When we think about digital pollution, we usually picture mountains of discarded smartphones or the electricity that keeps endless video streams flowing. But Rui Mendes, a systems thinker drawn to the invisible costs of modern infrastructure, has been looking at a different kind of environmental toll: the staggering amount of energy it takes to train a single, massive neural network from scratch. What he found is a story of resource consumption that quietly challenges our assumptions about the weightlessness of the digital world.

The numbers are out there, but they tend to hide in academic appendices or the footnotes of corporate sustainability reports. Training a single large-scale language system can emit as much carbon dioxide as five average American cars do over their entire lifetimes—including the fuel they burn and the energy used to build them. And that is not a worst-case projection; it is a measurement from models already built and running. The energy isn’t just for the final training sprint, either. It covers countless trial runs, hyperparameter tweaks, and discarded versions that never see the light of day. The real cost is the sum of all those failed attempts, not just the polished result.

Why Training Is So Much Heavier Than Running a Model

To grasp the environmental cost, you have to separate two very different phases: training and inference. Inference is what happens when you ask a question and get an answer back—it is relatively light, often handled by a single chip. Training is the process of building the model in the first place. It means pushing petabytes of data through billions of parameters, over and over, for weeks or months, using thousands of power-hungry processors running in parallel. All that computation generates tremendous heat, which then requires even more energy to remove. It is an industrial process, not a casual one.

Rui points out that the public conversation often lumps these two phases together, which makes the whole thing seem less consequential. A single query might feel ephemeral, but the model that answers it was born from a sustained, energy-intensive burn. Think of training as a one-time carbon debt. Every subsequent query slowly chips away at that debt over the model’s operational life. The uncomfortable question is whether the debt ever gets fully paid off, especially when models are frequently retrained or replaced by even larger successors.

Mapping the Energy Supply Chain

The electricity that feeds a training run does not come from a single, clean source. Its carbon intensity depends entirely on the regional grid mix at the time the computation happens. A training cluster plugged into a grid dominated by coal plants will have a dramatically higher footprint than one drawing from hydroelectric dams or nuclear reactors. Yet data center locations are usually chosen based on tax breaks, land prices, and network latency—not on how clean the local grid is.

Rui sees this as a classic systems problem: the metrics we optimize for are misaligned. A company might proudly announce that its data centers are carbon-neutral thanks to renewable energy certificates, but the physical electrons powering the GPUs still come from the local grid. If that grid is dirty, the training run causes real emissions in real time, even if the accounting looks clean on paper. The gap between when energy is consumed and when renewables are generated is a mismatch that few reports bother to address.

The Lifecycle of a Training Run

To really grasp the scale, you have to look at the physical infrastructure. A state-of-the-art training cluster can contain tens of thousands of interconnected processors. Each one is a marvel of engineering that required mining rare earth minerals, purifying silicon, and shipping components across the globe. The embodied carbon of manufacturing these chips and servers is a significant upfront cost, often comparable to the operational energy used over the hardware’s lifespan. When a new generation of chips arrives every two to three years, the churn of hardware accelerates, locking in even more embodied emissions.

Then there is the cooling. The heat generated by a training cluster is immense. Data centers use chilled water, powerful fans, and sometimes even submersion in dielectric fluids to keep temperatures in check. In arid regions, water consumption for cooling competes with local agriculture and drinking supplies. A single large training run can evaporate millions of liters of water—a hidden cost that rarely makes it into environmental assessments.

The Geography of Computation

Where a training cluster sits dictates its carbon profile. A model trained in Quebec, where electricity is almost entirely hydroelectric, will have a fraction of the emissions of an identical model trained in Virginia, where the grid still leans heavily on natural gas and coal. Yet the decision of where to place these clusters is often opaque. Cloud providers may shift workloads between regions for load balancing, making it difficult for researchers to know, let alone control, the carbon intensity of their experiments.

Rui notes that this geographic lottery creates a strange ethical landscape. A research team in one country might inadvertently produce a much dirtier model than a team in another, simply because of the default region selected in their cloud console. Some researchers have started to advocate for “carbon-aware” computing, where training jobs are scheduled to run when and where the grid is cleanest. But this requires a level of transparency and flexibility that most cloud platforms do not yet offer.

The Dataset Factor

Energy is not the only resource. The datasets used to train these networks are themselves products of energy-intensive processes. Crawling the web, storing petabytes of text and images, cleaning and deduplicating the data, and then shuttling it to the training cluster all consume electricity. The larger the dataset, the more storage and network infrastructure is needed. Some training sets are so vast that they cannot be stored in a single location, requiring distributed file systems that add their own overhead.

There is also a less visible cost: the human labor of data annotation. While not a direct carbon emission, the global supply chain of annotators, often working in conditions with their own environmental and social footprints, is part of the system. Rui’s systems-minded approach insists on seeing the full picture, from the mines that extract the metals for the servers to the offices where labeling guidelines are written.

Measuring What Matters

One of the core challenges is simply measurement. Estimating the carbon footprint of a training run requires knowing the power draw of the specific hardware, the duration of the run, the Power Usage Effectiveness (PUE) of the data center, and the carbon intensity of the grid. Some of these numbers are proprietary. Hardware manufacturers may not disclose the full energy profile of their chips. Cloud providers may not reveal real-time grid mix data. Researchers are left to make rough estimates based on public information, which can vary by an order of magnitude.

Several tools have emerged to help, such as open-source calculators that estimate emissions based on hardware type, cloud region, and runtime. But these tools rely on average grid intensities that may not reflect the actual moment-by-moment mix. A training run that spans weeks will inevitably include periods of high and low carbon intensity, but without real-time data, the true impact remains fuzzy. Rui sees this measurement gap as a fundamental barrier to accountability. Without clear numbers, it is too easy to ignore the problem.

The Efficiency Paradox

There is a common counterargument: as hardware becomes more efficient, the energy per computation drops. This is true. The number of floating-point operations per watt has improved dramatically. But the total energy consumed by training runs has not decreased; it has grown. This is Jevons paradox in action: as efficiency improves, the demand for computation increases even faster, swallowing up all the gains and then some. The ambition to build ever-larger models, with ever-more parameters, outpaces the efficiency improvements of the hardware.

Rui finds this dynamic particularly troubling because it suggests that technological progress alone will not solve the problem. Without a conscious effort to prioritize efficiency over scale, or to question whether the largest models are always necessary, the environmental cost will continue to climb. The field is caught in a Red Queen’s race, running faster and faster just to stay in the same place.

Who Bears the Cost?

The environmental burden of training is not distributed equally. The data centers that host these computations are often located in regions with lower land and energy costs, which frequently overlap with lower-income communities. These communities may experience increased air pollution from the fossil fuel power plants that feed the data centers, as well as noise pollution and water stress. Meanwhile, the economic benefits of the technology accrue largely to corporations and users in wealthier regions.

This geographic displacement of harm is a pattern that repeats across many industries, but it is particularly stark in the digital space because the product feels so intangible. A user in Stockholm streaming a video or interacting with a language model has no visible connection to the coal plant in West Virginia that might be powering the backend. Rui argues that making these connections visible is the first step toward a more honest accounting of the technology’s true cost.

Can Transparency Help?

Some organizations are pushing for greater transparency. Proposals include mandatory reporting of training energy consumption, standardized efficiency benchmarks, and “energy star” style labels for models. If a model card included not just accuracy metrics but also the total carbon emitted during training, downstream users could make more informed choices. A small, efficient model might be perfectly adequate for a task, avoiding the need to invoke a massive, energy-hungry one.

Rui is cautiously optimistic about this direction but notes that transparency alone is insufficient. It must be paired with incentives. Cloud providers could offer discounts for training in low-carbon regions or during off-peak renewable hours. Funding agencies could require environmental impact statements for large-scale training projects. The goal is not to stop progress but to bend the curve of resource consumption downward while still reaping the benefits of the technology.

Rethinking Scale

The dominant narrative in the field has been that bigger is better. More data, more parameters, more compute. But a growing body of research suggests that this is not always true. Carefully curated datasets, efficient architectures, and techniques like transfer learning can achieve comparable results with a fraction of the resources. The environmental cost of training a massive model from scratch might not be justified if a smaller, fine-tuned model can perform the same task nearly as well.

Rui sees this as a design choice, not a technical limitation. The pressure to build ever-larger models comes from a culture that equates size with progress. Shifting that culture requires new success metrics that reward efficiency and parsimony alongside raw performance. Some academic conferences now ask authors to report the computational cost of their experiments, a small step that could nudge the field toward more sustainable practices.

The Water Footprint

Beyond carbon, water consumption is an underappreciated aspect of training large models. Data centers use water for cooling, and in many regions, this water is evaporated and lost to the local watershed. A single training run can consume millions of liters, enough to fill an Olympic-sized swimming pool. In water-stressed areas like the southwestern United States, this can exacerbate local shortages. The water footprint is rarely disclosed, yet it is a critical part of the environmental equation.

Rui points out that water and energy are deeply intertwined. Thermoelectric power plants, which provide much of the world’s electricity, themselves consume vast amounts of water for cooling. So the water footprint of a training run includes both the direct water used in the data center and the indirect water used to generate the electricity. This double counting makes the true impact even harder to measure but no less real.

FAQ

Why does training a large model consume so much energy?

Training involves repeatedly processing enormous datasets through networks with billions of parameters, requiring thousands of specialized processors to run for weeks or months. The computation itself draws massive power, and the hardware generates heat that demands energy-intensive cooling systems. The scale of these operations, often involving entire data center halls, is what drives the high energy consumption.

Can renewable energy solve the carbon footprint problem?

Renewable energy can significantly reduce the carbon intensity of training, but it is not a complete solution. Many data centers purchase renewable energy certificates to offset their consumption, but the physical electricity they use may still come from fossil fuel plants, especially when the sun is not shining or the wind is not blowing. Additionally, the manufacturing of hardware and the construction of data centers carry their own carbon costs that renewables do not address.

What can be done to reduce the environmental impact of training?

