The Cognitive Revolution · 2025-12-14 · 124 min

AI's Energy & Water Demands: Sorting Fact from Fiction with Andy Masley

Erik TorenbergNathan LabenzAndy Masley

YouTube ↗

TL;DR

Personal chatbot use is environmentally immaterial at today’s scale: Andy Masley estimates roughly 0.3 grams of CO₂ and about 0.3–0.6 watt-hours per median prompt. That is around one second of microwave use, and it would take roughly 1,000 prompts in a day to raise an average person’s emissions by 1%. His blunt conclusion: “Poking around on chatbots, even if you use them a lot throughout the day, is just not going to add a meaningful amount to your carbon or water footprint.”
The right comparison is bits versus atoms: a 20-mile car trip, one hamburger, or one hot shower each lands near 5,000–10,000 median chatbot queries. A full tank of gasoline represents roughly 250,000–500,000 prompts, while a transcontinental flight can reach 1–2 million. If AI prevents even one physical trip—as Nathan Labenz says it recently did for his family—the avoided real-world activity can cover years of ordinary prompting: “Computing is just so wildly efficient.”
The aggregate buildout is enormous for the power industry but modest against the global energy system. A 1-gigawatt data center draws about as much electricity as 1 million US homes and would require roughly 10 square miles of solar capacity; the hypothesized $7 trillion, 80-gigawatt buildout would need about 800 square miles of panels, under 1% of Nevada, and add roughly 1–2% to global energy use. Masley nevertheless calls gigawatt facilities “mind-blowing” because concentrated loads can overwhelm particular grids even when the global percentage looks manageable.
The viral claim that one prompt consumes a bottle of water is wrong by roughly two orders of magnitude. Masley’s estimate is about 2 milliliters per median prompt, or approximately 200 prompts per bottle; broader figures often mix water evaporated in data centers with power-plant withdrawals that are quickly returned. Water can still matter locally, but the decisive distinctions are consumptive use, non-consumptive withdrawal, pollution, and whether a facility is competing for genuinely scarce freshwater.
The most material local externality may be air pollution, not climate change or water consumption. New gas turbines or extended coal generation can concentrate immediate health damage in already vulnerable communities; Masley therefore ranks air pollution “far and away” above the other environmental risks, while remaining cautious about contested details surrounding xAI’s Colossus facility in Memphis. For decision-makers, siting, permitting, clean generation, transmission, and community compensation look more consequential than concern over individual prompt footprints.
Electricity-price pressure is real in some jurisdictions, but the national story is frequently overstated. US residential bills rose about 35% after 2020, yet the Lawrence Berkeley National Laboratory analysis Masley cites attributes the national increase mainly to inflation and supply-side disruptions, not data-center demand; utilities have, however, explicitly blamed data centers for some local rate increases. The grid can eventually exploit economies of scale, but continually accelerating loads may keep infrastructure spending—and political resistance—elevated for years.
AI’s second-order effects will dominate its direct footprint, in both directions. Masley cites an International Energy Agency projection that AI applications might avoid roughly four times the emissions caused by data centers by 2035, through areas such as logistics, building efficiency, antibiotics, photovoltaics, batteries, and material science; he also concedes that automated shopping, cheaper production, and autonomous vehicles could stimulate more consumption. His governing analogy is sharp: judging AI by data-center electricity is like watching Amazon emerge and focusing mainly on “how much energy the website’s using.”

Digest · the substance, structured for research

1. Excitement about current AI and fear of future AI comfortably coexist

Masley starts from the effective-altruist view that machines able to replicate “a lot or even most or all” of the human mind are more likely than not within his lifetime, though with “massive error bars.” He leans toward a lower P(doom) because he distrusts highly specific philosophical stories about misaligned systems rising up, but assigns a high “P weird” to rapid, dangerous social disruption.
That caution does not make him anti-technology: “Opus 4.5, basically my new best friend. It’s so cool. Love it.” He compares his position to supporting nuclear power while recognizing nuclear tail risks—the same belief in AI’s capability explains both his enthusiasm for present tools and his concern about what stronger systems could do.
Nathan’s experience of effective-altruist circles is similar: people working on extreme risks routinely ask how their organizations can use current AI better. Masley adds that the Washington, DC community is intellectually heterogeneous, and he has felt no pressure to endorse “the absolute craziest things” about AI risk.

2. The strongest correction is not “no harm,” but “measure the trade-offs”

Masley does not dismiss uncertainty around the data-center buildout, which he calls a “huge new industrial project.” His objection is to treating any additional energy or water use as automatically harmful while excluding taxable investment, utility revenue, infrastructure upgrades, and AI’s potential value from the ledger.
The claim he wants to “demolish” is narrower and firmer: personal chatbot use does not meaningfully change an individual carbon or water footprint. He has encountered influential institutions considering restrictions because individual prompts supposedly cause unacceptable environmental harm, despite uncertainty ranges too small to support that conclusion.
One educational institution, he says, wanted to provide chatbots to low-income students but met environmental objections strong enough to halt the idea. At the relevant scale, Masley argues, the same reasoning could prohibit buying books; denying useful computing over tiny marginal footprints becomes “a massive tragedy.”
He locates the narrative less in professional environmentalism than in guilt-heavy climate social media. Environmental organizations have largely moved toward grid-level reform, but the public discourse still treats every new emission as a moral failure—even when “every day we all wake up and cause new emissions” through ordinary life.

3. Relative claims conceal how tiny the baseline is

The old line that an AI query uses 10 times the resources of a Google search sounds alarming only because the baseline is omitted. Even 1,000 searches are a small personal load, so multiplying that tiny quantity does not suddenly make 100 chatbot prompts environmentally significant.
Digital work also feels strangely sinful because its physical infrastructure is invisible. Water appears to be a dwindling substance burned by an ephemeral “2D thing on your screen,” particularly when the observer already considers AI useless; Masley describes an almost “religious sense of sin” around spending physical resources on disliked software.
Data centers reinforce the misconception by aggregating millions of tiny computations inside conspicuous facilities. Masley’s analogy: if society concentrated every toaster or microwave in one place, their load would also appear immense relative to one household, without implying that each use had become individually consequential.

4. A median prompt is approximately one second of microwave use

Masley’s all-in working estimate is about 0.3 grams of carbon per average prompt, including inference, cooling, training, and embodied hardware under assumptions he considers conservative. That is around one one-hundred-thousandth of daily personal emissions, requiring roughly 1,000 prompts to add 1%.
Generating and reading 1,000 responses might occupy 10 hours. Masley’s inversion is that an activity consuming only 1% of daily emissions while absorbing most of the day probably lowers the user’s footprint, because the displaced alternatives—driving, taking a walk, watching a show, or other physical activity—usually consume more.
For median requests, the conversation settles around 0.3–0.6 watt-hours, or approximately one second from a 1,000-watt microwave. Another mnemonic is about 60 prompts per full phone charge, though both speakers warn that people routinely classify appliances merely as using “little,” “normal,” or “a lot” of energy.
Long coding or reasoning jobs consume more, but Masley treats this as running the metaphorical microwave longer. If a minute produced “a very thorough research overview” or valuable code, he would not regard that minute as profligate—especially when ordinary microwaves are accepted for heating vegan chicken nuggets.

5. Household physics makes the scale legible

Nathan anchors the orders of magnitude: a human body runs near 100 watts, the brain around 20 watts, and an average US home around 1,000 watts. Heating dominates many household loads, with microwaves and electric kettles each drawing roughly 1,000 watts while operating.
A phone charge uses around 20 watt-hours—roughly one-third of a cent at Nathan’s Detroit electricity rate. Over a 12–24-hour cycle, that is about the same energy the brain uses in one hour; smaller language models already run locally on phones and laptops.
Masley supplies the unit discipline from his seven years teaching physics: a watt is a rate, while a watt-hour is that rate multiplied by time. The distinction matters because a large power rating used for a second can still represent little energy.

6. ChatGPT’s subscription price creates a hard upper bound

At $0.20 per kilowatt-hour, Nathan’s $20 ChatGPT subscription could purchase at most 100 kilowatt-hours—about four days of average household electricity or 10% of one month’s home use—if OpenAI spent every cent on power.
That deliberately absurd ceiling ignores chips, researchers, networking, facilities, and every other cost. Masley further notes that electricity produces only roughly one-quarter to one-third of total emissions, with cars and other direct fossil-fuel uses accounting for much of the remainder.
Even the impossible assumption that all subscription revenue buys electricity therefore raises total personal emissions only around 3%. Actual usage is far below that, unless AI companies are quietly giving users “huge amounts of free energy,” something their economics strongly discourage.

