Which AI Is Most Sustainable?
by Sustainability Tracker 04/11/2026
Guide, Thought Starters
by Sustainability Tracker 04/11/2026
Guide, Thought Starters
This article was contributed by Sustainability Tracker.
Last updated: 02 November 2026
Contributors: Sustainability Tracker; Anne Kehlhofer – Director, ESG, Xero; James Tilbury – Partner, Linden Sustainability; Sydney Straver – Director, &BLOOM Sustainability & ESG
On current public evidence, the five major AI platforms can’t be ranked on environmental footprint, because they disclose very different numbers. Google publishes the most detail for Gemini, including per-prompt energy, carbon and water. Microsoft reports full corporate emissions and has published a per-query energy range for large models. OpenAI’s CEO has shared an unverified per-query estimate. Anthropic and xAI publish the least. How you use AI is likely to change your footprint more than which platform you pick. Reasoning or “thinking” modes, and image and video generation, use far more energy than standard text answers. AI also has social impacts that sit outside the emissions numbers, from air quality and local power prices near data centres to the working conditions of the people who label training data.
Start with a conversation inside your organisation: who is using which AI tools, how often, and whether AI belongs in your sustainability and procurement policies. Include the AI built into software you already pay for. That’s an easier first step than estimating emissions, and it gives you facts to act on. If you run a smaller business outside mandatory reporting, the prompt is more likely to come from a larger customer asking for your emissions data for its own reporting, and the same conversation gives you a starting point for that answer. The comparison below shows what each provider discloses, so you know what to ask.
At Sustainability Tracker, we make sustainability information visible so people can make more sustainable choices. With AI, the gaps in that information are the main finding right now.
Providers measure different things. Google’s own August 2025 technical paper prices the identical median Gemini prompt at either 0.10 Wh or 0.24 Wh, a 2.4x gap, depending only on what gets counted. One text prompt uses about as much energy as an LED light left on for a minute or two. The problem is scale: billions of prompts a day, and the data centres built to handle them. The impact comes from scale, with billions of prompts a day and the data centres built to serve them. Independent benchmarks add further spread: the model and the prompt length change the answer by orders of magnitude, so a single per-prompt figure hides a spread of more than 65x once reasoning models are included.
A fair comparison starts with a simpler question: what has each company told the public, and can anyone check it?
We reviewed each provider’s public disclosures through our FACTS lens (Facts, Appearance, Context, Transparency, Substantiation), and each column of the comparison table tests part of it. The per-prompt, corporate emissions and water columns cover Facts: whether the provider has published these figures at all. The evidence type shown for each per-prompt figure covers Appearance and Transparency: whether the number comes from a published method, a research study or an executive’s statement. Independent verification covers Substantiation, and energy sourcing covers Context. Each provider’s data centre record, such as permits and lawsuits, is in its own section below. We used primary sources and reputable reporting, checked as at 2 November 2026.
How Sustainability Tracker uses AI: Tracker AI uses a mix of Relevance AI and OpenAI models. Relevance AI manages the workflow, and OpenAI generates the answers. To save resources, it picks a model based on how hard the question is. Simple questions go to a smaller, faster model. Harder questions that need more thinking go to a larger, more powerful one. This way, the big models are only used when they’re needed.
This article was researched and drafted with AI assistance and reviewed by Rhiana A, Marketing Manager and Remi M, Head of Product.
We didn’t ask the providers to review this article before publishing. If you represent one of them and spot an error, contact us and we’ll correct it and record the change in the update log below.
Platform (company) | Per-prompt energy figure (evidence type) | Corporate emissions report | Water data | Independent verification | Energy sourcing reported |
| Gemini (Google) | Yes: first-party, published method | Yes, annual | Yes | Yes, for selected corporate metrics | Yes: 100% of electricity matched with renewable purchases |
ChatGPT (OpenAI) | Executive statement only | No standalone report | Executive statement only | No | No |
| Copilot (Microsoft) | Research estimate: peer-reviewed, not Copilot-specific | Yes, annual | Yes, corporate and per-query range | Yes, for selected corporate metrics; per-query study peer-reviewed | Yes: 100% of electricity matched with renewable energy |
Claude (Anthropic) | None | No | No | No | No |
Grok (xAI) | None | No | No | No | No |
Google has published the most detailed first-party measurement of any provider. Using its comprehensive methodology, Google estimates the median Gemini Apps text prompt uses:
The full methodology is in Google’s technical paper and blog post.
