- One-time carbon cost per model that varies dramatically based on data center electricity source.
- Continuous emissions that can exceed training costs within months for popular models with millions of users.
- ChatGPT queries produce 0.4 to 4 grams CO2 equivalent, roughly 2-20 times more than Google searches.
- AI could reach 0.5-1% of global electricity consumption within a decade, comparable to streaming video today.
- Use 5-20 times more energy per query than standard models due to extended internal processing.
- AI is in exponential growth phase unlike mature industries, making current decisions about energy efficiency critical.
Training vs Inference: Two Different Problems
The carbon footprint of AI has two distinct components. Training is the process of building a model, running a large compute job that may last weeks or months and consume enormous quantities of electricity. InferenceInferenceThe process of running a trained AI model to generate outputs — what happens when you send a prompt and receive a response.Learn more → is the process of using a model to answer user requests, running continuously as long as the model is deployed. Both contribute substantially to AI's carbon footprint, but in very different patterns.
Training is a one-time cost per model. It produces concentrated, intense emissions over a defined period. The decision about where to run a training job, which data center, with what electricity source, has a large impact on the carbon cost of the resulting model. A model trained in a data center powered by coal-heavy electricity may have a dramatically higher carbon footprint than the same model trained in a data center powered by hydropower, even if the compute consumed is identical.
Inference is a continuous cost that grows with the number of users. Once a model is deployed, every query generates a small emissions footprint. The aggregate inference emissions from a popular model used by hundreds of millions of people daily can exceed the training emissions within months of deployment. This makes inference efficiency as important as training efficiency from a carbon perspective.
The ratio of training to inference emissions varies enormously by application. A highly specialized model deployed to a small number of users might have training emissions that dwarf its lifetime inference emissions. A general-purpose model used by millions of people daily might accumulate more inference emissions than training emissions within its first year. Understanding which component dominates for a given application is essential for prioritizing emissions reduction efforts.
Reasoning models complicate the picture further. These models, which generate extended internal chain-of-thought before producing a response, use substantially more compute per query than standard models. A reasoning modelReasoning ModelA class of LLMs trained specifically to 'think' through problems step by step using extended chain-of-thought reasoning before producing a final answer, excelling at math, coding, and complex logic.Learn more → might consume five to twenty times more energy per inference than a standard model of similar size. As reasoning models become more widely deployed, inference emissions per query increase. The total inference emissions from the AI industry may grow faster than user numbers suggest.
Carbon Per Query: What a ChatGPT Message Actually Costs
Estimating the carbon cost of a single AI query requires several assumptions. The electricity consumption per query depends on the model size, the hardware used, and how efficiently the hardware is utilized. The carbon intensity of that electricity depends on the energy source: coal-heavy grids produce roughly 800 to 1,000 grams of CO2 per kilowatt-hour, while hydropower or nuclear produces close to zero.
Researchers from the University of Washington published estimates in 2023 based on disclosed efficiency metrics and publicly available grid carbon intensity data. They estimated that a text-based ChatGPT query using GPTGPTGenerative Pre-trained Transformer — the model architecture and family name behind OpenAI's most famous models, from GPT-2 to GPT-5.Learn more →-3.5 consumed approximately 0.001 to 0.01 kilowatt-hours per query, producing roughly 0.4 to 4 grams of CO2 equivalent depending on the electricity source. For reference, a Google search produces approximately 0.2 grams of CO2. An AI query uses roughly two to twenty times more energy than a traditional search.
These figures sound small at the individual level. But at scale, they accumulate. ChatGPT processes over 100 million queries per day. At an average of 2 grams per query, that is 200 metric tons of CO2 equivalent per day, or roughly 73,000 metric tons per year. This is the emissions equivalent of driving a medium-sized car for approximately 30 million miles.
