The headlong rush into artificial intelligence comes with a general awareness that AI chatbots consume vast amounts of power, much of it generated by burning fossil fuels. What’s harder to know is precisely how much planet-warming carbon dioxide is released each time ChatGPT, Claude or DeepSeek leap into action.

The biggest US AI companies haven’t published detailed accounts of their energy or emissions performance. That would require sharing technical details they would rather keep under wraps. AI’s emissions depend not just on which service you are using, but where your prompt gets processed. The data center handling the task may be powered by coal, or by wind turbines. And AI server farms use a variety of semiconductor and cooling technologies, adding more variables to the equation.

Yet researchers have been finding ways to make something better than an educated guess so that companies, governments and users can better understand AI’s environmental impact.

“People are super intimidated by this,” said John Bistline, head of science at Watershed, which helps AI companies to measure and report their environmental impact. But they also recognize “that companies and regulators and auditors are all asking for these AI emissions estimates that actually hold up.”

The query is sent to a data center to be processed on semiconductors, typically graphics processing units adapted for AI workloads. The chips burn through electricity as the AI breaks the written prompt into smaller chunks, or “input tokens,” that its neural network can process. It then performs billions of calculations multiple times to deliver the ideal response. This intense data crunching creates a lot of heat that needs to be removed so the chips don’t malfunction.

As much as 40% of the energy used by a data center is used for cooling, according to a report by Sasha Luccioni, co-founder and chief scientific officer of the research and advisory firm Sustainable AI Group, and AI startup Hugging Face.

The proliferation of AI data centers is leading to a surge in energy demand that threatens to slow or even arrest the decline in carbon emissions brought about by the growth in renewables and the electrification of industry, heating and transportation.

In a 2024 whitepaper, the Electric Power Research Institute estimated that for earlier models, a single ChatGPT query consumed 10 times more power on average than a traditional Google web search.

Text-based tasks will vary in their power consumption. A question that requires a short and simple response such as “what is the average temperature in Los Angeles” is likely to be less energy intensive than a query about quantum physics. The type of AI model that’s processing the request matters as well.

AI’s appetite for energy is exploding as more people tap into the technology and developers roll out larger, more capable AI platforms that can handle increasingly complex workloads. Asking ChatGPT to make an image or a video will require far more energy than posing a question. The adoption of autonomous “agents” that can perform complex, multi-step tasks is adding to AI’s power consumption. An analysis from benchmarking company Vals AI found that using AI to build a software application can have an environmental impact 10,000 times greater than that of a simple query.

Calculating an AI model’s emissions also means factoring in everything that needed to happen for the chatbot to answer it. So researchers have begun to study how much energy is consumed during an AI service’s entire life cycle — from extracting raw materials to make chips and data center cooling plants to the operation of the AI service itself and the eventual scrapping of redundant chips. The cement alone used to construct US AI facilities through 2030 could generate as much greenhouse gases as 415,000 cars emit in a year, according to environmental nonprofit RMI.

Nidhal Jegham, a researcher at Sustainable AI, built an open-source dashboard that compares the estimated energy consumption and carbon emissions of various AI models. Its baseline is a US data center using a dedicated, gas-fired power plant. Many data center developers in the country are turning to natural gas to power the facilities.

Jegham and his team found the greenhouse-gas emissions of individual models depend on several variables, including batch size (how many prompts are being handled together), the number of chips used to process a query, the specific type of chip and energy use per token (the basic unit of text or other data that an AI model reads and generates).

US frontier models such as ChatGPT 5.6 Terra appear to consume more energy than Chinese counterparts such as Alibaba Group’s Qwen. This is because the latter generally use a “reasoning” architecture that breaks complex math, logic, and coding tasks into smaller, sequential steps instead of activating the entire neural network for every single token.

As a general rule, the most complex, frontier models tend to burn through the most energy. That’s because generating a more sophisticated response to a query involves running more calculations. So Anthropic’s “lightest” AI model, Claude Haiku, consumes far less energy to generate its responses compared with the company’s more cutting-edge platform, Claude Fable 5.

In practice, smaller models often burn through more tokens to generate a response than their larger counterparts to compensate for a lower level of intelligence. This will narrow the gap in energy consumption.

“If you look at Claude Sonnet, it’s way cheaper than Opus and it’s way smaller. However, it generates almost 60% more tokens to compensate for being smaller or for being less intelligent,” Jegham said. “The model emits less carbon per token, but needs way more tokens to get to the same answer.”Even factoring in verbosity, however, most lighter models still appear to require less power for a comparable task than frontier models. Prompting a chatbot to perform an “agentic” task that involves multiple steps consumes many times more energy than inputting a simple text query.

Here again, the choice of platform will make a big difference to the energy consumption and — depending where the work is handled — the carbon emissions. The estimated 76.2 watt-hours sucked up by Claude Fable for an agentic task would produce about the same emissions as four smartphone charges.

That may not seem like a lot in isolation. But AI platforms are being prompted billions of times per day, and the energy usage and carbon emissions quickly add up. US AI companies say they are working hard to limit their emissions.