That figure has limits. Like every per-prompt figure in this article, it leaves out the energy used to train the model and to build the data centres it runs in. Chip manufacturing is only partly covered. Google’s carbon figure includes a share of the emissions from making its hardware, but its energy and water figures don’t, and no other provider says its figure includes hardware at all. These are the costs growing fastest: Google names data centre construction and its supply chain as the main drivers of its emissions growth, and Microsoft names its data centre expansion. The report covers text prompts only, leaving out the higher energy costs of image and video generation, and Google hasn’t revealed its total daily Gemini query volume.
At company level, Google’s 2026 Environmental Report shows the pressure AI is creating. Electricity demand rose 37% in a year, while operational emissions fell 2% and Google matched 100% of its electricity with renewable purchases. Supply-chain (Scope 3) emissions grew 25%, with data centre construction and suppliers’ electricity use cited as the main drivers, including suppliers in the Asia-Pacific region on grids that still rely heavily on fossil fuels. Google’s report states its AI infrastructure buildout is accelerating faster than the grid is decarbonising.
Google’s report raises a bigger question: can clean electricity be added as fast as AI demand grows? The answer depends heavily on where data centres are built.
The Paris-based International Energy Agency, of which Australia is a member, expects renewables to meet nearly half of the extra electricity demand from data centres worldwide to 2030, with gas and coal covering most of the rest. The mix varies sharply by country. China is adding clean power faster than anywhere else, with a record 315 GW of solar installed in 2025, and an analysis for Carbon Brief found its CO2 emissions were flat or slightly down that year as renewables met rising demand. Coal still plays a role there: the IEA’s 2025 projections had coal and renewables each supplying about the same amount of China’s new data centre electricity. In the US, natural gas is expected to be the largest source. Connecting a new data centre to the US grid can take five years or more, so developers are increasingly building their own gas generation on site. About 101 GW has been announced, though some projects may not go ahead. xAI’s turbines near Memphis, covered below, are an early example.
This is where annual renewable matching, the measure Google and Microsoft both report, has limits. It balances a year’s total electricity use against a year’s renewable purchases. A data centre can still draw gas-fired power at night, or sit on a fossil-heavy grid, while the company reports 100%. Google and Microsoft have both set 2030 goals to match their electricity with carbon-free energy every hour, which is a much tougher test. The other three providers in this comparison don’t report their energy sourcing at all.
For anyone buying AI, this comes down to a practical question: where does my AI run, and what powers it there? Ask your provider whether it reports carbon-free energy by the hour, or emissions based on the grids its data centres actually use.
OpenAI’s main disclosure is a figure from its CEO. In a personal blog post, Sam Altman said the average ChatGPT query uses about 0.34 Wh of electricity, which he compared to a high-efficiency lightbulb running for a couple of minutes, and about 0.32 millilitres of water. OpenAI hasn’t published a methodology for either figure or defined an “average” query, so neither can be independently verified.
At company level, OpenAI had no standalone sustainability report as at November 2026. A spokesperson has said the company has a sustainability steering committee, pointing to its Stargate Community commitment to build infrastructure that benefits local communities, including by minimising water usage. That plan commits OpenAI to funding the extra energy generation and grid upgrades each data centre site requires, so local electricity costs don’t rise.
Microsoft publishes comprehensive corporate data in its 2026 Environmental Sustainability Report. Total emissions (Scopes 1, 2 and 3) rose 25% on the prior year, driven by the AI infrastructure buildout and a shift away from non-additional renewable energy certificates toward developing new carbon-free energy. Microsoft says it matched its entire electricity consumption with renewable energy, and for the first time replenished more water than it consumed. It is still working toward carbon negative, water positive and zero waste goals by 2030.
