Your AI Carbon Footprint: What Every Query Really Costs

Every time you ask an AI chatbot for a recipe, have it summarize an article, or draft a report, a cluster of GPUs in a data center draws power from the electrical grid, generates heat that must be cooled with water, and produces carbon emissions tied to whatever fuel mix supplies that grid. A single query seems trivial. Multiplied by the billions of daily AI interactions now occurring worldwide, the cumulative impact is anything but. Here is the uncomfortable truth: no major AI provider publishes complete, verifiable, per-query energy and emissions data. The figures that do exist come from a handful of company disclosures, academic estimates built on reverse-engineering, and third-party benchmarks, each using different assumptions, boundaries, and methodologies. Estimates for a single AI query range from 0.03 to 68 grams of CO₂, a spread so wide it borders on the meaningless without extensive context. The Hugging Face AI Energy Score, which offers a standardized benchmarking initiative co-led by Hugging Face and Salesforce, was created precisely to address this gap, rating AI models on energy efficiency across common tasks. But its leaderboard is populated almost entirely by open-source models, because most major commercial providers have declined to participate. Every AI response begins as tokens, the basic units AI models use to process and generate text, roughly equivalent to three-quarters of a word. A short reply might use 200 tokens; a detailed explanation can run 1,000 or more. Tokens matter for energy accounting because AI systems consume electricity proportional to the number of tokens they generate: more tokens require more computational cycles, more time on energy-intensive chips, and more heat to dissipate in data centers. Researchers use token counts as a standardized measure to compare energy consumption across models, it’s a rough approach similar to how miles per gallon lets you compare cars with different engines. AI technology is advancing rapidly, so we took a snapshot in time of its environmental impact based on the best available evidence, which is often already outdated. But this is an active and important debate, so we want to be transparent about the limits of any estimate of the impact from AI or any cloud service: the following is based on evidence and designed to help understand your own AI carbon footprint. As MIT Technology Review documented in its Power Hungry investigation, the factors that determine the carbon cost of your AI query—which data center processes your request, what energy mix powers it, how efficient the hardware is—are treated as trade secrets by every major provider. OpenAI, Anthropic, Google, xAI, Microsoft, Apple, and Perplexity all operate what researchers call “closed” models, in which the operational details are closely held by the companies that build them. Only two companies have disclosed specific per-query energy figures. In June 2025, OpenAI CEO Sam Altman stated that an average ChatGPT query uses about 0.34 watt-hours of electricity. In August 2025, Google published a detailed methodology showing the median Gemini text prompt consumes about 0.24 watt-hours and produces 0.03 grams of CO₂e. No other company in this guide has published comparable data. Anthropic, which offers Claude, has not disclosed per-query energy figures but states that it works with cloud providers that prioritize renewable energy and carbon neutrality. As of March 2026, Anthropic has not reported Scope 1, 2, or 3 emissions in any public filing. Perplexity, Microsoft Copilot, xAI’s Grok, and Apple have published no per-query environmental metrics. This opacity is the story. Without standardized disclosure, consumers cannot make informed choices, regulators cannot set evidence-based policy, and companies face no accountability for the environmental cost of scaling AI services. The best available data comes from multiple sources that approach the problem differently. Epoch AI developed bottom-up estimates of ChatGPT’s impact using model architecture and hardware specifications. The academic team of Jegham et al. (2025) benchmarked 30 models across infrastructure-aware frameworks. Google published its own measured data. We have synthesized these into the most complete picture currently possible, while flagging critical caveats. Table 1: Estimated Energy and Carbon Per Standard AI Text Query by Provider Provider Model Wh/Query g CO₂e Energy Mix Google Gemini (text) 0.24 0.03 64% Carbon-free* OpenAI GPT-4o 0.30–0.43 0.13–0.19 Mixed† OpenAI o3 (reasoning) 3.9–33+ 1.7–14.5 Mixed† Anthropic Claude Sonnet ~0.5–1.0‡ ~0.2–0.4‡ Mixed† Anthropic Claude Opus ~4.05‡ ~1.8‡ Mixed† xAI Grok Not disclosed ⚠ Contested Natural gas§ Microsoft Copilot Not disclosed Not disclosed Mixed† Perplexity Perplexity AI Not disclosed ~4.0‡ Mixed† Apple Apple Intelligence Not disclosed Not disclosed 100% renewable¶ * CFE = Carbon-Free Energy; Google reports 64% globally, with 10 regions at 90%+. † “Mixed” = Azure/AWS grid mix; varies by location and time. US data center grid carbon intensity averages 48% higher than the national average. ‡ Third-party estimate, not company-disclosed. Treat as approximate. xAI’s Memphis Colossus facility powered substantially by 35 methane gas