7. Chip prices reveal little about their carbon composition

An eight-GPU H100 server node costs approximately $300,000, while four years of electricity total around $35,000—only about 10% of the purchase price. That does not mean manufacturing dominates emissions; much of the price reflects Nvidia’s premium, margins, supply-chain complexity, and market structure.
Nathan applies Nvidia’s roughly 80% margin to show that marginal production cost might be near $60,000 before TSMC, transport, and other inputs. Translating dollars directly into carbon is therefore misleading: market power and engineering value are not quantities of burned fuel.
Nvidia’s published lifecycle figures suggest operational electricity may create around 20 times the carbon of manufacturing the chip. Masley’s analogy is a wire designed to carry current until it wears out: nobody treats the embodied carbon in household wiring as more important than decades of electricity flowing through it.
The larger hidden cost is model training, which Masley guesses might roughly double total prompt energy from around 0.3 to 0.6 watt-hours, with wide uncertainty. Embodied hardware may add only 5–10%, while electricity could be dirtier than average because reliable, inexpensive grids still often lean on fossil fuels.

8. One physical trip overwhelms years of ordinary prompting

Driving a sedan for 20 miles emits roughly 3–4 kilograms of CO₂, equivalent to about 10,000 median chatbot prompts. Nathan independently derives the same order of magnitude from gasoline mass: a 20-gallon tank becomes roughly 150 kilograms of exhaust carbon dioxide, yielding approximately 250,000–500,000 prompts per fill-up.
A transcontinental flight can represent 1–2 million prompts, probably more than Masley expects to submit in his lifetime despite being a power user. A long solo car trip can be surprisingly comparable to a long flight for a passenger; flying’s main effect is enabling much longer distances, not necessarily being dramatically worse per mile.
Nathan’s background calculations put a hamburger and a hot shower in the same broad 5,000–10,000-prompt range. Masley’s rule follows: skip one car journey and “be good for at least a year of prompting,” even before counting any journey AI itself helps avoid.
These comparisons expose the difficult part of decarbonization: most emissions come from useful activities people resist giving up. Focusing on supposedly useless prompts avoids confronting transport, heating, food, and industrial systems—the large, valuable loads where nearly all meaningful reductions must occur.

9. Substitution matters more than the meter inside the data center

Masley expects almost all environmental consequences of AI to come from changed behavior, not inference electricity. Assessing Amazon through the power used by its website would miss shipping, retail substitution, warehouses, and purchasing patterns; AI deserves the same system boundary.
Nathan offers a personal example involving his son’s Burkitt lymphoma/leukemia. Hundreds of queries across ChatGPT, Gemini, Claude, and occasionally Grok helped the family evaluate whether to relocate for care and whether suspected basement mold required renting and heating another house, ultimately helping him choose a lighter response using HEPA filters.
That experience converted the abstract substitution argument into physical avoided activity: a possible family move and a second heated home were more resource-intensive than the prompting involved. The energy used by several hundred or even 1,000 prompts was negligible beside one of those decisions, before considering the medical and financial value.
Masley contrasts individual guilt with collective leverage: helping open a solar plant, battery facility, transmission line, or preserve nuclear generation can have hundreds of thousands of times the impact of trimming prompts. By the chatbot level, personal austerity resembles “pausing YouTube a few seconds early for the sake of the climate.”

10. The bottle-of-water claim fails basic accounting

Masley calls a bottle per prompt perhaps the most popular falsehood in the debate. He traces it to a Washington Post calculation that assumed 10–20 prompts for a 100-word email, frozen chip efficiency, particular training allocations, and hydroelectric evaporation; his preferred median estimate is about 2 milliliters, roughly one bottle per 200 prompts.
A prompt’s all-in water footprint may be around one eight-hundred-thousandth of daily consumption once food, electricity, and supply chains are included. People see drinking and shower water but not the hundreds of bottle-equivalents embedded elsewhere in an ordinary day.
Avoiding one additional pair of jeans might save water equivalent to roughly 1 million prompts because irrigated cotton incorporates substantial water into physical production. Even a one-watt digital clock may cause about 3 liters of monthly off-site power-plant water use—an invisible footprint far larger than intuition suggests.

11. “Water use” combines three economically different activities

Consumptive use removes water from a local source through evaporation or another pathway that prevents prompt return. This is the category Masley worries about most in high-stress basins, where withdrawals can outrun replenishment and gradually dry the source.
Pollution is separate: cooling systems may add chemicals to keep equipment clean, then return water requiring treatment. Masley considers this a real but generally smaller concern than agricultural pollution, while declining to claim that data-center discharge is harmless.
Non-consumptive withdrawal takes water, uses it, and returns it—sometimes warmer—to approximately the same source. Much of the water attributed to electricity generation falls here, so combining gross withdrawal with consumption can inflate the apparent scarcity impact by an order of magnitude.
The fourth distinction is freshwater versus potable water. Data centers often prefer highly treated potable water to prevent buildup, but Masley says treatment capacity can exhibit economies of scale; he repeatedly qualifies this discussion as “some guy” reporting months of research, not a credentialed water expert.

12. The “half the United Kingdom” statistic is mostly returned water

One widely repeated projection says AI could use 50% as much water as the United Kingdom by 2027. Masley decomposes it: around 90% is power-plant withdrawal that is returned, another 5–7% is consumed at power plants, and only roughly 3% occurs inside data centers.
The headline invites an image of half Britain’s water flowing irreversibly through server racks. Once the accounting categories are separated, Masley estimates the relevant quantity closer to a small single-digit percentage of UK water involved—still nonzero, but a fundamentally different planning problem.
For US perspective, his best 2023 estimate is that AI used eight to 10 times the water used by his 15,000-person hometown. Spread across America, that resembles adding 10 small towns: meaningful enough to plan for, not remotely a national water emergency.
Even aggressive near-term data-center growth might add water demand comparable to roughly 1% of US irrigated corn use. Electricity demand is orders of magnitude more consequential relative to its system, which is why Masley’s confidence about harmless individual prompts does not extend to every facility-level power decision.

13. Gigawatt facilities are small globally and immense locally

The H100 draws around 700 watts, while newer accelerators plus supporting equipment make 1,000 watts per chip a workable mnemonic. That conveniently matches average US household electricity demand: one accelerator, roughly one home.
A 1-gigawatt facility therefore maps to approximately 1 million chips or 1 million homes. Stargate’s discussed 5-gigawatt target maps to 5 million homes, perhaps the residential scale of a 10–12-million-person metropolis; Nathan estimates that 5-gigawatt buildout near 1% of current US emissions.
Solar provides another ruler: about 10 square miles can supply one gigawatt, implying approximately 800 square miles for the hypothesized 80-gigawatt, $7 trillion buildout. That is under 1% of Michigan or Nevada, so land availability alone does not appear to be the binding national constraint.
Against roughly 6,000 gigawatts of global energy use, 80 gigawatts adds around 1.3%, consistent with the episode’s 1–2% range. General development in poorer countries may add more over the same period, though both speakers stress that longer reasoning, efficiency gains, and new generation make any forecast provisional.

14. Air pollution outranks water in Masley’s risk hierarchy

Once the discussion moves from prompts to sites, Masley’s “single thing” of greatest concern is air pollution—far above water and above direct climate impact. Data centers may be compact, but supplying them with coal or gas can impose immediate health costs on neighboring communities.
Carbon is globally fungible: emitting one unit here and preventing 10 elsewhere produces a net reduction of nine. Air pollution is not; Masley’s deliberately extreme analogy is a coal plant piping exhaust into his house while claiming aggregate pollution fell somewhere else.
Indoor air pollution already kills millions globally, largely through household cooking and heating, while US estimates for total air-pollution deaths range roughly from 30,000 to 100,000 annually. Masley emphasizes the causal uncertainty but notes that even the low end rivals other highly visible public-health problems.
By contrast, he has not found a case where normal data-center operations clearly reduced local water access; alarming examples often involve construction instead. He treats that as a provisional research finding, not proof that no case exists, and invites correction.

15. Memphis illustrates both the danger and the evidentiary limits

Around xAI’s Colossus facility in Memphis, residents reported smelling gas, and a recording reportedly showed more gas turbines operating than permits allowed. Masley is explicitly reluctant to adjudicate the episode: subsequent city testing reportedly did not show an ongoing problem, while the earlier facts remain contested.
The surrounding area already had “grade F air,” making the distributional concern unavoidable. Even normal grid expansion can shift health costs toward poorer neighborhoods located near undesirable industrial infrastructure; a rushed facility can intensify that existing pattern.
Forecasts through roughly 2030 include substantial gas and coal supporting AI loads, alongside contracts for new renewable power. Renewables could become cheaper through scale and improve the climate balance, but that possibility does not compensate a nearby community for immediate combustion pollution.

16. Data centers have not driven the national 35% bill increase

Masley leans on Lawrence Berkeley National Laboratory’s work suggesting that the roughly 35% rise in average US electricity bills since 2020 came mainly from inflation and supply disruptions, including temporary gas-market effects from the war in Ukraine. Nationally, data centers appear to have contributed little or none of that increase.
Local evidence is different: some utilities have explicitly raised rates because of data-center infrastructure. Highly concentrated loads require generation, substations, and transmission, creating legitimate disputes over who pays.
US electricity consumption was broadly flat from around 2008 through the early 2020s as efficiency improved and the economy shifted more toward services while industry declined. Demand is now rising again, driven chiefly by data centers plus electrification, but projected growth remains slower than the rates routinely managed between 1950 and 2000.
Higher regional demand does not mechanically lock in permanently higher prices—otherwise large cities would always be costlier than rural grids. Scale can lower unit costs after construction, but relentless data-center expansion might deny utilities enough catch-up time, prolonging elevated bills and capital requirements.