Microsoft hasn’t published a figure for Copilot specifically. In June 2026, Microsoft researchers published a peer-reviewed study in the journal Joule estimating that a typical query to large models in its data centres uses 0.16 to 0.60 Wh of electricity and up to 0.067 mL of water. The study covers large models in general and doesn’t give a carbon figure.
Copilot also runs on other companies’ models. Microsoft says Copilot continues to be powered by OpenAI’s latest models, with Anthropic models also available. Claude models are enabled by default in commercial tenancies and off by default in the EU, EFTA and UK. So Microsoft’s report covers its own operations, and part of your Copilot footprint depends on OpenAI and Anthropic infrastructure.
This is common. Many people never choose a model at all, because AI comes built into tools they already use, such as Microsoft 365, Google Workspace or a CRM. In that case, the software vendor decides which models run behind its features, so the questions in this comparison are best put to that vendor.
Alongside xAI, Anthropic discloses the least of the providers in this comparison. As at November 2026, it had not reported Scope 1, 2 or 3 emissions in any public filing, or per-query energy figures. It has not made a net-zero pledge or published a sustainability report, though it did join the Frontier carbon-removal coalition. A company spokesperson has said Anthropic is working with carbon accounting platform Watershed to measure its company-wide footprint and comply with reporting law.
Anthropic has made public commitments on energy. In February 2026 it pledged to cover electricity price increases that consumers face from its data centres. It says it is investing in systems that cut power use during peak demand, and in water-efficient cooling. These commitments apply directly where its own data centres run its workloads; for leased capacity, the company said it is exploring further ways to address price impacts.
In May 2026, Anthropic announced it would use all the compute capacity at SpaceX’s Colossus 1 data centre, giving it more than 300 megawatts of new capacity. When building Colossus, xAI installed and operated dozens of gas turbines, claiming no federal permit was needed because they were for temporary use. According to the Southern Environmental Law Center, xAI removed its unpermitted turbines at Colossus 1 after a notice of intent to sue and obtained permits for its remaining 15.
xAI publishes no emissions data. It has no sustainability report, no Scope 1, 2 or 3 figures and no climate target, and SpaceX’s S-1 acknowledges its gas turbine technology only qualitatively. xAI also publishes no per-query energy figure for its current flagship.
The main environmental questions relate to its Memphis-area data centres. Training of newer Grok models shifted primarily to Colossus 2, which draws power from turbines across the state line in Southaven, Mississippi. The NAACP’s filing alleges that the 33 unpermitted gas turbines have the potential to emit 2,508 tons of nitrogen oxides each year. xAI and state regulators argue in court filings that the turbines are exempt because they are mobile and meant to operate on site for less than a year. The case is ongoing.
Federal rules have since tightened. In January 2026 the EPA clarified that trailer-mounted turbines can’t be classed as non-road engines, and companies must obtain Clean Air Act permits before installing them.
Several issues in this comparison are social as well as environmental:
Other issues sit further along the AI supply chain:
We haven’t scored providers on these issues, because there’s no common disclosure to compare them on yet. They’re worth raising with your vendor alongside emissions.
Your usage habits are likely to change your footprint more than your choice of platform. Many AI tools let you switch on a “reasoning” or “thinking” mode, often labelled “Thinking” in ChatGPT, Claude and Gemini. In this mode, the AI works through a problem in several steps before it answers, which suits complex tasks like analysis or planning. In standard mode, it answers straight away, which is enough for most everyday questions. Hugging Face’s AI Energy Score found reasoning models used on average 30 times more energy than models without reasoning. Image and video generation also sit well above text.
A few practical habits help:
On-device AI is another option. Apple Intelligence and small open models you can run on a laptop handle simpler tasks on your own device, so those requests don’t use data centre capacity. Apple sends heavier requests to its Private Cloud Compute servers. There’s little published data comparing the footprint of on-device and cloud AI, and the trade-off is contested. In October 2026, OpenAI CEO Sam Altman said “it’s quite terrible to have everybody have a bunch of the world’s limited supply of DRAM sitting in their own computer”, arguing that RAM in personal computers is used about 5% of the time and that data centres use hardware more fully. OpenAI’s business runs in data centres, so weigh the comment with that in mind.
