When I first started digging into the energy demands of modern computing, I expected big numbers. But I didn’t expect them to feel so tangible. Training a single massive model can draw as much electricity as a small town uses in a month. That’s not a metaphor—it’s a measurable, billable quantity of kilowatt-hours pulled from a grid that might still be burning coal or gas. And yet, most conversations about these systems fixate on their smarts, not their appetite.

I’m not here to argue against progress. I’m here because I’m curious about the systems that make these models possible, and what they cost beyond the sticker price. Every training run has a supply chain: rare earth minerals clawed from the earth, water evaporated in cooling towers, carbon pumped into the sky. If we’re going to build and use these things, we should at least know the bill.

What Does “Training” Really Consume?

Training a massive model means running tens of thousands of specialized chips—GPUs or TPUs—at full throttle for days, sometimes weeks. A single high-end GPU can guzzle 300–400 watts. Multiply that by 10,000 units, add the networking gear, the storage, the cooling, and you’re looking at a constant megawatt-level draw. That’s just the electricity you can see on a meter.

But there’s a deeper, dirtier story. Before a server ever lights up, it’s already racked up an environmental debt. The silicon, copper, gold, and rare earths inside it were mined, refined, and shipped across the globe. A GPU’s life starts in a pit mine, not a data center. And after three to five years of service, it often ends up as e-waste, dismantled in places with few environmental protections.

Carbon Footprints: More Than Just a Number

When researchers tally the carbon footprint of a training run, they usually multiply the energy used by the carbon intensity of the local grid. That gives you a figure in tonnes of CO₂. But the real number can swing wildly depending on location. A run in Quebec, with its hydro-heavy grid, looks much cleaner than the same run in a region powered by coal.

Some companies buy renewable energy certificates to offset their consumption. But that’s often an accounting trick, not a physical solution. The electrons flowing into the servers still come from whatever mix the grid is using at that moment. If it’s a coal-heavy night, the carbon is still being emitted—just somewhere else on the balance sheet. Real transparency would mean publishing the actual grid mix during training, not just the annual average.

Water: The Overlooked Thirst

Electricity grabs the headlines, but water is just as critical. Data centers use it for cooling—either by evaporating it directly or through the water-intensive process of generating electricity. One 2021 study estimated that training a single large model can consume hundreds of thousands of liters of water, depending on the cooling setup and energy mix.

In places already struggling with water scarcity, this creates real friction. Data centers compete with farms, households, and ecosystems. And unlike carbon, which spreads globally, water use is intensely local. A data center in Arizona has a very different impact than one in Norway.

Direct vs. Indirect Water Use

Direct water use is what flows through the cooling towers. Indirect water use is what’s consumed at the power plant to generate the electricity. Often, the indirect footprint is larger, especially if the grid leans on thermoelectric plants—coal, nuclear, natural gas—that use steam turbines. So even a data center with air cooling still has a significant water footprint through its power draw.

Hardware Lifecycle: From Mine to Landfill

The GPUs and accelerators that power training runs are engineering marvels. But they carry a heavy material backpack. Manufacturing a single GPU requires mining and processing dozens of minerals: silicon, copper, gold, tantalum, rare earths. The extraction is energy-hungry and often involves toxic chemicals that can poison local water sources.

Then there’s the end of the line. Hardware moves fast. A GPU that was top-of-the-line three years ago might be too slow or memory-starved for today’s biggest training jobs. E-waste is the fastest-growing waste stream on the planet, and only a fraction gets properly recycled. The rest piles up in landfills or gets processed in informal yards, where workers handle hazardous materials with little protection.

Why Scale Changes the Game

It’s not just that models are getting bigger. The whole approach to training has shifted toward scale. A decade ago, a top-tier model might train on a single GPU in a few hours. Now, the largest runs use tens of thousands of accelerators for months. Energy consumption scales roughly with the number of processors and training time, but the trend is superlinear: each new generation isn’t just larger, it’s trained longer on more hardware.

This creates a compounding effect. A model twice as large might need four times the compute to train, because it also needs more data and more steps to converge. And if the hardware is replaced more often to keep up with the latest chips, the embodied energy and e-waste grow right alongside.

Where’s the Transparency?

