Training a single AI model last year generated roughly as much carbon as 16,000 round-trip flights from San Francisco to New York. Running GPT-4o may consume enough water annually to meet the drinking needs of every person in Los Angeles and San Francisco combined. We talk endlessly about AI ethics in terms of bias and jobs, but the conversation almost never turns to the water bill. The numbers are getting harder to ignore. And they raise a question that the industry would rather not answer: what is the true environmental cost of intelligence on demand?

The numbers nobody wants to talk about

Stanford's 2026 AI Index report paints a stark picture. Grok 4's estimated training emissions reached 72,816 tons of CO₂ equivalent, roughly the same greenhouse gas output as driving 17,000 cars for a year. AI data center power capacity has risen to 29.6 gigawatts, comparable to what it takes to power the entire state of New York at peak demand. The cumulative power demand of all AI systems now rivals the national electricity consumption of Switzerland or Austria. And the trajectory is accelerating, not flattening. Recent models show a dramatic increase in training emissions over time: GPT-3 in 2020 produced approximately 588 tons of CO₂, GPT-4 in 2023 reached about 5,184 tons, and Llama 3.1 405B in 2024 hit roughly 8,930 tons. The latest frontier models dwarf all of these. Water is the other hidden cost. A study published in Cell Reports Sustainability estimated that the carbon footprint of AI systems alone could be between 32.6 and 79.7 million tons of CO₂ emissions in 2025, while the water footprint could reach 312.5 to 764.6 billion liters. For perspective, that water figure is in the same range as the entire global annual consumption of bottled water. North American data centers used nearly 1 trillion liters of water in 2025, roughly equivalent to the annual demands of New York City. U.S. data centers now account for about 4.4% of national electricity consumption, up from 1.9% in 2018. By 2028, that number could climb to 12%.

The sustainability report paradox

Here's what makes this uncomfortable: the same companies publishing glossy sustainability reports are the ones scaling compute exponentially. Alphabet's 2026 capital expenditure guidance is $175 to $185 billion, more than double what it spent in 2025. Amazon is guiding close to $200 billion. Meta sits in the $115 to $135 billion range. Microsoft runs at roughly $145 billion annualized. Combined, the four hyperscalers expect spending approaching $700 billion in 2026 alone. Meta's 2025 environmental report showed water usage at sites it owned rose 51%, from 3,726 megaliters in 2020 to 5,637 megaliters in 2024. And that figure doesn't even include leased facilities or those under construction. Meanwhile, investors are starting to push back. Reuters reported in April 2026 that shareholders are pressing Amazon, Microsoft, and Google for more detailed data on water and power use in their data centers. The disconnect between sustainability pledges and infrastructure spending is becoming too visible to paper over.

The invisibility problem

The environmental cost of AI is invisible because it's distributed. A data center in Iowa uses the water. A user in Singapore gets the ChatGPT response. Out of sight, out of mind. Large data centers can consume up to 5 million gallons of water per day, equivalent to the water use of a town of 10,000 to 50,000 people. But the person asking an AI to summarize a meeting or generate an image never sees that. There's no label on the response that says "this answer cost 0.26 milliliters of water and 0.24 watt-hours of energy," even though Google has now measured exactly those figures for a median Gemini text prompt. Stanford's Foundation Model Transparency Index found that AI companies now average just 40 out of 100 on transparency, a significant decline from the previous year. Very few jurisdictions mandate data center or AI-specific energy or environmental transparency, creating what researchers at a recent study called "a global blind spot." The Artificial Intelligence Environmental Impacts Act, introduced in the U.S. Congress, proposed requiring the EPA to study environmental impacts and NIST to develop a voluntary reporting system. But voluntary reporting for an industry spending $700 billion on infrastructure is like asking the fox to count the chickens.

Right-sizing the model to the task

The answer isn't to stop using AI. That ship sailed a long time ago. The answer is to build and use AI more responsibly. Smaller models are a genuine path forward. UNESCO has recognized small language models as a "cheaper, greener route into AI." They require less processing power, use less energy, can run on local devices, and may not need a cloud connection at all. Training a model with fewer than 10 billion parameters uses a fraction of the computational power required by frontier models, with corresponding reductions in carbon and water. The efficiency gains are real when companies prioritize them. Google reported a 33x reduction in energy and 44x reduction in carbon for the median Gemini prompt compared to 2024. The Stanford AI Index noted that carbon emissions from models with the least efficient inference are over 10 times as high as those with the most efficient inference. DeepSeek's V3 models consume around 23 watts per medium-length prompt response, while Claude 4 Opus uses about 5 watts. The gap between efficient and inefficient is enormous. The "one agent, one job" philosophy, matching model capability to task complexity, isn't just a cost optimization. It's an environmental one. There's no reason to route a simple text classification through a trillion-parameter model when a smaller, fine-tuned model will do the job with a fraction of the energy.

What individual developers can do

If you're building with AI, you have more leverage than you think:

Researchers at Cornell found that locating data centers in regions with lower water stress and improving cooling efficiency could slash water demands by about 52%. Combined with grid and operational best practices, total water reductions could reach 86%. Where AI runs matters as much as how it runs.

The bigger question

Should AI companies be required to disclose environmental costs per API call? Google has already done the math internally. They know a median Gemini text prompt uses 0.24 watt-hours of energy, emits 0.03 grams of CO₂ equivalent, and consumes 0.26 milliliters of water. The measurement methodology exists. The question is whether the rest of the industry will follow, voluntarily or otherwise. Other competitive industries manage transparency just fine. Food has ingredient labeling. Healthcare has side-effect disclosures. Cars have fuel economy ratings. There's no fundamental reason AI can't have an equivalent: a carbon and water cost per inference, disclosed publicly, enabling users and businesses to make informed choices. The AI industry has positioned itself as the great optimizer, capable of solving everything from climate modeling to drug discovery. But an industry that can't optimize its own environmental footprint has a credibility problem. And at $700 billion in annual infrastructure spending with carbon emissions rivaling New York City, the footprint is no longer small enough to wave away. The numbers are in. The question is whether we'll read them.

References