Image generation queries are significantly more energy-intensive than text queries. Generating a high-resolution image from a text promptPromptThe input text sent to a language model — the question, instruction, or context that triggers a response.Learn more → requires running a diffusion modelDiffusion ModelA generative AI model that creates data by learning to reverse a noise-adding process, iteratively denoising random inputs to produce high-quality outputs.Learn more → for many denoising steps, each of which consumes compute. Estimates suggest that image generation is five to ten times more energy-intensive per request than text generation. As multimodalMultimodalCapable of processing and generating multiple types of data — such as text, images, audio, and video — within a single model.Learn more → AI becomes more common and image and video generation become mainstream features, the per-query emissions from AI will increase.
The trend in AI query volume is steeply upward. As AI features are integrated into more products and more users adopt them, the total number of AI queries globally is growing rapidly. Even if individual query efficiency improves, the growth in total queries means that aggregate AI inference emissions are likely increasing. The efficiency gains from better hardware and algorithms are being outpaced by growth in usage, a pattern familiar from the history of other digital technologies.
How AI Compares to Other Industries
Placing AI's carbon footprint in context requires comparing it to other industries and technologies. The global data center industry as a whole accounts for approximately 1 to 1.5 percent of global electricity consumption, and approximately the same share of global carbon emissions. AI is a rapidly growing component of this, but it is not yet the dominant contributor. At current growth rates, AI data centers could account for 0.5 to 1 percent of global electricity consumption within the decade.
Aviation accounts for approximately 2.5 percent of global CO2 emissions and a larger share of total climate impact when non-CO2 effects are included. Global shipping accounts for approximately 2.9 percent of global CO2 emissions. Steel production accounts for approximately 7 to 9 percent. Agriculture accounts for approximately 10 to 12 percent. In absolute terms, AI's current carbon footprint is smaller than these sectors.
The comparison to streamingStreamingStreaming returns a model's tokens incrementally as they're generated rather than waiting for the complete response to finish.Learn more → video is instructive. Streaming video accounts for approximately 0.3 percent of global carbon emissions. AI training and inference combined are already comparable to or larger than this, despite video streaming having a far larger established user base. The energy intensity of AI per unit of data processed is higher than for video, which requires less computation.
The cryptocurrency comparison is also relevant. At its peak in 2021, Bitcoin mining consumed approximately 130 to 150 terawatt-hours of electricity per year, comparable to the entire electricity consumption of Argentina. This generated approximately 65 to 70 million metric tons of CO2. The AI industry is not yet at this scale but is growing faster. The structural difference is that cryptocurrency mining has a clear value-per-unit calculation (the value of Bitcoin mined), while the value of AI computation is harder to quantify but arguably more diverse and broadly distributed.
One important difference between AI and many other industries is the trajectory of its growth. Steel, aviation, and shipping are mature industries with relatively slow growth. AI is in an exponential growth phase. This means that even if AI's current carbon footprint is small relative to other industries, its trajectory matters. Decisions made now about the energy sources and efficiency standards for AI infrastructure will determine whether AI's carbon footprint remains manageable or becomes a significant climate concern.
Google's Broken Carbon Pledge
In 2020, Google announced that it had achieved carbon neutrality for the calendar year 2019, making it the first major company in the world to do so. It pledged to run entirely on carbon-free energy by 2030. These announcements were widely celebrated as evidence that a major technology company could achieve ambitious climate targets while continuing to grow.
In its 2024 Environmental Report, Google disclosed that its total greenhouse gas emissions had increased by 48% between 2019 and 2023. The primary driver was data center energy consumption for AI workloads. The company stated that its 2030 carbon-free energy goal was now 'extremely ambitious' given the pace of AI growth. The honesty of the disclosure was notable. The substance was a direct contradiction of the path the company had projected in 2020.
Google's situation is not unique. Microsoft disclosed that its emissions had increased by 29% between 2020 and 2023, directly attributed to data center expansion for AI. Amazon Web Services emissions have also increased despite substantial renewable energy investment. The pattern across major cloud providers is consistent: absolute emissions are growing faster than renewable energy procurement and efficiency improvements can offset.