turbines. Natural gas emission intensity: ~0.49 kg CO₂/kWh. ¶ Apple reports 100% renewable energy matching across data centers since 2014; 2.5 billion kWh consumed in 2024. On-device processing has near-zero cloud footprint. A widely reported but undisclosed 2025 study by TRG Datacenters ranked Grok as the most eco-friendly chatbot at just 0.17 grams of CO₂ per query. That claim deserves serious scrutiny because the infrastructure behind Grok tells a very different story. The study relied on a low per-query figure for Grok, which likely reflects the smaller model architecture rather than the full infrastructure uses. xAI’s Colossus facility in Memphis, Tennessee, which runs the primary supercomputer powering Grok, operated for months with 35 methane gas turbines running without proper air pollution permits. According to the Southern Environmental Law Center, the turbines could produce 1,200 to 2,000 tons of nitrogen oxides annually, making the facility one of the largest industrial emitters of NOx in the Memphis area. The facility sits adjacent to Boxtown, a historically Black neighborhood that already carried disproportionate pollution burdens before xAI arrived. Epoch AI estimated the emissions intensity of training Grok 4, which consumed 310 GWh of electricity, cost $490 million, used approximately 750 million liters of water, and produced emissions equivalent to roughly 150,000 tons of CO₂. That training-phase footprint is equivalent to the annual carbon output of more than 10,000 Americans. When a “clean” chatbot runs on an unpermitted gas plant in an environmental justice community, the per-query carbon number is the wrong metric to focus on. Recent generations of AI—OpenAI’s o3, Anthropic’s extended thinking mode, Google’s Gemini with Deep Research, and so forth—represent a fundamental shift in energy consumption. Standard models predict the next word in a response, effectively mimicking it training rather than “thinking.” Reasoning models generate thousands of hidden tokens to consider a question before producing a visible response, which dramatically multiplies energy costs. According to Jegham et al.’s benchmarking study of 30 models, o3 and DeepSeek-R1 consumed over 33 watt-hours for a single long prompt, more than 70 times the energy of GPT-4.1 nano for the same task. Standard models averaged an additional 37.7 tokens per question, while reasoning models generated an additional 543.5 tokens on average, even for simple multiple-choice questions. The University of Rhode Island’s AI lab estimated that GPT-5, which integrates reasoning capabilities, consumes an average of over 18 watt-hours per medium-length response. With extended reasoning mode enabled, energy consumption can increase five- to tenfold, potentially exceeding 40 Wh per query, or roughly the energy needed to charge two smartphones. This matters because the industry is moving aggressively toward reasoning models as the default. OpenAI researchers have publicly stated ambitions for models that “think for hours, days, even weeks.” The energy implications of that trajectory should be part of the public conversation. Every AI interaction is different, but we can build reasonable estimates based on model type, token count, and the energy data available. A short query might involve 50–100 input tokens and 200–500 output tokens. A long research session can involve hundreds of thousands of tokens. Output tokens cost roughly three to five times more energy per token than input tokens because the model must generate each word sequentially. The following table estimates the carbon footprint of common AI tasks using the best available data. We use a blended average of 0.3–0.5 Wh per standard query (based on OpenAI and Google disclosures) and scale by token volume and model type. CO₂ estimates use the US average grid intensity of 0.39 kg CO₂/kWh, which is lower than the 48%-higher data center average identified by Harvard researchers.Record the prompts you use, compare them to the table below, and with a bit of math you can estimate your daily CO2 emissions due to your AI use. You may need to use AI to help, alas. Table 2: Estimated Carbon Cost of Common AI Tasks AI Task ~Tokens ~Wh ~g CO₂e Everyday Equivalent Asking about the weather ~300 0.2–0.3 0.08–0.13 Running a microwave for 1 second Summarizing a 2,000-word article ~3,500 0.4–1.0 0.16–0.4 Watching 5–10 seconds of television Drafting a 500-word email ~1,500 0.3–0.5 0.12–0.2 Running a fridge for 6 seconds Generating a page of code with debugging ~5,000 1.0–2.5 0.4–1.0 Charging a phone to 10–15% Deep Research / Extended Thinking report 50K–200K+ 10–40+ 4–18+ Charging 1–2 smartphones fully; streaming Netflix for 15–45 min Generating a single AI image N/A 0.3–2.9 0.1–1.3 Running a laptop for 2–10 minutes on standby AI agent monitoring 100 stocks daily 500K–1M+/day 50–200+/day 20–88+/day Driving a gasoline car 0.5–2 miles per day; 7–30 kgCO₂/month Full-day coding agent session (Claude Code, Cursor) 5M–10M+ 50–600+ 20–260+ Driving a gasoline car 1–15 miles; comparable to a home Wi-Fi router running 24/7 Generating a 5-second AI video N/A ~944 ~414 Riding an e-bike 38 miles; running a microwave for over 1 hour Note: Estimates use US average grid carbon intensity (0.39–0.44 kg CO₂/kWh). Actual emissions vary significantly