17. Local governments must price the benefits and externalities together

A data center can be a large taxable industry, a major utility customer, and a source of funds for aging pipes or grid upgrades. Masley’s line is memorable: “Nothing’s quieter than a ghost town.” Noise matters, but so did noise from the factories whose disappearance many communities now lament.
Race-to-the-bottom dynamics remain real: jurisdictions offer longer tax holidays and fewer obligations because companies can move elsewhere. Nathan argues for higher-level standards protecting air quality and ensuring local compensation, since aggregate surplus does not guarantee that affected residents receive it.
Masley declines to prescribe a universal municipal deal. His practical request is a full accounting of tax revenue, utility revenue, water conditions, air pollution, generation sources, and infrastructure obligations—followed by a willingness to say no where the trade is bad.
He favors stronger controls on pollution and pressure for renewable generation, batteries, and transmission, while warning that poorly designed environmental review can itself obstruct clean-grid construction. The policy target is “rigorous environmental regulations” compatible with the infrastructure required for electrification.

18. Desert siting can improve water economics if it displaces worse uses

Masley’s counterintuitive example is Maricopa and the greater Phoenix area: golf courses around Phoenix—or golf courses statewide, by another estimate—may use roughly 30 times as much water as the state’s data centers. A green course in the desert is his visual shorthand for water allocated to a low-revenue use.
By his estimate, data centers generate around 50 times as much tax revenue per gallon as golf. Replacing courses with server facilities could therefore create billions in revenue without increasing total water consumption; he supports substitution, not layering new demand onto an already stressed basin.
Irrigated alfalfa is larger still, possibly consuming 1,000 times the water AI used nationally in 2023, much of it ultimately feeding livestock. Masley also points to irrigated corn for ethanol and lawns, while admitting uncertainty over how much golf-course water later returns underground and whether every comparison uses identical consumption definitions.

19. AI’s indirect climate effects could swamp the data-center footprint

Masley cites an International Energy Agency projection that by 2035 AI applications might avoid about four times the emissions generated by data centers. He treats it as highly uncertain and notes that much of the benefit may come from specialized deep-learning systems rather than giant chatbots.
A single application can already operate at system scale: if Google Maps reduced car emissions by even 1%, that saving would represent a huge share of global data-center emissions. Building controls, routing, industrial optimization, and behavior changes matter more than the electricity meter serving the model.
Nathan recalls Jim Collins’s antibiotic work using models he believes had only millions of parameters and tens of thousands of data points, running on a few computers for days, though he explicitly says he does not remember the exact figures. Similar small systems might unlock materials, batteries, photovoltaics, or energy-efficiency gains without anything resembling frontier-model resource requirements.
Masley’s “goofy” historical calculation imagines a 1950 computer needing half of US energy to run Minecraft, versus a modern personal machine doing it casually. After seven decades of optimization, computing produces so much output per unit of energy that its applications are likely to dominate its direct footprint.

20. Rebound effects keep the net outcome genuinely uncertain

AI might make shopping more persuasive, manufacturing cheaper, and consumption easier, increasing total energy use even while each task becomes more efficient. Autonomous vehicles could reduce waste yet induce far more travel—for example, letting a five-year-old ride alone to a friend’s house with parental approval.
Masley separates wealthy-country restraint from development: he wants affluent users to decarbonize but considers much greater energy access for the world’s poorest people essential “basically by any means.” Rising energy demand can represent welfare improvement rather than policy failure.
At the far edge, advanced AI could destabilize geopolitics or contribute to war, which would plainly be environmentally destructive. He brackets those “goofy sci-fi scenarios” because his probabilities are unclear and they obscure the tractable near-term questions.
His closing assignment for environmentalists is to shape how AI is deployed, clean the grid, and police serious local harms—not make data-center energy the whole story. Adding “a drop of water” to stronger objections about surveillance or dangerous capabilities dilutes the critique: “People notice very fast when you’re just kind of reaching for any tool that you can.”

Nathan Labenz

Today, my guest is Andy Masley, a blogger and thinker who's done some of the best recent independent analysis of the energy and water demands associated with AI. Andy is the director of Effective Altruism Washington, DC, and this conversation served as a great reminder of why I appreciate the EA community. Their emphasis on identifying causes that are neglected, tractable, and large in scale seems right on to me, and the culture of epistemic humility and earnest truth-seeking is genuinely admirable.

What's more, Andy is a great example of how EA thinkers are generally very pro-progress, even as they worry about extreme risks from AI—a consistent pattern in my experience that contradicts the popular doomer caricature. In all sincerity, the main reason I've never gone around calling myself an EA is simply that I don't consider myself virtuous enough to deserve the label. Regardless of affiliation, this conversation also demonstrates how a genuinely curious, truth-seeking, numerate person can, with AI help, make a meaningful contribution to the discourse on a complicated topic, even without formal credentials or deep experience.

In this conversation, we tackle the prevailing narratives around AI's consumption of energy and water. We acknowledge that there can be local issues associated with large-scale data centers that really do matter to specific communities, such that local leaders should be careful about the deals they strike with data center companies. The bottom line, however, is that AI is not a huge deal when it comes to global emissions or water use. The main reason is simply that bits really are that much less massive, and therefore easier to manipulate, than atoms.

For the purposes of quick mental math, I find it helpful to remember a few key heuristics that we develop in this conversation. A single ChatGPT query uses roughly as much energy as running a microwave for 1 second. A single cross-town car trip, a hamburger, and a hot shower each use roughly as much energy as 10,000 ChatGPT queries. That means that if you can save just 1 car trip with AI use—which I've done a number of times in just the last couple of months—you've more than offset your AI use for a year.

Meanwhile, a 1-gigawatt data center uses as much electricity as 1 million American homes. That 1 gigawatt of power requires roughly 10 square miles of solar panels to produce. The full $7 trillion buildout, which is estimated to use something like 80 gigawatts of power over time, would represent a 1% to 2% increase in global energy usage. That's less than the expected increase due to general global economic development over the same period of time.

Finally, that full 80 gigawatts of power, if it were all powered with solar panels, would require 800 square miles of area—less than 1% of the size of the state of Nevada. The water-usage analysis is very similar, albeit a bit more complicated for reasons that you'll hear. Obviously, there is some uncertainty associated with all of these numbers. Of course, things will continue to change, with forces such as increased usage and longer thinking times pushing AI resource intensity up, even while increased efficiency and new energy sources push it down.

Nevertheless, I feel very good about the back-of-the-envelope calculations and the memorable comparisons in this conversation, and I hope that it gives everyone a confident, grounded sense of how resource-intensive AI really is. With that, let's get into the fundamentals of AI energy and water use with the director of EA DC, Andy Masley. Andy Masley, director of Effective Altruism DC and internet truth warrior online at andymasley.substack.com, welcome to The Cognitive Revolution.

Andy Masley

Yeah, so great to be here. Longtime listener, first-time caller. It's a really crazy privilege to be here. Thank you, Nathan.

Nathan Labenz

My pleasure. Thank you. That's very kind.

So today, we're going to get into AI's use of energy and a little bit of other resources, focusing on water, probably. This has been something that, for me, has really surprised me over the last couple of months. In particular, I went and did a talk at an AI-for-public-school-educators event in October.

There are a lot of questions on everybody's mind across society, and certainly in the education space right now: What is this going to mean? Aside from the very education-focused questions they were asking, I think the number-one question that kept coming up over and over again was, “What about energy? Is it going to ruin the environment? Is it going to blow up our electricity bills? What's going to happen with this energy thing?”

It's clearly something that people have understood is a potentially big problem. I think you and I are both, spoiler, generally on the same page that a lot of the fears circulating out there are pretty overblown relative to what we actually have to worry about—which is not nothing, but hopefully is something that we as a society can manage.

For starters, why don't you introduce yourself a little bit? I think one thing that is striking is that you may surprise some people with your profile, as I hope to do sometimes as well, with a mix of opinions that aren't super easily bucketed into doomer or accelerationist. Maybe give us your big-picture worldview and your relationship to AI as it exists today, and then we'll really dig into the resources question.

Andy Masley

Oh my, my full theory of the world? Yeah, we'll take a little while.

The basic place I'm coming from is the general EA perspective that it seems very likely that machines that can replicate a lot, or even most or all, of what can happen in the human mind might not be too far away. They seem more likely than not to happen in my lifetime, with massive error bars there, obviously, and that can come with a lot of risks.

I'm pretty agnostic about P(doom) stuff. I would probably lean in the direction of a lower P(doom), purely because I'm a little bit wary of the very specific philosophical arguments that get made around advanced AI rising up and having misaligned goals and stuff like that. That's kind of a separate conversation. But I have very high P(weird). I do expect the world to get very weird, and potentially dangerous, as more and more powerful AI systems come online.

This is all downstream of my believing that AI systems, as they currently are and as they will be in the future, just seem very likely to be very capable in general. As a result of AI being capable, I'm also pretty excited about AI tools as they currently exist. I think this is a big disconnect that actually just makes complete sense if you think about it from this perspective of AI being capable.

