Is AI as bad for the environment as people say it is?

A lot of the AI-environment writing on LinkedIn and in mainstream press, while correct when it was written, has been overtaken by new data. The per-query energy and water numbers that anchored the 2024 panic narrative have come down by an order of magnitude as first-party disclosures from Google, OpenAI, and Mistral replaced 2023 best-guesses. I was curious what the current truth is and I feel it is important to get a more clear picture of the real environmental impact AI is having, and is projected to have in the future.

First of all, it is worth noting that previous AI environmental discussions rarely consider whether AI adoption is net-positive for an operator’s overall footprint (you, me, anyone using AI). To really understand the impact, we need to think about:

• The class of AI model being used
• What workflow AI is replacing
• Whether a user can even avoid AI inference (now that AI Overviews run on roughly half of Google searches)
• What kind of infrastructure backs the compute, and
• Whether efficiency gains get spent on more output or more work

In this article I will cover:

• What the per-query energy numbers actually are in 2026, and why they fell so fast
• Why reasoning models break the math, and what that means for which model you pick
• The missing counterfactual: the workflow AI replaces is not zero
• Why opting out by going back to Google search is no longer an option
• Four pre-commercial infrastructure bets where the math gets rewritten
• What responsible AI use in 2026 looks like

The AI-environment narrative in 2026

The loudest version of the AI-environment story runs roughly like this: a single ChatGPT query uses ten times the energy of a Google search; a viral statistic says it drinks 500 mL of water per prompt; data center buildout is straining grids in Virginia and Arizona; the individual user, by reaching for ChatGPT instead of search, is participating in a climate harm. Some of that read is correct. Data center growth is a real systemic concern, reasoning models do change the math in ways most coverage hasn’t caught up to, and the deployment timeline for compute infrastructure is running ahead of the clean-energy timeline meant to power it.

The part that’s harder to hold in mind is that the numbers underneath the narrative moved fast and downward in late 2025 and into 2026. Hannah Ritchie’s August 2025 update at Sustainability by Numbers is the cleanest single read on what changed; Andy Masley’s individual-use essay built the counterfactual frame that the panic narrative tends to skip. Both writers were working in good faith with the numbers they had when they wrote. The numbers are only getting better.

The per-query numbers came down by orders of magnitude

Four independent estimates published in 2025 and into 2026 converge in a tight band. Google’s August 2025 paper on Gemini inference puts a median text query at 0.24 Wh of electricity. Sam Altman casually mentioned in mid-2025 that a standard ChatGPT text query uses about 0.34 Wh. Epoch AI’s independent estimate for GPT-4o landed at roughly 0.3 Wh, which they explicitly framed as “ten times less than the older estimate” of around 3 Wh that was circulating in 2023. The May 2025 arxiv benchmark “How Hungry is AI?” measured a GPT-4o short query at 0.43 Wh. The most rigorous figure published since is also the most recent: an April 2026 peer-reviewed study in Joule from Microsoft Research measured a median query at 0.31 Wh (interquartile range 0.16 to 0.60) across optimized frontier-scale inference, and found that the widely circulated public estimates overstate real-world energy use by four to twenty times. It lands squarely in the same band, from a stronger source.

While the numbers are not identical, they sit in a band of roughly 0.24 to 0.43 Wh, and the most recent first-party disclosure is at the low end. Energy use is dropping and also converging, only a year and a half ago, they didn’t agree to within a factor of ten.

Studies lag. Rigorous per-query measurement runs, on average, six to twelve months behind a model’s release. The figures above are the most recent measured generation, the models the benchmarks could actually access, and the trend line points down from there. OpenAI, for example, declined to publish a first-party energy figure for GPT-5, so some of the newest flagship numbers are third-party estimates rather than direct vendor disclosures.

So why did things improve? More efficient hardware (the H100-class accelerators are an enormous step over the A100-class that anchored 2023 estimates), more efficient model architectures, and an honest correction to overly pessimistic token-count assumptions in the original estimates. Epoch AI names that last factor explicitly. And while Altman shared no methodology for his figure, the independent estimates from Google, Epoch, and the arxiv benchmark corroborate it.

