Is AI Bad for the Environment? Energy, Carbon and E-Waste Explained

Is AI bad for the environment?

AI is bad for the environment mainly through scale, not through any single conversation. One text prompt uses a very small amount of electricity, but billions of prompts per day, plus the construction of new data centers and the chips inside them, add up to a measurable share of electricity demand and corporate emissions. The environmental question is therefore less about whether one person should use a chatbot and more about how fast the industry is growing and what powers that growth.

The evidence points in 2 directions at once. Per-prompt efficiency has improved sharply, and Google reports that the energy per median Gemini text prompt fell 33-fold between May 2024 and May 2025. At the same time, total electricity use keeps rising because demand is growing faster than efficiency gains. The most sensible reading is that AI’s footprint is real and growing, but it is concentrated in infrastructure decisions made by a small number of companies and utilities rather than in individual use.

This article covers electricity, carbon, e-waste and what companies disclose. Water is summarized briefly because TechJournal covers it in detail in our guide to how much water AI uses.

How much energy does one AI prompt use?

One AI text prompt uses roughly a quarter to a third of a watt-hour, based on the 2 figures published by major providers. Google’s August 2025 technical paper reports that the median Gemini Apps text prompt consumes 0.24 Wh, emits 0.03 grams of carbon dioxide equivalent and uses 0.26 milliliters of water. OpenAI CEO Sam Altman wrote in a June 10, 2025 blog post that the average ChatGPT query uses about 0.34 Wh, which he compared to an oven running for a little over one second.

These figures matter because they replaced older estimates that were several times higher. Google’s methodology counts the AI chips, the host servers, idle capacity kept ready for traffic spikes and data center overhead such as cooling. Google notes that counting only the active AI chips would produce a lower figure of 0.10 Wh, which shows how much the measurement boundary changes the answer.

There are 3 important limits to these numbers. First, both are self-reported by the companies, and Altman’s figure came without a published methodology. Second, both describe typical text prompts, not image generation, video generation or long reasoning tasks, which require more computation. Third, a per-prompt figure excludes the energy used to train the model in the first place.

What do the per-prompt figures look like in context?

The illustration in the last row shows why individual text use is a minor part of a household’s electricity. A heavy user sending 100 median text prompts a day would use about 24 Wh per day by Google’s figure. The practical conclusion is that personal chatbot use is not where most of AI’s environmental impact comes from, although heavier workloads such as video generation change that math.

How much electricity do AI data centers use in total?

Data centers used about 485 terawatt-hours of electricity worldwide in 2025, according to the International Energy Agency’s April 2026 report Key Questions on Energy and AI. The IEA says that demand grew 17% in 2025 and that electricity use in AI-focused data centers grew 50%. The agency projects total data center demand will roughly double to 950 TWh by 2030, or around 3% of global electricity demand.

The United States carries the largest share of this growth. The IEA’s original Energy and AI report estimated that the US accounted for 45% of global data center electricity use in 2024 and that data centers represent nearly half of US electricity demand growth through 2030. A Department of Energy report prepared by Lawrence Berkeley National Laboratory, released December 20, 2024, found that US data centers used 176 TWh, or 4.4% of US electricity, in 2023, and projected 325 to 580 TWh, or 6.7% to 12%, by 2028.

This matters because power plants and transmission lines take years to build. When new data center demand arrives faster than clean generation, utilities often keep fossil fuel plants running longer or add natural gas capacity. The IEA estimates that 15 to 27 gigawatts of onsite natural gas generation could power data centers by 2030, mostly in the United States. The cost side of that growth is covered in our explainer on AI data centers and electric bills.

What is AI’s carbon footprint?

AI’s carbon footprint depends on the electricity grid that powers each data center and on the emissions from building the hardware. The IEA estimated in its 2025 report that data center electricity use produced about 180 million metric tons of CO2 in 2024, which it described as below 1.5% of total energy sector emissions. That share is small compared with sectors such as transportation or heavy industry, but it is one of the faster-growing sources.

Company reports show the trend more directly. Google’s 2026 Environmental Report, released June 30, 2026, says the company’s electricity consumption rose 37% in 2025. Google reported a 2% drop in operational emissions, but ESG Dive’s review of the report notes that Google’s total emissions rose 18% from 2024 and stood 81% above its 2019 baseline, driven largely by supply chain emissions from data center construction and chip manufacturing. Microsoft said in a July 9, 2026 post that its total Scope 1, 2 and 3 emissions increased 25% year over year, “driven primarily by the expansion of our datacenter infrastructure.”