“It’s quite terrible to have everybody have a bunch of the world’s limited supply of DRAM sitting in their own computer”
Sam Altman on the Cheeky Pint podcast
Several reporting regimes will add information over the next few years. They differ in who they cover and whether the public gets to see the results.
California (SB 253). US-organised companies with more than US$1 billion in annual revenue that do business in California must report Scope 1 and 2 emissions by 10 November 2026, with Scope 3 from 2027. That includes US subsidiaries of overseas groups. All five providers in this article are US companies, so the law is likely to capture them. It’s the nearest-term rule that could produce emissions figures from Anthropic, OpenAI and xAI, which currently publish none. In the first year, companies that weren’t collecting emissions data by 5 December 2024 can file a statement explaining why in place of figures, so the November filings may not include numbers for every provider. First-year reports won’t need independent assurance, so treat early figures as self-reported.
European Union (AI Act). Developers of general-purpose AI models must document their models’ energy consumption. This information goes to regulators on request, so it won’t help you compare providers directly. Models released before August 2025 have until August 2027 to comply.
European Union (CSRD). Revised in 2026, the directive now captures non-EU groups with more than €450 million in EU turnover and a large EU subsidiary or branch. First reports for those groups cover FY2028.
Australia (mandatory climate reporting). Australia’s mandatory climate reporting regime under the Corporations Act uses AASB S2 as its climate standard. It can apply to Australian subsidiaries of global companies that meet the size thresholds. These reports cover Australian operations, so they won’t show the global footprint of an AI model.
We’ll update this comparison as new filings are published, starting with the California reports due on 10 November, and record each change in the update log at the end of this article.
Google has published the most specific first-party figure: 0.24 Wh for a median Gemini Apps text prompt. Microsoft’s researchers estimate 0.16 to 0.60 Wh for a typical query to large models generally. OpenAI’s CEO has cited about 0.34 Wh for an average ChatGPT query, though OpenAI hasn’t published a methodology, so the figure can’t be verified. Anthropic and xAI have not published per-prompt figures. The figures use different methods, so a like-for-like comparison isn’t possible. Model choice and prompt type change energy use far more than the brand does.
A single text query likely uses a small amount of energy. OpenAI’s CEO puts it at about 0.34 Wh, an unverified figure with no published methodology. The larger concerns are scale and transparency. OpenAI has not published a standalone sustainability report or audited emissions figures, and its Stargate data centre buildout adds significant new energy demand.
Not as at November 2026. Anthropic has not published Scope 1, 2 or 3 emissions or per-prompt energy figures. It has said it is working with carbon accounting platform Watershed to measure its footprint and meet reporting requirements, and it has made public commitments on electricity costs and water-efficient cooling.
xAI publishes no emissions data or per-prompt figures. The Colossus 2 data centre used for newer Grok models is powered by gas turbines that are the subject of a Clean Air Act lawsuit brought by the NAACP. xAI and Mississippi regulators argue the turbines are exempt from federal permits. The case is ongoing.
Often, yes. Copilot runs on OpenAI models, and for most commercial customers outside the EU, EFTA and UK, Anthropic’s Claude models are also switched on by default. Microsoft publishes its own corporate emissions, which rose 25% year on year in its 2026 report, and its researchers have published a per-query energy range for large models, though not a figure for Copilot specifically.
Choose based on what matters most to you. If you need data for emissions reporting, providers with published reports give you the most to work with. If local pollution matters, check the provider’s data centre record. Whichever platform you use, choosing smaller models and limiting reasoning, image and video generation to tasks that need them reduces your footprint. If you want to reward transparency, favour providers that publish verifiable data. If AI comes built into software you already use, ask that vendor which models it runs.
In an October 2026 interview on the Cheeky Pint podcast, OpenAI CEO Sam Altman said it’s “quite terrible” for so much of the world’s limited supply of DRAM (computer memory) to sit in personal computers, where he said it’s used about 5% of the time. He argued data centres use hardware more fully. It’s a fair point about making the most of scarce hardware, though there’s little public data comparing the full footprint of on-device and cloud AI, and OpenAI’s business depends on cloud use.
| Date | Update |
| 02 November 2026 | First published |