Rui Mendes here. I spend a lot of time thinking about the physical backbone of our digital lives—the data centers. We stream, we scroll, we store, and behind every click there’s a warehouse humming with servers, pulling electricity at a scale that’s genuinely hard to picture. The industry’s big answer to its own energy appetite has been a pivot toward renewables. Power purchase agreements for wind and solar are now standard fare for the major players. On paper, it looks like a win. But when you follow the actual electrons and the material flows, the picture gets messier. The question isn’t whether renewable energy is good—it’s whether plugging it into a wasteful system really solves the problem.

The Intermittency Blind Spot

Data centers need power that never wavers. A 99.999% uptime requirement doesn’t leave room for a cloudy, windless afternoon. When a hyperscale facility signs a contract for renewable energy, it’s rarely a direct wire from a nearby solar farm. The electrons feeding the servers still come from the regional grid, which balances supply and demand in real time. At night, or when the wind dies, that grid leans on whatever is available—often natural gas, sometimes coal. The renewable energy certificates (RECs) the data center buys are meant to bridge this gap, but they’re an accounting tool, not a physical solution. In many markets, RECs are so cheap and abundant that they don’t drive new clean generation. The data center still pulls from a grid that fires up fossil plants when renewables fall short. The carbon math might look tidy in a sustainability report, but the smokestacks don’t lie.

The Water You Don’t See

Electricity grabs the headlines, but data centers are also thirsty. A single hyperscale facility can guzzle millions of gallons of water a day, mostly for cooling. In arid regions—think Arizona, Chile, or parts of Spain—that’s a direct competition with farms and households. Many data centers use evaporative cooling because it’s energy-efficient, but that efficiency comes at the cost of water. You can run on 100% renewable power and still drain an aquifer that took millennia to form. The water cycle itself is energy-intensive: pumping, treating, and distributing water burns power, often from fossil sources. So even the water footprint has a carbon echo. Sustainability reports tend to fixate on electricity, but a facility that’s carbon-neutral on paper can still leave a community dry.

The Concrete and Silicon You Can’t Offset

Building a data center means pouring thousands of tons of concrete and erecting steel frames. Cement production alone coughs up about 8% of global CO₂ emissions. Then there are the servers themselves: manufacturing semiconductors is an energy-hungry, chemical-heavy process. A single server carries a hefty carbon debt before it ever processes its first packet. When we talk about “green data centers,” we usually mean operational energy use. But the upfront emissions—the concrete, the steel, the silicon—can rival or even exceed a facility’s lifetime operational footprint. If a company builds a new data center every quarter to keep pace with demand, those embodied emissions pile up fast. Renewables don’t touch the carbon baked into the building and the hardware.

The Rebound Effect

Here’s a systems quirk that doesn’t get enough airtime: making data centers more efficient or powering them with renewables can actually increase total energy consumption. It’s the rebound effect, a classic paradox. When a service becomes cheaper or feels less environmentally damaging, demand for it tends to swell. Greener cloud computing invites more cloud computing. More efficient streaming leads to higher-resolution video and more hours watched. The efficiency gains get eaten by growth. We’re watching this in real time. Despite impressive improvements in power usage effectiveness (PUE) and a flood of renewable contracts, data center electricity consumption keeps climbing steeply. The green label can become a permission slip to expand, not a brake that forces genuine reduction.

The Evening Ramp and Grid Strain

Data centers aren’t like other industrial loads. They run 24/7 at a near-constant draw. That flat, unyielding demand profile is a headache for grids with lots of solar. Solar generation peaks at midday and drops off a cliff at sunset, creating the famous “duck curve” where net load ramps steeply in the evening. A data center powered by solar RECs might look spotless on a spreadsheet, but physically it’s demanding power when the sun is gone. That evening ramp is almost always met by fossil fuels—usually natural gas peaker plants. In some regions, data centers are now the single biggest driver of new gas plant construction. The renewable energy they buy doesn’t erase the need for firm, dispatchable backup; it just shifts the accounting. The grid has to absorb that mismatch, often at a high carbon cost.

Geography Matters More Than We Admit

Renewable energy isn’t spread evenly across the map. The sunniest, windiest spots are often far from the fiber optic backbones and population hubs that data centers need. Building a data center in a remote area with abundant renewables means stringing new transmission lines, which face years of permitting battles and community pushback. The alternative—building where the grid is already strong but renewables are scarce—means leaning on RECs that don’t reflect physical reality. Either way, the data center’s location locks in a certain energy profile. In Northern Virginia, home to the world’s densest cluster of data centers, the local utility has proposed new gas plants to meet demand, even as the data center operators ink renewable deals. The geography of the grid simply doesn’t match the geography of renewable potential.