Several strategies can help: using more efficient model architectures, training on smaller but higher-quality datasets, scheduling training jobs in regions with cleaner grids, and improving hardware efficiency. On a systemic level, greater transparency about energy consumption and carbon emissions would allow researchers and companies to make more informed decisions. Some also advocate for prioritizing the reuse of existing models over training new ones from scratch.

Is the environmental cost of training large models justified?

This depends on the application. For some critical uses, such as medical diagnosis or climate modeling, the benefits may outweigh the costs. For others, like generating entertainment content or marginally improving a chatbot, the trade-off is harder to defend. A more careful assessment of whether a large model is truly necessary for a given task could help reduce unnecessary environmental impact.

The Hidden Price Tag of Giant Computing Projects

By Rui Mendes

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.

When we picture the digital world’s environmental footprint, we usually think of the physical stuff: sprawling server farms, blinking racks of hardware, cooling towers humming away. But there’s a quieter, more abstract layer of consumption that’s ballooning just as fast. It’s the sheer computational effort needed to train enormous neural networks—the kind that parse language, generate images, or predict how proteins fold. Rui Mendes, a systems-minded observer, has been following the energy trails behind these computational giants, and what he’s uncovered suggests we need a much broader conversation about what a sustainable digital future actually looks like.

Just How Much Energy Are We Talking About?

Training a single large-scale model can easily match the annual electricity use of several hundred homes. That’s not a rough guess—it’s a measurable fact. When researchers set out to build systems that can sift through billions of parameters, they’re not just writing code. They’re running thousands of specialized processors nonstop for weeks or months. Each chip draws power, and all that power eventually turns into heat that has to be aggressively managed.

To put it in perspective, a 2019 study from the University of Massachusetts Amherst found that training a single large transformer model could emit over 626,000 pounds of carbon dioxide equivalent. That’s roughly five times the lifetime emissions of an average car, including its manufacturing. And since that study came out, the scale of these systems has only grown. Today’s largest models are orders of magnitude bigger, with parameter counts stretching into the hundreds of billions or even trillions.

What makes this so striking is the concentration of impact. A car spreads its emissions over years of driving. A training run packs all that energy use into a few intense weeks. During that time, the power draw can rival that of a small town, placing sudden strain on local grids. Depending on the energy mix, the carbon output can be enormous.

Why Training Is So Power-Hungry

To get a feel for the energy demand, it helps to peek under the hood. Training a neural network means repeatedly adjusting millions or billions of internal knobs until the system reliably gets things right. Each adjustment requires a forward pass—pushing data through the network—and a backward pass to calculate errors and tweak those knobs. This back-and-forth, called backpropagation, is computationally brutal.

Modern networks lean on specialized hardware like GPUs or TPUs, chips built to handle massive parallel number-crunching. A single high-end GPU can pull 300 watts or more under full load. Cluster thousands of them together, and the power consumption skyrockets. Then you add the energy for cooling, networking gear, and storage systems. The total can easily reach megawatt scale.

The hardware itself has a hidden energy backstory. Manufacturing a single GPU involves mining rare earth minerals, precision fabrication in energy-intensive cleanrooms, and global shipping. That upfront cost gets spread over the chip’s lifetime, but the relentless pace of hardware upgrades means chips are often swapped out every few years. It’s a cycle of continuous material and energy investment that rarely shows up in the headline numbers.

It’s Not Just How Much Energy, but Where It Comes From

The carbon footprint of a training run isn’t just about total kilowatt-hours. It’s about the source. A run powered by a coal-heavy grid looks completely different from one tapping into hydro or nuclear. This geographic piece of the puzzle often gets lost in discussions that fixate on raw energy totals.

Cloud providers have made real progress in buying renewable energy and improving power usage effectiveness (PUE), a metric for data center efficiency. But renewable energy certificates and carbon offsets don’t always match real-time consumption. A data center might claim to be “100% renewable” on an annual basis while still pulling from fossil fuel plants during peak training hours. The timing of these massive computational jobs relative to when the sun shines or the wind blows is a subtlety that rarely makes it into public reports.

Rui Mendes notes that this opens the door to a kind of geographical arbitrage. Training runs can be scheduled in regions with cleaner grids or during times of high renewable generation. But that takes transparency and deliberate planning that aren’t yet standard. Without clear reporting on the carbon intensity of the energy used during specific training windows, the true environmental cost remains fuzzy.

The Rebound Effect: When Efficiency Feeds Bigger Appetites

Energy economics has a well-known quirk called the rebound effect. Make something more efficient, and people tend to use more of it, often wiping out the initial savings. The same pattern seems to be unfolding in large-scale model training.

Hardware makers have delivered impressive efficiency gains. Floating-point operations per watt have improved dramatically over the past decade. But instead of using those gains to shrink total energy consumption, the field has largely used them to train ever-larger models. The result? Absolute energy use keeps climbing, even as each individual operation gets cheaper.

That’s not automatically a bad thing—bigger models can unlock capabilities that smaller ones couldn’t touch. But it does mean that efficiency improvements alone won’t get us out of this. Without a deliberate effort to cap or reduce total energy use, the trend line points in one direction: up.

Measuring What Actually Matters

One of the biggest headaches is that we lack standardized, transparent reporting on the environmental impact of training runs. Some research papers now include carbon emission estimates, but they’re often calculated with different methods and assumptions, making apples-to-apples comparisons nearly impossible.

A handful of tools have popped up to help researchers estimate their carbon footprint, including calculators that factor in the local grid’s energy mix. But these tools are only as good as the data they’re fed, and they don’t capture the full lifecycle—from hardware manufacturing to eventual disposal. Then there’s the question of what to measure. Just the carbon from electricity? Should we include the embodied carbon in the hardware? What about the water used for cooling?

Rui Mendes suggests the field needs a standardized framework, something like a nutrition label for training runs. Such a label might list total energy consumed, the carbon intensity of the energy source, hardware lifecycle estimates, and maybe even a comparison to the energy a human would use to perform a similar task over a lifetime. That kind of transparency would let us make more informed trade-offs about whether a particular model’s benefits justify its environmental cost.

The Overlooked Water Connection

Energy isn’t the only resource on the table. Water plays a quiet but critical role in cooling data centers. A mid-sized facility can go through hundreds of thousands of gallons a day, mostly for cooling. In regions already facing water scarcity, that can create real tension between computational needs and community water supplies.

The water footprint varies a lot depending on cooling technology and local climate. Evaporative cooling, which is highly energy-efficient, drinks more water. Closed-loop systems sip water but gulp more energy. It’s a tricky optimization problem with no universal answer. In arid regions, water consumption might be the bigger worry; in areas with plenty of water but carbon-heavy grids, energy use takes the lead.

Some data centers are trying out creative cooling approaches, like dunking servers in non-conductive fluids or building facilities in cold climates to use outside air. These solutions can slash both energy and water consumption, but they’re not a fit everywhere.

Steering Toward More Sustainable Practices

Tackling the environmental impact of large-scale training means pulling multiple levers at once. On the technical side, researchers are exploring ways to make training less wasteful. Techniques like model pruning, quantization, and knowledge distillation can shrink computational requirements without gutting performance. Sparse models, which activate only a fraction of their parameters for any given input, offer another promising path.

There’s also growing interest in “green” scheduling—timing training runs to line up with periods of high renewable energy availability. This doesn’t cut total energy use, but it can meaningfully lower carbon emissions by shifting the load to cleaner times. Some cloud providers already offer tools to help customers schedule workloads based on carbon intensity forecasts.

On the policy side, calls for greater transparency and accountability are getting louder. If the environmental cost of training were more visible, it might nudge more thoughtful decisions about which models are worth the investment. Not every problem demands the largest possible model, and sometimes the marginal performance gain doesn’t justify the exponential jump in resource consumption.

Redefining What Progress Looks Like

Maybe the deepest shift needed is in how we define progress. The current playbook equates bigger with better, measuring advancement mostly through benchmark scores and parameter counts. But that narrow lens ignores the wider system these models live in.

What if we measured success not just by accuracy on a test set, but by the ratio of benefit to environmental cost? A model that hits 95% of the performance using 10% of the resources might be a more impressive achievement than one that squeezes out an extra percentage point at enormous expense. This kind of thinking is already standard in other engineering fields, where efficiency and resource constraints are baked into the design process from day one.

Rui Mendes observes that the conversation is starting to turn. More researchers are acknowledging the environmental dimension of their work, and some are actively working to minimize it. But there’s still a long road ahead before sustainability becomes a first-class consideration in the design and deployment of large-scale neural networks.

Frequently Asked Questions

How much energy does training a large model actually use?

The energy consumption varies widely depending on the model size, architecture, and training duration. A large transformer model with hundreds of billions of parameters can consume several thousand megawatt-hours of electricity during training—equivalent to the annual electricity use of hundreds of U.S. households. The exact figure depends on the hardware efficiency, the number of GPUs or TPUs used, and the length of the training run, which can span weeks or months.

Why don’t companies just use renewable energy for training?

Many companies do purchase renewable energy credits or contract for renewable power, but the reality is more complicated. Data centers are physically connected to regional grids, and they draw whatever mix of energy is available at the time of use. If a training run occurs during a period of low wind or solar generation, the actual electricity may come from fossil fuel plants, even if the company has offset those emissions on paper. True 24/7 carbon-free energy matching is a goal that few have achieved.

Can efficiency improvements solve the problem?

Efficiency improvements are important but insufficient on their own. The history of computing shows that as hardware becomes more energy-efficient, demand for computation tends to increase, often outpacing the gains. This is the rebound effect. To meaningfully reduce environmental impact, efficiency must be paired with conscious choices about model size, training frequency, and the necessity of each training run. Without that, we risk running faster while staying in the same place.

What can individuals do about this issue?

While the decisions about large-scale training are made by organizations, individuals can influence the direction through their choices as consumers, employees, and citizens. Supporting companies that prioritize transparency and sustainability, advocating for better reporting standards, and questioning whether the largest models are always necessary are all meaningful actions. On a technical level, practitioners can choose to work with smaller, more efficient architectures and report the environmental impact of their experiments.