A lot of people are thrown off because I'm very visibly a massive chatbot user. Opus 4.5 is basically my new best friend. It's so cool. Love it. But I'm also quite wary and worried about AI's overall impacts on society in the long term.

I think that combination is actually pretty common behind the scenes. A lot of people probably identify with both being excited about AI as it exists and being scared about the future. But I think a lot of people have this really simplified version of the debate where you're either a hardcore doomer who hates all technology, including AI, and wants to stop it, or you're an accelerationist who wants to go full speed ahead and loves AI because you think nothing bad will ever happen.

There's an obvious middle ground here—a pretty easy Venn diagram intersection where it's like, “Yeah, I don't know. For the same reason that I think nuclear power is really good, but also that nuclear stuff comes with some tail risks, machines that mimic a lot of what happens in the human mind are very useful for me personally, and I'm very excited about a lot of what they're doing. But, man, this could get really crazy really fast, and there are a lot of potential risks that come with that.” That's the general background attitude that I'm coming in with.

Nathan Labenz

Yeah, I think that is pretty representative of the EA perspective. It's obviously hard to paint a movement with such a simple descriptor, but in my personal experience, when I've attended EA events, most of what people seem to want to talk to me about is, “Hey, could you maybe help advise my organization on how we can make better use of AIs as they exist today?” At the same time, they may be trying to raise awareness of, or figure out how to better wrap their heads around, some of these big-picture, long-term tail risks in any number of ways.

In my experience, that fear is very often packaged with an appreciation for what the technology can do for us today. Certainly, I'd say very few people in the EA movement, as I've encountered it, fail to recognize that.

Andy Masley

Very much. I think behind the scenes, EA is much more heterogeneous than a lot of people might be aware of. There are wildly different takes about the basic AI x-risk case. There are definitely different scenes where you can maybe receive social capital or punishment for believing different things, and I think that's bad, obviously.

But at least where I am in DC, which is one of the largest EA communities anywhere, I've never personally experienced any kind of social pressure to believe the absolute craziest things about AI risk. I feel very confident being very agnostic about that, actually, which is very nice.

So I'm a fan there. And yeah, I think most EAs I know are coming at this from a general perspective: we're very excited about a lot of technology. In some ways, it wouldn't really make sense if everyone were saying, “AI is going to be so capable in 20 years that it could kill everyone, but it's not currently capable enough to be useful in answering my emails or something like that.” That combination is becoming pretty rare pretty fast. Usually, a lot of skepticism about current AI models is also paired with skepticism about future risks from misuse or misalignment and stuff like that.

Erik Torenberg

Yeah. The coalitions are a little odd. The scrambling is well underway, which is probably healthy. I think better that than total polarization.

Nathan Labenz

Yeah.

Erik Torenberg

But it does mean people sometimes incorrectly project certain beliefs onto their debate partners.

Nathan Labenz

Yeah. Oh, sorry. Go ahead.

Erik Torenberg

No, please.

Nathan Labenz

Oh, yeah. Without calling anyone out, every now and then I'll bump into someone saying, “Oh, this guy's an EA. He's a communist-lite who wants to destroy all technology and just level everything because he hates freedom and stuff.” This doesn't really describe the profile of the people I'm bumping into.

I think there's just a lot of AI doom in the air right now from a bunch of different directions, including climate doom, which we're going to get into here. In most cases, there are just a lot of examples of worry about AI that a lot of EAs don't really share. I think my takes on AI and the environment are shared not just in EA, but with a lot of other people who are knowledgeable about the technical aspects of this. But there's a huge disconnect between that and broader society, and a lot of circles that are much more worried.

I tend to say that, but basically, part of this has just been presenting: “Hey, look, I'm an EA. I can say what I believe very directly about this stuff.” I definitely don't believe a lot of the standard environmental worry about AI as it's presented in a lot of the media right now.

Erik Torenberg

Yeah. So let's get into that. How would you describe what people are broadly hearing from the media? What is the baseline mainstream narrative that you're setting out to correct?

Andy Masley

Yeah, there are different levels of confidence I have about different claims here. The media is covering the general data-center build-out a lot, and there is a huge amount of uncertainty and concern that is definitely warranted because this is just this huge new industrial project. We don't really know how it's going to go or what the consequences are going to be.

So there, the main thing I'm trying to correct is that there are a lot of assumptions that this will automatically, definitely be bad for the environment, or that it's already causing these huge cataclysms. I want to adjust people more toward observing what the actual facts on the ground are and what the potential trade-offs are here. Let's not just assume automatically that using more energy or water is always necessarily bad.

There are a lot of places where this can potentially be good for communities that receive a new massive taxable industry that's using a normal amount of water and providing a lot of new revenue for utilities and stuff like that. I'm not saying this doesn't also come with problems. I'm just asking that the positives be weighed against the negatives on a case-by-case basis.

I do worry that a lot of the coverage, especially there, is framing it as, “Any environmental harm cannot be traded off against the positives. There's nothing good that's coming from this.” I just want to push against that extreme case while still leaving a lot of uncertainty.

The main thing that I feel most confident about, separately, that I do want to demolish—honestly, I'm just so tired of this talking point bouncing around—is the idea that your personal chatbot use is significantly contributing to your emissions or your personal water footprint. We can have a debate about whether personal footprints don't even matter and all that matters is big systemic stuff. I actually think that's pretty healthy as a way of thinking.

But if you are worried about this, I'm more convinced than I can be of most things that poking around on chatbots, even if you use them a lot throughout the day, is just not going to add a meaningful amount to your carbon or water footprint, and in many cases might actually reduce it. We can get into that more later.

I still bump into so many people with such wild misconceptions about this, including people who are very influential in big institutions and stuff, who are making decisions saying that the people in their institution cannot use chatbots because the individual environmental harm of those individual prompts is too much, basically. I think that's ridiculous. We just have the numbers. It's looking pretty good.

There's uncertainty, but the range of uncertainty isn't enough to justify this right now. That's a big thing I'm pushing back against.

Erik Torenberg

Yeah, that's crazy. I haven't heard of any organization leaders giving dictates along those lines.

Andy Masley

Yeah, I'm sworn to secrecy on who, unfortunately. I recently interacted with a large educational institution where the people in charge of the technology there were just completely convinced that this isn't actually so bad and that we should really be able to pay for this for our low-income students, especially, to have access to chatbots and stuff.

But we're mainly getting a lot of pushback from a lot of different places, basically being told, “This will be bad for the environment if we do this,” to the point that we can't do this. Using that logic, it also wouldn't make sense to ever purchase books for any students and stuff like that. There's just a lot of stuff to say about that.

But this is happening, and I think a lot of people who are focused on AI on the ground are pretty shielded from this because I think they're much more aware of what the actual numbers look like. I think people might be blind to just how common this is and how many people have not touched chatbots at all specifically because they're completely convinced that this is a significant addition to their personal emissions.

There's a lot more to say about that. Some people are not touching them because they're worried about contributing to a systemic thing. I can say more about that. But the general misconception here, I think, is just so widespread still, and it's a massive tragedy.

I think blogging about this kind of feels like shooting fish in a barrel. It's a little goofy when you actually look at the numbers. But yeah, I tend to get into that. That's the other main thing that I'm trying to address.

Erik Torenberg

Yeah. One strategy for rising to prominence online, building your online voice, is to pick an area where people are dramatically and consistently wrong, and then you can be very consistently right.

Okay. So where do you think—what is the move that's happening that's distributing this bad information, these incorrect understandings, in the first place? One thing I've observed in general is just that the world is really huge, and so when you aggregate numbers associated with almost anything, you get to these really big numbers.

I think one pattern that maybe is happening is people are saying, “Here's how much energy AI is going to use. It's, you know, whatever number of watts, and isn't that insane?” People are like, “That is insane,” and then there's just no reference to anything else. This very big number hangs out in a vacuum and intimidates people because they don't see the other, even much bigger numbers that maybe they should be comparing it to.

How much of it is that? What do you think is going on that's creating this broad misconception in the first place?

Andy Masley

Yeah, it's coming from a lot of different directions. I think there was a lot of coverage of this early on. There was a lot of individual news coverage about the cost of individual prompts, and there's an outdated talking point about how it's 10 times as much as a Google search.

I think for a lot of people, this was the first time they had ever seriously thought about their internet activity using resources in general. I think there's actually something that just feels really wrong to people about that, because digital goods are ephemeral and they're just this 2D thing on your screen.

Especially if you don't like AI or if you think AI is all silly or whatever, it's this ridiculous, meaningless use of these valuable resources, especially water. Actually, I think the idea of AI burning through water really sticks with people because they imagine it as this dwindling resource that we might eventually run out of, and we need to conserve it and protect ourselves from that.

There's this ridiculous new thing that they don't like burning through a huge amount of it for silly reasons. I think there's almost a religious sense of sin about that that really gets to people at some deep level, just because they might not have thought about how data centers have used water for a while for other things.