The E-Waste Afterlife

Servers live fast and die young—three to five years, typically. Then they’re swapped for newer, more efficient models. The old hardware enters a waste stream that’s notoriously hard to trace. Some gets recycled properly; a lot gets shipped to developing countries where informal recyclers burn circuit boards to salvage precious metals, releasing dioxins and heavy metals into the air and soil. The renewable energy feeding the new servers does nothing about the toxic legacy of the old ones. A genuinely systems-minded approach would demand circularity: design for disassembly, closed-loop material recovery, and extended producer responsibility. But the industry’s fixation on operational energy and renewable procurement leaves the back end of the hardware lifecycle mostly unexamined.

What a Systems Approach Would Actually Demand

If we’re serious about fitting digital infrastructure within ecological limits, renewable energy procurement is just one piece of the puzzle. A systems approach would require:

• Time-matched, location-matched clean energy. Not just annual RECs, but hourly matching of consumption with carbon-free generation, ideally within the same grid region. This forces data centers to own the intermittency problem instead of offloading it onto the grid.
• Water-neutral or water-positive operations. In water-stressed basins, data centers should use closed-loop cooling or dry cooling, even if it costs more energy. The trade-off between water and carbon needs to be made explicit, not swept under the rug.
• Embodied carbon budgets. Construction and hardware manufacturing emissions should be tracked, reported, and capped. This means valuing existing facilities and extending server lifespans, not just chasing the newest, shiniest hardware.
• Demand-side thinking. Instead of asking “how do we power this growing load cleanly?” we should ask “does this load need to exist in the first place?” That means questioning the necessity of energy-intensive applications, from cryptocurrency mining to certain AI training runs, and designing software for efficiency, not just speed.

Frequently Asked Questions

Why can’t data centers just use batteries to store renewable energy?

Battery storage is improving, but the scale needed for a typical data center is staggering. A 100-megawatt facility would need hundreds of megawatt-hours of storage to cover a single windless night. Lithium-ion batteries come with their own supply chain headaches, including mining impacts and limited lifespans. While batteries can help with short-term grid services, they’re not yet a full substitute for firm, dispatchable generation. The materials and energy needed to build that much storage also carry a significant carbon footprint that’s rarely accounted for.

Are there any data centers that actually run on 100% renewable energy 24/7?

A few small-scale projects have achieved 24/7 carbon-free energy matching, often by combining local renewables with storage and backup from hydroelectric or geothermal sources. But for the vast majority of hyperscale and colocation facilities, true 24/7 matching remains aspirational. The data and tracking infrastructure to verify hourly matching is still being developed, and in most grid regions, the physical supply simply isn’t there. What’s more common is “100% renewable” on an annual net basis, which masks the hour-by-hour reliance on fossil fuels.

What can a regular person do about data center energy use?

Individual actions matter, but they’re limited. You can choose cloud providers and services that are transparent about their energy practices and that pursue 24/7 matching. You can reduce your own data footprint—delete unused files, stream at lower resolutions, keep devices longer. But the real power lies in policy and collective action. Support local opposition to new gas plants being built for data centers. Advocate for transparency laws that require companies to report not just REC purchases but actual hourly energy sources, water use, and hardware lifecycle impacts. The systems that shape data center growth are political and economic, not just technological.

When a hyperscale cloud provider announces a new solar farm or a long-term wind contract, the press release practically writes itself. The numbers are huge, the intentions sound progressive, and the charts all point toward a net-zero future. But if you trace where the actual electrons go—and where the materials come from—a much messier picture starts to form. The problem isn’t the renewables. It’s that we’ve started treating a single procurement checkbox as the whole story, while the physical reality of digital growth keeps churning away, largely unexamined.

I’ve spent years looking at infrastructure systems, and one pattern shows up everywhere: we optimize for one metric and quietly externalize the rest. In data centers, that metric is the power usage effectiveness ratio and the percentage of renewables on the balance sheet. But a data center isn’t a lightbulb. It’s a node in a sprawling, material-hungry network that stretches from rare earth mines in Inner Mongolia to water tables in Arizona. If we only talk about matching kilowatt-hours with wind and solar, we miss the actual shape of the problem.

The Matching Game and Its Limits

Most large operators now claim to be “100% renewable.” What that usually means is they buy enough renewable energy certificates or sign enough power purchase agreements to cover their annual electricity consumption on paper. It’s a meaningful step—it channels capital into wind and solar projects that might not otherwise get built. But it’s also a statistical abstraction. A data center in Virginia might be burning natural gas and fissioning uranium at 3 a.m., while its certificates are satisfied by a wind farm in Oklahoma that generates most of its power at night. The grid doesn’t see a matched load. It sees a fossil plant ramping up and down to follow the data center’s flat, relentless demand.

