 |
VOOZH | about |
|
VOOZH | about |
Yes, AI is bad for the environment, and the numbers are no longer up for debate. Global data-center electricity use is set to jump from 415 TWh in 2024 to roughly 945 TWh by 2030 per the International Energy Agency, mostly because of AI. Cornell researchers project AI servers alone will emit 24 to 44 million metric tons of CO₂ and consume 731 to 1,125 million cubic meters of water per year by 2030. That is equivalent to adding 5 to 10 million cars to U.S. roads and matching the household water use of 6 to 10 million Americans.
That is the bad news, and it is real. The longer answer is more useful: AI’s footprint depends enormously on which model you use, where it runs, and what you ask it to do. A single text prompt to Gemini burns about 0.24 Wh and 0.26 mL of water. The exact same kind of prompt to DeepSeek-R1 can burn roughly 29 Wh, more than 100 times as much. In this article you will get a clear verdict, the five reasons AI is environmentally harmful in 2026, and a per-model comparison table. You will also see practical steps you can take if you still want to use AI without making the problem worse.
- AI is environmentally harmful, but the per-prompt cost is smaller than the viral “bottle of water” headlines. A single ChatGPT prompt uses about 0.34 Wh and 0.000085 gallons of water, roughly one fifteenth of a teaspoon, per OpenAI CEO Sam Altman.
- Data-center electricity demand is on track to more than double by 2030. The IEA’s Energy and AI report projects 415 TWh in 2024 → 945 TWh in 2030, with AI accounting for nearly half the growth.
- Model choice matters more than you think. The gap between the most and least efficient AI systems is over 65 times per peer-reviewed research from Jegham et al.
- Generative video is in a different league. A single 10-second Sora 2 video uses about 1 kWh of electricity and 4 liters of water, roughly 2,000 times more per second of output than a text prompt.
- AI can also help the environment. The Grantham Institute estimates AI could cut global emissions by 3.2 to 5.4 billion tonnes by 2035 if deployed well, while Boston Consulting Group puts the upside at 5 to 10 percent of all greenhouse-gas reductions by 2030.
The short answer is bad enough to matter, but not yet catastrophic on a per-user basis. Data centers today consume around 1.5 percent of global electricity, rising to nearly 3 percent by 2030 per the IEA. They are responsible for about 0.5 percent of global CO₂ emissions. Those numbers are small compared to transport or steel, but they are growing faster than almost any other sector.
The real worry is the trajectory. In the United States, datacenters are projected to add 240 TWh of new electricity demand by 2030, a 130 percent increase. In China, 175 TWh, a 170 percent jump. The IEA expects roughly 40 percent of that new demand to still come from gas and coal. That is why the regional picture looks ugly even though the global average looks manageable. In Dublin, data centers already consume 79 percent of city electricity. In Virginia, they hit 26 percent of state-wide consumption in 2024.
Water is the second pressure point. A single hyperscale data center can use up to 5 million gallons of water per day, the same as 16,000 average U.S. households. Training GPT-3 alone evaporated an estimated 700,000 liters of clean freshwater in Microsoft’s U.S. facilities according to UC Riverside research.
The bottom line: AI is not destroying the planet today, but the next five years will determine whether it becomes a major climate problem or a manageable one. The siting decisions, model efficiency choices, and grid mix being locked in right now will lock in the outcome.
For the full breakdown of where the water goes, we built the full breakdown of AI’s water consumption and how much electricity AI really consumes in 2026 as deeper companion guides.
The damage breaks down into five distinct vectors. Skim the table, then read the section that worries you most.
| Reason | 2030 projection | Comparison anchor | Primary source |
|---|---|---|---|
| Electricity demand | 945 TWh global data-center load | Roughly 2.3× Japan’s annual electricity use | IEA Energy and AI, 2025 |
| Water consumption | 731 to 1,125 million m³ for AI servers per year | Annual home water use of 6 to 10 million Americans | Cornell, Nature Sustainability Nov 2025 |
| Carbon emissions | 24 to 44 MMT CO₂ added per year from AI | Like putting 5 to 10 million cars on the road | Cornell, Nature Sustainability Nov 2025 |
| E-waste | 16 million tonnes by 2030 | Toxic mercury, lead, arsenic in landfills | Research Square 2024 |
| Critical minerals | Copper demand roughly doubles | Strain on rare earths, water-intensive mining | UN Environment Programme |
Data centers were a 1.5-percent slice of global electricity in 2024. By 2030, the IEA expects them to hit nearly 3 percent, with AI workloads driving almost half of the entire growth. Accelerated servers running AI grow at 30 percent annually; everything else grows at 9 percent.
The grid cannot absorb that without burning more fossil fuel. The IEA forecasts that around 40 percent of the new demand will still come from gas and coal by 2030. Gas generation for data centers is on track to more than double from 120 TWh in 2024 to 293 TWh by 2035 according to Carbon Brief’s analysis.