by provider, data center location, time of day, and model. Ranges reflect uncertainty across sources. Token counts are approximate. If the energy cost of AI varies enormously, so does the intelligence-per-watt return. The Jegham study introduced an “eco-efficiency” score using Data Envelopment Analysis to balance model performance with environmental costs across 30 models. Anthropic’s Claude 3.7 Sonnet scored highest in eco-efficiency (0.886), combining strong reasoning performance with efficient infrastructure use when running on Amazon Web Services. OpenAI’s o4-mini (0.867) and o3-mini (0.840) also performed well, demonstrating that smaller reasoning models can deliver solid results at a fraction of the environmental cost of their larger counterparts. At the opposite end, DeepSeek-R1 (0.058) and DeepSeek-V3 (0.060) scored lowest, reportedly reflecting high energy consumption compounded by infrastructure inefficiencies in its data centers. OpenAI’s GPT-4.5 (version 5.4 was released just a week before this article), also ranked among the least efficient, confirming that newer does not automatically mean greener. The practical takeaway: choosing the right model for the right task is one of the most effective ways to reduce your AI carbon footprint. Using a frontier reasoning model to draft a grocery list wastes 10 to 100 times more energy than a smaller model that handles the task just as well. Individual query footprints, even at the high end, pale in comparison to the infrastructure buildout now underway. US data centers consumed 4.4% of all national electricity in 2024, with the share potentially tripling to 12% by 2028. A Harvard study found that the carbon intensity of electricity used by data centers was 48% higher than the US average, because data centers disproportionately draw from fossil-fuel-heavy grid segments. The investment scale is staggering. SoftBank, OpenAI, Oracle, and Emirati investment firm MGX intend to spend $500 billion on new US data centers over four years through the Stargate initiative. Apple has committed $500 billion to manufacturing and data centers. Google plans to spend $75 billion on AI infrastructure in 2025 alone. Anthropic has suggested the US build an additional 50 gigawatts of dedicated power by 2027. A December 2025 study in the journal Patterns estimated that AI systems running in data centers could produce between 32.6 and 79.7 million tons of CO₂ in 2025, comparable at the low end it’s closer to Norway’s annual emissions, and, at the high end, it exceeds New York City’s emissions by 50%. The study’s author urged that further disclosures from data center operators are urgently required to improve the accuracy of these estimates and to responsibly manage the growing environmental impact of AI systems. Efficiency improvements are real. Google reported a 33x reduction in energy per median prompt over one year, but historically, efficiency gains in computing have been overwhelmed by growth in usage. The industry is betting that reasoning models, agents running continuously, AI-generated video, and AI embedded in every app will drive exponential growth in total compute demand. Whether efficiency can keep pace with that demand is the central climate question of the AI era. Right-size your model to the task. If you need a quick answer, use a smaller or faster model. Reserve reasoning modes and frontier models for tasks that genuinely require them. The energy difference can be 70x or more. Write efficient prompts. Output tokens cost three to five times more energy than input tokens. Asking for a three-sentence summary instead of an open-ended response can cut energy use significantly. Avoid unnecessary follow-up queries by being specific upfront. Factor in the energy source. Google’s infrastructure currently produces the lowest published per-query emissions, partly because of its 66% carbon-free energy rate and specialized AI hardware. Apple’s on-device processing for simpler Apple Intelligence tasks avoids cloud computing entirely, so your source of energy will influence the total environmnetal impact of asking Siri a question. If provider energy sources are available to you, look into it. Audit AI agent usage. Always-on AI agents, such as ones monitoring stocks, scanning inboxes, or running continuous analysis, can consume orders of magnitude more energy than conversational use. If you deploy agents, evaluate whether continuous operation is necessary or whether periodic batch processing achieves the same outcome at a fraction of the energy cost. Demand transparency. The single most impactful action may be pushing AI providers to disclose standardized environmental metrics. The EU is already moving toward mandatory disclosure. Consumers, enterprise buyers, and developers should ask providers for per-query energy data, data center energy sourcing, and water consumption figures. The Hugging Face AI Energy Score leaderboard and ML.Energy are independent resources that benchmark model efficiency across tasks. Skip AI when you don’t need it. A traditional Google search uses about 0.3 Wh. Opening a weather app that pulls cached data uses almost nothing. Not every question requires a large language model. The most sustainable AI query is the one you did not have to make.

View Original Article