When we picture the environmental toll of our digital lives, we usually think of sprawling server farms baking under the sun or the scarred landscapes where rare minerals are mined for our gadgets. But there’s another kind of consumption, quieter and harder to visualize, that’s starting to rival those concrete images. I’m talking about the brute-force computational sprint needed to teach a massive neural network how to spot a cat, write a sentence, or fold a protein. Not the everyday queries—the training. The foundational, pre-launch grind. Rui Mendes here, and I’ve been digging into the physical resources that go into creating a state-of-the-art model. The numbers are weirder and more sobering than I expected.

The Physics of a Digital Brain

To grasp the environmental side, you have to shake off the cloud metaphor. Training a large model is a physical, industrial process. Racks of specialized chips—GPUs or TPUs—run at full throttle for weeks or months, drawing enormous amounts of electricity and radiating heat. A single high-end GPU can pull 300 to 400 watts, and a training cluster might pack thousands of them. The total power draw can match that of a small town. But kilowatt-hours are only part of the story. That heat has to go somewhere, so data centers pair their chips with cooling systems that are just as thirsty. For every watt spent on computation, a significant fraction of another watt is spent fighting thermodynamics—often with water-hungry cooling towers or energy-intensive chillers.

Then there’s the question of where the electricity comes from. A training run plugged into a hydro-rich grid in Quebec has a fundamentally different carbon shadow than one relying on a coal-heavy mix. That distinction gets flattened in most headlines, but it’s the difference between a mild sunburn and a scorching. The source matters as much as the amount.

Beyond the Plug: The Lifecycle View

A systems-minded look pushes us past the electricity meter. Every GPU arrives with a history—mining, refining, manufacturing, shipping. The embodied carbon in a rack of servers, the emissions baked in before they ever spin up, is a debt that training runs only partially repay. When we talk about the cost of a single training session, we’re really amortizing a slice of that hardware’s full lifecycle. It’s like calculating the fuel for a road trip while ignoring the car’s manufacturing footprint. The trip is short, but the car’s existence is a long-term environmental commitment.

Water: The Quiet Partner

Electricity grabs the headlines, but water is the silent twin. Data centers can be shockingly thirsty. Cooling towers evaporate millions of liters during a large training run—water that doesn’t return to the local watershed. In arid regions, that’s water that won’t irrigate a field or flow from a household tap. A single, massive training session can consume enough water to fill several Olympic swimming pools, and that’s not a metaphor. The siting of a data center becomes an environmental justice question: do you put it where the power is green but the water is scarce, or where water is plentiful but the grid is dirty? There’s no easy answer, only trade-offs.

Why Bigger Models Bite Harder

The relationship between model size and energy use isn’t a straight line. It’s a curve that bends upward, sharply. When researchers chase better accuracy, they often double or triple the number of parameters—the internal knobs the system tunes. That jump demands more data and many more computational steps. A model ten times larger than its predecessor might need a hundred times the compute. It’s a Red Queen’s race: incremental performance gains demand disproportionately larger environmental budgets. We’re sprinting just to stay in the same place, and the track is getting longer.

And then there’s the ghost of experimentation. The final model that makes the press release is the lone survivor of hundreds or thousands of failed prototypes. Researchers tweak architectures, fiddle with learning rates, and run ablation studies. The cumulative energy burned on those dead ends can dwarf the cost of the final, polished run. That R&D overhead is a shadow footprint, almost never included in public estimates. It’s the unseen bulk of the iceberg.

The Carbon Accounting Fog

Transparency is a mess. Some organizations publish energy and emissions data for their biggest projects, but the methods are all over the map. One might report only the GPU draw, ignoring the data center’s overhead (the power usage effectiveness, or PUE). Another might buy renewable energy certificates to offset their consumption, a move that can paper over the actual grid mix they relied on. A rigorous, standardized lifecycle assessment is what we need, but we’re nowhere close. Right now, it’s a patchwork of voluntary disclosures, making apples-to-apples comparisons nearly impossible.

Efficiency: A Double-Edged Sword

A common pushback is that hardware keeps getting more efficient. And it’s true: each new chip generation does more calculations per watt. But that efficiency gain tends to get swallowed by the appetite for bigger models. It’s a classic rebound effect—as the cost of computation drops, demand rises to meet it, and total energy use climbs. The treadmill speeds up, and we run faster just to keep pace. Specialized hardware like TPUs can lower the energy per operation, but those savings are usually reinvested into scaling up, not shrinking the absolute footprint. Breaking that cycle would take a deliberate constraint: a self-imposed size limit, a preference for leaner architectures, or an external shove like carbon pricing.

Geographic Lottery: Where You Train Matters

Location is destiny for a training run’s environmental impact. Train in Quebec, where the grid is almost entirely hydro, and your carbon profile is a whisper. Train in a coal-dependent region, and it’s a shout. Some researchers have floated “carbon-aware” scheduling—shifting workloads to times and places where renewables are plentiful. It’s a clever operational fix, but it doesn’t touch the absolute energy consumption. It just changes the color of the electrons.

Water stress adds another dimension. A data center in Arizona might run on low-carbon solar during the day but still drain scarce aquifers. Balancing carbon, water, and land use is a multi-dimensional puzzle that few organizations tackle publicly. The geographic lottery means two identical training runs can carry wildly different environmental price tags, depending on where the plug meets the socket.

Rethinking the Scoreboard

The environmental cost forces an uncomfortable question: what are we actually optimizing for? If the only metric is accuracy on a benchmark, then any amount of energy is fair game for a marginal gain. But if we broaden the definition of performance to include resource efficiency, the leaderboard shifts. A model that hits 95% of the accuracy with 10% of the energy might be the real engineering triumph. That requires a cultural shift in how research is judged and celebrated.

Some corners of the community are pushing for “Green” benchmarks that report energy and emissions alongside accuracy. That transparency would let practitioners make informed trade-offs, picking a model that fits their operational limits and environmental values. It also nudges innovation toward efficient architectures—sparse models that activate only a fraction of their parameters for a given input, or training algorithms that converge in fewer steps.

The Data Quality Lever

Another lever is the data itself. Training on massive, noisy datasets scraped from the web is energetically wasteful. Curating smaller, higher-quality datasets can yield comparable or even better results with a fraction of the compute. It shifts the burden from brute force to thoughtful data engineering. It’s the old “garbage in, garbage out” rule, but with an environmental sting: noisy data doesn’t just hurt performance, it burns energy. Investing in data quality is a direct investment in energy efficiency.

FAQ: Unpacking the Energy Debate

How much energy does it actually take to train a single large model?
Estimates vary, but training a very large model can consume hundreds of megawatt-hours of electricity—enough to power dozens of average homes for a year. The exact figure depends on model size, hardware efficiency, and data center PUE. Some published figures for the largest models exceed 1,000 megawatt-hours, with associated carbon emissions comparable to the lifetime emissions of several cars.

Does the energy consumption stop once the model is trained?
No. Training is a one-time, intensive burst, but deploying the model for millions of users—a process called inference—also consumes energy continuously. While a single inference query is cheap, the aggregate energy use of a popular service can quickly surpass the training cost over its lifetime. Efficient inference hardware and model compression techniques are essential to manage this ongoing load.

Can renewable energy solve the problem entirely?
Renewable energy is a critical part of the solution, but it’s not a silver bullet. Even if a data center is matched with 100% renewable energy certificates, the physical infrastructure still consumes water and materials. Additionally, the intermittent nature of solar and wind requires grid-scale storage or backup generation, which have their own environmental footprints. A truly sustainable approach must reduce absolute energy demand, not just green the supply.

What can a regular person do about this?
Individual actions have limited direct impact on the training practices of large organizations, but collective pressure can shift norms. Supporting transparency initiatives, choosing services that publish their environmental metrics, and advocating for research funding that prioritizes efficiency can all contribute. On a personal level, being mindful of the energy behind digital tools—and using them intentionally rather than wastefully—is a small but meaningful practice.

Toward a More Honest Accounting

The environmental cost of training large models isn’t an argument against progress. It’s a call for a more honest accounting. We need to measure what matters, report it transparently, and design systems that respect physical limits. The curious, systems-minded observer will see that the true cost isn’t just in kilowatt-hours or carbon tons, but in the choices we make about what kind of intelligence we value. A lighter, more efficient model that serves a community’s real needs may be a far greater achievement than a bloated giant that wins a benchmark but burdens the planet.

As we keep building ever-larger digital constructs, we have to remember they’re tethered to the physical world by cables, pipes, and smokestacks. The electrons that animate them come from somewhere, and the heat they generate goes somewhere. Ignoring that connection isn’t just sloppy engineering; it’s a failure of systems thinking. The most elegant solution is the one that achieves its purpose with the least harm—a principle that applies as much to a neural network as it does to a bridge or a building.

When we talk about the environmental toll of our digital lives, the conversation usually lands on data centers packed with humming servers or the billions of smartphones sucking power from the grid. But there’s a quieter, less visible cost that’s been ballooning in the background: the staggering amount of energy it takes to train massive neural networks. I’m Rui Mendes, and I’ve spent years tracing the systems that keep our digital world running. What I keep bumping into is a story of runaway demand, clever engineering, and a question we’re only starting to ask—what does it actually cost the planet to teach a machine to spot a cat, translate a sentence, or spit out a paragraph of text?

This isn’t a tidy good-versus-evil tale. It’s a messy tangle of hardware, geography, and the very bones of modern computing. The numbers are eye-popping, but they’re also deeply human, tied to our stubborn drive for more capable systems. Let’s walk through the lifecycle of a single training run—from the silicon in the chips to the cooling towers baking in the desert—and see what we’re really burning.