Grid engineers have known about this temporal mismatch for years, but it rarely makes it into the glossy sustainability reports. A facility that runs at near-constant load requires a near-constant supply of power. Wind and solar are anything but constant. The gap gets filled by whatever is on the grid at that moment—usually gas, sometimes coal, occasionally hydro. The certificates make the accounting look tidy, but the physical electrons flowing into the servers tell a different story. This isn’t a failure of renewables. It’s a failure of how we talk about them. We’ve confused annual net-zero accounting with real-time decarbonization, and the two are not the same.

The Water-Energy Blind Spot

Even if a data center could run entirely on on-site solar and battery storage—a physical impossibility for most facilities—there’s another resource that rarely makes it into the sustainability report: water. Cooling towers evaporate millions of gallons a year, often in regions already facing water stress. A single large data center can consume as much water as a small city. In places like Phoenix, Arizona, or Loudoun County, Virginia, this creates a quiet competition between server racks and residential taps.

The water footprint usually gets reported as a separate metric, if it’s reported at all. But it’s deeply tangled up with the energy question. Different cooling technologies come with different energy-water trade-offs. Evaporative cooling uses less electricity but more water. Dry cooling saves water but bumps energy consumption by 5-10%. In a world where we’re racing to electrify everything—cars, heating, industrial processes—that extra energy demand cascades back into the grid, requiring more generation, more transmission lines, and more land. The system is coupled. Optimize one variable in isolation, and you often degrade another.

The Circularity Gap in Hardware

Servers have a lifespan of three to five years. After that, they’re decommissioned, and the industry’s standard practice is a mix of resale, recycling, and landfill. The metals inside—copper, aluminum, gold, palladium—require enormous energy to extract and refine in the first place. A single server’s embodied carbon, from mine to factory to rack, can rival its operational carbon over its entire useful life. Yet most renewable energy commitments cover only the operational phase. The supply chain, which accounts for the majority of a tech company’s total carbon footprint, sits outside the boundary.

This is where the systems thinking gets uncomfortable. If we’re truly concerned about the climate impact of data centers, we should be asking not just “how is it powered?” but “how often do we replace it, and what happens to the old one?” Extending server life from four to six years, or designing for component-level upgrades instead of full-box swaps, could reduce embodied carbon more than any power purchase agreement. But that would require rethinking depreciation schedules, supply contracts, and the performance-obsessed culture of IT procurement. It’s easier to buy a wind contract and call it done.

Land Use and the Spatial Footprint

Renewable energy requires land—lots of it. A 100-megawatt solar farm covers roughly 500 to 700 acres. A data center campus can easily demand 300 megawatts or more. When a tech company signs a power purchase agreement for a new solar installation, that land is removed from other potential uses: agriculture, habitat, or community development. In rural areas, this can create tension between clean energy goals and local land-use priorities. In some cases, solar farms are sited on prime farmland, raising questions about food security and soil health that never appear in the data center’s sustainability report.

Wind turbines have a smaller direct footprint but require spacing that fragments landscapes and can affect wildlife corridors. The point isn’t that renewables are bad—they’re essential. The point is that “100% renewable” claims obscure the physical reality that digital infrastructure is consuming not just electricity, but territory. As data centers proliferate to support streaming, cloud computing, and the insatiable demand for real-time everything, the land requirements of their energy supply chains will become a geopolitical issue. We’re already seeing pushback in Ireland, the Netherlands, and Singapore, where data center moratoriums have been imposed due to grid constraints.

The Utilization Paradox

Here’s a statistic that should make us pause: the average server utilization in many data centers hovers around 12-18%. That means 80-85% of the computing capacity is sitting idle, drawing power, generating heat, and aging toward obsolescence. Virtualization and cloud computing were supposed to fix this by pooling resources, but the gains have been offset by the sheer growth in demand. We keep building more data centers, filling them with servers that mostly wait, and then congratulating ourselves for powering them with renewables.

If we doubled average utilization to 30-40%—still leaving headroom for spikes—we could theoretically halve the number of physical servers needed for the same workload. That would reduce embodied carbon, water use, land use, and grid strain simultaneously. It’s a point of intervention that addresses multiple problems at once. But it requires coordination across competing cloud providers, changes to service-level agreements, and a cultural shift away from over-provisioning as a risk management strategy. The technical solutions exist; the organizational incentives don’t align.