People also watch Netflix shows that I don't think are very good, and that uses water. Maybe I would be more upset about that if I thought it was a serious problem, but even there, not too much water is being used.

I think there's just been a lot of really weird and bad media coverage of this that's definitely caught on in social media, too. There are a lot of TikToks and tweets giving, I think, very silly explanations—grimly talking about how a chatbot uses 10 times as much energy as a Google search and saying, “You're multiplying your emissions by 10 times.” They never stop to mention, “By the way, 10 times a Google search is still not that much.”

If I searched Google 1,000 times throughout the day, that's still not going to add too much to my emissions, and so I shouldn't expect that 100 ChatGPT prompts will either. There's a lot of hyperfocus on these relative numbers because almost no one—including me, before I actually started looking into this—actually knows where a lot of the energy they use goes or how much it is.

There are a lot of different things in the mix. There's a general anti-AI attitude, which I otherwise understand people worrying about. There's a sense that this is using all these valuable resources.

The very last one is this very strong sense that, because we're in a climate crisis, it's never acceptable to add a new source of emissions. This is very common. I bump into this all the time, where even if AI is adding so little, it's still new emissions that didn't exist before.

This is actually a really strange category when you think about it, because every day we all wake up and cause new emissions in everything we do. Every time I drive my car, I'm adding new emissions to the atmosphere that will stick around for 1,000 years or something. The source doesn't really matter, in my opinion. It definitely matters to an extent, but I think a lot of people are really hung up on the idea that this is a new thing or a new way of emitting and are thinking less about other stuff.

The very last thing I'll say is that the data center buildout is objectively huge, and I think people have a lot of trouble computing how such a huge new use of energy can relate to such small amounts on a personal level. I think a lot of this is just that data centers are these really weird, paradoxical buildings that aggregate huge amounts of very tiny emissions in one place.

They all look really huge together, but this would all appear that way if we were able to aggregate anything else we do in society as well. If all toasters or all microwaves were in one spot, that would also look very large by the standards of an individual person. A lot of stuff is floating around in the background here, but all of this has congealed into this very broad sentiment that using ChatGPT is bad for the environment. In my opinion, it is not.

Erik Torenberg

Yeah, one of my general principles is not to psychologize other people's AI takes because I think the facts are confusing enough, and the fog of AI forecasting is pretty thick in many cases. So I really do try to avoid that. I guess I'll break that rule very slightly, very briefly, by asking: Do you think this is primarily coming from the environmental movement and is just bad tactics or a misunderstanding?

Another corner it could be coming from is an anti-AI motivation that's trying to find things that are resonant with the public, even if those arguments aren't necessarily at the core of the motivation. What would you say you've encountered in your online battles?

Andy Masley

Yeah, I actually mostly don't think this is coming from the actual environmental movement, like on-the-ground environmentalists. Maybe a few of them get mad about this in the way that other people do, but it's not really originating there at all. I think it's mostly coming from a much broader swath of people who are fired up and sometimes a little bit confused about climate, who engage with a lot of doomy climate social media.

I just want to flag throughout all this that I am very worried about climate change. I don't want to say that this is not a problem or whatever, but there's this separate category of person who isn't actually involved in actual environmental work but just engages with a lot of guilt-ridden media about very specific things that they do in their personal lives.

I think the environmental movement, for the most part, has correctly moved on from policing personal lifestyle stuff so much and into focusing on systemic changes to the energy grid. Huge win there. That's great. But the public discourse hasn't really followed along.

I used to be a physics teacher for 7 years, which is also important context here. I remember a lot of my students actually believed really strange things about climate, mostly because of TikTok. They'd be like, “Oh, we're all going to die by 2024,” or society's going to fall apart because of climate change.

I think this is mostly coming from people who have this cluster of beliefs about climate and who have kind of memed themselves into this idea that any tiny amount of new emissions is always a catastrophe. Then this silly thing, AI, is causing new emissions, and so that's bad, basically. That's my basic guess about the main culprit here. There's a broader social media sense of climate doom that sometimes causes people not to think so much about the actual numbers involved.

Erik Torenberg

Yeah. Okay, great. Well, let's get into some of the actual numbers involved. How do you want to attack it? I think we can come at it from multiple angles. There's the unit of 1 ChatGPT request and what that uses. There are also all sorts of comparisons we could use—to a microwave, to heating one's home, or to driving one's car.

There's also kind of top-down economic stuff, which I think is often way too quickly brushed past. ChatGPT is $20 a month. That sort of puts a pretty clear limit on how much energy it could really be using. Of course, they could be subsidizing it, but they are subsidizing the free users, presumably in part with the revenue from the paid users.

We'll probably at least dabble in all 3 of those, and maybe even more different ways of thinking about it. But what do you find to be the most intuitive or resonant way to start to break this down for people?

Andy Masley

For a while, I was mostly making those kinds of individual comparisons, and I was finding that in a lot of places they didn't hit. I was trying to figure out why. It's this much to charge your phone, and a microwave can add up to hundreds or even more if you use it for long enough. I'll limit it to hundreds for now. I realized that they weren't getting this other big intuition I had.

So my elevator pitch, if I meet someone at a party now, is that, to the best of my knowledge, based on all the resources that we have on this, once you add up all the ways that a chatbot can cause emissions—not just the inference, but the cooling, the training, and the embodied cost of the hardware—my best guess right now is that lands you at about 0.3 g of carbon emitted per prompt on average. I'm making a lot of assumptions and assuming that's actually kind of high, and that's about 1/100,000th of your daily emissions.

And so what that means is that you would have to prompt ChatGPT something like 1,000 times to increase your emissions by 1%. If you did that—if you spent the entire day doing nothing but prompting ChatGPT and reading all the responses—that would probably take you something like 10 hours, at least. And if you think about it, spending 10 hours a day on something that only uses 1% of your emissions actually implies that your emissions will be way lower than they would otherwise be, if that makes sense.

Because most other things that we do in our lives—we're not just going to be sitting and not doing anything all day. We would be taking a walk, driving our car, or something like that, or even watching a show on TV. And so, if ChatGPT is this small, it actually probably means that anything you replace it with will emit more, except for other things online. Computing is just so wildly efficient that almost nothing else we do uses as little energy, basically.

My elevator pitch is basically that ChatGPT uses so little energy that, on net, it actually seems much more likely to reduce your emissions. In general, probably the single most climate-optimized way you can live your life is living in a big city, being vegan, and spending literally all day on your computer. I check 2 of 3 of those boxes. I try to get out, [laughter] but I don't spend all day on my computer.

Computing is just so efficient compared to most other things that we do that it's not a good thing to trade off if that's what you're worried about for the climate. So that's my basic pitch, and I can go into a lot of the other comparisons. I like making the comparisons to microwaves and stuff like that, but the basic, ground-level thing is that this is going to reduce your emissions if you trade it off against anything that's not on the computer specifically.

It's just so small that it would be kind of sad if we were focused on this in any other context. It's like taking a little dropper from your pot of spaghetti and saying, “Oh, I need to save a few drops for later.” I think we would correctly call that a distraction for someone so focused on climate and environmental issues. So that's my basic pitch in a nutshell.

Nathan Labenz

Yeah, I think the bits-versus-atoms distinction is a pretty strong place to start. I often think about the energy that it takes to charge a phone. This is actually a little bit of GPT-4 lore: when I was doing the red-teaming of the original GPT-4, which was basically 3 years ago—a little more than 3 years ago at this point—it really took me through some of these energy analyses very early on. It was one of the Eureka moments where I thought, “Man, this thing can really help me educate myself in all these different areas.”

But I was shocked to learn that a typical cell phone, to charge it fully, takes something like 20 watt-hours of energy, which in my jurisdiction in Detroit, Michigan, is about a third of a cent worth of electricity. We've also heard a lot recently that the brain itself runs on 20 watts. So, depending on how intensively you use your phone over a 12- or 24-hour cycle, that's the same amount of energy that your own brain is using in just 1 hour.

Already, all the things that your phone is doing—and obviously we've seen language models that can run on phones—we've not seen the frontier ones, of course, but you can run Gemma models on your phone, you can run things on your laptop. Those are literally running on less energy than your own brain, which is an interesting stark comparison. And it really just goes up from there.

Andy Masley

Yeah. And this is actually a reason why I've mostly moved away from comparisons to individual objects, because I've realized that most people—including me, before I started—don't actually have much of a solid idea of how much energy individual objects use. I've come to think that most people seem to have a scale in their head where something either uses a little amount of energy, a normal amount, or a lot. And it's one of those 3, basically.

So they might think, “Oh, ChatGPT uses”—this isn't actually true, but let's say ChatGPT uses as much energy as it takes to charge your phone. They might think, “Oh, that's a normal amount. So ChatGPT is adding this new, significant thing to my carbon emissions,” and they're not thinking about just how big the scale is.

Without getting into Twitter stats too much, I just had a guy respond to me the other day saying, “Oh, ChatGPT uses as much as charging your phone or running a microwave for an hour.” I wanted to hit the brakes and say, “Well, those are 2 very different quantities.” Almost nobody seems to notice just how little energy it takes to charge your phone specifically. This is a comparison that gets made a lot.

