Where we stand on AI and natural resources
:quality(80))
In 2025, AI was responsible for roughly a third of all data center electricity use worldwide. That's around 0.5% of global electricity consumption, according to the IEA. That share is set to grow fast: the agency projects data center consumption overall will climb from roughly 485 TWh today to around 945 TWh by 2030, a bit more than Japan's entire annual electricity use — with AI-focused data centers driving about two-thirds of that growth. [1]
Water tells a similar story: researchers estimate AI systems specifically could account for somewhere between 312.5 and 764.6 billion liters (83 to 202 billion gallons) of water globally in 2025 — from roughly the world's annual bottled water consumption to more than double it. [2]
Without fuller disclosure from data center operators, no one can pin down the exact global number with confidence. But still, gulp.
Like every industry, AI isn't environmentally neutral. But that doesn't mean it has no place in a sustainable future.
Even by 2030, AI-focused data centers are projected to account for about 1.5% of global electricity, [1] and data centers overall will add less new demand between now and then than air conditioning or electric vehicles. [3] At the level of a single request, the IEA reports that a simple AI text query now typically uses less electricity than running a television for the same stretch of time. [1]
When people talk about AI's environmental footprint, electricity use usually comes up first, which makes sense. But not all the electricity data centers consume comes from fossil fuels.
Today, roughly 27% of the electricity used by data centers worldwide already comes from renewable sources (solar, wind, hydro), a share expected to keep growing. A recent IEA report estimates that close to half of the future growth in data center electricity demand could be met by renewables. [4]
The technology has scaled faster than the industry's ability to manage water responsibly, but solutions are emerging fast.
Nvidia, for instance, is developing hybrid direct-to-chip and immersion cooling as part of a federally backed US effort to re-engineer data center cooling. [5] Both keep coolant in a closed loop, recirculating it instead of evaporating water, the design Microsoft says its next-generation AI data centers now use for zero-water cooling. [6]
AI isn't just a new digital tool. It's a technological shift on the scale of the arrival of the internet.
The real question is: what is the energy we're using actually going toward? Is it worth the water and electricity to help a struggling student catch up in math? For us, the answer is yes.
At Ed.ai, we believe the future isn't a choice between rejecting AI outright and adopting it blindly. It's about finding uses that are proportionate and beneficial.
That's why, when we build our tools, we ask ourselves the same questions every time. Does this AI actually help teachers support their students? Does it advance inclusion? Does it deliver real pedagogical value?
The goal is using the right AI, for a specific need. But we want to push further.
Grading a handwritten assignment and turning it into individualized feedback is complex. Behind one result, hundreds of tasks run in the background. These tasks are like a stack of walnuts, and right now, the only tool we can reach to break them is a jackhammer.
The most capable AI models available today, what the industry calls frontier models, are built to do almost everything, from reasoning through quantum physics to summarizing a book. As we develop the product, these models are extraordinarily useful: they let us solve a problem before building the perfect tool for it. Yet they are oversized for many of the jobs we ask them to do.
We have already started replacing some of them.
When Ed.ai began, we used a relatively large, energy-hungry model to process photographs of handwritten assignments. It worked, but it was a jackhammer.
So we built a nutcracker: a much smaller model for that one task, under 1 MB, running directly on a smartphone instead of a data center.
Every time a frontier model performs a specific task for us, we learn more about what that task requires. Enough passes, and a leaner model that does one job extremely well comes within reach.
That is where our economic interest and our environmental interest line up.
A general model charges for the raw compute it consumes and the premium the market puts on intelligence that broad. A specialized model costs less because it draws less. And it doesn't always need a big data center to run.
We believe the answer is to build AI that is appropriately sized for the work it does. The scanning model was the first tool we traded. It won't be the last.
Like anything new, AI is unsettling, especially when the numbers show up with no decoder.
AI's environmental footprint is still poorly documented, and it tends to get inflated by quick comparisons or alarmist projections. It's not a myth, and it's not the apocalypse either: it's a complex subject still taking shape.
And we believe the impact depends less on the machines than on what we choose to do with them.
Sources Cited
[1] International Energy Agency, "Key Questions on Energy and AI," Executive Summary (2026). iea.org/reports/key-questions-on-energy-and-ai/executive-summary
[2] de Vries-Gao, A. (2025). "The carbon and water footprints of data centers and what this could mean for artificial intelligence." Patterns, 7, 101430. doi.org/10.1016/j.patter.2025.101430
[3] Carbon Brief, "AI: Five charts that put data-centre energy use – and emissions – into context" (2025), based on IEA data. carbonbrief.org/ai-five-charts-that-put-data-centre-energy-use-and-emissions-into-context
[4] International Energy Agency, "Energy supply for AI," Energy and AI (2025). iea.org/reports/energy-and-ai/energy-supply-for-ai
[5] Nvidia's hybrid direct-to-chip and immersion cooling system, funded under the US Department of Energy's ARPA-E COOLERCHIPS program (cited via Data Center Frontier, 2023).
[6] Microsoft, "Inside Microsoft’s two-decade push to cut water intensity while scaling for growth," The Official Microsoft Blog (June 24, 2026). blogs.microsoft.com/blog/2026/06/24/inside-microsofts-two-decade-push-to-cut-water-intensity-while-scaling-for-growth