Data Centers With a Conscience: How an AI Company Should Be Held Accountable (Starting With Us)

Accountability Starts With Us

Here is what accountability looks like for an AI company that serves food brands.

Use the smallest model that works. Ingredient matching, nutrition calculation and compliance checks are structured problems. Research by Luccioni and colleagues found general-purpose generative models can use far more energy per task than task-specific ones, roughly 30 times more for extractive question answering. Mistral AI's lifecycle study found impacts scale roughly with model size: a model 10 times bigger has about 10 times the impact for the same output.
Design for efficiency. Keep ingredient data structured, reuse work instead of recomputing it and save the heavyweight models for the problems that truly need them.
Measure what we can, and say what we cannot. Shared benchmarks like the AI Energy Score from Hugging Face and Salesforce rate model efficiency on a 1 to 5 star scale. Public yardsticks like this are how the industry gets honest.
Invest in carbon removal. Journey contributes a share of revenue through Stripe Climate, which directs contributions to durable carbon removal through Frontier. Frontier is an advance market commitment that pools demand from buyers including Stripe, Google, Shopify, Salesforce and JPMorgan Chase to help early carbon removal technologies scale. Carbon removal does not replace efficiency. It is one piece of a responsible stack.
Stay accountable to the research. We read the science on what data centers mean for the bodies and farms next door, and we let it shape our choices. More on that below.

The Footprint Is Real, and Manageable

Lawrence Berkeley National Laboratory estimates US data centers used 176 TWh in 2023, about 4.4% of national electricity, with 6.7% to 12% projected by 2028. Globally, the International Energy Agency puts data center use at about 415 TWh in 2024, roughly 1.5% of world electricity.

Per task, the numbers can be small. Google reported that a median Gemini Apps text prompt uses 0.24 Wh of energy and about 0.26 mL of water, roughly five drops, based on May 2025 data in its technical paper. Mistral estimated 1.14 g CO2e and 45 mL of water for a 400-token response, including upstream impacts like server manufacturing. The methods differ, which is exactly why transparency matters. Small numbers add up at scale, and they land in specific places.

Listening to the Neighbors

Those places often grow food. A 2026 study in Environmental Science & Technology of 9,347 data centers worldwide found 28% overlap with irrigated agriculture water demand. The National Wildlife Federation notes that about two-thirds of planned US data centers are proposed in rural areas, where rezoning can raise land costs for farmers.

Air matters too. A UC Riverside and Caltech study estimated that pollution from data center power, including backup diesel generators, could cause up to 1,300 premature US deaths a year by 2030. Communities are paying attention: in a June 2026 New York Times column, sociologist Tressie McMillan Cottom described rural residents in Utah and North Carolina organizing to stop data center construction.

For food brands, a sourcing region's water table, air quality and land prices are part of ingredient risk. For an AI company, they are part of the job.

What Smart Data Center Building Looks Like

Cooling is where most of the water goes. Many facilities use evaporative cooling towers, which save electricity but lose water to the air. Closed-loop and chip-level liquid systems recirculate the same water. The standard yardstick is Water Usage Effectiveness (WUE), defined by The Green Grid as annual site water use divided by IT equipment energy, in liters per kilowatt-hour. Its energy cousin, PUE, divides total facility energy by IT energy.

Less water. Microsoft says all its new data center designs since August 2024 evaporate zero water for cooling, avoiding more than 125 million liters a year per site with what it calls a nominal increase in energy. Pilots start in 2026.
Smarter water. Google's Douglas County, Georgia site cools with recycled wastewater instead of drinking water.
AI that runs the building. DeepMind's machine learning cut energy used for cooling at Google data centers by up to 40%. AI measuring and tuning its own infrastructure is AI done right.
Honest numbers. Per-prompt disclosures from Google and Mistral show the industry can publish its footprint.

Disclosure Is Raising the Bar

European Union. Under the Energy Efficiency Directive, data centers of 500 kW or more report energy and water data each year, including total and potable water input, under Delegated Regulation 2024/1364.
California. AB 2469 and AB 2619, signed September 21, 2026, take effect January 1, 2027, requiring water supply assessments before permits and annual water reporting at business license renewal.
Minnesota. HF 16 (2025) adds state review for data centers proposing more than 100 million gallons of water use a year.
Local voices. Tucson's council unanimously ended Project Blue in August 2025 after residents raised water concerns.

We see this as good news. Clear rules reward the builders doing it right.

Ask Us These Too, and Ask Every AI Vendor the Same

Bring these to your next AI review, starting with us:

Where does our workload run? Regions and cloud providers.
What are the WUE and PUE of those facilities? Reporting year, and site-level or fleet average.
What cooling do they use? Evaporative, closed-loop or air, and potable, reclaimed or recycled water.
What is the energy and water per task? The method, the boundary and whether it is third-party reviewed.
What model size runs our use case? And whether a smaller model was tested.
How is backup and onsite power handled? Diesel, gas or batteries, and whether permits are in place.
What do you invest in beyond efficiency? Clean power, durable carbon removal, community commitments.

Good partners will welcome the questions. For a more personal take on bodies, neighborhoods and founders, read Every Cloud Has a Water Bill: Data Centers, Our Bodies, and the Farms Next Door.

FAQs

How much water does an AI query use?
It depends on the model, data center and method. Published estimates range from Google's 0.26 mL per median Gemini prompt to Mistral's 45 mL per 400-token response, which includes upstream impacts. Always ask how a figure was measured.

What is WUE?
Water Usage Effectiveness is annual site water use divided by IT equipment energy, in liters per kWh. Lower is better, but check whether it is site-specific or a fleet average.

Can AI help a brand fix the recipe without a big footprint?
Yes. Right-sized models working on structured ingredient data can handle nutrition, cost and compliance checks with far less compute than a general chatbot.

Is carbon removal a substitute for efficiency?
No. Efficient models and responsible siting come first. Durable carbon removal addresses emissions that remain.

Want to fix the recipe with AI that is built to be accountable? See how Journey AI works.

Sources

About the Author

Journey Foods

Journey Foods Team News