One of the most maddening parts of this topic is the lack of public data. Most companies don’t disclose the energy, carbon, or water use of their training runs. When they do, the numbers are often aggregated or presented in ways that make comparisons a headache. Without standardized reporting, it’s hard to know if the industry is getting more efficient or just better at hiding its footprint.

There have been efforts to create reporting standards—like the ML CO2 Impact calculator and various research papers estimating the carbon footprint of specific models. But these are voluntary and often rely on assumptions that may not hold. For example, the carbon intensity of the grid is usually taken as a regional average, not the actual marginal emissions at the time of training.

Efficiency Gains vs. Rebound Effects

Hardware gets more efficient every year. Today’s GPUs do more operations per watt than those from five years ago. But efficiency gains don’t always mean lower total energy use. Often, they just enable bigger models and more training runs, which can push overall consumption higher—a classic rebound effect.

There’s also the question of utilization. A data center running at full capacity is more energy-efficient per operation than one at half capacity. But many training clusters are overprovisioned, meaning they have more hardware than is constantly in use. The idle hardware still draws power, even if it’s not doing useful work. This “stranded capacity” rarely shows up in efficiency claims.

What About Inference?

Training gets the headlines, but inference—the process of using a trained model to make predictions—can dominate the total energy cost over a model’s lifetime. A model trained once but used millions of times will have a much larger operational footprint than its training footprint. For popular services, inference can account for the vast majority of energy use.

This shifts the environmental math. If a model is trained inefficiently but used efficiently, the total impact might still be acceptable. But if a model is deployed at massive scale with little regard for energy optimization, the ongoing cost can be enormous. Unfortunately, inference energy use is even less transparent than training energy use.

Geographic Disparities and Environmental Justice

The environmental costs of training large models aren’t spread evenly. Data centers often land in regions with cheap electricity and loose environmental rules. That means the communities hosting these facilities bear the brunt of the pollution—whether from coal-fired power plants or diesel backup generators—while the benefits flow elsewhere.

Water use is another justice issue. In arid regions, data centers can compete with local communities for scarce water. Some facilities use evaporative cooling, which consumes water directly. Others pull electricity from thermoelectric plants, which also consume water. Either way, the local impact can be severe, especially during droughts.

What Can Be Done?

There’s no single fix, but a mix of approaches could help. First, transparency: companies should report the energy, carbon, and water footprints of their training runs using standardized metrics. That would let researchers and the public compare models and hold developers accountable.

Second, efficiency: better hardware, smarter algorithms, and improved data center design can reduce the resources needed. Techniques like model pruning, quantization, and knowledge distillation can make models smaller and faster without sacrificing much accuracy. And training can be scheduled to take advantage of times when the grid is cleaner.

Third, location matters. Siting data centers in regions with clean energy and abundant water can dramatically reduce the environmental impact. Some companies are already doing this, but it’s not yet the norm.

Frequently Asked Questions

How much energy does it take to train a large model?

Estimates vary, but training a single large model can consume hundreds of megawatt-hours of electricity—enough to power dozens of homes for a year. The exact number depends on the model size, hardware efficiency, and training duration. Some studies have put the carbon footprint at over 250 tonnes of CO₂ equivalent, comparable to the lifetime emissions of several cars.

Does using renewable energy solve the problem?

Renewable energy helps, but it’s not a complete solution. Even if a data center is powered by solar or wind, the manufacturing of the hardware still has a significant carbon and material footprint. And if the renewable energy is purchased through certificates rather than directly consumed, the actual grid mix at the time of use may still include fossil fuels. It’s a step in the right direction, but not a cure-all.

Why isn’t this talked about more?

The environmental costs of training large models are often hidden behind corporate sustainability reports that aggregate data across all operations. The specific impact of a single training run is rarely disclosed. There’s also a tendency to focus on the benefits of the technology rather than its costs. But as these models become more widespread, the conversation is starting to shift.

Can smaller models be just as good?

In many cases, yes. Research has shown that smaller, well-optimized models can match or even outperform larger ones on specific tasks. The key is to invest in efficient architectures and training techniques rather than simply scaling up. This not only reduces environmental impact but also makes the technology more accessible to those without massive computing resources.

The environmental cost of training large models isn’t a reason to stop building them. But it is a reason to build them thoughtfully, with full awareness of the trade-offs. The systems that power these models—electrical grids, water supplies, mineral supply chains—are not infinite. They are shared resources, and how we use them says a lot about what we value.