Rethinking the Metric of Success

What if we measured data center sustainability not by the percentage of renewables purchased, but by the total resource throughput per unit of useful computation? This would combine energy, water, materials, and land into a single efficiency metric that reflects the physical reality of operating digital infrastructure. It would expose the trade-offs that current reporting hides. A facility in a water-scarce region might score poorly despite its solar panels. A cloud region with high server utilization would outperform one with low utilization, even if both buy the same renewable certificates.

This kind of metric is harder to calculate and harder to communicate. It requires transparency that most operators are unwilling to provide. But it would shift the conversation from “are we green?” to “are we doing this efficiently?”—which is a more honest question. Efficiency, in the thermodynamic sense, is about minimizing waste across all inputs. Renewable energy addresses only one input. The rest of the waste stream—heat, water vapor, decommissioned hardware, transmission losses—continues to grow in proportion to our digital appetite.

The Role of Demand Management

There’s another factor that gets almost no attention: reducing the demand for data center services in the first place. Not through austerity or turning back the clock, but through smarter design of the digital services that run on these servers. Video streaming at resolutions the human eye can’t distinguish. Cryptocurrency mining that performs no useful work. AI training runs that are repeated hundreds of times to tune hyperparameters. Redundant backups of backups. These are not essential services; they’re artifacts of a system that treats compute as infinite and free.

Pricing doesn’t reflect the true cost. Cloud storage and compute are so cheap that there’s no incentive to delete anything or optimize code. If the environmental cost of data—the water, the land, the embodied carbon—were internalized into the price of cloud services, behavior would change. Developers would write more efficient queries. Companies would clean up their data lakes. Consumers might choose lower-resolution streaming when it makes no visible difference. The market would allocate digital resources more carefully, and the pressure to build new data centers would ease.

Where Do We Go From Here?

The renewable energy transition for data centers is necessary but insufficient. It’s a first step that has been mistaken for the entire journey. The next steps require uncomfortable conversations about growth, efficiency, and the physical limits of the planet. They require data center operators to report not just power usage effectiveness and renewable percentage, but water usage, server utilization, hardware lifecycle, and land footprint—in a standardized, auditable format. They require cloud customers to see the environmental cost of their workloads, not just the dollar cost. And they require all of us to question whether every bit of data we generate and store is worth the resources it consumes.

I’m not arguing against renewable energy. I’m arguing against the complacency that comes with it. The danger of a “100% renewable” claim is that it makes the problem seem solved, when in fact we’ve only addressed one dimension of a multi-dimensional challenge. Data centers are physical systems embedded in ecological and social contexts. Until we treat them that way—with the same rigor we apply to power purchase agreements—we’re just rearranging deck chairs on a ship that’s still taking on water.

Frequently Asked Questions

Why isn’t buying renewable energy enough to make data centers sustainable?

Purchasing renewable energy through certificates or power purchase agreements addresses only the operational electricity consumption. It doesn’t account for the timing mismatch between when renewables generate power and when data centers consume it, nor does it cover the embodied carbon in server manufacturing, water use for cooling, or land-use impacts of renewable installations. A data center can be “100% renewable” on paper while still relying on fossil fuels during calm, cloudy periods and consuming scarce water resources in drought-prone regions.

How does server utilization affect the environmental impact of data centers?

Most servers run at only 12-18% utilization, meaning the vast majority of their computing capacity sits idle while still drawing power and generating heat. Higher utilization rates would allow the same computing workload to be handled by fewer physical servers, reducing the need for manufacturing new hardware, the energy to run and cool them, and the land required for the facilities themselves. Improving utilization is one of the most effective ways to reduce the total resource footprint of digital infrastructure.

What can cloud customers do to reduce their data center footprint?

Cloud customers can optimize their code to run more efficiently, delete unnecessary data and redundant backups, choose cloud regions powered by cleaner grids with lower water stress, and extend the life of their virtual machines instead of constantly provisioning new ones. They can also ask their cloud providers for transparent reporting on the environmental impact of their specific workloads—not just the provider’s overall sustainability claims—to make more informed decisions about where and how to run their applications.

You’ve seen the headlines. A tech behemoth signs a flashy new deal for a solar farm, or a cloud provider boasts that its data center is now “100% renewable.” It sounds like a neat, tidy solution. The electrons zipping through server racks in northern Virginia are, on paper, matched by clean ones generated somewhere else. But if you follow the actual copper wires, you bump into a much messier truth: the physical grid doesn’t care about your accounting. That data center is still gulping down whatever the local power plant is serving, and right now, that’s often a stiff mix of gas, coal, and nuclear. We’ve gotten very good at offsetting the idea of dirty power without changing the physical reality of it.