The credibility gap between stated commitments and actual performance is a fundamental problem. When companies announce ambitious climate targets and then report increasing emissions year after year, the targets begin to function as marketing rather than accountability. Investors, regulators, and the public calibrate their expectations accordingly. The result is a kind of climate credibility erosion that undermines confidence in corporate sustainability commitments more broadly.
Some of Google's specific claims also deserve scrutiny. The 2019 carbon neutrality claim rested heavily on the purchase of renewable energy certificates, instruments whose effectiveness is contested. The company was purchasing credits for renewable energy generated somewhere on the global grid, not necessarily running on clean electricity in the specific locations where its data centers operate. This practice is legal and increasingly common, but it is not the same as actually operating on renewable energy.
The Rebound Effect: Efficiency That Makes Things Worse
The rebound effect is an economic phenomenon where efficiency improvements lead to increased consumption, partially or fully offsetting the environmental benefit of the efficiency gain. It was first described in the context of energy by Stanley Jevons in 1865, who observed that more efficient steam engines led to more coal consumption, not less, because efficiency made coal economically attractive for more applications. This is now known as the Jevons paradox.
The Jevons paradox applies directly to AI. More efficient AI models are cheaper to run. Cheaper AI enables more AI use cases. More AI use cases means more total compute consumed. The net effect of efficiency improvements is often an increase in total energy consumption, even as the energy consumed per unit of output falls.
A concrete example: Nvidia's H100 GPU is approximately three to five times more energy-efficient than the previous-generation A100 for AI workloads. This is a genuine and impressive efficiency improvement. But users have responded to the improved efficiency by buying more H100s, training larger models, and running more inference requests. Total data center electricity consumption has increased, not decreased, as a result of the efficiency improvement.
This is not an argument against improving efficiency. Efficiency improvements are genuinely beneficial: they reduce the cost of AI, increase its accessibility, and slow the growth of absolute consumption relative to what it would be without the improvements. The point is that efficiency improvements alone are insufficient to stabilize or reduce absolute emissions. Additional mechanisms, such as caps on total consumption, carbon pricing, or mandatory offsets from actual emissions rather than credits, are needed to ensure that efficiency gains translate into environmental benefit.
The rebound effect also applies to AI-enabled productivity. If AI makes workers more productive, and organizations respond by doing more work with the same number of workers rather than using fewer workers to do the same amount of work, total economic output increases. Higher economic output is generally correlated with higher energy consumption. AI-driven productivity growth, if it is broadly realized, will likely increase total energy demand, of which AI infrastructure is itself a part.
Carbon Offsets and Greenwashing
Carbon offsets allow companies to compensate for their emissions by funding activities that reduce emissions elsewhere: planting trees, funding renewable energy in developing countries, capturing methane from landfills, or paying communities not to deforest land. In principle, a carbon offset represents a genuine reduction in atmospheric CO2 equivalent to the offset's face value. In practice, the offset market has significant integrity problems.
A 2023 investigation by The Guardian, Zeit Online, and SourceMaterial found that more than 90% of the rainforest offset credits approved by Verra, the world's leading carbon standards body, were 'phantom credits' that did not represent genuine carbon reductions. The projects certified as offsetting carbon emissions turned out to be in areas that were not actually at risk of deforestation, or where claimed emissions reductions were greatly exaggerated. Companies that purchased these credits were not actually offsetting their emissions.
The additionality problem is fundamental to offset markets. An offset is only genuine if the carbon reduction would not have happened without the offset payment. Counting renewable energy that would have been built anyway as an offset is not additionality. Counting forests that were never at risk of logging as avoided deforestation is not additionality. But determining what would have happened without the offset payment is inherently difficult, creating opportunities for manipulation and error.
Permanence is another challenge. A forest planted today to offset today's emissions may burn down in a wildfire, or be converted to agriculture by a future government, releasing the stored carbon. Carbon emissions today persist in the atmosphere for hundreds of years. An offset that lasts twenty years does not balance emissions that persist for a century. The asymmetry between the permanence of emissions and the durability of many offset projects undermines the accounting.