I can make these other comparisons where it's something like 60 chatbot prompts per phone charge, or per minute of microwave use. A chatbot prompt is effectively like 1 second of a microwave. I will flag here that whenever I say “a prompt,” I'm talking about the median chatbot prompt. This is a whole separate thing we can get into with much longer prompts and stuff like that, but we need a place to start.

So, something like 60 prompts per phone charge, or about 1 prompt per second of using a microwave, is a useful comparison. These comparisons to household stuff can sometimes be less helpful than I expected, because a lot of people—including me, a lot of the time—don't have much of an idea about this.

Erik Torenberg

Yeah, that's helpful. Well, maybe let's dwell on it a little bit more and maybe help popularize a few of these memes. I guess orders of magnitude are always helpful, right? The human body runs on basically 100 watts, and 20 watts of that is the brain.

The microwave typically runs at about 1,000 watts. An electric kettle is also about 1,000 watts, and the theme here is obviously that heating things is energy-intensive. The microwave is heating, and the electric kettle is also heating.

Running my furnace here in the wintertime—there's a lot of energy going through that, and obviously it depends on how cold it is outside and all that kind of stuff. But it roughly shakes out to, at least for electricity, from what I understand, the average U.S. home is using about 1,000 watts of electricity on an averaged-out basis. Obviously, when you turn on your microwave, that spikes up relative to your base load.

The rough intuition is that a home is running on 1,000 watts, a body is running on 100 watts, a brain is running on 20 watts, and ChatGPT then comes in with—you said 1 prompt is roughly equal to 1 second of a microwave.

Andy Masley

Yep, just like that. And my inner physics teacher is coming out, so I just want to flag for the listeners: a watt is a rate of using energy, and a watt-hour is a unit of energy because it's that rate multiplied by a certain amount of time.

A thousand watts, if you run that amount of power for an hour, means you use 1,000 watt-hours specifically. A ChatGPT prompt, from the best I can tell, is between 0.3 and 0.6 watt-hours for the median prompt. That's effectively like running a 1-watt thing for about 0.3 to 0.6 hours.

A 1-watt thing is also like an alarm clock or something. If your digital clock is running normally, 0.3 watt-hours is about the same as 0.3 hours of running a digital clock specifically.

Erik Torenberg

Yeah, that's really helpful. And I guess it's also worth noting that a kilowatt-hour is the typical unit of electrical billing. So when I said it's 20 cents in my power jurisdiction, that 20 cents is for a kilowatt-hour.

That also leads to another interesting comparison: if my $20 ChatGPT subscription went entirely to energy, it would buy 100 kilowatt-hours, which would be basically 4 days of a typical home's electrical usage. That would assume, again, that they're taking all of my $20 and spending it on electricity. Obviously, we know they're not, because they've got chips to buy, researchers to train, and all sorts of other costs.

But that kind of gives you an absolute maximum. If you were to just spend $20 more on electricity, that's roughly increasing your typical home electrical bill by 10%. And so, again, it's way less than that. We seem to be coming down to these 1%-ish numbers, and it's probably still significantly less than that today.

You had said you have to get to 1,000 prompts a day to increase your overall footprint by 1%. I do think that's where the people at the model-development companies think we're going, and that's why they're planning such big build-outs. But it is hard to imagine usage that moves one's overall energy footprint all that much more than that.

Andy Masley

And I will flag, too, that when people talk about energy, obviously there's some concern about electric bills around data centers, but mostly when people talk about this, they're actually using it as a proxy for carbon emissions, right? Because we're worried about climate.

Something that doesn't get through to people a lot is that—I don't have the exact number, but it's something like 25% to a third of our carbon emissions come from the electricity that we use, and two-thirds come from other sources, like direct use of fossil fuels, driving your car as an example.

Even within that electricity, assuming that all of your money—all of the $20—is going to ChatGPT and it becomes 4 days of your household bill and 10% of your electricity use, because your electricity use is only about 25% of your actual emissions, that actually drops even more to something like maybe 3% of your actual emissions.

The absolute most ridiculous situation is where OpenAI is taking your money and spending all of it on electricity all the time, which is actually much smaller than that. Even there, the absolute most this could possibly be raising your emissions is something like 3%. And that would actually be significant. Something that raises your emissions by 3% is definitely not nothing.

But my claim is also that it's much smaller than that. And that's the absolute most it can be economically, unless, of course, ChatGPT is just giving everyone huge amounts of free energy, which just doesn't seem likely. AI companies just don't want to do that if they can avoid it.

Erik Torenberg

Yeah. Well, there's maybe a good way—I don't know if this is the next place we should go immediately—but another good angle of attack on that question is: What does it cost to buy chips, and then what does it cost to run chips? Do you think we should dig into that now?

Andy Masley

Sure.

Erik Torenberg

Yeah, let's do it.

Andy Masley

Yeah, I was looking into this a little bit, and it's actually pretty counterintuitive. The cost of a chip versus the cost of running it in money actually doesn't seem to tell you very much about the carbon costs involved, because most of the actual cost of the chip seems to be the NVIDIA premium and the cost of these incredibly complex supply chains all working so well together, and the market, basically, with NVIDIA running a lot of this stuff. There's this huge premium that you're paying.

As an example, an 8-GPU H100 server node costs about $300,000. Over a 4-year lifespan, the server will consume about $35,000 in electricity, which is only about 10% of its purchase price. So, for this chip, as an example, 90% of the cost is in purchasing it, and only 10% is in actually running it.

You might think that if that's the case, maybe most of the carbon cost is also in producing it. But that price is mostly reflecting other stuff.

One intuition I have that's pretty useful for this—and I'll talk about more of the proportions in a second—is that ultimately, when you think about it, an AI chip is just designed to have electricity flowing through it as much as possible over its entire lifespan. If I designed a wire and electricity just started to flow through the wire until the wire burned out, and I was like, “Okay, what took more carbon: making the wire itself, or making all that electricity flow through the wire to the point that it eventually blows out?” I don't know. It's kind of similar to—I’ve never once thought about the carbon emissions of the wires in my home, but I have thought about the carbon emissions of my electricity bills a lot.

And so, in the same way, because so much electricity flows through AI chips, we should basically expect most of their carbon cost to be in the electricity itself rather than in the embedded cost of making the chip in the first place. NVIDIA conveniently shares the carbon cost of all its chips. It seems like one positive aspect of people focusing more on the environmental stuff is that a lot of companies are just doing this more.

Based on what I can tell, the ratio is something like 20:1 for the carbon of the electricity versus the actual embodied cost of making these chips themselves. So it's wildly disproportionate. It's something like at least 10:1 in the other direction for the cost of purchasing the chips in money versus the cost of the electricity. But the carbon cost is completely flipped, where most of the carbon cost is in the electricity.

Erik Torenberg

Yeah. So another way to think about that, or kind of validate that, would be: If it cost $300,000 to buy this pod of 8 chips and then $35,000 to run it over a lifespan of 4 years, then we all know right off the bat that NVIDIA's margins are something like 80%. So that takes down their cost of marginal production to something like $60,000.

And that still includes the relationship with TSMC, shipping the things all around the world, and all that sort of stuff. So you have at least a governor there on the top of, well, it can't be more than that, because we know what their margins are.

And then you're also highlighting that they're just straight-up publishing an analysis of what their energy is and saying it's like—did you say 1% of total, or—

Andy Masley

Oh, it's 10%. The energy cost of running it, in dollars, is about 10% of the cost of actually purchasing the chip in the first place. And then the carbon cost is flipped, where it's something like 20:1. The carbon from the emissions of actually using the electricity is about 20 times as much as that.

And it could get even higher, because a lot of AI data centers are built with grids with reliable, cheap energy. Unfortunately, right now, in the reality we live in, that tends toward more fossil fuels being used. As a result, their electricity might actually be even dirtier than the grid average. That's another thing that's actually raising the carbon cost even more.

So out of a $300,000 pod of 8, you can then expect, call it, a $30,000 energy bill over the lifetime of the chip. And then if you flip back to, well, what was the energy cost of that on the production side? It's another 20:1 reduction, meaning something like $1,500 spent on energy, or, in other words, basically half a percent of the original cost of the chip.

It's always interesting to flip these things back and forth between money and emissions. But because the markets are so weird here, it just doesn't track very closely. Money tells you very little about how much was emitted at different points.

And I think that, in general, this is a really funny area where sometimes when I talk about the carbon cost of a median prompt, people will be like, “Oh, but you're not including these hidden costs,” like the embodied cost of the hardware itself. They'll never actually try to make a guess at what that actually is.

When you poke at the embodied cost of chips, that anyway seems pretty marginal. It's like 10%, 5% sometimes. So that doesn't concern me super much. It doesn't add too much to the cost.

The real secret cost that I do think significantly raises it is the cost of training the models, but that's a software question rather than a hardware question, obviously. That's something else we can talk about. But my best guess right now is that maybe it doubles the full energy or emissions cost. So it goes from 0.3 to 0.6 watt-hours or something like that.