Image generation lands in roughly the same range as text per query, by MIT Technology Review’s May 2025 reporting that Ritchie summarizes, which surprised most people who followed the early coverage. Video is the genuine outlier: a five-second clip runs two to three orders of magnitude above a text query, with measured figures landing anywhere from 30 to nearly 1,000 Wh depending on resolution and length, and that gap looks structural rather than something efficiency gains will close quickly. The cleanest early measurement came from OpenAI’s Sora, which OpenAI retired in 2026. The current video models (Google’s Veo 3, Kling, Runway) have not published comparable first-party numbers, but nothing about them changes the underlying physics: generating video is a far larger compute event than generating text. Mistral’s lifecycle assessment contributes a useful upper-bound figure for thinking about scale, with Ritchie’s caveat that Mistral’s methodology disclosure was light.

What about water?

Water flows along the same pattern. The viral 500 mL-per-prompt figure was actually a misreading of the original study, and the honest range works out to something like 10 to 30 mL per query depending on cooling architecture and data-center location. And some air-cooled facilities run a water usage effectiveness of zero. The water question hasn’t gone away (it concentrates regionally, which matters in drought-stressed siting decisions), but the per-query framing the panic narrative used was off by a couple of orders of magnitude.

Things are also trending in the right direction: first-party disclosures have replaced third-party best-guesses, efficiency is improving generation over generation, and the direction looks durable.

Reasoning models break the math

It’s easy to overlook that short-query numbers don’t apply to reasoning models. The April 2026 Joule study finds that long reasoning and agentic queries raise energy consumption by more than an order of magnitude, driven by the extra tokens generated and the reduced serving concurrency those workloads force.

Study data. The University of Rhode Island’s AI lab measured a medium GPT-5 response at roughly 18 Wh on average and up to 40 Wh under extended thinking, against a fraction of a watt-hour for a plain query. That is one model swinging by more than an order of magnitude depending on how hard it is asked to think, running about 8.6 times more power-hungry than GPT-4 on a medium response. Earlier cross-model work found the same pattern: the May 2025 “How Hungry is AI?” benchmark measured o3 and DeepSeek-R1 at over 33 Wh per long prompt, more than 70 times a lightweight model on equivalent work. Running a reasoning model on a one-line question is not the same energy event as running a small model on it. The gap is large, and it scales with how much reasoning the model is asked to do.

That same 2025 benchmark also found that, among the models it could test at the time, Claude 3.7 Sonnet ranked highest in eco-efficiency, a useful counterweight to the simple “bigger model, worse number” framing. The specific model is a generation old now (we are on the 4-series), but the point that survives is the general one: energy efficiency is not strictly a function of parameter count or reasoning depth, architecture and training choices matter at least as much.

So in other words: matching model class to task complexity is not just a cost optimization, it is a real energy decision. Don’t run a frontier reasoning model (GPT-5 Pro, Gemini 3 Deep Think, Claude Opus in extended thinking) on a one-line lookup. Don’t reach for the most capable model when a smaller one will produce the same output. The discipline we walked through in our AI cost optimization framework (dispatcher-first architecture, model-tier matching, agent decomposition) is the same discipline that reduces energy consumption at scale. Thankfully, cost discipline is energy discipline.

Is it better on the environment to do the same work without AI?

Most published coverage of AI’s energy use compares an AI query against itself: Wh per query, mL per query, grams of CO2 per response. That framing leaves out the comparison that actually matters: what workflow does the AI query replace, and what did that workflow cost (in terms of energy, effort, water etc.)?

Andy Masley and Hannah Ritchie built a counterfactual model. Masley’s water comparisons show that a ChatGPT query is materially less water-intensive than streaming music, an hour of social media browsing, or an hour on Zoom (the specifics live in the FAQ below). Ritchie’s data shows that AI’s share of global electricity sits well under 1%.

So what about not using AI, would that be better for the environment? Let’s call this the slow-path workflow: Doing it the old way (searching Google, opening ten pages, reading them, synthesizing the findings, writing it up) burns display energy, network round-trips, and server calls, plus a lot of human time. Those all add up. A single AI call compresses them into one event, and that compression is real. The old way also draws power and water, arguably more of both, plus more of your time, which is the resource people most often forget to count.

That doesn’t make AI free. It does mean the honest question is always “compared to what?”

So how much of our total water does AI actually use?

Per query the water cost is tiny, as we saw above. Zoom out to national water use and AI barely registers. You get clicks talking about AI draining our water, but here is the actual breakdown:

• 41.3% Thermoelectric Power
• 36.6% Agriculture
• 12.1% Public use
• 4.6% Industry
•  2.3% Aquaculture
• 1.0% Mining
• less than 1% Data centers

So for all the water we are measuring, we are barely using any water for data centers right now at all. Source

If we really wanted to reduce water use, we should be looking at agricultural practices of growing food crops that are excessively water intensive and rethinking how we generate power (which, by the way, is advancing).