”

The Scale of the Machine

To wrap your head around the environmental cost, you first have to grasp the sheer bulk of a modern training cluster. We’re not talking about a rack of servers humming in a closet. A top-tier training run for a large model might rope in tens of thousands of specialized processors—GPUs or custom accelerators—churning nonstop for weeks or months. Each of those chips can pull hundreds of watts. Multiply that by, say, 50,000 units, and the power draw starts to rival a small town. But the energy isn’t just for the chips. It’s for the whole ecosystem keeping them alive: networking switches, storage arrays, and the cooling gear that stops a multi-million-dollar cluster from turning into a puddle of slag.

Take a single GPU like the NVIDIA A100, a real workhorse of modern training. Under full load, it can suck down around 400 watts. A cluster of 10,000 of those, running flat-out for 30 days, would chew through roughly 2.88 million kilowatt-hours. That’s enough to power about 270 average U.S. homes for an entire year. And that’s just the processors. Toss in the cooling overhead—often another 30–40%—and the total energy footprint swells. The physical footprint is just as wild: these machines live in purpose-built warehouses, floors reinforced to carry the weight, power feeds thick as a wrestler’s arm.

Image source: Pexels

The Lifecycle of a Training Run

A single training run isn’t a one-and-done affair. It’s an iterative grind. Researchers fiddle with hyperparameters, tweak model architectures, and restart training dozens—sometimes hundreds—of times before landing on a final version. Every one of those experiments carries its own energy bill. A paper out of the University of Massachusetts Amherst estimated that training a single large transformer model can belch out over 626,000 pounds of CO₂ equivalent. That’s roughly five times the lifetime emissions of an average American car, manufacturing included. And that’s just one successful run—not the graveyard of failed attempts that came before it.

The carbon intensity leans hard on the energy mix of the grid where the data center sits. A cluster sipping hydroelectricity in Quebec will have a fraction of the emissions of one plugged into a coal-heavy grid in Virginia. But location often gets decided by latency, tax breaks, and land prices—not environmental math. Some operators buy renewable energy credits to offset their draw, but those certificates don’t always mean fresh clean power on the grid. Sometimes they’re just accounting tricks that paper over the physical reality of burning fossil fuels.

The Water That No One Sees

Electricity isn’t the only thing getting consumed. Large training clusters throw off enormous heat, and the most common way to shed it is through water-based cooling systems. Evaporative cooling towers—which spray water over heat exchangers—can guzzle millions of gallons a year. In drought-prone spots like the southwestern United States, that sets up a direct competition with farms and households. A 2023 study from the University of California, Riverside figured that training a single large model could evaporate up to 700,000 liters of freshwater. That’s enough to fill an Olympic swimming pool halfway. And that water doesn’t come back; it’s gone, lost to the atmosphere.

Some facilities are shifting toward closed-loop liquid cooling, where coolant snakes through pipes clamped right onto the chips, then dumps heat through outdoor radiators. That can slash water consumption, but it means ripping out old infrastructure and swallowing upfront capital costs. The choice between water and air cooling is often a trade-off between local environmental strain and energy efficiency, with no clean answer.

Image source: Pexels

The Hardware Supply Chain

Before a single watt ever flows into a data center, the chips themselves have already racked up an environmental bill. Semiconductor fabrication is one of the most resource-hungry manufacturing processes on Earth. A single advanced processor might go through hundreds of steps involving toxic chemicals, ultra-pure water, and energy-guzzling lithography tools. The silicon wafers get etched in cleanrooms that keep out particles with constant air filtration, chewing through enormous amounts of electricity. A modern fab can pull as much power as a mid-sized city, and the water needed to produce a single chip can top thousands of gallons once you count the repeated rinsing between layers.

Then there’s the global supply chain. Raw materials—silicon, copper, gold, rare earth elements—get mined, refined, and shipped across oceans. The embodied carbon in a single GPU, before it ever crunches a byte of data, is pegged at around 150 kg of CO₂ equivalent. Multiply that by the tens of thousands of accelerators in a training cluster, and the upfront carbon debt is staggering. This cost often gets amortized over the hardware’s lifetime, but when gear gets swapped out every 3–5 years to keep pace with performance demands, that debt never really gets paid down.

The Geography of Power

Data centers don’t float in a vacuum; they plug into regional grids with wildly different carbon profiles. A training run in Sweden, where the grid leans on hydro and nuclear, might cough up 10 grams of CO₂ per kilowatt-hour. The same run in West Virginia, where coal still holds sway, could spit out 900 grams per kilowatt-hour—a 90-fold gap. This geographic lottery means two identical experiments can have radically different environmental footprints based purely on where the servers sit.

Some operators are starting to weigh carbon intensity when picking sites, but it’s rarely the main driver. Latency to end-users, tax sweeteners, and cheap land usually elbow out environmental concerns. There’s also the headache of grid transparency: real-time carbon intensity data isn’t always available, and even when it is, the scheduling algorithms that kick off training jobs rarely pay it any mind. A training run that could be slid to a cleaner time of day or a different region often stays put, because the scheduling systems weren’t built to optimize for carbon.

Image source: Pexels

The Efficiency Paradox

Here’s where the systems thinking gets twisty. As hardware gets more efficient—more computations per watt—total energy consumption doesn’t necessarily drop. Often, it climbs. This is a textbook case of Jevons paradox: when a resource gets cheaper or more efficient to use, demand for it swells enough to wipe out the savings. In the world of large-scale training, each new generation of chips delivers more performance per watt, but researchers answer by training bigger models on more data, shoving total energy consumption higher.

Look at the trend over the past decade. The computational resources poured into the largest training runs have been doubling every 3.4 months, leaving Moore’s Law in the dust. That means even as individual processors get thriftier, the aggregate energy appetite of bleeding-edge experiments keeps climbing. The efficiency gains are real, but they’re getting swallowed by the hunger for scale.

There’s also a rebound effect in cooling. More efficient chips can be packed tighter, which jacks up the heat density of server racks. That, in turn, demands more aggressive cooling, which can cancel out the initial efficiency wins. It’s a tangled knot of feedback loops that shrugs off simple fixes.

Measuring What Matters

If we want to trim the environmental cost of training large models, we first need to measure it straight. That’s harder than it sounds. The energy draw of a single training run can be guessed from hardware specs and runtime, but that misses the overhead of cooling, networking, and storage. It also ignores the embodied carbon baked into the hardware itself. A full lifecycle assessment would tally everything from mining to manufacturing to operation to disposal, but those assessments are rare and pricey.

Some researchers have floated standardized metrics—like carbon per training run or carbon per inference query—to make comparisons easier. But those metrics are only as solid as the data behind them, and plenty of organizations keep their detailed energy numbers close to the chest for competitive reasons. There’s also the allocation puzzle: if a data center juggles multiple workloads, how do you fairly pin its total energy consumption on a specific training job?

The Cooling Conundrum

Cooling is the hidden multiplier in data center energy equations. Traditional air cooling leans on powerful fans and chillers, which can tack on 30–50% to the total energy draw. Evaporative cooling trims that overhead but drinks water. Direct-to-chip liquid cooling is more efficient but demands expensive infrastructure overhauls. Immersion cooling—where whole servers get dunked in dielectric fluid—offers the best thermal performance but stirs up new headaches around fluid handling and hardware compatibility.

Each path has its own environmental trade-offs. Air cooling in a region with a carbon-heavy grid might leave a bigger carbon footprint than water cooling in a drought-prone area, but the water consumption creates a different kind of environmental pinch. There’s no one-size-fits-all answer—it hinges on local conditions, the specific hardware getting cooled, and the values we decide to put first.

Frequently Asked Questions

How much energy does training a single large model actually consume?

Estimates bounce around a lot depending on model size, hardware efficiency, and data center location. A 2019 study found that training a large transformer model can burn through over 650 megawatt-hours of electricity—about the same as the annual energy use of 60 average U.S. homes. More recent models likely chew through several times that, though exact figures often stay under wraps.

Does the carbon footprint of training outweigh the benefits of the resulting model?

This is a knotty question with no clean answer. The carbon cost is a one-time hit for training, while the model might get used millions of times for inference, spreading its usefulness over years. But if the model gets swapped out fast for a newer version, or if its applications don’t lead to real energy savings elsewhere, the net environmental impact could tip negative. Lifecycle analysis is what’s needed to make fair comparisons.

Can renewable energy solve the problem?

Renewable energy can slash the carbon emissions of training, but it doesn’t mop up all environmental impacts. Water consumption, hardware manufacturing emissions, and land use for data centers still nag. Plus, buying renewable energy certificates doesn’t always mean the actual electrons feeding a data center are carbon-free, especially if the local grid still runs heavy on fossil fuels.

What can be done to reduce the environmental cost?

A handful of approaches can help: picking data center spots with cleaner grids, designing thriftier model architectures that need less computation, reusing existing models instead of training from scratch, and getting more open about energy consumption. Hardware breakthroughs like more efficient chips and smarter cooling also play a part, but they have to be paired with conscious choices about scale to dodge rebound effects.

Looking Upstream

The environmental story of large-scale training isn’t just about electrons and water. It’s about the materials that make the machines possible. Rare earth elements like neodymium and dysprosium are essential for the magnets in hard drives and the capacitors on circuit boards. Their extraction—often bunched in a few countries—leaves behind toxic tailings and radioactive waste. The semiconductor industry’s appetite for these materials is swelling, and with it, the ecological scars of mining.

Then there’s the question of e-waste. Training clusters have a lifespan of maybe three to five years before they get shoved aside by faster, more efficient hardware. The old gear doesn’t just vanish. It gets torn down, shipped out, and often processed in informal recycling yards where workers face hazardous materials. The full environmental cost of a training run includes a slice of this downstream burden, though it’s rarely counted.

We’re building systems of extraordinary capability, but we’re doing it on a planet with hard limits. The challenge isn’t to stop building—it’s to build with eyes open to the whole system, from the mines to the cooling towers to the recycling yards. That awareness is the first step toward making different choices.

As I trace these connections, I’m not left with despair. I’m left with a nagging curiosity about what comes next. The same ingenuity that dreamed up these models can be pointed at measuring and shrinking their impact. The question is whether we’ll choose to look at the whole picture, or keep our eyes locked on the screen.