The technology companies that rely heavily on carbon offsets to claim carbon neutrality are exposed to these integrity problems. When their offset portfolios consist substantially of credits that do not represent genuine reductions, their carbon neutrality claims are misleading, even if they are made in good faith. The appropriate response is to prioritize direct emissions reductions over offset purchases, use only high-quality, independently verified offsets, and be transparent about the distinction between actual emissions reductions and offset-based accounting.
Policy Responses: What Governments Are Doing
Governments are beginning to address the carbon footprint of AI through several policy mechanisms. The European Union's Energy Efficiency Directive requires large data centers to report energy and water consumption and efficiency metrics. The EU is also considering whether data centers should be required to meet sustainability standards to access certain government contracts or financial instruments.
The United States has taken a more indirect approach. The Securities and Exchange Commission's climate disclosure rules, finalized in 2024, require large public companies to disclose their Scope 1 and Scope 2 greenhouse gas emissions. Data center-heavy companies will need to disclose the emissions associated with their electricity consumption, creating more public information about the carbon footprint of cloud computing and AI.
Some states have been more proactive. Virginia, the state with the highest concentration of data centers in the world, has enacted legislation requiring data centers to meet certain energy efficiency standards to qualify for tax exemptions that have historically been broadly available. California has been considering requirements for data centers to use recycled water for cooling and to disclose their environmental impacts in more detail.
China has been developing data center energy efficiency standards and has encouraged the siting of new data centers in regions with lower-carbon electricity grids. The government has identified data centers as a priority sector for energy efficiency improvement and has included them in China's carbon market, requiring facilities above certain thresholds to manage their carbon footprint.
International coordination on AI's environmental impact is nascent. The OECD has published principles for trustworthy AI that include environmental considerations. The G7 has discussed AI governance, including environmental aspects. The International Energy Agency has published analysis of data center energy trends that informs policy discussions. But binding international agreements specifically addressing AI's carbon footprint do not exist. The regulatory gap means that environmental performance varies widely across jurisdictions and that competitive pressure discourages individual companies from taking unilateral action that increases their costs.
What Users and Developers Can Do
Individual users have limited but non-zero influence over AI's carbon footprint. Choosing AI providers that publish transparent, detailed, and independently verified environmental data sends a market signal. Avoiding unnecessary AI queries, particularly for tasks that do not require AI capability, reduces demand for energy-intensive inference compute. Supporting regulatory frameworks that require environmental accountability from AI companies is a meaningful political action.
Developers make more consequential choices. Selecting the smallest model that is sufficient for a task is one of the highest-impact decisions a developer can make. A task that can be completed adequately by a small, efficient model should not be completed by a frontier modelFrontier ModelThe most advanced and capable AI models at the cutting edge of the field, typically developed by well-funded research labs with massive computational resources.Learn more → that consumes ten times more energy per query. This is both economically rational and environmentally responsible. The two goals align.
Model selection also matters at the infrastructure level. Some model providers run their inference on electricity grids with lower carbon intensity than others. Some providers are more transparent about their environmental practices than others. Choosing providers based partly on environmental criteria is increasingly practical as more information becomes available.
Batching inference requests reduces energy consumption per query. Rather than sending each request individually and waiting for a response, batching multiple requests together allows the hardware to be utilized more efficiently, reducing the compute and energy required per response. For applications where real-time response is not required, batching can reduce inference emissions substantially.
The most significant thing developers and organizations can do is advocate for environmental accountability in AI governance. The regulatory frameworks that will determine whether AI's carbon footprint is managed responsibly are being designed now. Technical community input on what is feasible, what should be measured, and how performance should be evaluated is genuinely valuable for policymakers who often lack detailed technical knowledge. Engagement with this policy process, rather than treating it as someone else's problem, is both a civic responsibility and an opportunity to shape an outcome that matters.