What about power use?

As of June 2026 the best estimate I can find is that we are using 2 to 4% of our national power grid for AI data centers, and 4 to 6% overall once you count all data centers (not just AI ones).

One scope note while we are counting: these are operational numbers, the cost of running the models. The embodied carbon of building the hardware (chip fabrication, data-center construction, the minerals in a GPU) is a real and separate part of the footprint that this piece does not try to tackle.

I won’t pretend AI’s environmental number is small and therefore fine. But the displacement direction matters more than the absolute number. AI is actually net-positive for footprint when it substitutes for a higher-energy workflow, neutral when it’s additive at the margin, but net-negative when it scales total output proportionally to whatever efficiency gain it produced.

So if we did the same amount of work, and did it all with AI, we should be using less energy, less water and winning more free time. (But let me get back to this at the end of the article.)

You can’t opt out of AI…

The activist read in 2024 sometimes ended with a clean recommendation: just use Google search instead. In 2024 that recommendation was correct, or at least available. In 2026 it has been functionally retired by the search product itself. BrightEdge data from February 2026 finds that AI Overviews now trigger on roughly 48% of tracked search queries, up sharply year-over-year. Other trackers put the US-specific figure higher.

There is no permanent product-wide opt-out. Search Labs toggles are feature-specific. Browser extensions can hide the AI Overview surface but do not reduce the underlying inference. Whether you click into the AI Overview or not, the inference ran. That part of the activist advice has been overtaken by product reality, not refuted by argument.

The way I’d put it: “you can’t avoid it” is a statement of fact, not a defense. The alternative to using AI for an information task in 2026 is the slow-path workflow described above, and that workflow has a real resource cost too. The question shifted from “should I use AI” to “which AI workflow uses less energy for the work I’m doing.” That’s a different conversation, and a more useful one. It maps to the same shift we covered in our piece on making your business visible to AI search: the restructuring of discovery is restructuring the energy question alongside it.

Where we are going

Everything above is the demand side: what a query costs to run today. The environmentally-friendly supply side is moving too, and it is where the long-term picture gets decided. Below are a few early bets on where the energy to power AI comes from next:

Panthalassa. The Oregon-based startup raised a $140 million Series B in May 2026, led by Peter Thiel with participation from John Doerr, Marc Benioff’s Time Ventures, and Mike Schroepfer’s Gigascale Capital. The technology: floating autonomous nodes that convert wave energy directly to compute, cooled by surrounding seawater, transmitting via LEO satellite, with a 2027 commercial target.

Musk’s 100 GW claim. At WEF Davos on January 22, 2026, Musk said Tesla and SpaceX were each separately working to build 100 GW per year of US solar manufacturing capacity this decade, framing energy as “the bottleneck of the AI revolution.” For context, China’s annual PV production capacity is in the same order of magnitude.

Aikido Technologies. The middle path between speculative and tactical. Aikido’s AO60DC integrates an offshore wind turbine in the 15 to 18 MW range with roughly 10 to 12 MW of IT capacity on a single floating platform, designed for farms scaling toward 1 GW-plus of IT load. A small proof-of-concept is operating in Norway; first commercial project targets UK waters by 2028.

Space-based solar. The long-horizon bet. Feasibility analyses (per pv-magazine’s coverage) are real engineering rather than slideware, but the production timeline is probably decades out.

What responsible AI use in 2026 looks like

Match model class to task. Don’t run a frontier reasoning model (GPT-5 Pro, Gemini 3 Deep Think, Opus in extended thinking) on a one-line question. The reasoning-model energy gap, an order of magnitude on average and up to 70 times at the extreme, is the data point that makes this not just a cost recommendation but an energy one. Pick the smallest model that does the task at quality. Reserve reasoning models for tasks that actually reason.

Optimize the dispatcher layer. Cost discipline is energy discipline. Reductions in API spend tend to flow through to roughly proportional energy reductions on the same workload. Spend the engineering time on prompt compression, output budget tuning, and tier routing. The savings compound.

Use AI to replace higher-energy workflows, not layer on top of them. AI that substitutes for an hour of Zoom-and-document-review is net-positive on the math. AI that gets added to existing workflows without removing anything is additive.

