Environmentally Responsible AI

The Commitment

CRAI is committed to accounting for and neutralizing the carbon and freshwater footprint of our AI usage and applying the same standard to every project we run with partners. Our approach to internalizing these external environmental costs has two parts.

We fund the permanent retirement of carbon allowances in the Regional Greenhouse Gas Initiative (RGGI) market through the Carbon Lighthouse Association, reducing fossil-fueled power-sector emissions by an amount at or above what our AI use is responsible for. We also contribute to the Colorado Water Trust's myRiver Balance program, returning an equal or greater volume of water to Colorado rivers than our AI use consumes elsewhere.

We publish the calculations and the purchase records, so anyone can check the claim.

Why This Fits Our Mission

Rural and Western Slope communities carry a disproportionate share of the physical cost of AI. Data-center water withdrawals concentrate in specific watersheds, and Colorado's rivers already run under drought and over-allocation.

CRAI's work puts rural communities in a position of ownership over AI, and the environmental footprint of AI is one of the costs that ownership has to reconcile with. Accounting for it belongs inside our working definition of AI for the public good — and this program lets us teach the practice by doing it first.

Part One

Our Own Footprint

We estimate organization-wide AI use at roughly ~50 million tokens per month across all platforms. Run through the Tokenomy energy estimator with average assumptions for electric-grid mix and AI model type, that use produces an estimated 1.25 tons of carbon dioxide and ~4,000 gallons of water per month — about ~15 tons of CO₂ and ~48,000 gallons of water per year.

These are estimates with wide error bars, so we purchase above the estimate rather than at it.

In Everyday Terms

What keeping four tons of CO₂ out of the air and returning roughly 20,000 gallons of water looks like each quarter.

Carbon method

Allowance Retirement, through the Carbon Lighthouse Association and RGGI

The Carbon Lighthouse Association is a nonprofit that buys carbon allowances at Regional Greenhouse Gas Initiative (RGGI) auctions and holds them permanently on a donor's behalf. RGGI is the first mandatory market-based program in the United States to cap power-sector CO₂ emissions, covering a group of Northeastern and Mid-Atlantic states. Every regulated power plant must surrender one allowance for each ton of CO₂ it emits, and the total number of allowances is capped. When the Association buys an allowance and retires it, that allowance leaves the supply available to power plants and the cap effectively drops by one ton.

We chose this method for three reasons:

• 1.Retiring an allowance keeps a ton of CO₂ from being emitted in the first place, rather than compensating for a ton already released.
• 2.Allowance retirement sidesteps the verification problems that weaken many offsets, because a retired allowance is a government-issued, serial-numbered instrument.
• 3.The mechanism acts on the system rather than around it, tightening the cap and raising the cost of emitting for regulated polluters.

Water method

Returning Water to Colorado Rivers, through the Colorado Water Trust's myRiver Balance

myRiver Balance is a program of the Colorado Water Trust, a Colorado nonprofit that has restored water to rivers since 2001. A contributor gives based on the volume of water they want to return, and the Water Trust uses the contribution to purchase or lease water and put it back into Colorado rivers and streams as environmental flows — on the Western Slope where possible. At the Water Trust's published rate, a $15 contribution returns roughly 20,000 gallons.

Water is local in a way carbon is not. A gallon returned to a Colorado river does not cancel a specific gallon withdrawn by a data center in another state. So we choose to return an equal or greater volume of water to the rivers and watersheds of the region we most directly serve.

The Public Record

This only works if the record can stand up to outside scrutiny. Every figure on this page is a bounded estimate, not a precise measurement — so we state the bounds and buy above the estimate, so the commitment holds even if the true footprint runs somewhat higher.

For each quarter we keep a running, shareable record of the token estimate, the assumptions behind it, the allowance serial numbers from the Carbon Lighthouse Association, and the myRiver Balance contribution receipts.

Part Two

Holding Partner Projects to the Same Standard

For every project we run with a partner, we estimate the project's AI use and the cost to at least neutralize its carbon and water footprint through these same two mechanisms — either as a line item in the project budget, or as a direct contribution where a budget line isn't available.

The estimate and the purchase records stay with the project, so a partner can see what their project's AI use cost and account for it. Over several projects this builds a documented practice we can point to when we advise rural institutions on adopting AI.