Training one large model needs water for cooling and even more water indirectly, because thermal power plants that generate the electricity also need cooling water. A megawatt-hour of data-center power uses about 1,900 liters of water directly and another 4,540 liters indirectly per the Environmental and Energy Study Institute.
Cornell researchers project AI servers alone will pull 731 to 1,125 million cubic meters of water per year by 2030, the household water use of 6 to 10 million Americans. Over 50 percent of new data centers built since 2022 sit in water-stressed regions per SNHU’s reporting.
If you want the ChatGPT-specific numbers, our deep dive on how much water ChatGPT actually uses per prompt covers the 0.000085 gallons per query figure and the viral “bottle of water per email” claim that turned out to be wildly overstated.
By 2030, the same Cornell study projects AI servers will add 24 to 44 million metric tons of CO₂ per year in the U.S. alone, equivalent to 5 to 10 million extra cars on the road. Amazon’s total emissions rose 6 percent in 2024 versus 2023, and 42 percent of executives told Capgemini they are re-examining climate goals because of AI growth.
Worst-case projections from Morgan Stanley put generative AI emissions at 2.5 billion tonnes of CO₂ by 2030, roughly 40 percent of total U.S. emissions today. Researcher Alex de Vries Gao calculated in the journal *Patterns* (2025) that the AI boom in 2025 released as much CO₂ as the entire city of New York.
AI servers run on cutting-edge GPUs that are replaced every two to three years to keep up with model demands. Research Square estimated in 2024 that AI-related e-waste could hit 16 million tonnes by 2030. That hardware contains mercury, lead, arsenic, and other toxic substances that contaminate soil and groundwater when dumped.
E-waste is the most underreported part of the AI footprint. Most environmental coverage skips it entirely because it does not show up in the daily prompt-by-prompt accounting.
AI infrastructure needs copper for power delivery, lithium and cobalt for backup batteries, and a long list of rare-earth elements for the chips themselves. The UN Environment Programme warns that AI demand could nearly double global copper consumption. Rare-earth mining is water-intensive and often happens in countries with weak environmental protections, exporting AI’s costs to communities that gain almost none of the benefit.
This is where headlines get the most distorted. A single text prompt is small. The aggregate is enormous. The table below shows the latest peer-reviewed and first-party numbers per model.
| AI Model | Energy per prompt (Wh) | Water per prompt (mL) | CO₂ per prompt (g) | Source |
|---|---|---|---|---|
| Google Gemini (median) | 0.24 | 0.26 | 0.03 | Google, August 2025 |
| ChatGPT (avg) | 0.34 | ~0.32 | ~0.17 | Sam Altman blog, 2025 |
| GPT-4o (long prompt) | 2.88 | n/a | varies | Jegham et al., arXiv 2025 |
| DeepSeek-V3 (Azure, long prompt) | 3.70 | n/a | varies | Jegham et al., arXiv 2025 |
| Mistral Le Chat | >3 | n/a | 1.14 | Mistral disclosure |
| Claude 3.7 Sonnet (long prompt) | 5.67 | n/a | varies | Jegham et al., arXiv 2025 |
| DeepSeek-R1 (Azure, long prompt) | 7.41 | n/a | varies | Jegham et al., arXiv 2025 |
| LLaMA-3.1-405B (long prompt) | 25.20 | n/a | varies | Jegham et al., arXiv 2025 |
| DeepSeek-R1 (self-hosted, long prompt) | 29.08 | n/a | varies | Jegham et al., arXiv 2025 |
| GPT-5 (heavy query) | ~18, up to 40 | a few to tens of mL | varies | University of Rhode Island AI Lab |
Three things jump out. First, the spread between the most and least efficient systems is over 65 times per Jegham et al., comparing the 29 Wh DeepSeek-R1 peak against the 0.44 Wh figure for LLaMA-3.1-8B. Second, reasoning models like DeepSeek-R1 and OpenAI’s o-series can use 50 to 100 times more energy than a standard query because they think before answering. Third, cheap to use does not mean clean to run: DeepSeek’s low API price hides one of the highest per-prompt energy footprints in the table.
For the deepest breakdown, see per-prompt energy figures for ChatGPT and how ChatGPT and Gemini compare on efficiency.
Text is the cheap end. Image and video generation are in a different league. A single AI image uses about half a smartphone charge per ACM Digital Library research. A 10-second Sora 2 video uses about 0.936 kWh of electricity, just over 4 liters of fresh water, and emits roughly 466 grams of carbon per Reclaimed Systems’ analysis. Energy use quadruples when video length doubles, so a 20-second clip is not twice as expensive, it is four times.
If Sora 2 hit even a fraction of TikTok-scale usage, the projections get ugly. One analysis estimated Sora 2 could emit roughly 1.9 million tonnes of CO₂ per year, around 23 percent of Meta’s entire 2024 corporate carbon footprint, just for short AI videos.
