EcoCost labeling for AI inference
How we make the environmental cost of AI visible.
We believe all AI consumers and organizations should be able to easily understand the environmental impact of their AI use. Hence, EcoCost is a public repository and API to provide estimates of energy use, emissions, and water consumption of frontier AI models and the datacenters that run them, allowing any AI tool to display environmental costs alongside model price and performance. The open-source Python library, ecocost, is available now.
The problem
Every AI query consumes electricity and water, but AI users and builders lack a common resource for comparing environmental performance across models and server regions. Without this information, we cannot readily factor environmental impact into our AI model and provider choices. Gathering this information isn't easy either: models are released constantly, datacenters often don't publish their ecological emissions, and data quality estimates from across the ecosystem can vary widely.
Theory of change
If AI services displayed their environmental impact, users, organizations and governments could consider ecological costs when choosing models and providers, alongside accuracy, latency, sovereign-deployability and price. We hope that, like other forms of product labeling of externalities (think "organic" or nutrition information), displaying eco-costs will shape market demand, push consumers to demand cleaner AI, and create greater economic incentives for frontier labs and datacenters to lower their water, carbon, and power consumption. Additionally, by combining our efforts into a shared collaborative knowledge source - a code repository that can be updated by AI agents and vetted by human maintainers - we hope to aggregate relevant, up-to-date, high-quality research around AI eco-costs that the entire industry can rely on.
How it works
The hard part is aggregating worldwide research and estimates of emissions. This public repository is an ecosystem effort to collate emissions data into standardized, easy-to-integrate per-token estimates of energy, emissions, and water use. We hope that data scientists, researchers, hyperscalers, and AI research agents contribute high-quality, vetted data. Repo maintainers and AI agents will continually curate the information to improve confidence in the estimates. Data centers, telcos, energy grids, and hyperscalers can also upload their best estimates. Connected AI platforms that leverage this repo inform consumers and organizations of the eco-cost of their AI usage, let them compare models and server regions, and track the impact of their AI usage over time.
Building the coalition
Research and data
Researchers, data scientists, and organizations develop evaluation and estimation methods, publish data, and perform monitoring.
Products and platforms
Like-minded product and platform providers publish data and make eco-cost information accessible to users.
Funders and philanthropies
Funding orgs establish the repository, evangelize the effort and API and support research, especially when direct estimates from model builders and/or datacenters are difficult to directly ascertain.
Hyperscalers and providers
Participating providers contribute vetted inference energy and emissions data, increasing the confidence of the estimates.
EcoCost in an AI product
Our hope is that AI tools - like Gooey.AI itself - show environmental information within their respective interfaces so that consumers and users can consider their impact, alongside other factors including a model's cost, intelligence, or speed. Below we show how we've integrated EcoCost information inside our own user interface in our agent builder. Here, a compact label beside the AI run's price surfaces emissions, and an expanded view explains resource use, compute time, and the location and energy mix behind the estimate.
The figures in the mockups come from the calculations of the ecocost repository. The run used GPT OSS 20B on a Norwegian data center provider, with 2,100 input and 219 output tokens. It came to 0.97 mg CO₂e (likely range 0.28–3.4 mg), about 0.15 ml of water (almost all of it at the power plant), and 22 mWh of electricity, about 8 seconds of a 10 W LED bulb.
The request ran in Norway, where ecocost's grid data shows about 90% hydro and 8% wind, at 28 g CO₂e per kWh (range 10–45). A link explains how the estimate is made, in this case with low confidence given that the data hasn't been calculated from data verified by the Norwegian hyperscaler.
Estimate your AI's EcoCost
This playground runs the real ecocost Python package in your browser, so nothing is sent to a server. Pick a model and provider, set the token counts, and the estimate updates as you go.
Loading the playground…
System architecture
The proposed architecture connects source data to estimates through five stages. Its shared principles are transparency, confidence, documented methodology, energy data that accounts for region, and eco-cost breakdowns for individual models.
- Transparency
- Confidence
- Documented methodology
- Region-aware energy data
- Per-model eco-cost breakdowns
-
01
Source data
Inputs include academic research, industry reports, websites and blogs, operational data and efficiency metrics from data centers, and public disclosures from hyperscalers and model providers. Example sources include Epoch AI, nScale, Google, Anthropic, and OpenAI.
-
02
Curation and contributions
Human contributors can review data, update methodologies, and validate results. Research agents run daily and automate collection and analysis across the web. Providers can directly submit environmental data, regional information, and infrastructure details.
-
03
Open source repository and estimation engine
The first version is live: GooeyAI/ecocost on GitHub, installable with
pip install ecocost. It is community-maintained and open to pull requests, and it contains:- Methodology documentation covering the calculation, the sources behind every default, and calibration against published measurements. Similar to epoch.ai, we estimate each model's token usage in seconds consumed on an H100 GPU equivalent. From this, we calculate the water, energy, and carbon required to produce a token at a datacenter, using the nation's energy grid as an approximation when first-hand data is not available.