Prefer hyperscalers with published WUE and PUE. Hyperscaler data-center efficiency tends to run meaningfully better than typical on-prem deployments (specifics in the FAQ below). For workloads without regulatory or data-residency reasons to stay local, the embodied-carbon argument also runs in favor of cloud.

Are we just working harder?

If the same work gets done in 20% of the time but the working hours don’t change, total resource use goes up roughly 5x, not down. Efficiency that doesn’t change how we choose to spend the gains multiplies total resource use rather than reducing it. That’s not a uniquely AI problem (Jevons named the pattern in coal in 1865), but it is the question every operator has to answer now. AI adoption is net-positive for footprint when it’s substitutional, neutral when it’s additive at the margin, and net-negative when it scales total output proportionally to the efficiency gain. The decision is not the technology. The decision is: what we do with the time the technology gives back?

Where the energy picture goes from here looks more open than the panic narrative allows. Fusion is always 20 years away. Geothermal at scale, solar manufacturing at the scale Musk is gesturing at (even discounted appropriately), offshore wave and wind compute integration, and space-based solar all have shipping timelines in 2027 to 2030. The future shape looks energy-rich rather than energy-poor, if compute discipline can keep pace with capacity.

Frequently asked questions

How much energy does one ChatGPT query actually use in 2026?

The per-query band is covered in detail in the body above. The short version: a standard text query lands roughly an order of magnitude lower than the figure widely cited in 2023, and the revision comes from more efficient hardware and models plus a correction to overly pessimistic token-count assumptions in the original estimates. Reasoning models (GPT-5 in extended thinking, DeepSeek-R1, and similar) are a different category and use materially more energy per long prompt, an order of magnitude or more in several 2026 measurements.

Is AI worse for the environment than a Google search?

Per query, the gap is much smaller than the “10x” claim that circulated in 2024. A Google search and a Gemini text query land in the same per-query range on Google’s own August 2025 disclosure. The question is mostly moot now, because AI Overviews already trigger on roughly half of tracked Google searches, which means most search queries include AI inference whether the user clicked into it or not.

Can I opt out of AI in Google search?

Not in any complete way. There is no permanent product-wide opt-out. Search Labs toggles are feature-specific. Browser extensions can hide the AI Overview surface, but the inference still ran on Google’s infrastructure when the search happened. The framing has shifted from “should I use AI” to “which AI workflow uses less energy for the work I’m doing.”

Does generating an AI image use the same energy as charging my phone?

No — image generation runs roughly two orders of magnitude lower than charging a phone, per MIT Technology Review’s May 2025 reporting. Video generation is the genuine outlier: a short AI-generated clip lands closer in energy cost to charging a phone or running a microwave for a minute. The image-versus-video gap is the most important distinction in the consumer AI energy conversation.

How much water does a ChatGPT query use?

The honest range works out to something like 10 to 30 mL per query depending on cooling architecture and data-center location, with some air-cooled facilities running effectively zero on-site water. For comparison, Andy Masley’s counterfactual finds streaming a song uses roughly 250 mL, an hour of social media browsing 430 mL, and an hour on Zoom 1,720 mL.

Should I stop using ChatGPT to help the environment?

For meaningful climate impact, individual AI use is not the leverage point. Data centers account for roughly 1.5% of global electricity by IEA figures, and AI specifically is under 0.2%. Systemic decisions — grid mix, hyperscaler siting, model architecture choices, workflow substitution — are where the math moves. Where individual choice does matter: don’t reach for reasoning models on tasks that don’t need reasoning, and use AI to replace higher-energy workflows rather than as a layer added on top.

What percentage of US electricity goes to data centers?

Approximately 4 to 6% of US electricity, by Hannah Ritchie’s summary of IEA data. AI specifically is a fraction of that. Data center electricity is concentrated regionally — Virginia, Texas, the Pacific Northwest, parts of the Southeast — which is where the grid-pressure conversations are most active. The national figure understates the local-cluster strain in those regions.

Is cloud AI more efficient than running AI locally on my laptop?

Generally yes, by a meaningful margin. Hyperscaler data centers run a power usage effectiveness (PUE) between roughly 1.08 and 1.25 (Google’s published range), versus typical on-prem deployments that run 1.4 to 2x worse. The embodied carbon of laptop and consumer-GPU manufacturing also tends to push the comparison further in cloud’s favor, particularly for workloads that don’t run continuously. Where local makes sense is regulated data-residency workloads, intermittent inference on hardware that already exists for other reasons, and edge applications where round-trip latency matters more than efficiency.