The Fable of the Flawless Match

Most corporate clean energy claims rest on a system of certificates—Renewable Energy Certificates (RECs) in the U.S., Guarantees of Origin in Europe. A facility in Ashburn, Virginia, can buy credits from a wind farm in Oklahoma and call itself green. It’s a financial transaction that supports renewable projects, and that’s genuinely useful. But it’s also a spatial and temporal sleight of hand. The actual, physical load of that data center is being met by whatever generators are spinning on the PJM interconnection at that exact moment. On a hot July afternoon, that’s a lot of natural gas. The electrons don’t carry labels; the grid just balances supply and demand with whatever is available.

This isn’t just an accounting quirk. It’s a fundamental mismatch between a financial product and a physical system. When a data center campus in Loudoun County ramps up its compute load, a generator somewhere nearby has to ramp up to meet it. If the wind isn’t blowing in Oklahoma, that generator is almost certainly a fossil fuel peaker plant. The certificate system has successfully funneled money into building more wind and solar, which is a real win. But it has also let us pretend that a data center’s operational carbon footprint is solved, when its physical footprint on the local grid remains as carbon-heavy as ever. We’re greening the portfolio, not the power lines.

The Grid’s Concrete Ceiling

Data centers aren’t just passive energy consumers; they’re hulking, hyper-concentrated loads dropped onto specific nodes of a creaky electrical grid. A single hyperscale campus can demand as much power as a midsize city, and it expects that power with “five nines” of reliability—99.999% uptime. This creates a physical strain that no amount of remote wind farms can ease. The grid’s transmission lines, transformers, and substations have hard limits. You can carpet a desert with a gigawatt of solar panels, but if there aren’t high-voltage arteries to carry that power hundreds of miles to the data center, it might as well be on the moon.

The U.S. interconnection queue is a perfect snapshot of this bottleneck. Lawrence Berkeley National Laboratory data shows over 2,000 gigawatts of generation and storage projects waiting to connect, the vast majority of them renewable. Wait times routinely stretch past five years. A data center can be built in 18 months. The physical infrastructure to deliver clean power simply can’t keep up with the pace of digital expansion. So even a data center with the best intentions and a signed PPA for a new wind farm will, for years, be physically powered by the existing, fossil-heavy grid. The bottleneck isn’t ambition; it’s steel, copper, and permitting.

The Duck Curve Meets the Flat Line

The problem gets sharper when you look at the timing of energy use. Data centers are the ultimate flat-line consumers; they run 24/7 at a near-constant hum. Solar power, by contrast, is a temperamental, intermittent resource that peaks at midday and vanishes at night. This creates the infamous “duck curve,” where net grid demand plummets during sunny hours and then rockets up as the sun sets. A data center that claims to be solar-powered is, in physical reality, a data center that leans heavily on the grid’s fossil fuel or storage resources for most of the day.

This leads to a stubborn fallacy: that you can “baseload” a data center on renewables. The only way to physically pull that off is through massive overbuilding of generation and an even more staggering deployment of energy storage. The math is sobering. To run a 1-gigawatt data center on solar alone, you wouldn’t just need 1 GW of panels. You’d need perhaps 5 GW of solar capacity and tens of gigawatt-hours of battery storage to cover nighttime, cloudy weeks, and seasonal dips. This isn’t a matter of buying more RECs; it’s a fundamental re-engineering of the grid’s physical assets, a process that takes decades and trillions of dollars.

The Storage Gap

Battery storage gets trotted out as the silver bullet, but the numbers don’t add up yet. The current global installed base of grid-scale batteries is measured in gigawatts, while the need for a fully renewable-powered data center industry would be in the terawatts. Lithium-ion, the dominant tech, is great for short bursts—two to four hours—but it’s not economically viable for the long-duration storage (12+ hours, or even seasonal) needed to bridge the gaps in wind and solar generation. Flow batteries, compressed air, green hydrogen—these are still in their awkward teenage years of deployment. A data center operator can sign a PPA for “24/7 carbon-free energy,” but the physical reality is that the technology to deliver it at scale simply doesn’t exist on the grid yet.

The Thirst Nobody Talks About

Another dimension often left out of the “100% renewable” victory lap is water. Many data centers rely on evaporative cooling systems that guzzle millions of gallons of water a day, especially in water-stressed regions. A single large facility can consume enough to supply a small town. This creates a direct, physical strain on local aquifers and municipal water supplies, a strain that has nothing to do with the carbon content of the electricity. In some communities, data center water use competes directly with agriculture and residential needs, sparking a localized resource conflict that renewable energy certificates do absolutely nothing to address.

