Every time someone types a prompt into ChatGPT, a server somewhere draws power, pulls water from a cooling system, and adds a fraction of a gram of CO₂ to the atmosphere. Multiply that by billions of daily queries, and you get a problem the AI industry spent three years pretending wasn’t there.
The numbers are landing now. Alex de Vries-Gao’s research at Vrije Universiteit Amsterdam puts AI’s carbon footprint at somewhere between 32.6 and 79.7 million tonnes of CO₂ for 2025 — a range that wide because the industry still won’t disclose basic operational metrics. The low end already rivals New York City’s annual emissions. The high end exceeds it.
Training Is the Headline. Inference Is the Problem.
Here’s what most coverage gets wrong: the conversation about AI’s footprint tends to fixate on model training — the dramatic, one-time energy expenditure that produces a GPT or a Gemini. GPT-4’s training run consumed roughly 50 gigawatt-hours. That’s a lot. It’s also a one-time event.
Inference is different. Inference is every response, every image generated, every autocomplete, running continuously across thousands of servers, every hour of every day. MIT Technology Review’s analysis found that 80 to 90 percent of AI computing power now goes to inference, not training. McKinsey projects that inference will grow at a 35 percent compound annual rate through 2030, reaching over 90 gigawatts of data center capacity. The training conversation is about a match being struck. The inference conversation is about a furnace that never turns off.
This distinction matters for how we regulate and report AI’s environmental footprint. Training happens once. Inference scales with every new user, every new product integration, every API call buried inside an app you didn’t know was using AI.
The Ghost Hardware Problem
There’s a second crisis nobody is talking about loudly enough: e-waste.
The AI hardware cycle moves fast. H100s are already being displaced by B200s. Entire server racks cycle out on timelines measured in years, not decades. De Vries-Gao’s newest research, published in Resources, Conservation and Recycling in early 2026, revised earlier e-waste projections downward — AI servers may generate 131 to 224.8 kilotonnes of waste annually by 2030, not the five million tonnes some alarmist estimates suggested. That’s the good news. The less good news: 224 kilotonnes of discarded AI hardware still matches the annual e-waste output of Norway or Denmark. Countries, not companies. And only about 17 percent of global e-waste is formally recycled.
The Jevons Paradox is lurking here, too. NVIDIA’s Blackwell architecture is significantly more energy-efficient per computation than the H100 generation it’s replacing. The industry treats that as a win. It isn’t — not automatically. When efficiency drops, costs and demand rise to absorb the savings. More efficient chips running more queries at lower marginal cost could easily produce higher total energy consumption than the less-efficient systems they replaced. This pattern held in every prior computing generation. There’s no particular reason AI breaks the trend.
Google’s Inconvenient Report Card
Google at least publishes numbers, which is more than most. Its 2024 environmental report confirmed a 13 percent year-over-year rise in total greenhouse gas emissions, with data centre energy consumption listed as the primary driver. The company that built its brand on “don’t be evil” and once claimed 100 percent renewable energy coverage is now one of the more visible examples of AI’s accelerating environmental cost.
To be fair, the same Brookings analysis found Google reduced the median energy consumption per Gemini prompt by a factor of 33 between May 2024 and May 2025 — driven by model quantization, custom TPU chips, and more efficient architectures. That’s real progress. The problem is that the denominator keeps growing. Per-query efficiency gains that get swallowed by volume increases aren’t net improvements; they’re treadmills.
A Roadmap That Works — If Anyone Uses It
Cornell’s November 2025 study in Nature Sustainability didn’t just project bad outcomes — it published a genuine pathway out. Smart geographic siting (co-locating data centres with low-carbon grids), faster renewable integration at AI build-out sites, and aggressive cooling efficiency improvements could cut worst-case carbon projections by 73 percent and water projections by 86 percent. The math is tractable. Current AI growth trajectories, absent those interventions, push net-zero targets out of reach by 2030.
The regulatory picture is sharpening, too. The EU AI Act now includes environmental disclosure requirements. Arizona and other US states with significant data centre footprints are beginning to explore water regulation specific to AI infrastructure. The Green AI Institute’s Green AI Index — a standardised metric set covering carbon, energy, and water — launched in late 2024 specifically because voluntary disclosure wasn’t working.
What to Watch in 2026
The real signal this year isn’t which model is most capable. It’s whether the companies building and deploying AI start treating environmental metrics with the same rigour they apply to benchmark leaderboards.
If inference is now 80 to 90 percent of the footprint, and inference scales with product adoption, then every new AI feature in every new app is an environmental decision whether its builders acknowledge it or not. The bill keeps accumulating. The question is who’s keeping the ledger.
Related: AI’s Hidden Heat Map: How Data Centers Are Quietly Raising Temperatures Around You
