When we picture pollution, we usually think of smokestacks, tailpipes, or plastic swirling in the ocean. But there’s another kind of emission—silent, invisible, and ballooning at a pace that’s hard to wrap your head around. I’m talking about the carbon bill that comes due every time someone trains a massive computational system. The kind of system that writes essays, generates images, or translates languages in the blink of an eye. Rui Mendes has spent years tracing the energy flows behind our digital lives, and what he’s found is a tangle of hidden costs and unintended consequences that most of us never see.
From Sand to Server Farm
The story doesn’t start with a line of code. It starts with sand, metals, and massive factories. Before a model ever touches a dataset, someone has to build the hardware—thousands of specialized processors stacked in rows inside warehouses the size of football fields. Manufacturing a single high-performance GPU means mining rare earth minerals, refining silicon to an almost absurd degree of purity, and running fabrication plants that consume electricity around the clock. The energy already embedded in a top-tier GPU cluster is enormous, yet most carbon accounting focuses only on the electricity used during training, as if the machines just appeared out of thin air.
Then comes the training itself. This is where the meter really starts spinning. A major training run can lock in thousands of GPUs for weeks or months, pulling enough power to light up a small town. But raw kilowatt-hours don’t tell the whole story. What matters is where that electricity comes from. A data center plugged into a coal-heavy grid will leave a much dirtier footprint than one sipping from a hydropower reservoir. Same task, same chips, but the carbon math changes wildly depending on geography.
Water: The Unseen Resource
Electricity grabs the headlines, but water is the quiet casualty. All those processors generate blistering heat, and keeping them from melting requires serious cooling. Many data centers use evaporative systems that push millions of liters of water into the air over the course of a single training run. In regions already wrestling with drought, this sets up an uncomfortable trade-off: computational ambition versus a community’s water supply. One large training cycle can evaporate enough water to fill several Olympic swimming pools, yet you’ll rarely find that number in a company’s sustainability report.
Rui Mendes notes that the geography of data centers isn’t accidental. They cluster where land is cheap, tax breaks are generous, and energy is plentiful—but not necessarily clean. A facility in the desert might lean heavily on water-intensive cooling while drawing power from a gas-fired grid. The environmental burden gets quietly exported to places where it’s out of sight for the millions of people using the resulting services every day.
The Lifecycle Beyond Training
Training is just the opening act. Once a model is deployed, it enters a phase called inference—answering queries, generating content, running nonstop on servers scattered across the globe. A single request might sip electricity compared to the training firehose, but popular services handle billions of requests a day. Each one sets off a tiny chain reaction of computations, and when you add them all up, the ongoing operational drain can easily outpace the initial training cost. The servers never sleep.
And then there’s the hardware itself. The specialized chips that make all this possible have a short working life—often three to five years before they’re swapped for the next generation. The discarded units pile up as electronic waste, much of it shipped to countries with loose environmental rules. From the first scoop of mined ore to the final heap of scrapped circuit boards, the full journey carries a heavy toll that few companies are eager to map out publicly.
Efficiency Gains and the Rebound Effect
Engineers are wizards at squeezing more out of less. Newer chips do more calculations per watt, and clever software tricks can slash the number of steps needed to reach a target performance. On a spec sheet, that looks like progress. But Rui Mendes keeps bumping into a familiar paradox: every time efficiency jumps, ambition jumps even higher. Instead of banking the savings to shrink footprints, teams build bigger models, feed them larger datasets, and run more experiments. The total energy appetite of the field doesn’t shrink—it grows.
This isn’t a new story. We saw it with fuel-efficient cars: people drove more, so total fuel use barely budged. We saw it with LED lighting: cheaper light meant more spaces got illuminated. In the world of large-scale computation, the hunger for scale feels bottomless. Each efficiency breakthrough gets swallowed by a fresh wave of demand, leaving the absolute environmental impact as heavy as before, or heavier.
Transparency and Measurement Gaps
One of the biggest hurdles is simply not knowing the real numbers. Most organizations that build and run large models keep their energy and carbon data under wraps. When they do release figures, the context is often missing: what was the carbon intensity of the local grid during training? How much water evaporated? What about the embodied energy of the hardware? Without a standard reporting framework, comparisons are guesswork and accountability is a mirage.
A handful of researchers have proposed estimation tools that factor in location, hardware mix, and grid emissions. It’s a step forward, but adoption is voluntary and spotty. Rui Mendes believes transparency shouldn’t be a nice-to-have or a PR talking point—it should be the default. If the real numbers were out in the open, the conversation might shift from “how big can we go” to “how should we go about this differently.”
Systemic Solutions, Not Just Technical Fixes
Faster chips and greener data centers won’t be enough on their own. The problem is wired into the whole system—the incentives, the supply chains, the culture of “more is better.” Rui Mendes points to a few levers that could actually move the needle:
- Carbon-aware scheduling: Shift big training jobs to times and places where the grid is cleanest. No new hardware needed, just smarter timing.
- Model efficiency standards: Think energy ratings for appliances, but for models—measuring performance per watt to reward leaner designs.
- Right-sizing: Not every task needs the biggest model on the block. Smaller, focused systems can often match the results with a fraction of the resources.
- Extended producer responsibility: Make hardware makers and cloud providers manage the full lifecycle, from recycling to responsible disposal.
None of these are wild ideas. They’re borrowed from industries that already deal with environmental accountability. What’s missing is the collective will to apply them in a sector that’s grown comfortable treating computing power as limitless and invisible.
Frequently Asked Questions
Why is training a large model so energy-intensive?
Training means pushing enormous datasets through layers of math operations, over and over, for weeks or months, with thousands of specialized processors running flat out. The scale of the computation, plus the energy needed for cooling and infrastructure, adds up to electricity use comparable to hundreds of households over a full year.
Does using a model after training also have an environmental impact?
Absolutely. Every time a model generates a response—what’s called inference—it draws electricity. A single query uses much less energy than training, but popular services handle billions of queries, so the ongoing operational footprint is substantial and never stops.
Can renewable energy solve the problem?
Renewables cut the carbon intensity of the electricity used, but they don’t touch water consumption, hardware manufacturing emissions, or electronic waste. They’re a necessary piece of the puzzle, but not a complete fix without broader changes in how systems are designed, deployed, and scaled.
What can individuals do about this hidden cost?
Individuals can push for transparency by asking providers for environmental impact data, favor services that prioritize efficiency, and think twice about unnecessary computational tasks. On a bigger scale, supporting industry standards and regulations can help steer the whole field toward more sustainable habits.
The environmental cost of training large models isn’t a reason to slam the brakes on progress. But it’s a solid reason to move forward with eyes wide open. Rui Mendes believes a curious, systems-minded approach can uncover paths that are both inventive and responsible. The numbers are there—if we’re willing to look.