When we talk about pollution, we usually picture smokestacks, traffic jams, or plastic swirling in the ocean. But there’s another kind of environmental toll that’s harder to see—it hums inside anonymous warehouses, travels through fiber-optic cables, and surges every time someone decides to train a really big neural network. I’m Rui Mendes, and I’ve always been drawn to the hidden wiring of complex systems: the energy flows, the feedback loops, the side effects nobody planned for. So when I started digging into what it actually takes to build the massive models behind so many modern tools, I found a story that doesn’t get told nearly enough.

This isn’t about blaming technology. It’s about understanding the real resource demands of computation at an industrial scale. If we’re serious about a future that’s both smart and sustainable, we need to measure what matters—and right now, the environmental cost of training large models is one of those things that quietly slips through the cracks.

Why Training Runs Are So Thirsty for Power

To get why the energy numbers are so big, you have to look at what’s actually happening inside those server racks. Training a massive model isn’t just a laptop running hot overnight. It’s thousands of specialized processors—GPUs or TPUs—working in parallel for weeks or months. Each chip performs quadrillions of tiny calculations, and every one of those operations pushes electrons through silicon. The joules pile up fast.

But the computation itself is only part of the story. Moving data back and forth between processors and memory often burns more energy than the math. And then there’s the heat. All that electricity eventually becomes thermal energy, and getting rid of it requires industrial cooling: chillers, fans, and sometimes evaporative systems that gulp water as well as power. Researchers also rarely train a model just once. They tinker with architectures, tweak settings, and run dozens—sometimes hundreds—of test runs before the final big push. The total energy bill for all that trial and error can easily overshadow the cost of the final model.

The Carbon Roulette of Location

Energy use is only half the equation. The carbon footprint depends just as much on where the electricity comes from. A data center plugged into a coal-heavy grid will leave a much dirtier mark than one drawing from hydro or nuclear, even if both use the exact same number of megawatt-hours. It’s a kind of geographic lottery: two identical training runs can have wildly different climate impacts based on nothing more than where the servers happen to sit.

A few organizations have started paying attention to this, timing their big jobs for when renewables are plentiful or even shifting workloads to cleaner regions. But it’s far from standard practice. Most training happens wherever the hardware is available, and the energy source is rarely disclosed. As someone who thinks in systems, I find that opacity maddening. Without open data on location, duration, and grid mix, we can’t hold anyone accountable—or even have a proper conversation about the trade-offs.

Water: The Overlooked Ingredient

Electricity gets the headlines, but water is the quiet partner in this story. Data centers use massive amounts of it for cooling—sometimes directly, through evaporation, sometimes indirectly through the power plants that feed them. In places already struggling with drought, this creates real friction. One study from 2021 estimated that training a single large model can consume tens of thousands of liters of water, both on-site and off-site. Multiply that by the number of models being trained worldwide, and the freshwater footprint starts to look alarming, especially in arid regions where data centers cluster for cheap land and tax breaks.

This is a classic systems trap: optimize for one variable (say, electricity cost) while quietly offloading another (water scarcity). The feedback loops are long and tangled, so the consequences don’t show up on a quarterly earnings call. But they do show up in depleted aquifers and dropping reservoir levels.

The Lifecycle Nobody Talks About

Energy and water are the day-to-day costs, but the hardware itself carries a heavy backpack of embedded carbon. Manufacturing GPUs and other specialized chips means mining rare earth minerals, running chemical-intensive processes, and operating fabrication plants that are themselves energy hogs. A single high-end processor can represent dozens of kilograms of CO₂ before it ever flips a single bit.

And then there’s the lifespan problem. In the race for speed, data centers swap out their fleets every three to five years. Some of the retired gear finds a second life in other markets, but a lot of it gets shredded or dumped. The environmental toll of e-waste—toxic metals seeping into soil, unsafe recycling practices in developing countries—is well documented, yet it almost never comes up when people talk about model training.

When you step back and look at the whole chain—mineral extraction, chip fabrication, operational energy, disposal—the picture gets sharper. Training a large model isn’t just a blip on a power meter. It’s a node in a global supply chain with environmental consequences at every link.

Why Better Efficiency Isn’t a Magic Fix

It’s easy to point at technological progress and say, “See? Chips get more efficient every year, and new training tricks cut the number of computations.” But there’s a stubborn pattern in environmental economics called Jevons paradox: when something gets more efficient, we often end up using more of it, not less. Cheaper, faster training leads to more models, bigger models, and more experimentation. Total resource consumption can actually climb even as per-unit efficiency improves.

We’re watching this happen right now. The cost to train a model of a given capability has dropped, but the frontier of capability keeps moving, and the number of teams chasing it has exploded. The net result is a sector whose energy appetite keeps growing. Data center electricity consumption is projected to rise steeply through this decade, and model training is a big piece of that pie.

The Transparency Gap

One of the most striking things I’ve noticed while researching this is how little information is actually out there. Energy consumption, carbon emissions, water usage, hardware lifecycle details—these are rarely shared in any consistent way. A few companies release selective numbers for flagship projects, but there’s no industry-wide reporting standard. Without consistent, auditable data, it’s nearly impossible to compare approaches, track progress, or even grasp the true scale of the problem.

This opacity isn’t just an environmental issue; it’s a governance failure. When the costs are hidden, they’re easy to ignore. And when they’re ignored, they pile up. I’m convinced that transparency is a prerequisite for responsible development. If we can’t measure it, we can’t manage it—and we certainly can’t have an honest public conversation about what we’re trading off.

What a Systems View Brings Into Focus

Looking at this through a systems lens, the environmental cost of training large models isn’t a standalone problem. It’s tangled up with energy policy, hardware supply chains, water management, and even land use. A data center doesn’t float in a vacuum; it competes for resources with other human needs and with natural ecosystems. When a new facility goes up in a drought-prone region, it nudges local water tables. When it pulls from a coal-heavy grid, it contributes to air pollution and climate change that affect communities far beyond its fences.

These interconnections mean that solutions have to be just as systemic. Better processor efficiency is nice, but it’s not enough. Switching to renewables is better, but it doesn’t touch water or e-waste. A genuinely responsible approach would look at the full lifecycle, from design to decommissioning, and would include transparent reporting, location-aware scheduling, and serious hardware reuse programs.

There’s also a deeper question about priorities. Not every model needs to be the biggest. Not every problem demands a brute-force computational assault. Sometimes a smaller, more targeted model trained on a cleaner grid can deliver most of the benefit at a fraction of the cost. The trick is knowing when “good enough” is actually good enough—and having the discipline to stop there.

FAQ: Environmental Costs of Large-Scale Training

How much energy does training a single large model actually use?

Estimates vary a lot depending on model size, hardware efficiency, and training duration, but some published figures put the electricity consumption for a single large training run in the range of hundreds of megawatt-hours—roughly the annual electricity use of dozens of U.S. households. The associated carbon emissions depend on the local grid mix, ranging from relatively low in regions with clean power to hundreds of tonnes of CO₂ in coal-dependent areas.

Why don’t companies just use renewable energy for all their training?

Some companies do buy renewable energy credits or build data centers near clean power sources, but it’s far from universal. Challenges include the intermittent nature of solar and wind, the need for steady baseload power for 24/7 operations, and the simple fact that the cheapest electricity often comes from fossil-fuel-heavy grids. Plus, renewable energy procurement doesn’t automatically address other impacts like water consumption or hardware waste.

Is there a way to compare the environmental impact of different models?

Right now, there’s no standardized reporting framework, so direct comparisons are tough. Some researchers have proposed metrics like “carbon per training run” or “carbon per inference,” but these aren’t widely adopted. Without consistent disclosure of energy consumption, grid mix, water usage, and hardware lifecycle data, any comparison remains partial and often leans on estimates rather than measured values.

What can be done to reduce the footprint?

Several approaches can help: using more efficient hardware and algorithms, scheduling training on grids with lower carbon intensity, designing models that achieve required performance with fewer parameters, extending hardware lifecycles through reuse and refurbishment, and improving transparency through mandatory reporting. The most effective strategy, though, is to question whether the largest possible model is truly necessary for the task at hand.

As I wrap up this exploration, I’m left with a sense of cautious optimism. The conversation about sustainable computing is gaining momentum, and the tools to measure and reduce impact are improving. But the pace of growth in model size and training frequency is relentless. Without a parallel commitment to transparency and lifecycle thinking, we risk building intelligence on a foundation of hidden environmental debt. And in any system, debt has a way of coming due.

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.

When we talk about the digital revolution, we often picture sleek data centers humming quietly in remote locations, invisible streams of data flowing through fiber-optic cables, and a world made more efficient by software. What we rarely picture is the plume of smoke from a coal-fired power plant, the water rushing through a hydroelectric dam’s turbines, or the rare minerals ripped from the earth to build specialized hardware. Yet these physical realities are the bedrock of every major computational achievement, and nowhere is that more apparent than in the training of large-scale machine learning systems.

Rui Mendes has spent years thinking about systems—how they interact, where their boundaries lie, and what we miss when we focus only on the shiny outputs. When a new model breaks records for language understanding or image generation, the headlines celebrate the benchmark scores. But Rui’s curiosity pulls him toward the less glamorous question: what was the material cost of that breakthrough? How much water, how much electricity, how much rare metal had to be extracted and transformed to make a single training run possible?

This article is not a condemnation of technology. It is an attempt to map the invisible supply chain that powers modern machine learning, to trace the environmental toll from the mine to the motherboard, and to ask whether the way we measure progress needs a deeper recalibration.

The Energy Appetite of a Single Training Run

Let’s start with the most direct metric: electricity consumption. Training a state-of-the-art model is not like running a laptop for a few hours. It involves thousands of specialized processors—GPUs or TPUs—operating in parallel for weeks or even months. Each of these chips draws hundreds of watts, and a typical cluster might contain tens of thousands of them.