It has pretty wide error bars. But yeah, I can go into that separately. But at least for the hardware, I think it makes about as much sense to worry a lot about the embodied carbon cost of AI chips as it does to worry about the embodied carbon cost of the wires in your home that electricity flows through. These things are just so hyperoptimized.

There’s also just a common theme here: so many people are being paid to think in really complicated ways about optimizing the energy use of these chips as much as possible. There’s a huge amount of incentive to do that. We should expect that, yeah, there’s a lot of electricity flowing through them, but they’re also relatively optimized.

There’s just a lot of stuff to say, basically, but I expect that most of the cost in general of a chip is in running the energy through it rather than in the energy used to make it.

Speaker 1

Yeah, okay. I think that’s quite helpful.

Andy Masley

Oh, I mean, I love sending them. Do you want to do one about driving a car or taking a flight?

Speaker 1

Yeah. So, like, this is actually—let’s see—the best I can tell is that if you’re driving a car for 20 miles or so, that’s probably going to emit 3–4 kg of CO2 if it’s a sedan. And that’s something like 10,000 chatbot prompts.

Andy Masley

I sometimes think about it like this: if I didn’t want to feel guilty about chatbot prompts, if I was worried about that, avoiding a single car trip would give me 10,000 free chatbot prompts to use however I want.

This is also an interesting argument for ways that chatbots might be able to net reduce someone’s emissions in other ways. In these conversations in general, something else I try to flag is that almost all of the climate impacts of AI, and environmental impacts in general, are almost definitely going to come from how AI is used rather than the actual energy used in data centers.

For the same reason, all other computing works the same way. If we’re trying to think about Amazon’s impact on the environment, there’s shipping and other things that it affects, but looking at the energy cost of running the Amazon website in data centers isn’t going to tell us very much about its total impacts. And so, if AI is actually changing your behavior at all, that’s just where most of the impact is going to be.

If one of those 10,000 chatbot prompts that equals a 20-mile car ride prevents you from taking that trip—you realize, “Oh, I looked into this, and I don’t actually need to drive out to do this thing,” or, “I can get it somewhere else”—that’s at least going to modify your behavior enough that it’s going to significantly change your emissions.

Potentially not for the better, either. Maybe it will cause you to do something else that is silly. But the comparisons here get kind of comical. Flights can get up to the millions at times.

A single transcontinental flight can, I think, go over at least 1 million and potentially 2 million chatbot prompts, which is more than I’ll probably do in my lifetime. I consider myself to be a chatbot power user, but I’m probably not going to hit that. If I make a very extreme decision like, “Oh, I won’t take that European vacation this year, but at least now I can use chatbots basically forever without any guilt,” that’s a pretty funny comparison.

Water comparisons get even more ridiculous because I think the average person doesn’t actually know where most of their own water footprint is. The main way we experience water in our lives is in the water we drink, maybe a shower, or watering our lawns if we have one. Most of the actual ways that our lifestyles consume water are far away from our homes. It’s in the food that we eat, in the power that generates or is used to power our homes, and stuff like that.

I’m pretty sure the average water cost of a prompt, if you include everything, including the upstream cost, rises to something like 1/800,000th of your daily water consumption. The numbers get really silly there. I’m pretty sure that if someone avoids buying a single additional pair of jeans, that’s worth a million chatbot prompts’ worth of water or something, just because jeans might require irrigated cotton to grow.

This makes intuitive sense to me because, if I think about how water is being used in the jeans case, huge amounts of water are flowing to grow plants, and they become part of the plant itself. Whereas with AI, the water is flowing in the vicinity of hot computers, basically, and it’s capturing some of the heat that my individual prompt generated. I’m like, “Well, how would I expect the proportions here to look?” Ultimately, it makes sense that there can be these huge order-of-magnitude differences between those things.

A single second of your microwave is probably about 1 ChatGPT prompt, or even 3 or so. It just gets pretty comical pretty fast. If you’re willing to use your microwave for a few additional seconds, you should be okay with a few additional chatbot prompts, too.

This is actually a good comparison for longer prompts because some people will be like, “Oh, what Andy’s saying is about median prompts, but obviously there are all these other things, these long coding tasks and reasoning tasks and stuff like that.” It’s true that those use a lot more energy, but it seems kind of like just running your microwave for longer.

If I had a friend with a microwave that, if you used it for a minute, could produce a very thorough research overview of a field, or if it heated up a ton of new useful code for whatever I’m working on, I wouldn’t actually be super bothered if they were using the microwave for that. I use it to heat vegan chicken nuggets, and I don’t feel bad about that.

In comparison, when these larger models start to get talked about, and either the prompt is much longer or the reasoning time is much longer, that definitely adds to the cost. But I think anyone who’s used a deep research prompt or a coding prompt also knows how valuable the outputs are compared to even a normal prompt.

The microwave comparison is really useful here because it’s basically like, “Oh, this thing is like a microwave, and however long it takes is kind of like running a microwave that long,” with error bars, obviously. But if I had a microwave in my home that produced a deep-research-level response to something, I would be using that quite a bit and not feel super bad about the results personally.

So, yeah, that’s all I have to say about that. The comparisons can be useful here in a lot of ways once you get the context down.

Speaker 1

So let’s—let me just rederive that one again, just for my own and everybody’s benefit, to get the reps on stepping through it. Let’s say I have a 20-gallon tank in my car. That’s one way I like to think about it: the unit of the fill-up, right?

Gasoline weighs about 5 lb per gallon. Water is 8 lb per gallon. Gasoline is lighter. So, roughly, you fill up your car with a 20-gallon tank, and you have put 100 lb of fuel into the tank.

Now that actually becomes, somewhat surprisingly or weirdly, a lot more massive as it gets converted to exhaust because your hydrocarbon is basically just a chain of carbons with hydrogen atoms alongside it. Each carbon atom becomes CO2 in full combustion. You can have CO, I guess, in there as well, but obviously we’re going to simplify away a lot of details, so we’ll call it CO2.

Oxygen weighs more than carbon, so you’ve more than—roughly speaking, you’ve tripled the mass. For every carbon atom that gets combusted, you have a CO2: triple the mass. So I’ve gone from 100 lb of fuel into the car to 300 lb of emissions, which, if I convert that to kilograms, we’ll call it 150 kg.

Then that becomes 150,000 g. Compared with 1 chatbot prompt at a third of a gram, my full tank of gas gets into the range of roughly half a million median chatbot prompts.

Then you had scaled that down by, let’s say, taking just 1 20-mile trip. That would be like 1 of my 20 gallons. So if I divided half a million by 20, I get down to 20,000–25,000, like 10,000. We came at it from a different angle, but we ended up within 50% of each other.

Andy Masley

Yeah, yeah. I should double-check the numbers, obviously, because it gets a little hand-wavy at the extremes here. But yeah, exactly.

This brings me back to when I was first learning about climate change. I remember hearing the fact that, oh, if you see a huge pile of coal and it all gets emitted, the CO2 weighs so much more than the pile because oxygen is being added to it. I remember just having this deep sense of evil. Coal is such the ultimate bad guy in climate conversations that it’s a really funny fact that the stuff that you use will end up adding way more to the atmosphere than it itself weighs. It’s pretty crazy.

But yeah, that’s a really good intuition. A full tank of gas can use at least hundreds of thousands of chatbot prompts, and we do that pretty regularly. This leads into another interesting general comment on climate stuff, where the main way that we emit—and the main way that most of our emissions happen—is actually with stuff that is otherwise very useful to us.

I think people sometimes have this idea that they should focus on chatbots a lot because they think chatbots are useless, and it's a useless form of emissions. I think something like 99% of the actual emissions reductions that we need to make are going to be in the useful category. Highlighting these tiny little useless, quote-unquote, emissions makes sense to worry about. Obviously, we want to emit less if we can, and if something's useless, we want to cut it.

But in terms of the amount of attention it should actually take up, the big challenge of climate change that's actually preventing us from making change is how much of our emissions are actually very useful. Driving a car is very useful to me. Another part of this debate that I find comes up a lot is that no matter how small I can express the chatbot's emissions, people will be like, “Oh, yeah, but it's useless, and it's just this silly plagiarism machine.” So we should somehow focus on it more than all these useful emissions.

I want to get it through to them that the useful emissions are actually so large and so much more difficult to change. That's what needs the most of our attention. So, there's a lot to say about that, but the car example just brings up a lot of different associations there. Just to repeat it one more time.

Nathan Labenz

Yeah. Yeah. Yeah.

Andy Masley

One cross-town 20-mile trip is worth on the order of 10,000 chatbot prompts. One tank of gas, 20 times more than that, is worth something like a quarter to a half million chatbot prompts. And then if we go up to a cross-country flight, we take basically another couple of orders of magnitude up and get to basically a million, which again makes sense if you think about just the cost of the flight and how much of that goes to energy.

From what I've seen, it's a nontrivial amount. Unlike AI companies, the airlines are actually turning around and spending a pretty significant share of your money on jet fuel. So, you might be spending north of $100 per passenger on jet fuel to get cross-country.