Both companies also say they matched 100% of their annual electricity use with renewable energy purchases. That claim is accurate in accounting terms, but matching purchases over a year is not the same as running on clean power every hour, and it does not cover emissions from concrete, steel and chip fabrication. The most important point for readers is that renewable matching reduces reported electricity emissions without eliminating AI’s growing climate impact.

Does AI create e-waste?

AI creates e-waste because AI servers and chips are replaced on short cycles as newer, more efficient hardware arrives. A 2024 study in Nature Computational Science estimated that generative AI could add 1.2 million to 5 million metric tons of e-waste by 2030, depending on how quickly the industry grows. The same researchers estimated that combining strategies such as extending server life, refurbishing and reusing components, and improving chip design could cut that waste by 86%.

E-waste matters because discarded electronics contain lead, mercury and other hazardous materials, and much of the world’s electronic waste is not formally recycled. A June 2026 UN University report summarized by UN News projected up to 2.5 million metric tons of AI-related e-waste annually by 2030. These projections depend on assumptions about hardware lifetimes and adoption, so they describe a plausible range rather than a fixed outcome.

How much water does AI use?

AI uses water mainly to cool data centers and, indirectly, at the power plants that supply them. Google’s per-prompt figure of 0.26 milliliters is small, but data center water use becomes significant in dry regions where facilities compete with local users. The UN University report warned that AI-related water consumption could equal the basic annual domestic needs of 1.3 billion people by the end of the decade. It also noted a trade-off: some renewable energy and cooling choices that cut emissions can raise water or land use. The detailed numbers, including why estimates disagree, are in our full AI water use guide.

What do AI companies disclose about their environmental impact?

AI companies disclose company-wide totals far more often than per-product figures. Google and Microsoft publish annual environmental reports with electricity use, emissions by scope and water data, and Google published a per-prompt methodology for Gemini in August 2025. OpenAI has shared a per-query estimate through its CEO’s blog but has not published a comparable technical methodology, and many other AI providers publish no per-query figures at all.

This gap matters because readers and policymakers cannot compare models on energy use without consistent measurements. There are 4 details worth checking in any AI environmental claim:

• The measurement boundary, such as chips only or the full data center.
• The task type, such as a text prompt, an image or a video.
• The emissions method, such as location-based or market-based accounting with renewable certificates.
• Whether training, hardware manufacturing and construction are included.

A claim that leaves out those details describes only part of the footprint, whether the number looks reassuring or alarming.

What can you do to reduce your AI footprint?

To reduce your AI footprint, focus on the heaviest tasks rather than cutting ordinary text questions. Individual text prompts have a small footprint, so the practical gains come from avoiding unnecessary heavy workloads and from keeping hardware in use longer. The following 5 steps are the most useful for most people:

1. Use a standard search or a direct website for simple facts, such as opening hours, that do not need a generated answer.
2. Choose a smaller or faster model setting for routine tasks, and reserve extended reasoning modes for problems that need them.
3. Generate images and video deliberately, because those tasks require more computation than a text reply.
4. Keep phones and laptops longer, and repair them when possible, because manufacturing is a large part of any device’s footprint.
5. Recycle old electronics through a certified program, following the EPA’s electronics donation and recycling guidance, and wipe your personal data first.

These steps have a limited effect on their own because the largest decisions, such as where data centers are built and which power sources they use, are made by companies, utilities and regulators. Readers who want to influence those decisions can follow state and local decisions on data center approvals and grid commitments, such as Pennsylvania’s new AI data center order.

Frequently Asked Questions

Is using ChatGPT bad for the environment?

Using ChatGPT for ordinary text questions has a small environmental footprint per query. OpenAI’s CEO estimated the average query at about 0.34 Wh, though that figure is self-reported and the larger impact comes from total demand and data center construction.

How much energy does an AI prompt use compared with a Google search?

An AI text prompt uses roughly 0.24 to 0.34 Wh based on figures published by Google and OpenAI. Neither company has published a current per-search figure for standard web search in the same format, so direct comparisons rely on older estimates.

Is AI training or AI use worse for the environment?

AI use, also called inference, now accounts for most of AI’s energy demand because models answer billions of requests daily. A June 2026 UN University report estimated that day-to-day use makes up roughly 80 to 90 percent of total AI energy demand.

Do tech companies use renewable energy for AI?

Google and Microsoft both report matching 100% of their annual electricity use with renewable energy purchases. That annual matching does not mean every data center runs on clean power every hour, and both companies still reported rising total emissions in their 2026 reports.

Will AI’s environmental impact get better or worse?

AI’s total environmental impact is projected to grow through 2030 even as per-prompt efficiency improves. The IEA projects data center electricity demand will roughly double from 485 TWh in 2025 to about 950 TWh in 2030.