?
Complex. AI has environmental costs (energy, water, hardware manufacturing) and potential environmental benefits (enabling better climate modeling, material discovery for clean energy, grid optimization). Net impact depends on use cases and the energy mix powering the compute. The environmental concern is legitimate and often underdiscussed; it’s also less catastrophic than some coverage suggests relative to other high-energy industries.
Are AI companies doing anything about energy use?
Significant renewable energy purchase agreements, investments in nuclear power (Microsoft contracted for Three Mile Island restart), efficiency improvements in hardware and software, and research into more efficient architectures. Whether this is sufficient is contested. Some companies have moved away from pledging carbon neutrality on specific timelines as data center demand has grown beyond projections, which suggests honest acknowledgment that the commitments were aspirational rather than guaranteed.
AI energy consumption, ChatGPT electricity use, AI data center water usage, AI carbon footprint, GPU energy efficiency, AI environmental impact, data center power consumption, AI renewable energy