An environmental defense of AI
Making the Earth wealthier without making it hotter
Techno-optimists see AI as the solution to all our problems, while many environmentalists view it as an unmitigated disaster. AI is terrible for the environment, study finds. AI has an environmental problem. Here's what the world can do about it. Fixing AI’s energy crisis. Critics point to the growing energy, emissions, and water use of data centers cause for grave concern.
Environmental critiques of AI are usually based on weird comparisons. Take one example, that “AI-related infrastructure may soon consume six times more water than Denmark” (UNEP). Denmark is a small country with remarkably low per-capita water consumption, contributing only 0.024% of global freshwater use1. Or take the headline that a ChatGPT query uses 10x more energy than a Google search. This objection should be reversed if a ChatGPT query saved you 10 Google searches!
Practical environmentalism is about clear-eyed comparisons. Every action has an environmental cost, but how costly is the alternative? Palm oil causes deforestation, but if we replaced it with any other oil crop we’d need to deforest 4x more land! Nitrogen fertilizer causes ~5% of global emissions, but allowed us to feed 3.5B additional people without clearing millions of hectares of habitat. The cost of banning nuclear power far outweighs any tail risks.
If we expect AI agents to become capable members of the workforce, then we should be comparing their environmental impact to that of a human worker. And if you think data centers are an environmental disaster, just wait until we look at the resource footprint of a skilled knowledge worker in a developed country.
If trends hold, we can expect an AI worker to use hundreds of times less energy, land, and water than a human worker. Even using dirty grid electricity, the emissions of an AI worker will be an order of magnitude lower than a human. My argument is not that we should get rid of the humans – it is that environmentalists should get excited about AI. Digital minds will help decouple economic growth from physical resource consumption, multiplying the effective workforce without multiplying its environmental footprint.
Powering a digital worker
In order to estimate the environmental impact of an AI worker, we need to estimate its required compute. Once we know the compute budget, we can forecast energy use, water use (for cooling), land use (for renewables), and emissions (if electricity comes from the grid). If you just want to see the environmental impact assessment, feel free to skip this section.
The brain-centric view of AI
A reasonable reference point to start from is the human brain, the existence proof that inspired initial work on neural networks over 60 years ago. The range of estimates for the computational power of a human brain is vast, from 10^12 to 10^18 FLOPS (floating-point operations per second) or more. A widely used estimate is that a human brain is doing about 10^15 FLOPS.
Conveniently, one petaflop (10^15 FLOPs) is roughly the computational power of an NVIDIA H100 GPU2! So to help anchor your intuition, you can imagine that a human brain and a single H100 are similar in raw horsepower3.
The problem with the brain-centric view is that we’re comparing inputs (FLOPS) and not outputs (real-world tasks). Currently, no one can make an H100 GPU do all of the things that a human brain does, despite their equivalent processing power. But we are trending in that direction. The breakthroughs of deep learning, the transformer architecture, and algorithmic improvements from leading AI labs have helped us figure out how to elicit more and more intelligence from raw compute.
The task-centric view of AI
A more practical way to compare human an AI workers is on tasks, ideally ones that map to real-world economic value. For example, you could have an AI and human both solve the same software engineering problem, and record how long the human took and how much compute the AI used. Then you’d know roughly how much compute it takes to “keep up” with the pace of a skilled human.
Unfortunately, such comparisons are few and far between. Most AI benchmarks report scores, but not compute, energy, or cost. The one exception I’ve found is the ARC-AGI benchmark, where AIs and humans both solve visual puzzles. These puzzles are designed to be intuitive for humans (you can try them), but hard for AI. Indeed, while a panel of paid humans got a perfect score, most AI models are in the single digit percentage range. To make things more difficult, AI models need to match human accuracy while costing less than $1 per task to win the grand prize.
ARC-AGI seemed impenetrable until OpenAI’s o3 model achieved4 scores of 76% and 88%, depending on how much compute was used. We also got a glimpse at how much compute was involved. In its “low compute” configuration, o3 generated 33 million reasoning tokens and took about 1.3 minutes per task. This is truly an insane amount of thinking to solve a puzzle, equivalent to writing ~191 novels per minute5. The “high compute” configuration was hundreds of times more expensive still. While o3 only scored 75.7%, it proved a step change in capability.
There are two intuitions that o3 surfaces:
Scaling up inference-time compute leads to significant performance gains. Put simply, models can gain a lot of effective intelligence by thinking really, really hard.
On challenging (for AI) reasoning tasks, current models are extremely inefficient. o3 had to do 191 novels worth of thinking each minute to solve puzzles at roughly human speed, which would require a data center of 337 H100s to achieve6!