Even the renewable energy sources themselves can worsen water stress. Concentrated solar power plants can be water hogs. But more broadly, the land use required for massive solar and wind farms creates its own ecological pressures—habitat fragmentation, the mining of rare earth minerals for batteries and turbines. A data center running on 100% renewable energy is still a massive industrial facility with a physical footprint that extends far beyond its electrical meter.

The Rebound Effect and the Hunger for More

Here’s the most uncomfortable twist: the availability of “100% renewable” energy can actually accelerate the growth of data center demand, creating a classic rebound effect. When a company believes it has solved the carbon problem, it may feel licensed to expand digital services without restraint. The result is a net increase in total energy consumption, even if the percentage of renewable energy is high. The cloud, AI, streaming, and the Internet of Things are not static industries; their energy appetites are growing exponentially. A 2023 report from the International Energy Agency projected that data center electricity consumption could double by 2026, driven largely by AI and cryptocurrency. If the grid’s total renewable capacity isn’t growing at the same pace, then all this new demand is effectively being met by fossil fuels, regardless of what the certificates say.

Thinking in Systems, Not Spreadsheets

So, if renewable energy procurement is not enough, what is? The answer requires a shift from a carbon-accounting mindset to a systems-thinking one. A data center is a node in a complex network of electrical, hydrological, and ecological systems. Its sustainability must be measured by its actual, physical impact on those systems, not just by its ledger of certificates.

1. Location and Grid Integration

The most consequential decision a data center operator can make is where to build. Placing a facility in a region with an already-clean grid—like the Pacific Northwest with its hydropower, or France with its nuclear—has a far greater immediate physical impact than building in a coal-heavy region and buying offsets. In addition, data centers can be designed as active grid participants, not just passive loads. This means investing in on-site generation, battery storage, and demand-response capabilities that actually help balance the local grid. A data center that can reduce its load during peak hours or provide frequency regulation services is a better physical citizen of the grid than one that simply buys RECs.

2. Radical Efficiency and Circular Design

The cleanest megawatt-hour is the one never consumed. The industry’s focus on PUE (Power Usage Effectiveness) has been a success story, but it only measures the overhead of cooling and power distribution, not the efficiency of the computing work itself. We need to move toward metrics that measure useful computation per unit of energy, incentivizing more efficient code, better server utilization, and a shift away from wasteful “always-on” architectures. Beyond energy, a circular approach to hardware—extending server lifespans, reusing components, and designing for disassembly and material recovery—can dramatically reduce the embodied carbon and resource extraction associated with the relentless churn of IT equipment.

3. Transparent, Granular Accounting

The current system of annual REC matching is too coarse. The industry is moving toward hourly matching of carbon-free energy, a concept known as “24/7 CFE.” This requires data centers to match their electricity consumption with local or grid-delivered carbon-free sources on an hourly basis. While still an accounting framework, it forces a much closer alignment between the data center’s demand profile and the actual generation profile of clean resources, driving investment in the storage and firm clean generation that the grid physically needs. This transparency must also extend to water use, land use, and supply chain impacts, creating a full-system picture of a facility’s footprint.

FAQ

If a data center buys enough renewable energy to cover its annual consumption, isn’t it carbon neutral?

Not in a physical sense. The annual matching system allows a data center to claim carbon neutrality by purchasing certificates from a renewable project that may generate power at different times and in a different location. The actual electricity consumed by the data center at any given moment comes from the local grid mix, which often includes fossil fuels. The certificates represent a financial investment in renewables but do not change the physical source of the electrons powering the servers. True physical carbon neutrality would require the data center to be directly connected to a dedicated renewable source with sufficient storage to cover its 24/7 demand, a configuration that is currently rare.

Why can’t we just build more solar and wind farms to power all data centers?

The limitation is not just the number of solar panels or wind turbines, but the physical capacity of the grid to transmit that power and the ability to store it for when the sun isn’t shining and the wind isn’t blowing. The transmission grid is a major bottleneck, with new high-voltage lines taking a decade or more to plan, permit, and build. Additionally, the variability of solar and wind requires massive amounts of energy storage to provide a stable, 24/7 power supply. Current battery technology is insufficient for long-duration storage, and the scale of deployment needed to firm up a fully renewable grid for a large data center industry is decades away.

What is the single most effective thing a data center operator can do to reduce its real-world environmental impact?