This is the question almost nobody answers directly. Based on first-party disclosures and peer-reviewed measurements published through May 2026, here is our greenest-to-dirtiest ranking for text prompts. Numbers are per query, lower is better. For the tool-level picture, our guide to the most eco-friendly AI ranks the greenest chatbots and dedicated green tools. Among the dedicated tools, our GreenPT review covers a privacy-first chatbot built around lighter, energy-efficient models.
The takeaway is uncomfortable for the cheap-equals-clean assumption: the lowest-priced API on the market is one of the dirtiest per query.
The most common comparison turns out to be the most misleading. The viral “AI uses 10 times more energy than a Google search” claim came from a 2023 estimate that compared early ChatGPT (around 3 Wh per query) against Google’s own 2009 figure of 0.3 Wh per search. Two years later, the numbers look very different.
A modern ChatGPT prompt at 0.34 Wh is roughly equal to a Google search at 0.3 Wh. A Gemini prompt at 0.24 Wh is actually lower than a Google search by Google’s own measure. The honest version of the comparison is closer to 1× or 1.1×, not 10×, for current text models. The 10× multiplier still applies for image and video generation, and it gets dramatically worse for reasoning models.
We covered this in detail in our breakdown of common myths about AI’s energy use, debunked.
Yes, and this is the part most doom-takes leave out. A 2024 Boston Consulting Group analysis estimated that AI applied thoughtfully could mitigate 5 to 10 percent of global greenhouse-gas emissions by 2030. The Grantham Institute put the savings at 3.2 to 5.4 billion tonnes of CO₂ by 2035 if AI is used to optimize energy grids, accelerate materials science, improve weather modeling, and cut industrial waste.
Real examples already exist. AI is mapping illegal deepwater dredging, charting methane plumes from satellite imagery, optimizing wind-turbine placement, and improving the efficiency of fertilizer use in agriculture. The European Commission published research in March 2026 showing that AI data-center waste heat is being redirected to power water purification and carbon capture systems.
The honest verdict is that it depends on how AI gets used. The infrastructure side is unambiguously costly. The applications side has real potential, but only if the climate-positive use cases actually scale faster than the climate-negative ones.
You do not have to give up AI to reduce your footprint. Most of the gains come from smarter use, not abstinence. Read our EcoGPT, the carbon-aware AI chatbot we reviewed for a deeper look at sustainability-first AI tools.
1. Pick the lighter model when you can. A factual lookup does not need a frontier reasoning model. Switching from a heavy reasoning model to a standard chat model can cut your per-prompt energy by 20 to 100 times. Apps that let you switch between models in one interface make this trivial, and search-first tools like Ecosia’s green AI search default to leaner models.
2. Skip the AI video and image generation for casual use. A single Sora 2 video burns roughly 3,000 times more energy than a text prompt. If the same idea works as text, write it.
3. Batch your prompts. Most energy goes to spinning up context. One long, well-thought-out prompt that gets you the answer in one shot is much more efficient than ten back-and-forth iterations.
4. Use multi-model apps. Fello AI is a good example: one app, one subscription, and you can pick between ChatGPT, Claude, Gemini, Grok, and DeepSeek for any task. Routing a lookup query to Gemini at 0.24 Wh instead of running it on a heavy reasoning model can cut your prompt footprint by an order of magnitude. The app costs $9.99 a month and supports Skills for Excel, Word, PowerPoint and PDF outputs, so you can also batch a whole task instead of many small prompts.
5. Push providers to disclose. Only 12 percent of executives measured their gen-AI environmental impact in 2025 per Capgemini’s survey. Opacity is the single biggest obstacle to fixing this. When you sign up for a tool, look for first-party emissions reporting. Reward the ones that publish.
AI in 2026 is environmentally harmful, but not equally so across the board. Per-prompt costs are small enough that personal abstinence is the wrong solution. What matters is the structural picture: a sector growing at 30 percent a year, locking in 40 percent fossil-fuel-powered new demand, and racing ahead of grid decarbonization. The IEA’s 945 TWh projection, Cornell’s CO₂ and water numbers, and Jegham’s 65× efficiency spread all point to the same conclusion. AI’s footprint will keep growing fast, and the only realistic mitigation is on the production side: cleaner grids, smarter siting, and more efficient models.
For you as a user, the highest-leverage move is picking the right model for the job and avoiding casual AI video. If you want to dig deeper into the numbers behind each pillar of AI’s environmental impact, our companion pieces on how much electricity AI really consumes in 2026, the full breakdown of AI’s water consumption, and our analysis of AI’s role in 2026’s electricity-demand crunch go a layer deeper than any single ranking can.
Stay ahead with expert AI insights trusted by top tech professionals!
Join thousands of AI fans & professionals benefiting from exclusive tips and insights from industry leaders.