- Calculator logic that converts tokens into compute, energy, carbon, and water.
- An open knowledge base of models, providers, grid regions, and hardware. Every value has a source and an evidence tier: published, derived, or assumed.
- A likely range and a worst case for every figure.
- Versioning. Every result records the method version that produced it.
Planned additions include research agents that gather new data and automated checks for anomalies.
-
04
API contract
A standard interface lets tools submit information about an AI request and receive environmental estimates. See the contract below.
-
05
Product interfaces
AI products such as Gooey.AI, environmental dashboards, chat interfaces, and partner applications display the estimates. Users could explore and compare impact, track usage, and expand a label to see model-level details such as tokens, usage, cost, eco-costs, and confidence.
API contract
The ecocost library has one function, estimate(). It
takes the details of a single request and returns eco-cost
estimates. The example values below come from ecocost 0.2.0
(method 0.3.0) for GPT-4o on OpenAI's API with 700 input and 300
output tokens. To estimate a run that uses several models, call it
once per model and add up the results. See the
full API reference, or try it above.
Inputs from AI tools
| Field | Example or purpose |
|---|---|
| Model |
gpt-4o. Provider-specific model ids resolve to
the same model.
|
| Token counts | Input tokens (700), output tokens (300), and cached input tokens (optional) |
| Provider |
openai. Optional: ecocost can infer it from the
endpoint URL or, for closed models, from the model's vendor.
|
| Grid region |
Optional. A grid id such as US-VA or
FR, if you know where the request ran.
Otherwise ecocost uses the provider's known sites.
|
Estimates returned to tools
| Field | Example or purpose |
|---|---|
| Carbon | 0.099 g CO₂e (likely 0.025–0.37 g). Split into operational emissions, from generating the electricity, and embodied emissions, from manufacturing the hardware. |
| Energy | 0.26 Wh at the data center meter (likely 0.075–0.88 Wh) |
| Water | 0.95 ml consumed (likely 0.15–6.1 ml): 0.06 ml for data center cooling and 0.89 ml at the power plant |
| H100-equivalent compute time | 0.79 seconds (likely 0.25–2.2 s) |
| Region and grid energy mix | US, 373 g CO₂e per kWh (195–591 across OpenAI's candidate sites). Largest source: natural gas (40%). |
| Worst case | The same figures with every input set to its extreme at once |
| Provenance | Review status of each model, provider, region, and hardware record used |
Displaying the results
The compact example label combines carbon, water, energy, and compute time. An expanded view would add the assumptions and context needed to interpret those figures, including region, grid energy mix, and the likely range, with optional detail by model.
Where we need help
-
01
Better AI research monitors
We leverage AI research agents to scour the internet for new EcoCost reports to update our estimates. This is an early implementation with likely significant logic and security holes that may still be present. These processes need to be hardened, stress-tested, red-teamed, and expanded to include every data center offering private and open-source inference on earth.
-
02
Human and AI moderation systems
We take inspiration from collective knowledge aggregation systems like Wikipedia, but acknowledge they only work well with dedicated human moderators. With a research effort this large and controversial, we need systems for human moderators to adjudicate conflicts in human- and AI-sourced research, and improved AI systems to replicate this moderation.
-
03
First-hand data outreach and integration
Telcos, datacenters, hyperscalers, and neo-clouds have the first-hand data required to create high-confidence EcoCost estimates. We want these providers to publish and update their ecological impact, and to see this effort as a mechanism for creating market demand for cleaner AI.
-
04
Environment research community outreach
In their effort to win the AI race, many AI providers are turning to dirty, readily available energy sources. Many may choose to keep their ecological impact data private. We must continuously absorb the environmental research community's estimates when reliable first-hand data is not available.
-
05
Evangelism to AI tool makers
Users and organizations can make cleaner choices when they can see the impact of their AI usage. EcoCost labeling must be broadly available in the AI tools people use every day, from mainstream providers to enterprise and consumer AI products.
-
06
Outreach to institutional AI users
Companies, universities, and public-sector organizations with board-level commitments to reduce their ecological footprint should be able to determine their model token usage and accurately estimate its EcoCost.
-
07
Policy collaborations
Building a reliable code repository would be complemented by policy interventions requiring AI model makers, datacenters, and providers to audit and publish their ecological emissions.
Get involved
EcoCost is at an early stage. The first version of the library is public, and we're looking for researchers and data scientists, product and platform teams, funders, and hyperscalers to help improve its methodology and data, and to add more models and providers.
Open-source repository
Browse the code, methodology, and data. You can open a pull request to correct a figure or add a model or provider.
github.com/GooeyAI/ecocost