In theory, a single H100 should be sufficient to match a human brain, but right now we need hundreds of them. The digital brains we’ve trained so far seem to be orders of magnitude less computationally efficient than our biological brains that evolved over millions of years. They need to exhaustively think through every reasoning step in excruciating detail, and attempt a task dozens or hundreds of times before they can come up with the right answer.
This is more a reflection of the algorithmic progress left to go than of any fundamental limitation. Over the next few years, I expect algorithm improvements to bring us towards the “1 H100 = 1 brain” equivalence or even past it. At the risk of over-generalizing, this is the story of any new technology: first, we demonstrate that something is possible, then we figure out how to make it cheap and efficient, then it exceeds our wildest expectations.
The AI 2027 Forecast
Finally, what are the superforecasters saying? The recent AI 2027 project lays out a median scenario for how they expect artificial intelligence to develop between now and 2027. It’s worth reading in its entirety, but I’ll highlight some key points.
By 2027, the pseudonymous company “OpenBrain” develops Agent-3. Agent-3 is as skilled as the best human programmers, but still requires some complementary humans skills to manage it and set research direction. By July 2027, companies begin hiring Agent-3 to join their remote workforce. Let’s mark this as the arbitrary line where “AGI” is achieved, defined as the ability for AI to substitute for a remote human worker.
How much compute does Agent-3 require? Even as OpenBrain’s models get more capable, they also get smaller and more compute-efficient! In the project’s compute forecast, these frontier models start at 10T active parameters, but distillation and algorithmic improvement allow them to shrink to 5T and then 2T parameters7. This is made possible in large part by Agent-3’s ability to assist with internal research.
What we don’t yet know is the amount of thinking that an AI like Agent-3 would be doing. Humans might only type at ~100 words per minute, but our brains are also rapidly retrieving information, using our five senses, navigating interactions with other people, and making other complex judgments that we don’t vocalize, write down, or even consciously realize. If you had to write down every thread of consciousness in our brains as text, it seems plausible that we’re generating pages and pages of information each minute. What we say or write is only the tip of the conscious iceberg.
To capture this, I plotted the number of H100s required to run different models at different reasoning speeds, expressed in tokens8 per minute. Larger models or more reasoning requires more compute.
My best guess is that a very careful and systematic AI knowledge worker will need to process somewhere between a report (10 pages) and a novel (200 pages) of information per minute. This would involve a mix of reading, generating intermediate thoughts, tool use, self-critique9, and writing down findings or deliverables. Based on the chart, we see that a ~2T parameter model would need roughly ~0.5–5 H100 GPUs to think at these speeds.
The chart above also shows the tradeoff between model size (quality) and reasoning speed (quantity) of thinking. If the number of GPUs is fixed, then you could either have a huge model thinking slowly (i.e, wise zen master) or a small model thinking really fast (i.e, dumb but enthusiastic intern).
Prediction: A digital worker will use 0.5 to 5 H100s
To summarize, we have three compute estimates:
The brain-centric view suggests that a single H100 should be able to match a human, given the right algorithms.
The task-centric view suggests that it might take hundreds of H100s for current models to match humans on hard tasks. We can consider this an upper bound, since AI models are getting about 3x more efficient at a given task each year.
The AI 2027 scenario suggests that a capable remote worker (Agent-3) will have roughly ~2T parameters. Depending on how fast this agent must reason to match a human, it falls in the ~0.5–5 H100s ballpark.
Humans are not optimized for work
To argue that AI workers are an environmental problem, you must also argue that human workers are more sustainable. As we’ll see below, this claim doesn’t hold up. There are a few fundamental reasons why human workers have a large resource footprint.
Humans only work for 12% of their life
One big problem with human workers is that they spend a lot of time consuming resources while not working. The average person in the US lives 77.5 years but only works for ~80k hours, which is about 12% of their lifetime. For every hour of work that a person does, there are 8.3 hours of idle resource consumption.
Humans spend the first quarter of their life being trained to join the workforce, and (hopefully) the last decade or more enjoying retirement. In between, we have mostly weekends, holidays, leisure time, and sleep. Many people aim to minimize the percentage of their life working.
The advantage of AIs are that they only consume resources when they’re working. My point isn’t that humans are lazy or a waste of resources – it’s that we aren’t efficient working machines and shouldn’t aspire to be.
Biological food is extremely inefficient
Almost all of the land and water we consume comes indirectly from our food. Globally, about 70% of freshwater usage is for agriculture. The space required for your house, car, office, etc is dwarfed by the ~6,000 meters per person10 of pasture and cropland required to produce your food!
This is all a consequence of trophic levels and the inefficiency of agriculture. While we can feed the AIs directly with electricity from ~20% efficient solar panels, the biological food we eat is hundreds if not thousands of times less efficient at storing solar energy.
