Introduction
As AI models grow larger and more capable, so does their demand for resources. The rapid expansion of data centers across the country to support training and inference has sparked fierce debate about the environmental cost of AI, but these conversations are largely proceeding without data on the true scale of impact.
With seven in 10 Americans opposing construction of data centers in their locality, pressure mounts on governments at every level to enact regulations. A number of municipalities, including Minneapolis, Denver, and Seattle, have placed temporary citywide restrictions on building or expanding data centers as they evaluate environmental outcomes and formulate long-term solutions. This March, Senator Markey and Representative Beyer introduced the Artificial Intelligence Environmental Impacts Act of 2026, an act which would require data centers to report on their environmental and energy-related impacts. California has gone further, requiring large companies to disclose their greenhouse gas emissions and climate-related financial risk.
While municipalities and impacted communities grapple with unease over AI’s environmental footprint, that concern hasn’t made its way into how we evaluate models. Labs offer little transparency, leaving reliable data on real-world environmental impact lacking. Meanwhile, LLMs have been benchmarked extensively on accuracy, speed, and cost, but not sustainability. This report aims to start filling this gap.
Methodology
To estimate environmental impacts, we collected token usage data for tasks from our Vals Index benchmark. The index is composed of long-horizon, multi-turn, agentic tasks from different industries such as finance, coding, healthcare, and law. These tasks are a good proxy for real-world work that a model would be expected to perform were it to be deployed in the economy. We use this data in combination with EcoLogits, an open-source library for estimating environmental impacts of AI inference.
EcoLogits takes in token usage along with a model’s total and active parameter counts, then estimates carbon, electricity, and water consumption based on assumed hardware and model specifications. Because this approach depends on knowing a model’s architecture, our study covers open-weight models only; we discuss this and other constraints in the Limitations section. You can read more about the EcoLogits methodology here.
Vals Index
Results
Small gains in capability can come with outsized environmental costs. Kimi K3 tops the open-weight leaderboard with 57% accuracy on the Vals Index, with the runner-up, DeepSeek V4 Flash, only 4% behind. This performance comes at the cost of a larger environmental footprint. The difference is the gap between charging your laptop once and twenty times, drinking a glass of water and flushing a toilet, or reparking your car and driving across San Francisco.
These estimates represent only the environmental impacts of an average task on the Vals Index. Effects could quickly add up if models were to be deployed to automate a significant portion of the economy. Our score-weighted estimate for running the entire index (roughly 2000 tasks) on Kimi K3 emits as much carbon as flying from San Francisco to Miami, uses enough water to fill the water tank of a fire truck, and enough energy to power the average US home for an entire month.
Running Vals Index
Impacts extrapolate benchmark-aligned mean output tokens across the canonical 2,157-task Vals Index. Code Migration follows its published 75/25 CLI/COBOL score weighting rather than literal task frequency, so this is a score-weighted estimate. Input, cache, and prefill compute are excluded.
Token pricing and cost per task are imperfect proxies for environmental footprint. On agentic tasks, cheaper models can be less token-efficient, driving up both resource demand and cost per task. DeepSeek V4 Pro, for instance, is nearly 15x cheaper per output token than Kimi K3 yet has roughly the same environmental footprint. Cost per task fares no better as a predictor, since DeepSeek V4 Pro is also roughly 6x cheaper. The relationship between pricing and environmental impacts is further distorted by market dynamics, as model providers may offer tokens at steep discounts while trying to capture market share.
Vals Index
The x-axis shows each model's published output-token price per million tokens; observed benchmark usage does not affect it.
Estimating environmental impact requires transparency about model architecture. Because this information is unavailable for closed-weight models, we could not produce estimates for them directly. Instead, we built a calculator that estimates environmental impact under user-specified assumptions about a model. Suppose a model had 6T parameters with 30% activation: running a single Vals Index task would use enough electricity to power a US home for one day, emit as much carbon as driving 50 miles, and consume enough water for a 10-minute shower. If model size continues to scale with capabilities, the environmental footprints will grow along with it.
Environmental Impact Calculator
We’ve only covered the environmental impacts of industry adoption, but AI usage is also growing among individuals. Using data from our upcoming Child Safety Benchmark, we estimated the environmental impacts of a 10-turn conversation a real user may have with a model. Kimi K3 has roughly 3x the environmental footprint of GLM 5.2, despite using half as many tokens, using enough energy to charge 14 phones, emitting as much carbon as driving a car for 0.2 miles, and consuming enough water to fill three glasses.
Environmental Impacts of One Conversation
| Model | Generated tokens | Energy | Carbon | Water |
|---|---|---|---|---|
| DeepSeek V4 Flash | 6.9k | 10.5 Wh | 5.7 gCO₂e | 46 mL |
| Kimi K3 | 6.4k | 169.5 Wh | 92.4 gCO₂e | 733 mL |
| GLM-5.2 | 12.3k | 50.4 Wh | 27.4 gCO₂e | 218 mL |
An estimated 66 million US adults now use AI chatbots daily. If each had one conversation on a relatively more efficient model like DeepSeek V4 Flash, the collective footprint would be roughly 690 MWh of energy, 375 tons of CO2, and 3 million liters of water.
Limitations
Estimating a model’s environmental impact requires knowing its architecture and serving infrastructure. Because this information is unavailable for closed-weight models, our analysis is scoped to open-weight models only. This leaves out many of the most widely used systems, though their absence is not indicative of a smaller footprint.
EcoLogits relies on several simplifying assumptions. Our estimates assume models are deployed on NVIDIA H100 GPUs with 16-bit weights and served with vLLM. In practice, models are often quantized to lower precision, which can reduce the hardware and energy required. Actual impacts can vary dramatically with hardware, data center efficiency, energy mix, and serving configuration. These results also do not consider the advent of closed-loop cooling systems, which may significantly reduce water consumption in the coming years. These results should be read as best-effort estimates rather than direct measurements. The full list of EcoLogits limitations can be found here.
Finally, carbon emissions, electricity use, and water consumption do not capture the full environmental and social cost of data centers. Nearby communities can face light, noise, and water pollution, urban heat-island effects, and strain on local power grids. None of these externalities are reflected in our estimates, so a model’s total impact may extend well beyond the three measures reported here.
What’s Next?
As model capabilities and deployment scale, their environmental footprints may grow as well, but the data needed to track that growth remains scarce. Better measurement will require greater transparency across the whole supply chain: clearer disclosure of model architectures, hardware utilization, and data center resource consumption. Standardized reporting of this kind exists in other resource-intensive industries, and AI should be no exception.
Raw metrics alone won’t paint the full picture either. Emission figures must be paired with detailed case studies on how data center construction affects local communities to connect numbers with lived outcomes.
Model development is advancing at a relentless pace and policies are scrambling to keep up. At Vals, we measure model capabilities to give policymakers and the public an evidence-based read on how and where to deploy this technology. We believe the same approach can inform sustainability policy. We are eager to collaborate with interested groups and welcome any feedback.