A widely cited 2019 study from the University of Massachusetts Amherst estimated that training a single large natural language processing model can emit over 284 tonnes of carbon dioxide equivalent—roughly the same as five average American cars over their entire lifetimes, including manufacturing. Since then, models have grown by orders of magnitude. The compute used in the largest training runs has been doubling every 3.4 months, a pace that far outstrips Moore’s Law. What was once a 284-tonne problem is now, for the largest experiments, likely in the thousands of tonnes.

But carbon emissions are only part of the story. The electricity that feeds these clusters has to come from somewhere. In regions where the grid is still dominated by fossil fuels, the carbon intensity is high. Even in areas with significant renewable penetration, the sheer scale of demand can strain local infrastructure, sometimes leading utilities to fire up peaker plants—often natural gas—to meet the load. The location of a data center matters enormously. A training run powered by hydroelectricity in Quebec has a vastly different footprint than one plugged into a coal-heavy grid in parts of the Midwest or Asia.

The Water That Cools the Cloud

Electricity is the most visible resource, but water is the silent partner. Data centers generate immense heat, and cooling systems are essential to prevent equipment failure. Many facilities use evaporative cooling, which consumes water directly. Others rely on electricity generated by thermoelectric power plants, which themselves withdraw vast quantities of water for cooling.

A 2021 study estimated that training a single large model can consume up to 700,000 liters of freshwater—enough to fill an Olympic-sized swimming pool more than a quarter of the way. This figure includes both on-site cooling and off-site water use at power plants. In regions already facing water stress, such as parts of the southwestern United States, the siting of large data centers has become a point of tension. Residents and local governments are beginning to ask whether the economic benefits of hosting these facilities outweigh the strain on aquifers and reservoirs.

Rui Mendes finds this systems-level view compelling because it reveals interdependencies that are easy to ignore. A model trained in a data center cooled by a closed-loop system still draws water indirectly if its electricity comes from a coal or nuclear plant that uses once-through cooling. The water footprint extends far beyond the server room.

The Hardware Lifecycle: From Mine to E-Waste

Beyond operational energy and water, there is the embodied cost of the hardware itself. The GPUs and TPUs that power training runs are marvels of engineering, but they are also products of extractive industries. Rare earth elements, cobalt, tantalum, and gold are mined, refined, and shipped across the globe to fabrication plants. Mining operations are water-intensive and often leave behind toxic tailings. The refining process requires high temperatures and chemical baths, adding further energy and water demands.

Once fabricated, these chips have a limited lifespan. The relentless pace of hardware improvement means that accelerators are often retired after three to five years, even if they remain functional. E-waste is the fastest-growing waste stream on the planet, and data centers contribute a significant share. While some components are recycled, the complex mix of materials makes full recovery difficult. Many end up in informal recycling operations in developing countries, where workers are exposed to hazardous substances without proper protection.

Rui Mendes sees this as a classic systems trap: we optimize for performance per dollar or per watt at the chip level, but we fail to account for the full lifecycle costs. A more efficient chip might reduce operational energy but increase embodied energy if it requires rarer materials or more complex manufacturing. Without a full view, we risk shifting the burden rather than reducing it.

The Scaling Race and Its Environmental Multiplier

The environmental cost of training a single model is significant, but the real concern is the scaling trend. The field has been driven by a simple empirical observation: bigger models, trained on more data with more compute, tend to perform better. This has led to an arms race where each new generation of models requires an order of magnitude more resources than the last.

Consider the compute used in notable projects over the past decade. Early breakthroughs required compute budgets measured in petaflop/s-days. By 2020, the largest training runs were consuming exaflop/s-days—a thousandfold increase. The trend shows no sign of slowing. If anything, the competitive dynamics of the industry encourage overprovisioning: teams will often train multiple versions of a model, discarding all but the best-performing one. The discarded runs still consumed energy, water, and hardware lifespan.

This scaling has a multiplier effect on environmental impact. A tenfold increase in compute does not necessarily mean a tenfold increase in performance, but it does mean a roughly tenfold increase in resource consumption. The question Rui Mendes finds most pressing is whether the marginal benefit justifies the marginal cost. At what point does the pursuit of a slightly better benchmark score become an irresponsible use of shared resources?

Carbon Accounting: What Gets Measured Gets Managed

One of the challenges in addressing this issue is the lack of standardized reporting. Unlike the airline or automotive industries, where emissions are regulated and publicly reported, the computational sector has no mandatory carbon accounting. Some research labs voluntarily disclose the energy and carbon cost of their experiments, but many do not. Even when they do, the methodologies vary widely, making comparisons difficult.

A 2022 paper proposed a standardized “model card” that would include not only performance metrics but also energy consumption, carbon emissions, and hardware details. The idea is to make environmental cost a first-class consideration in model development, alongside accuracy and speed. Some conferences have begun encouraging such disclosures, but adoption remains patchy.

Rui Mendes notes that this is a classic case of externalities: the benefits of a large model are captured by the organization that trains it, while the environmental costs are distributed globally. Without mechanisms to internalize those costs—whether through regulation, market pricing, or cultural norms—the incentive to minimize them remains weak.

Geographic Arbitrage and Carbon Accounting Games

One subtle but important factor is the geographic arbitrage of carbon intensity. A training run powered by a grid with high renewable penetration will have a lower operational carbon footprint than one powered by coal. Organizations can, and do, choose locations partly on this basis. But this raises questions about additionality: does siting a data center in a region with clean energy actually reduce global emissions, or does it simply displace other consumers onto dirtier sources?

The answer depends on the specifics of the grid. In some cases, new demand is met by bringing additional renewable capacity online, which can have a genuinely positive effect. In others, the clean energy is already fully utilized, and new demand is met by fossil fuels. Without careful accounting, location choice can become a form of greenwashing—claiming low emissions based on average grid intensity while ignoring the marginal impact of new load.

There is also a temporal dimension. Training runs often last weeks, and the carbon intensity of a grid varies hour by hour. Some researchers have proposed “carbon-aware” scheduling, shifting workloads to times when renewable penetration is high. This is technically feasible but requires coordination that is not yet common practice.

Water Use in Water-Stressed Regions

The geographic dimension also applies to water. Data centers are often sited in arid regions because of low land costs and abundant solar energy potential. But these same regions are frequently water-stressed. In places like Arizona, New Mexico, and parts of Chile, data center water consumption has become a contentious issue. Local communities question whether the economic benefits of hosting these facilities outweigh the strain on already scarce water resources.

Some facilities use air-cooled or closed-loop systems that minimize direct water use, but these systems typically require more energy, creating a trade-off between water and carbon. In a world where both carbon budgets and freshwater supplies are under pressure, optimizing for one at the expense of the other is a delicate balancing act.

Rui Mendes sees this as a classic systems optimization problem with no easy answer. The variables are interconnected, the constraints are local, and the objectives are global. A solution that works in Scandinavia may fail in the American Southwest. A one-size-fits-all approach is unlikely to succeed.

Hardware Embodied Carbon: The Overlooked Giant

While operational energy and water use are beginning to receive attention, the embodied carbon and resource extraction associated with hardware manufacturing remain largely invisible. Producing a single GPU involves hundreds of steps across multiple countries, each with its own energy mix and environmental regulations. The supply chain is opaque, and manufacturers are not required to disclose lifecycle emissions.

Some estimates suggest that embodied carbon can equal or exceed operational carbon over the lifetime of a server, especially when the server is replaced every few years. If the industry continues to shorten hardware refresh cycles in pursuit of performance gains, the embodied carbon problem will only worsen. Extending the useful life of accelerators, designing for recyclability, and improving manufacturing transparency are all necessary steps—but they require coordination across an industry that currently has little incentive to act.

Rethinking Efficiency Metrics

Current efficiency metrics in machine learning focus almost exclusively on performance per compute unit: how many operations per second, how many parameters per watt. These metrics drive hardware design, model architecture, and even research agendas. But they ignore the full environmental picture.

Rui Mendes argues that we need new metrics that account for the entire lifecycle. What is the total carbon emitted per unit of useful work over a model’s lifetime, including training, deployment, and inference? What is the water footprint per query served? How much rare material was extracted to build the hardware, and what was the social cost in the communities where mining took place?

These questions are not easy to answer, but they are essential if the field is to take responsibility for its environmental impact. Without them, efficiency improvements can become a shell game: reducing operational energy while increasing embodied energy, or shifting water consumption from one watershed to another.

Inference: The Long Tail of Environmental Cost

Much of the discussion focuses on training, but inference—the process of using a trained model to make predictions—also carries a significant environmental burden. A model may be trained once but used millions or billions of times. For large-scale deployed models, the cumulative energy and water cost of inference can far exceed that of training.

Consider a model that answers millions of queries per day. Each query requires a forward pass through the network, consuming energy and generating heat. Over the model’s operational lifetime, the inference cost can dwarf the training cost. Yet inference efficiency receives far less attention than training efficiency in both research and public discourse.

Rui Mendes points out that this is a classic systems blind spot: we focus on the one-time cost of creation and ignore the ongoing cost of operation. A full lifecycle assessment must include both, and for many widely deployed models, the operational phase dominates.

What Can Be Done? A Systems Approach

Addressing the environmental cost of large-scale computing requires action on multiple fronts. No single intervention will solve the problem, but a combination of technical, organizational, and policy changes could bend the curve.

1. Transparent Reporting

Mandatory disclosure of energy use, carbon emissions, water consumption, and hardware lifecycle data would create accountability and enable better decision-making. Standardized reporting frameworks, like those proposed in recent research, could make environmental costs visible and comparable across projects.

2. Carbon-Aware Scheduling

Shifting training workloads to times and places with lower carbon intensity can reduce emissions without sacrificing performance. This requires coordination between cloud providers, grid operators, and research teams, but the technical barriers are surmountable.

3. Hardware Longevity and Circularity

Extending the useful life of accelerators, designing for repairability and recyclability, and creating markets for refurbished hardware can reduce embodied carbon and e-waste. This requires changes in procurement practices and manufacturer incentives.

4. Efficiency Research Beyond FLOPs

Research into model compression, sparse architectures, and efficient inference can reduce the environmental cost per query. But these efforts must be guided by full-scope metrics that capture the whole picture, not just computational efficiency.