Examining the footprint of a digital worker
Below, I compare the resource use of an AI worker to a human knowledge worker, exemplified by the average American. The functional unit for this “LCA” is 1 year of continuous work (8,760 hours). Because humans only work ~12% of their lifetime, we must scale every resource by 8.3x to compare them fairly to a single AI working continuously. Over a lifetime, it takes 8.3 persons-worth of resources to get the equivalent output of 1 person working nonstop.
You can find supporting calculations here.
Energy
Humans use 77,028 kWh of primary energy each year, so a continuous work-year requires ~645 MWh (8.3x higher, see above)
H100s use ~700W at peak power. I include an additional 25% for cooling, and another (conservative) 25% on top of that11 to account for training energy.
It’s also worth knowing that leading ML hardware becomes 40% more efficient each year! I bake 3 years of efficiency gains into my model to get us to the end of 2027.
Under these assumptions, the energy required for one continuous year of human work could sustain ~185 H100s! The shaded gray box (0.5–5 x H100s) is my forecast for an AI remote worker in 2027 (Agent-3). The AI uses 37–370x less energy than a human for the same amount of work.
Land
To simplify land calculations, let’s assume that the AI agents will be powered with 100% solar energy (based on fixed-tilt data from NREL). Nuclear, geothermal (or fossil fuels) would require less land.
I also assume an additional 30% land use for batteries, data centers, and other infrastructure.
We compare against global average land use, adjusted for one continuous year of work (see above).
Results: The land required for 8,760 hours of human work could power hundreds of AI agents (903 H100s). Carrying along our assumption of 0.5-5 H100s per agent, each AI agents uses 180–1800x less land than a human.
Water
As a baseline, I assume the water use of an average person in the US (1,543 m3 / year) and adjust by the 8.3x factor as above.
Data centers use water for evaporative cooling. I conservatively assume that each H100 is running at peak power, creating 0.7kW of cooling load that is entirely dissipated via evaporation in a cooling tower. An additional 25% cooling load is added to account for training.
Results: The AI worker is vastly more water-efficient, using about 575–5748x fewer cubic meters of water to do its work. It isn’t eating food from irrigated farmland, nor is it taking long showers, washing its clothes, or watering its lawn.
Emissions
In the worst case, we’ll be powering the AI workforce with the current electricity mix from the grid, which is still pretty dirty. In the United States, 0.369 kg of CO2-eq emissions are produced for every 1 kWh of electricity.
As a baseline, let’s assume that a knowledge worker has the average emissions of someone in the United States (14.2 tons CO2-eq). As above, the emissions are adjusted by 8.3x to be equivalent to a continuous year of work.
Results: Once again, the AI worker wins by a large margin. Even on the dirty US grid, the AI worker would produce about 18–180x lower emissions to get its work done.
Even better, the grid is getting cleaner, and data center growth seems to be spurring renewable energy development. The biggest AI companies are investing heavily in renewables, and even signing power purchase agreements with nuclear companies, acting as critical buyers of first resort. Most of the cost to run AI models is actually the cost of burning through your GPUs, not the power itself. So AI companies actually have the profit margins and cost structure to pay a bit more for green power.
34 million environmentalists in a data center
Our environmental assessment reveals that a digital workforce could augment the human workforce while consuming orders of magnitude fewer resources. In my worst case analysis, the digital worker produces 18x lower emissions than a human while consuming 575x less water, 180x less land, and 37x less energy. In my optimistic case, resource efficiency is 10x better than that!
While the digital workforce has a tiny per-capita (per-machina?) footprint, what would the total resource footprint be? Let’s suppose that the US deploys a digital workforce equal in output12 to its human workforce of 171 million people, doubling the effective size of the labor force. If powered from the grid, the AIs would produce the total emissions of ~1.6–16 million additional Americans, increasing the country’s emissions by at most 4.7%. Additional land and freshwater use would be a rounding error compared to agriculture. Doubling the workforce for a 4.7% increase in emissions seems like an excellent trade.
The AI-driven growth in electricity use, on the other hand, can’t be ignored. While humans get their primary energy from a variety of sources (electricity, fossil fuels, foods), the AIs concentrate all of their load on the electricity grid.
In my worst-case scenario, a workforce of 34 million AIs (equivalent in output to 171 million humans) would add ~600 TWh of load to the US grid, a 14% increase in demand. To put this in perspective, it took almost 20 years for the US to build that much renewable generation. On the other hand, China might add ~600 TWh more solar energy alone in the next 4 years. So we have an existence proof that rapid buildout of renewable energy is possible if there is political will.
Everywhere but the environmental statistics
AI isn’t the first technology to decouple economic growth from population and resource growth. Mechanized farming, steam power, assembly lines, fertilizers, electrification, computers, and the Internet have all allowed us to do more with less. At the same time, many of these efficiency gains led to greater overall consumption (Jevon’s paradox). The average person today has a staggering amount of wealth with which to consume energy, goods, services, and natural resources.