In its 2026 environmental report, Google estimated that the median Gemini Apps text prompt used 0.24 watt-hour of energy — the same as watching TV for less than nine seconds. It said that between 2024 and 2025 it was able to reduce its text prompt energy use and carbon footprint by 33 times and 44 times, respectively, due to “research innovations and software and hardware efficiency improvements.”Generally, however, estimates published by the industry tend to be vague and hard to verify.

Google develops its own chips called TPUs, or tensor processing units, which the company says are more energy efficient and faster, making the final carbon footprint smaller than when using GPUs. Google doesn't let outside researchers purchase and study its TPU chips and has published limited data about them, according to Jegham. Anthropic, OpenAI, Google, DeepSeek and Alibaba didn’t reply to Bloomberg emails seeking comment on Sustainable AI’s findings.

Prompting ChatGPT in, say, Copenhagen could emit a lot less carbon than prompting it in Charleston.

That’s because in Europe, fossil fuels account for a smaller share of power generation than in the US. AI data centers in Scandinavia consume a lot of renewable energy sources like wind, solar and hydropower. The average data center in the US runs on a mix of fossil fuels and renewables that varies from region to region. About 24% of US utility-scale electricity generation was produced from renewable sources in 2025. Google announced in September that it was investing $15.1 billion in AI infrastructure in Finland, in part because of the Nordic country’s cold climate and carbon-free power.

Among the European countries building the most AI computing capacity is France, which gets most of its power from low-carbon nuclear and renewables. New data centers in the US are projected to emit more carbon than the entire country of Mexico by 2030 if all are built on time with current air-quality permits, according to an analysis by the Sierra Club, an environmental nonprofit.

The number of gas plants linked to US data centers that were proposed in the first half of 2026 was double that of full-year 2025, reaching 189 gigawatts, according to Global Energy Monitor, a research firm. The US has twice as much gas power under construction as China does. It’s difficult to compare the carbon footprints of US AI platforms to their Chinese equivalents.

China has been building out its renewable energy infrastructure, adding more wind and solar power, in part to give itself a competitive edge against the US when it comes to AI. But China’s grids are still dirtier overall than those of the US, partly because Chinese power stations burn large amounts of coal. On the other hand, many US AI systems require more energy than Chinese ones. In the US, “the traffic is more, the models are bigger, the models need larger compute to handle them, so I would say US companies are emitting more, but it also depends on the grid itself because that's very important,” said Jegham.

Even if a data center is run by a tech company claiming net-zero emissions, it doesn’t mean there won’t be any greenhouse gases released.

Many use renewable energy certificates to offset their emissions. Those certificates pay for clean energy produced elsewhere, but the company’s operations still rely on local grids that are powered in large part by fossil fuels. Consumers, investors, activists, corporations and regulators want more granular data to guide their choice of model, help measure community impact, do cost analysis, make investment decisions and develop regulations around AI.

Natural gas plants could pose local air-pollution hazards, and states, cities and counties have pushed back against data-center expansion due to concerns over energy costs.

Companies with sustainability goals need to understand how the AI they are rapidly integrating into processes impacts their own emissions and their ability to meet net-zero targets. Some countries have policies in place that require this information to be accounted for and reported. Investors who prioritize sustainability might not want to purchase stakes in AI companies whose operations harm the environment. Investment research firm Morningstar said in a report that the lack of disclosure from AI companies “limits investors’ ability to assess environmental performance or resource efficiency.”

Relatively small changes can go a long way to reducing AI’s energy consumption and costs. If the average power-supply efficiency improved by 0.5%, it could reduce data center electricity consumption by roughly 5 terawatt-hours per year, which would cut out 2 million metric tons of carbon dioxide emissions annually, according to Delta Electronics, a Taiwan-based global green-hardware manufacturer. Companies don’t need perfect data to reduce their climate impact by choosing models known to require less energy, and from regions with cleaner power sources, said Bistline at Watershed.

“Model selection can be one thing that drives up to a 30 times difference in energy use for the same AI task,” he said.

“And that's probably the single highest-impact lever that most companies have today.”

The US Securities and Exchange Commission has backed away from Biden-era plans to require large companies to disclose their emissions. But Democrats' sweeping victory in the midterm elections means stricter AI regulation may be on the way, with more pressure for transparency. In the meantime, the state of California is beginning to roll out its own disclosure requirements for so-called scope 1 and scope 2 emissions, with scope 3 going into effect next year. These will apply to companies that generate more than $1 billion in annual revenue and do business in the state — including big AI startups.

The European Union’s emissions disclosure framework, CSRD, is already impacting European-based companies and will expand to multinational corporations in the next few years. There are other frameworks rolling out across the globe, including ISSB. The evolving regulation will force AI companies to be more transparent about their impact on the environment.

The bewildering growth of AI energy use means researchers such as Bistline and Luccioni are kept busy developing robust ways of measuring its impact.

“For a while, companies just didn't think it was possible” to measure the emissions, said Luccioni, “and now they're getting pressure from their employees, from shareholders... I think it's really hard nowadays to just sweep AI use under the rug.”