Training GPT-3 consumed roughly 1,287 megawatt-hours of electricity and emitted 552 metric tons of CO2 equivalent, according to peer-reviewed research from Patterson et al. GPT-4's training run is estimated at somewhere between 50 and 70 gigawatt-hours, dramatically larger just three years later. But a June 2026 United Nations University report argues that focusing on training emissions alone, the number most public conversation still fixates on, is genuinely outdated framing. The real environmental footprint of training large AI models is only one piece of a considerably larger, three-part picture spanning carbon, water, and land, and training itself, it turns out, isn't even the dominant piece anymore. This guide breaks down what the actual, current, peer-reviewed data shows.
The Training Numbers Themselves, Precisely Stated
It's worth starting with the specific, documented figures for major models, since these numbers are frequently cited imprecisely elsewhere. Training GPT-3 emitted 552 metric tons of CO2 equivalent in a single run, while training BLOOM, a comparably sized open model, produced 433 tonnes of CO2e. A separate life-cycle assessment of BLOOM-176B found meaningfully different figures depending on what was actually counted: 24.7 tonnes when only dynamic electricity consumption was considered, rising to 50.5 tonnes once idle time and upstream manufacturing contributions were included as well.
This specific discrepancy matters enormously for understanding why so many circulating AI carbon figures genuinely disagree with each other. The same model can report meaningfully different emissions figures depending purely on methodology, whether a study counts only active training electricity, or also includes idle server time, cooling infrastructure, and the upstream carbon cost of manufacturing the hardware itself. This isn't researchers disagreeing about facts; it's researchers measuring genuinely different boundaries around the same underlying activity.
Why the "Training Emissions" Framing Is Now Considered Outdated
This is genuinely the single most important corrective in this entire topic, worth understanding directly. Public debate has focused heavily on the electricity needed to train large models, but a UN University report specifically argues that this framing is now outdated. Once a model is actually deployed, inference, the continuous, everyday running of that model to answer real user prompts, accounts for 80 to 90 percent of total AI energy use across a model's full lifecycle.
This matters because it fundamentally reframes what "the environmental footprint of training AI models" actually needs to include to be genuinely accurate. ChatGPT alone is estimated to handle roughly 2.5 billion prompts per day, consuming an estimated 383 gigawatt-hours of electricity annually for that single product's ongoing operation, a figure that dwarfs any individual training run precisely because inference happens continuously, for years, across hundreds of millions of daily users, while training happens once, however energy-intensive that single event actually is.
The Real, Full Picture: Carbon, Water, and Land Together
It's worth understanding the genuinely important expansion this UN report specifically makes to how this issue should actually be measured. AI's environmental cost is being systematically mismeasured; most existing assessments focus purely on the carbon emissions associated with training large models, when every kilowatt-hour used to train or run an AI system also carries a genuine water footprint, from cooling and power generation, and a genuine land footprint, from the energy infrastructure and supply chains required to actually support it.
Critically, these three footprints don't move in the same direction, worth understanding directly, since this complicates any simple, single-number summary of AI's environmental impact. A data center powered by nuclear or wind energy might have a genuinely low carbon footprint while still requiring genuinely significant water for cooling. A facility relying on air cooling instead might reduce its water footprint considerably while increasing its energy demand, and therefore its carbon footprint, correspondingly. Judging AI sustainability by carbon emissions alone hides genuine trade-offs that can shift real environmental burden specifically onto regions already water- or land-constrained, rather than actually solving the underlying problem.
The Genuinely Staggering Scale Projected by 2030
It's worth grounding this discussion in the actual, projected scale, since the numbers involved are genuinely difficult to hold in mind intuitively. By 2030, data centers powering AI could draw 945 terawatt-hours of electricity annually, according to the UN University's own analysis, a figure close to triple the combined annual electricity use of Pakistan, Bangladesh, and Nigeria together, home to more than 650 million people. UNU-INWEH puts this 2030 figure at roughly 3 percent of projected total world electricity consumption.
This matters because it reveals the genuine, sobering scale this issue is heading toward, distinct from any single model's individual training footprint. A single training run, however large, remains a discrete, bounded event; the aggregate, ongoing electricity demand of the entire AI data center ecosystem represents a continuously growing, compounding draw on global electricity infrastructure, one already large enough to be compared meaningfully against the combined national consumption of countries home to hundreds of millions of people.
A Real, Documented Case of This Scale Hitting Physical Limits
It's worth understanding a genuinely concrete, already-happening example of this projected scale creating real, physical constraints, rather than treating this purely as a future, abstract concern. In Ireland, data centers accounted for 21 percent of total metered electricity in 2023, exceeding the electricity consumption of all urban households in the country combined. The national grid operator has paused new data center approvals around Dublin specifically until 2028, making Ireland a genuine, concrete, documented example of what happens when AI infrastructure growth outpaces a country's own energy planning.
This matters because it offers a genuine preview of what other countries and regions may increasingly face as AI infrastructure continues expanding at its current pace. Ireland's specific situation illustrates that this isn't purely a theoretical sustainability concern to be debated abstractly; it's already produced real, binding regulatory action, a formal moratorium on new approvals, precisely because the underlying electrical grid genuinely couldn't support continued, unconstrained growth alongside existing residential and commercial demand.