5. Siting and Infrastructure Policy

Local governments can play a role by requiring environmental impact assessments for large data centers, including water use and grid effects. Siting decisions should consider not only economic factors but also the strain on local resources and communities.

Frequently Asked Questions

How much energy does training a single large model actually consume?

Estimates vary widely depending on model size, hardware efficiency, and grid carbon intensity. A 2019 study found that training a large natural language processing model emitted around 284 tonnes of CO₂ equivalent. Since then, models have grown significantly, and the largest training runs today likely emit several thousand tonnes. To put that in perspective, the average American car emits about 4.6 tonnes of CO₂ per year.

Why is water use a concern for data centers?

Data centers consume water both directly, through cooling systems that use evaporative methods, and indirectly, through the water used to generate the electricity they consume. In water-stressed regions, this can compete with agricultural, residential, and ecological needs. A single large training run can consume hundreds of thousands of liters of freshwater when both direct and indirect use are accounted for.

Does the environmental cost continue after training is complete?

Yes. Once a model is deployed, every query or prediction it makes consumes energy and water. For models that serve millions of users, the cumulative cost of inference can far exceed the cost of training. This ongoing operational impact is often overlooked in discussions that focus only on the training phase.

Can renewable energy solve the problem?

Renewable energy can significantly reduce the carbon footprint of training and inference, but it is not a complete solution. Water use, hardware lifecycle impacts, and land use for renewable infrastructure all remain concerns. Additionally, the availability of renewable energy varies by location and time, and new demand can sometimes displace other users onto fossil fuel sources if not carefully managed.

Conclusion: Seeing the Whole System

The environmental cost of training large models is not a simple problem with a single villain. It is an emergent property of a system that prioritizes performance above all else, that externalizes environmental costs, and that lacks the feedback loops necessary to self-correct. Rui Mendes believes that the first step toward a solution is simply to see the system clearly—to trace the connections from the mine to the data center to the cloud, and to recognize that every digital achievement rests on a physical foundation.

This is not an argument against progress. It is an argument for a more honest accounting of what progress costs, and for a more thoughtful approach to how we pursue it. The goal is not to stop building models, but to build them in a way that respects the finite resources of the planet we all share.

The next time a new record is set on a benchmark, Rui Mendes will still be curious. But his first question will not be about the score. It will be about the smoke, the water, and the earth that made it possible.

Every time a new, jaw-dropping model gets announced—one that writes poetry, generates photorealistic images, or translates languages on the fly—the headlines focus on what it can do. What they almost never mention is what it took to build it. Not the cleverness of the code or the elegance of the architecture, but the physical stuff: the electricity, the water, the metals, the heat. I started pulling on this thread a while back, and the more I learned, the more it felt like we’re all driving a fleet of invisible, smoke-belching trucks through the digital world without ever seeing the exhaust.

This isn’t a story about whether the technology is good or bad. It’s a story about systems—how they connect, what they consume, and what we choose to ignore when the machinery is out of sight.

What a Single Training Run Actually Demands

Let’s get concrete. Training a large model isn’t like leaving your laptop on overnight to render a video. It’s a coordinated assault on a problem, using thousands of specialized chips—GPUs or TPUs—running flat-out for weeks or months. Each chip might draw 300-400 watts. Multiply that by 10,000 chips, then by 24 hours, then by 90 days. The numbers get absurd fast.

One of the few public estimates, from a team at the University of Massachusetts Amherst, pegged the carbon emissions of a single large training run at over 284 tonnes of CO₂ equivalent. That’s roughly the lifetime output of five average cars, including their manufacturing. And that’s just the electricity for the run itself—not the trial runs, the failed experiments, the hyperparameter tuning that came before. The final model is the tip of an iceberg, and the hidden mass below the waterline is made of burned coal and natural gas.

It’s Not Just About Electricity

When people talk about making data centers “green,” they usually mean buying renewable energy. That matters, but it’s only one slice of the pie. The grid mix—how much of the local power actually comes from wind, solar, hydro, or nuclear versus coal and gas—varies wildly by region. A training run in a data center hooked to a coal-heavy grid leaves a much deeper carbon scar than one powered by a hydroelectric dam. Same code, same model, wildly different physical consequences.

Then there’s water. Data centers run hot, and keeping thousands of chips from melting themselves requires serious cooling. In many facilities, that means evaporative cooling towers that drink millions of gallons of freshwater. A 2023 study out of UC Riverside put the number at up to 700,000 liters for a single large training run. In places like Arizona or Spain, that water is pulled straight out of watersheds already strained by drought and agriculture. It’s a quiet competition for a resource most people don’t associate with software.

The Ghost Emissions in the Hardware

Before a server ever powers on, it’s already carrying a carbon debt. Manufacturing GPUs and other accelerators involves mining rare earth elements, smelting, chemical baths, and fabrication in energy-hungry foundries. The supply chain sprawls across continents. A single server might embody hundreds of kilograms of CO₂ before it’s even plugged in. And in this industry, hardware doesn’t age gracefully. The relentless pace of improvement means racks get swapped out every three to five years, not because they’re broken, but because they’re no longer competitive. That churn produces mountains of e-waste—circuit boards, batteries, and cables that are notoriously difficult to recycle safely. We upgrade our models and quietly landfill the old ones.

The Efficiency Trap

There’s a comforting story that goes: “Chips are getting more efficient, so the problem will solve itself.” It’s partly true. A modern GPU delivers far more calculations per watt than one from five years ago. But efficiency has a funny way of backfiring. When something gets cheaper or faster, we don’t use less of it—we use more. Economists call it the rebound effect, and it’s alive and well in the world of large-scale computing.

Every efficiency gain gets swallowed by ambition. The next model isn’t just a little bigger; it’s ten times bigger. The compute budget expands to absorb whatever headroom the engineers created. Total energy use doesn’t drop; it climbs. We’re optimizing the parts while the whole system grows hungrier. It’s a classic systems trap—one that feels familiar to anyone who’s watched traffic expand to fill a new highway lane.

Where the Data Centers Land—and Who Pays

Data centers don’t sprout randomly. They cluster where land is cheap, tax breaks are generous, and power is abundant. Often, “abundant power” means fossil fuels. Virginia’s data center alley, for instance, sits on a grid that’s only about 5% renewable. Compare that to Quebec, where hydro dominates, and the same workload can have a radically different carbon profile. The location decision—made by a cloud provider or a corporate IT team—is one of the biggest environmental levers, and it’s almost never visible to the people using the service.

Then there’s the local fallout. A data center in a desert doesn’t just use water; it takes water that might have gone to a farm or a household. Backup diesel generators cough out particulates that settle in nearby lungs. The burdens aren’t spread evenly across a spreadsheet; they land on specific communities, often ones with the least power to push back.

The Transparency Problem

Here’s where it gets genuinely frustrating. Most organizations that train large models don’t publish their energy numbers, their water consumption, or their grid mix. When they do share figures, the accounting can be creative. Buying renewable energy certificates (RECs) lets a company claim “100% renewable” even if the actual electrons powering the servers came from a gas plant. The certificates might fund a wind farm somewhere, but they don’t change the physical reality at the data center’s plug.

There are grassroots tools—CodeCarbon, ML CO2 Impact—that try to estimate emissions based on hardware type, runtime, and location. They’re useful, but they’re guesses. Without mandatory reporting, the real numbers stay locked inside corporate spreadsheets. Sunlight is scarce.

Paths Toward a Lighter Touch

If we’re going to keep building these systems—and it seems we are—the question shifts from “should we stop?” to “how do we do it with less damage?” The answers aren’t purely technical. They’re about choices, incentives, and who gets a seat at the table.

Time-shifting the workload. Training doesn’t have to happen right now. It can be scheduled for when the grid is cleanest—midday when solar is peaking, or windy nights when turbines are spinning. Some data centers already shift compute loads in real time to chase renewable availability. It requires flexible scheduling and a willingness to accept slightly longer training times, but the carbon savings can be substantial.

Keeping hardware alive longer. Extending server life through modular upgrades rather than wholesale replacement cuts embodied emissions. A few large operators are experimenting with designs where individual components—CPUs, accelerators, memory—can be swapped independently. It’s less glamorous than buying the latest rack, but it keeps materials out of the landfill longer.

Smaller, smarter models. A quiet revolution is happening in model efficiency. Techniques like distillation, pruning, and quantization can shrink a model to a fraction of its original size while keeping most of its capability. A distilled model might lose a point or two of accuracy but train in a tenth of the time and run on a fraction of the hardware. For many real-world tasks, that’s a trade worth making. It also opens the door for groups without massive compute budgets to participate.

Mandatory disclosure. The simplest, most powerful lever might be this: require any training run above a certain compute threshold to report its energy consumption, grid mix, and water usage. Make the numbers public. Once the environmental cost is visible, it becomes a factor in decision-making. Researchers might choose cleaner regions. Companies might compete on efficiency, not just capability. Transparency doesn’t fix everything, but it makes the trade-offs impossible to ignore.

Zooming Out: The Bigger Pattern

The environmental weight of training large models isn’t an isolated problem. It’s one expression of a broader habit: we build complex digital systems without accounting for their full physical lifecycle. You see the same pattern in cryptocurrency mining, in the explosion of IoT devices, in the endless expansion of streaming infrastructure. Every new service adds another layer of concrete, steel, copper, and lithium. The cloud isn’t weightless. It’s heavy, and it sits on real ground.

This isn’t a call to halt progress. It’s a call to make progress conscious. We should be asking not just “Can we build this?” but “Should we build it this way, at this scale, in this place?” Some models will be worth their footprint—the ones that accelerate drug discovery, improve climate forecasts, or design better materials. Others are trained for marginal gains on benchmarks that don’t translate to anything useful. The difference matters, and it should be deliberate, not accidental.

FAQ

Why is training a large model so much more energy-intensive than regular computing?