Digital workers are one of several key reasons why I’m optimistic that we can become wealthier13 while shrinking our environmental footprint. They’re a special technology because they (theoretically) act as a true substitute for intelligence, not just a complement to it. As a result, I would be quite surprised if doubling the US workforce via AI did not lead to a measurable acceleration in economic growth.
The degree of economic growth from AI is hotly debated by bearish economists and bullish AI researchers. Whether you expect GDP growth to go from ~3% to ~3.5% or ~30%, the directional impact is the same. In the chart above, I expect a digital workforce to accelerate the upward trajectory of GDP per capita, even as CO2 emissions peak and decline. The net effect would be a continuation of the decline in economic carbon intensity (emissions per dollar of GDP). If we plotted the land intensity or water intensity of the economy, they would trend in the same direction.
AI won’t fix our environmental problems for us, but it can make them more tractable. First, there’s the obvious benefit of having more minds (and robotic hands) helping us with R&D and eventually deployment. Even if there are diminishing returns to labor, having twice the workforce developing climate solutions would speed things up.
Furthermore, if GDP goes up while resource use stays flat or declines, many environmental problems look less daunting, for the simple reason that problems are less scary when you have more money. For example, climate scientists estimate we’ll need to scale atmospheric carbon removal to ~10 gigatons by 2050. At a target of $100/ton, this waste remediation effort would cost ~1% of current global GDP annually. If real GDP grows at 6% instead of 3%, we’ll be more than twice as wealthy by 2050. In other words, the carbon removal problem will be 2x cheaper in the world with AI workers.
Denmark used 976M cubic meters of freshwater in 2015, only 0.024% of the 3.95T global total (Source). Denominators matter!
Here is a data sheet for the H100. The FLOPs depend on the numeric precision being used and the types of operations. Most LLMs use FP16, with a peak rate of ~2 petaflops. Models like DeepSeek use FP8, which is twice as fast. Throughout this piece I’ll assume that the practical speed of a GPU is 40% of the peak value.
Neural networks are only a crude approximation of biological neurons and axons. For practical reasons, a digital neuron does some very simple mathematical operations (e.g multiply and add) whereas a biological neural might have much more complex and nuanced behavior. FLOPS are also tricky to compare since GPUs can do operations at different precisions; FP8 arithmetic (“quarter precision”) is 8x faster than FP64 arithmetic (“double precision”).
There was some controversy about the results, since o3 was able to fine-tune on examples of ARC-AGI tasks. Some argue that this prevents a true assessment of AGI. In any case, we now have the ARC-AGI-2 benchmark, which no model seems to be competent at yet.
If o3 generates 33M reasoning tokens to solve a 1.3 minute task, then it generates tokens at a rate of ~25.4 tokens per minute. Assuming a novel is 100,000 words and the English language has ~1.3 tokens per word, then the model is thinking at a rate of ~195 novels per minute!
See Supplemental Materials. I assume that o3 is using GPT-4.5 as a base model (rumored to have 600B active parameters). The H100 is running FP8 inference, and at 40% FLOP utilization (an assumption used throughout AGI 2027 also).
The 10T, 5T, and 2T parameters are in terms of an FP8, dense-equivalent model. Due to recent techniques like mixture-of-experts, frontier models don’t use all of their parameters when doing inference. A dense-equivalent model is adjusted so that we can compare sparse models to traditional LLMs on the same basis. FP8 is the lowest precision that GPUs can do floating-point operations at, but is also the fastest (2x the speed of FP16).
LLMs process tokens (parts of words) rather than whole words. As a rule of thumb, English words are about 1.3 tokens on average.
For each AI worker, you might even need another AI worker acting as its supervisor to ensure safe(r) operation. It really depends on how reliable they agents turn out to be. This is another reason why I expect AI workers to have a lot more reasoning going on than a human.
From Our World in Data.
The 25% cooling energy results from a COP of ~4, and the 25% additional training energy is based on my own conservative estimates from GPT-4 training/inference statistics. As AI models become more widely deployed, the training energy will be amortized over larger inference workloads, shrinking its importance.
To complicate the analysis slightly, I use a 5x factor here to convert AI workers to current human workers (humans work for ~20% of an 8,760 hour year). This factor is smaller than the 8.3x factor used in the lifecycle assessments because I’m no longer considering the years of training/retirement for the human workers; they’re already productive members of the workforce. With a 5x factor, it would take only ~34M digital workers to match the output of ~171M human workers.
GDP per capita measures average wealth, and does not capture how that wealth is distributed. Indeed, growth in GDP per capita has not translated to better lives to many people. In this essay I don’t address the (valid) concerns that AI might further concentrate wealth or cause economic disruption.