Why Efficiency Gains Haven't Kept Pace With Growth
It's worth understanding a genuinely important pattern worth naming directly, since it explains why growing efficiency per individual query hasn't actually reduced AI's total environmental footprint. The number of computational resources used to train state-of-the-art models has doubled roughly every three to four months, a genuinely staggering compounding growth rate that has consistently outpaced whatever efficiency improvements individual hardware and software optimizations have delivered over the same period.
This matters because it reveals a genuine, real tension between two simultaneously true facts. Individual AI queries have genuinely become more energy-efficient over time, with newer hardware and optimized models requiring less energy per unit of computation than earlier generations. But the sheer, compounding scale of overall AI usage, more models, larger models, more total queries, has grown even faster than these genuine efficiency gains, meaning total, aggregate environmental impact has continued climbing despite real, documented per-query improvement, precisely the pattern research specifically warns is expected to continue rather than naturally self-correct.
It's Worth Being Fair: AI Also Has Genuine Emission-Reduction Potential
It's worth presenting a genuinely important, balancing consideration directly, rather than treating AI's environmental story as purely, unambiguously negative. A 2025 study from the Grantham Research Institute found that AI could reduce global emissions by an estimated 3.2 to 5.4 billion tonnes of CO2-equivalent annually by 2035, if applied deliberately and wisely to help design climate policy, improve environmental monitoring systems, and optimize energy grids and industrial processes.
It's worth being equally honest about a genuine, documented risk on the other side of this same coin, though, rather than presenting only the optimistic scenario. If AI is adopted at similar rates to accelerate fossil fuel extraction specifically, rather than to reduce emissions, net global emissions could actually rise by an estimated 0.47 to 1.8 gigatonnes of CO2 annually, so-called "enabled emissions" from making fossil fuel extraction faster and cheaper, estimated to be 3.3 to 13.3 times larger than AI data centers' own current, direct emissions. This matters because it reveals AI's net environmental effect genuinely depends on how it's actually deployed and applied, not on some fixed, inherent property of the technology itself.
The Open-Source Angle: A Genuinely Underexamined Dimension
It's worth understanding a specific, more recent research direction worth knowing about, since it addresses a genuinely underexamined part of this broader picture. Recent research analyzing training emissions across more than 5,200 models hosted on Hugging Face, the most widely used open-source AI model repository, found that collective training emissions across this open-source ecosystem are genuinely comparable to meaningful real-world benchmarks, equivalent to roughly 42,000 cars' worth of annual emissions, or the annual greenhouse gas footprint of about 6,000 EU residents.
This matters because it reveals that AI's environmental footprint isn't confined purely to the handful of massive, headline-grabbing frontier models like GPT-4; it extends across a genuinely enormous, distributed ecosystem of smaller, open-source models as well. This research specifically argues that sustainable AI development requires tracking the genuine, cumulative footprint of these derivative models too, not just the small number of original, foundational models that receive the overwhelming majority of public attention and scrutiny.
What This Means for How You Should Think About AI's Environmental Cost
Understand that a single model's training footprint, however large the specific number, represents only a modest fraction of that model's full, lifetime environmental impact. Given how directly inference now accounts for 80 to 90 percent of total AI energy use, evaluating any specific model's environmental cost purely by its training emissions genuinely misses the larger, more significant part of the actual picture.
Treat any single-metric AI environmental claim with genuine, informed skepticism. Given how directly carbon, water, and land footprints can move in different directions simultaneously, a claim praising a specific data center's low carbon footprint alone, without addressing its water or land impact, offers a genuinely incomplete picture worth questioning further.
If you're evaluating AI providers or infrastructure specifically, ask about the full three-part footprint, not carbon alone. Given how directly the UN University's own research argues this narrower, carbon-only framing hides genuine trade-offs, a more complete evaluation considers water and land impact as seriously as carbon emissions.
Recognize that AI's net environmental effect genuinely depends on its specific application, not some fixed, inherent technological property. Given the documented, real potential for both significant emission reduction and significant emission increase depending on how AI actually gets deployed, the honest, evidence-based question isn't simply "is AI good or bad for the environment," but "toward which specific applications is this technology actually being directed."
Final Thoughts
The environmental footprint of training large AI models is genuinely significant, well-documented in specific, peer-reviewed figures: 552 tonnes of CO2 equivalent for GPT-3, an estimated 50 to 70 gigawatt-hours for GPT-4's training run alone, and a broader, industry-wide data center electricity demand projected to reach 945 terawatt-hours annually by 2030, nearly triple the combined national consumption of three countries home to over 650 million people. Ireland's own real, documented grid constraints, and the resulting moratorium on new data center approvals around Dublin, illustrate this isn't a purely abstract, future concern.
At the same time, the honest, complete picture requires understanding that training itself, however large any single number appears, represents the smaller half of a model's actual lifetime footprint; inference now accounts for the substantial majority of total AI energy consumption. And the honest, complete picture requires holding genuine, documented potential for both significant harm and significant benefit simultaneously, since AI's actual net environmental effect depends considerably more on how deliberately, and toward which specific applications, this technology actually gets deployed than on any fixed, inherent property of large-scale AI training itself.