Training means pushing enormous datasets through networks with billions of parameters, over and over, for weeks or months, across thousands of power-hungry processors. Each chip draws hundreds of watts, and the cluster adds up to a small power plant’s worth of demand. On top of that, the cooling systems needed to keep everything from overheating consume significant energy and water on their own.

Can switching to renewable energy fix the whole problem?

Renewables slash the carbon emissions from electricity use, and that’s a big deal. But they don’t erase the embodied carbon from manufacturing the hardware, nor do they solve the water consumption issue. Clean power is a necessary piece of the puzzle, not the whole picture.

What can an ordinary person do about something so far removed from daily life?

Individual choices won’t directly change how large organizations train their models, but collective pressure can. Supporting transparency efforts, asking cloud providers about their energy sources, and pushing for policies that require emissions disclosure all add up. On a personal level, favoring services that publicly commit to environmental accountability sends a market signal, however small.

Are smaller, more efficient models really as good as the massive ones?

In many practical situations, yes. Distillation, pruning, and quantization can produce models that keep most of the original’s capability while being dramatically cheaper to train and run. You might trade a tiny bit of accuracy for a huge cut in energy use. For a lot of applications, the lean version is more than enough—and sometimes the only version that makes sense.

The environmental cost of training large models isn’t a reason to walk away from the field. It’s a reason to bring the same rigor and systems thinking we apply to the models themselves to the infrastructure that births them. Measure it. Disclose it. Optimize not just for accuracy, but for the full footprint. The smartest systems ought to rest on a foundation of responsibility—to each other, and to the planet that hosts all this computation.

Why We Should Look at the Power Cord, Not Just the Output

When a new machine learning system nails a breakthrough—say, it suddenly gets uncannily good at recognizing faces or translating obscure languages—the headlines gush about the cleverness. But behind every one of those wins is a physical reality that rarely gets a headline: racks of servers humming for weeks or months, cooling systems gulping water and pushing heat into the air, and a steady, invisible drain on the electrical grid. Rui Mendes has spent years thinking about the systems that make digital life tick, and lately he’s been drawn to the material footprint of these computational feats. The question isn’t whether the results are impressive. It’s whether we’ve even started to tally what they actually cost.

Training a big model isn’t a one-click affair. It’s an industrial process, often spread across thousands of specialized processors running in parallel. The energy draw can rival that of a small town, and the water used to cool the data centers can strain local supplies. Yet when a new model is announced, these numbers are almost never part of the story. Rui thinks that if we’re serious about building responsible technology, we need to start measuring what matters—and that includes the environmental load that comes with every training run.

The Scale of a Single Training Run

To get a feel for the problem, let’s look at a concrete example. Take a large natural language model trained on a web-scraped corpus. The process might involve hundreds of graphics processors running flat-out for several months. Each of those chips draws power, and the facility housing them needs constant cooling to keep things from melting down. The electricity consumption for a single training run can easily reach hundreds of megawatt-hours—roughly the annual energy use of dozens of households.

But electricity is only part of the story. The hardware itself carries a carbon and resource cost. Specialized chips require rare minerals, and their manufacturing is energy-intensive. If we’re going to talk about environmental impact, we should include the embedded emissions of the equipment, not just the operational energy. Rui often points out that this is where systems thinking becomes non-negotiable: a narrow focus on one metric, like kilowatt-hours during training, can blind us to the bigger picture.

Water Use in the Data Center

Another factor that’s easy to overlook is water. Many data centers rely on evaporative cooling, which consumes huge amounts of fresh water. In regions already dealing with water stress, this can pit the needs of a data center against those of the surrounding community. A single large training run can evaporate millions of liters of water—water that might otherwise go to crops or drinking supplies. Rui finds this especially jarring because the digital world is so often imagined as weightless and clean, when in fact it’s thirsty and tethered to very material resources.

Location matters enormously. A data center plugged into a coal-heavy grid will have a much higher carbon footprint than one running on renewables. Similarly, water use in an arid region hits differently than in a water-rich area. But transparency about these factors is rare, which makes it tough for researchers—or anyone else—to assess the true environmental load of a given model.

Why Efficiency Gains Can Fool Us

There’s a common counterargument: as hardware gets more efficient, the energy cost per computation drops. That’s true, but it doesn’t necessarily shrink overall energy consumption. In practice, efficiency gains often lead to bigger models and more ambitious training runs—a phenomenon sometimes called the rebound effect. When it becomes cheaper to train a model, the incentive is to train an even larger one, chasing marginal performance improvements that demand disproportionately more resources.

Rui sees this as a classic systems trap. Without a deliberate effort to cap resource use or account for environmental costs, efficiency improvements can actually accelerate consumption. The real question isn’t how efficient a single operation is, but how much total energy and water a project consumes from start to finish—and whether that total is justified by the value it creates.

Comparing Training to Other Energy-Intensive Activities

To put the numbers in perspective, training a single large model can emit as much carbon dioxide as several cars do over their entire lifetimes. Some estimates peg the carbon footprint of a state-of-the-art training run at over 280 tonnes of CO₂ equivalent. That’s roughly the lifetime emissions of five average cars, including their manufacturing. And since many models are trained multiple times during experimentation, the cumulative impact grows fast.

Rui finds it useful to compare these figures to activities that get far more environmental scrutiny. Air travel, for instance, is constantly discussed in terms of its carbon footprint, and many people actively try to fly less. Yet the computational equivalent—training a massive model—rarely enters the same conversation, even though its emissions can be comparable to a transatlantic flight for each researcher involved. The gap in awareness is something Rui thinks we need to close.

The Lifecycle Perspective: Beyond Training

Training is only one phase of a model’s life. Once deployed, a model keeps drawing energy every time it’s used. For a popular service handling millions of requests a day, the inference phase can dwarf the training phase in total energy consumption. This is especially true for models that need powerful hardware to run, like those used in real-time video analysis or large-scale recommendation systems.

Rui emphasizes that a full lifecycle assessment is necessary to understand the true environmental cost. That means accounting for the energy and materials to manufacture the hardware, the energy for training, the energy for inference, and the energy for eventual decommissioning and recycling. Without this wider view, we risk optimizing one phase while ignoring bigger impacts elsewhere.

What Could Transparency Look Like?

One of the biggest roadblocks to tackling this issue is the lack of transparency. Most organizations don’t disclose the energy consumption, carbon emissions, or water use tied to their models. Rui argues this has to change. Just as some companies now publish environmental impact reports for their physical products, developers of large models should provide similar disclosures. That could include the total electricity used during training, the carbon intensity of the grid at the data center location, and the water consumed for cooling.

Such transparency would let researchers, policymakers, and the public make informed comparisons. It would also create an incentive for developers to minimize environmental impact, rather than simply chasing maximum model performance. Rui believes that measurement is the first step toward management, and without open reporting, the environmental costs will stay invisible.

Rethinking the Metrics of Success

The current culture of machine learning research rewards ever-larger models and incremental improvements on benchmark datasets. But those benchmarks rarely include environmental cost as a factor. Rui suggests the community should consider new metrics that balance performance against resource consumption. For example, a model that hits 95% accuracy with half the energy use of a model that reaches 96% might be the smarter choice from a systems perspective.

This shift would require changes in how research is funded, published, and celebrated. Conferences and journals could require submissions to include estimates of computational resources used. Funding agencies could prioritize projects that demonstrate efficiency alongside effectiveness. Rui sees this as a chance to align the values of the research community with broader societal goals.

Practical Steps for Developers and Organizations

While large-scale change needs systemic shifts, there are practical steps individual developers and organizations can take. Choosing data center regions powered by renewable energy is one straightforward option. Using more efficient hardware and optimizing training algorithms to require fewer computational steps can also make a difference. Rui points out that simply being mindful of the issue and measuring energy use is a meaningful first step.

Another approach is to ask whether a large model is truly necessary for the task at hand. In some cases, a smaller, more specialized model can achieve comparable results with a fraction of the resources. Transfer learning, where a pre-trained model is fine-tuned for a specific task, can also reduce the need to train from scratch. Rui encourages developers to ask: is the marginal improvement in performance worth the additional environmental cost?

The Role of Policy and Regulation

Individual and organizational efforts matter, but Rui believes policy interventions will eventually be necessary to drive widespread change. Possible measures include requiring environmental impact disclosures for large training runs, setting efficiency standards for data centers, or incorporating environmental costs into the pricing of cloud computing services. Such policies could help internalize the externalities that are currently ignored.

Regulation could also spur innovation in energy-efficient hardware and cooling technologies. If the environmental costs of computation were reflected in its price, there would be a stronger market incentive to develop greener alternatives. Rui sees this as a natural extension of existing environmental regulations that apply to other industries.

Why This Matters for Everyone

The environmental impact of large-scale computation isn’t just a concern for researchers or tech companies. As these models become embedded in everyday services—from search engines to healthcare diagnostics—their collective footprint grows. The energy and water used to train and run these models are drawn from shared resources. Rui emphasizes that this is a collective challenge that requires collective awareness and action.

For users, being informed about the hidden costs of digital services can lead to more conscious choices. For developers, it can inspire more efficient design. For policymakers, it can highlight the need for standards and regulations. Rui’s systems-minded perspective reminds us that everything is connected: the cloud is not weightless, and every query has a cost.

Frequently Asked Questions

How much energy does training a large model actually use?

The energy consumption varies widely depending on the model size, hardware efficiency, and duration of training. Some large models have been reported to consume hundreds of megawatt-hours of electricity during training—enough to power dozens of homes for a year. However, exact figures are often not publicly disclosed, making it difficult to generalize.

Does the environmental impact stop once the model is trained?

No. Deployed models continue to consume energy every time they are used for inference. For widely used services, the cumulative energy from inference can exceed the training energy over time. Additionally, the hardware lifecycle—from manufacturing to disposal—adds further environmental costs.

What can be done to reduce the environmental footprint of large models?

Several strategies can help: using data centers powered by renewable energy, improving hardware efficiency, optimizing algorithms to require less computation, and being transparent about resource consumption. On a systemic level, policies that require disclosure of environmental impact and incentivize efficiency could drive broader change.