Beyond siting new facilities in regions with already-clean grids, the most effective action is to focus on reducing total energy and resource consumption through efficiency at all levels. This means not just improving PUE, but also optimizing server utilization, investing in more efficient code and hardware, and extending the life of equipment. A data center that uses half the energy has half the physical impact, regardless of the grid mix. Coupling this with on-site generation and storage to actively support local grid stability creates a demonstrably smaller physical footprint than simply buying certificates for a remote renewable project.

When a tech giant announces a shiny new solar farm or a long-term wind contract to power its data centers, it’s easy to feel a little glow of progress. The electrons flowing into the servers are, on paper, clean. But follow those electrons back through the tangled wires of the real grid, and the story gets messy. The idea that a data center running on renewables is an environmental problem solved is not just optimistic—it ignores the stubborn physics of how electricity actually works. Rui Mendes has been poking at the seams of infrastructure and resource flows for years, and the more you dig, the clearer it becomes: matching kilowatt-hours on a spreadsheet is a long way from decarbonizing a round-the-clock industrial beast.

The Intermittency Gap and the Shadow Grid

Data centers don’t nap. They don’t take weekends off. They demand power with a flat, relentless hunger that wind and solar, by their very nature, can’t satisfy alone. A company might ink a deal for a wind farm in Oklahoma and claim its Virginia servers are running on that breeze. But electrons don’t come with name tags. The grid operator has to balance supply and demand in real time, and when the sun dips or the wind dies, that balance is kept by whatever’s spinning—usually a natural gas plant, sometimes coal. The renewable certificate gets filed away, the press release goes out, and the fossil fuel backbone stays right where it is, flexing to keep the servers humming through the night.

This isn’t a flaw in renewables. It’s a flaw in the story we tell about them. A hyperscale data center cluster—think Northern Virginia, where the load rivals a small city—can’t run on good intentions. The grid operator doesn’t care about your corporate sustainability report; it cares about frequency and voltage. When a cloud passes over a solar array or the wind drops, something has to ramp up instantly. That something is almost always a gas turbine. The renewable contract might make the accounting look tidy, but the physical system still leans heavily on fossil fuels, especially during the dark, windless stretches that happen more often than the brochures admit.

Water: The Thirst Nobody Talks About

Carbon gets all the attention, but data centers have another appetite that’s just as worrying: water. Cooling those endless racks of servers can guzzle millions of gallons a day, often in regions already grappling with drought. A solar-powered facility in Arizona might look virtuous on a carbon ledger, but it’s still pulling from the same overtapped aquifers as the local farms and communities. Evaporative cooling systems, which are common because they’re energy-efficient, literally send water vapor into the sky—water that’s gone from the local watershed for good. The renewable energy badge doesn’t cover that loss.

And it’s not just the operational thirst. The supply chain for renewables themselves is waterlogged. Manufacturing photovoltaic panels requires ultrapure water for rinsing silicon wafers, and the factories producing them often sit in industrial zones where water governance is, let’s say, less than rigorous. So a data center running on solar in a desert might be shifting its water burden to a river basin halfway around the world. The math gets uncomfortable when you zoom out.

The Hardware Churn and the Carbon It Leaves Behind

Servers don’t last. In a hyperscale facility, the refresh cycle is brutal—three to five years, then rip and replace. Each new generation of chips promises better performance per watt, and that’s real. But the old gear doesn’t vanish. It’s shredded, smelted, or shipped to a secondary market where it runs less efficiently. And the new gear? It starts its life in a semiconductor fab, one of the most energy-hungry and chemically intense manufacturing processes on the planet. The carbon baked into a single server—from mining rare earths to etching silicon to assembly—can rival years of its operational energy use. When a data center claims it’s “100% renewable,” it’s usually only talking about the electricity that keeps the lights on, not the mountain of embodied carbon that built the place and fills it every few years.

Then there’s the concrete and steel of the building itself. Cement production alone coughs up roughly 8% of global CO₂ emissions. Backup diesel generators sit there, ready to fire up during grid outages—which are becoming more common as climate change strains infrastructure. Those generators get tested regularly, burning fuel, and when a storm knocks out the lines, they roar to life. Renewable certificates don’t touch that reality. The irony is thick: data centers built to serve a digital economy are quietly deepening their reliance on diesel just as the grid they lean on gets wobblier.

The Map Doesn’t Match the Demand

Renewable energy is stubbornly place-based. The best wind whips across the Great Plains; the strongest sun bakes the Southwest. But data centers huddle around internet exchange points and population hubs—Northern Virginia, Frankfurt, Singapore. You can’t just pipe sunlight across three states without losing a chunk of it to transmission losses and without building new high-voltage lines that take a decade to permit and build. A data center in Loudoun County can sign a virtual power purchase agreement for a wind farm in Texas, but the actual electrons feeding its servers come from the local grid, which in Virginia is still cozy with natural gas. The renewable energy certificate is a financial tool, not a physical delivery mechanism. It was designed to spur renewable development, and it does that. But it doesn’t rewire the grid.

This geographic mismatch creates a timing problem, too. Data centers run around the clock, but solar panels clock out at sunset. Wind patterns are fickle. Without cheap, massive energy storage—which we don’t have at the scale needed—the gap between renewable generation and constant demand has to be filled by something. In most places today, that something is natural gas. And as data center energy appetite grows, driven by cloud services and machine learning workloads, the gap is widening. The International Energy Agency expects data center electricity consumption to double by 2026, topping 1,000 terawatt-hours. Even with record renewable buildouts, the sheer weight of new demand means more fossil fuel plants will stay online or get built to keep the lights on.

Jevons and the Efficiency Trap

Data center operators love to talk about efficiency gains—better power usage effectiveness, smarter cooling, higher server utilization. These are genuine improvements. But they also risk walking straight into a trap that William Stanley Jevons spotted back in 1865: when a resource gets more efficient to use, total consumption often goes up, not down. Cheaper computation invites more computation. The result is that total energy use climbs even as individual facilities get leaner. This isn’t a theory; it’s the plot of the last two decades. Despite all the efficiency wins, data center energy use has marched steadily upward because our hunger for digital services has grown even faster.

This dynamic pokes holes in the story that we can simply “green” the data center industry by buying more renewables. If the underlying demand keeps surging, even a sector running entirely on wind and solar would still chew up enormous amounts of land, transmission corridors, and mining for battery materials. The physical footprint of renewable energy isn’t zero. A systems-minded view forces a harder question: not just how we power data centers, but why we need so many of them, and which services genuinely earn that resource draw.

What a More Honest Accounting Would Look Like

If buying renewables isn’t enough, what would a clearer picture require? First, data center operators would need to report on a wider set of numbers: the hourly carbon intensity of the grid where they actually sit, not just annual certificates matched to total consumption. Google and Microsoft have started nudging toward 24/7 carbon-free energy goals, but those remain aspirational and technically heavy lifts for most operators. Second, it would mean transparent reporting on water consumption—not just in cooling, but across the whole supply chain. Third, it would force a real conversation about demand management. That could mean shifting computation to times when renewables are plentiful, or it could mean questioning whether always-on, low-value computation deserves a permanent place on the grid.

The uncomfortable truth is that a data center is an industrial facility, not a cloud. Its physical demands on land, water, and grid stability are real and growing. Treating renewable energy as a get-out-of-jail-free card hides those demands and postpones the kind of systemic thinking needed to align digital infrastructure with planetary limits. The goal shouldn’t be to make data centers look “green” on a spreadsheet, but to ensure they operate within the actual carrying capacity of the regions that host them. That means sometimes choosing not to build, or to build differently, or to locate where the resource trade-offs are less severe. It means accepting that efficiency and renewables are necessary but not sufficient tools. The real work is in questioning the growth itself.

Frequently Asked Questions

Why can’t data centers just use batteries to store renewable energy for nighttime use?

Battery storage at the scale a large data center needs is still wildly expensive and resource-heavy. A 100-megawatt facility would need hundreds of megawatt-hours of storage to get through a single night, which means a lot of lithium, cobalt, and other materials. The mining and manufacturing for those batteries carry their own heavy environmental and social price tags. For now, batteries are more practical for short-term grid balancing than for shifting a full data center load from day to night.

Doesn’t locating data centers in cold climates solve the cooling problem?

Cold climates cut the energy needed for cooling, but they don’t erase water use or other impacts. Many facilities in cool regions still use water-based cooling for efficiency, and building and running data centers in remote areas can disrupt local ecosystems and strain small-town infrastructure. Plus, the renewable energy available in cold, northern regions is often seasonal—plenty of hydropower in spring, but less solar in winter—creating a different kind of mismatch.

What is the difference between a power purchase agreement and actually using renewable energy?

A power purchase agreement is a financial contract that helps fund a renewable energy project, but the electricity from that project goes into the general grid, not directly to the buyer. The buyer gets renewable energy certificates that can be used to claim “100% renewable” status in carbon accounting. But the physical electricity powering the buyer’s facility still comes from the local grid mix, which may include fossil fuels. The agreement supports renewable development but doesn’t physically disconnect the facility from fossil fuel generation.

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.

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.