Nathan Labenz

Yeah. What it actually brings up is a funny intuition: driving solo actually seems to emit about as much as taking a long plane flight as a passenger, specifically, which I think people have some weird intuitions about. They think the plane might be 1,000 times worse, but driving is surprisingly bad. The main way that flying adds to our emissions is that it causes us to travel longer distances, rather than being worse per mile. So it makes sense that the plane scaling up is going to be about the same as an equivalent amount of distance traveled in a car.

And, yeah, this all just adds up to comical amounts of chatbot prompts. Even if you thought this was evil but saw any value in chatbots at all, you could skip a single car trip and be good for at least a year of prompting or something like that. I try to get across that all these cuts just seem very depressing to focus on in the first place.

Back when I was talking a lot more about climate change when I was in college, there was a really great general consensus that the name of the game is systemic change and changing the energy grid. There, you can have hundreds of thousands of times as much individual impact as anything you can do in your personal life. Even more extreme things—I’m vegan, I live in a city, and stuff like that—and that's good, but it's just not nearly as impactful as being one small part of a group of people who open a new solar plant or a battery facility, keep a nuclear power plant open, or something like that.

Within personal lifestyle stuff, if you're going to litigate anything, litigate a few big things. By the time you get to chatbot prompts, it's as if you're pausing YouTube a few seconds early for the sake of the climate or something. It honestly makes part of this debate quite depressing to me because I feel like the general climate conversation has receded or gotten worse since I was talking about it, with so much hyperfocus on these tiny things that just aren't going to matter compared to the big systemic stuff we need to do. Yeah, that's weird. I really [laughter] don't know what to say about that.

Andy Masley

It's upsetting, yeah.

Nathan Labenz

But I guess that's why we're here.

Andy Masley

Yeah.

Nathan Labenz

Okay, so 10,000 prompts per 20-mile car trip. A million prompts for a cross-country flight.

Andy Masley

One prompt is one second of a microwave. Mm-hmm.

Nathan Labenz

Hopefully these start to put people's minds a little more at ease.

Andy Masley

Yeah. Actually, one other comparison that I realized we hadn't made yet is that I want to make it completely clear that there's a really common talking point that I never want to hear mentioned again because it's so wrong and so far off: that a chatbot prompt uses a bottle of water.

If you look into where this came from, it was this really misleading article from The Washington Post that basically took this really wild estimate of how many chatbot prompts it would require to write a 100-word email. It seems to assume 10 to 20 prompts to write a single email. I'm a power user of chatbots, but I don't use that many.

Even there, it assumes all these other wild things, like chips hadn't gained any efficiency in the 5 years since they were commercialized, along with specific things about training and evaporating water from hydroelectric plants. All these things together tell what I think is a really misleading story. The actual amount is something like 1/200th of that. It's something like 2 mL of water.

This seems pretty conclusive based on a bunch of different estimates and different studies from people. The general comparison to water specifically—I always want to flag that maybe the single most popular wrong piece of information about these chatbots is that they consume a bottle of water per prompt. This is not the case. You'd have to prompt them about 200 times to use a bottle of water.

Also, you use thousands of bottles of water—or maybe not thousands, but at least hundreds of bottles of water—every day in your consumptive use that you don't see, because most of that is off-site in the electricity you generate and in the food that you eat.

Nathan Labenz

Yeah. I didn't pre-run the numbers on a hot shower, but I suspect there's an awful lot of prompts in a hot shower as well.

Andy Masley

Yeah. I could say all kinds of crazy things about hot showers, actually, where a lot of the water that goes into the hot shower is the water used to generate the electricity to heat the water. Basically, the way we use water is just very interesting and sometimes comical, because so much of the water we use is actually used to generate electricity. Even your alarm clock has a big off-site water cost.

My best estimate is that the average digital clock uses something like 3 L of water off-site at a power plant per month. Even there, that might be invisible to you, and you might not know that this thing is burning through 3 L at a time.

So, yeah, there's a lot of stuff to say about that. Part of the reason I've been writing about this so much recently is that it's honestly just pretty cool. There are a lot of counterintuitive ways that we use energy and water and stuff like that.

I think the chatbot conversation can honestly be really fun once you contextualize it in the counterintuitive ways that society uses water. We didn't even touch on all the goofy excesses in animal agriculture and stuff like that. That's its own separate topic.

Nathan Labenz

Yeah. Well, I think agriculture would definitely be worth doing a couple of little comparison exercises with, too. But maybe before we do that, how should we think about the relationship between energy and water, and what does it mean to use water?

I think there's conflation going on in many cases where people think this water is used—like it's destroyed or contaminated. Obviously, contaminating water is a big problem. My sense is that's not really what is going on, and that's not what people are referring to when they refer to water use, right?

This is more evaporative-cooling kind of processes, where the water is not removed from the water cycle entirely. It's just put into the air from wherever it was before that. But help me sharpen that up.

Andy Masley

Yeah. There are 3 different categories of using water that are useful to think about. First of all, the most important category to know about is consumptive use of water, which is water that is withdrawn from a very local source and either evaporates or does something else to it that causes it to leave that very specific source of water. That can be disruptive for high-water-stress areas because the water might not be returned at the same rate, and over time things can dry up. So consumptive use is basically the main harmful way that water can be used.

There's also water pollution, where water is returned polluted in some ways. AI data centers can contribute to this a little bit because they can use specific chemicals in the water to keep it very clean, but they might return some of it to a local system polluted with these chemicals. This doesn't seem to be nearly as big a problem as the pollution that might come from agriculture, but it's definitely not nothing.

That’s something to worry about a little bit. The other thing to focus on is non-consumptive withdrawal of water, which is where you take water, use it, and then basically give it back to the source. A lot of the ways that we generate electricity, as an example, mostly use non-consumptive water withdrawal, where water is taken and maybe returned a little bit warmer than it was before. That can sometimes create some problems, but for the most part, it’s just taken and put back.

This can lead to some really wild, confusing statements about how AI uses water. One very popular statement about AI water use is that by 2027, it’s projected to use 50% as much water as the entire United Kingdom. If you actually look at what that means, something like 90% of that 50% is water that power plants withdraw and then immediately put back. I think that paints a very different picture.

Another 5% or 7% of that water is consumed in the power plant and not in the data center itself. Only about 3% of that amount is going to be in data centers themselves, specifically. When people hear that, I think they imagine half of the UK’s water flowing through data centers. What’s actually happening is that, at most, maybe 5% of the UK’s water is involved, which is definitely significant—it’s not nothing—but it shows how easy it is to wildly and sometimes unintentionally mislead people about this, because the way we use water is so counterintuitive.

One other thing to note is that there’s an important difference between freshwater and potable water, which is water that’s been improved enough to basically be drinking quality. Data centers themselves rely a lot on potable water because it’s very important for the water to be very clean as it moves through the data center, so as not to gunk things up. From what I can tell, there’s not too much of a problem right now in converting freshwater to potable water. There are actually economies of scale where, if you’re converting more of it, that can make it cheaper rather than more expensive.

There’s a lot of stuff to say about that, but another common thing that gets said is that AI is using drinking water. That sounds like a big emergency because obviously people need water to drink. I think people have a really weird understanding of the relationship between drinking water and freshwater, which is basically just water that’s not salt water. In a lot of areas, the big constraint is how much access to freshwater people have, and not so much how much access they have to drinkable water, because there can be times when economies of scale kick in.

There’s a lot of uncertainty about this. I want to flag that I’m basically just some guy who’s done a lot of deep dives on this topic over the past few months. I’m reporting what I know to the best of my ability, but I’m not an expert. If you think I’m getting anything wrong, please email me right away. That would be very embarrassing. What I’m saying is based on a lot of reading and a lot of individual ideas that each aren’t actually super complicated, but can be confusing when you first bump into them.

So, there’s a lot more to say about water, but those are the really useful categories to understand.

Erik Torenberg

So let’s bottom-line that one more time. We’ve got the three categories. A lot of what we’re taking out to do this cooling can be put back. That’s not without any issues. What should I say at a party?

Andy Masley

Oh, so super quick. I will [laughter] go. Yeah. What should you say at a party? I’ll flag that most of the water that flows through data centers themselves actually does seem to be consumed, where it’s evaporated. Maybe 20% of it is returned to local sources, and sometimes a little bit is polluted and needs to be treated.

Okay, so my best guess right now is that, first of all, I would describe consumptive versus non-consumptive use. If someone says, “Hey, AI uses half as much water as the UK,” I would say that about 90% of that is withdrawn by power plants and then returned. So we’re actually talking about more like 5% of the water that’s actually consumed by the UK.

Consumption is the big problem. Withdrawal has some issues, but the main thing I’m worried about is water that’s not returned. Within that, AI is a relatively small percentage. My best guess right now is that in 2023—and obviously a lot has changed since then, but that’s the last time we have good data—AI in America used about 8 to 10 times as much water as my town used. I’m from a town of about 15,000 people.

That’s not nothing, but it’s basically like spreading 10 towns of 15,000 people each all across America. If you just ask how big of a water problem this would be for the country, I don’t actually think that would be a super big deal. It comes with some interesting planning problems, but I usually keep in the back of my pocket, “Oh, it’s about this much.” At least it was 2 years ago. Then I can talk about how much it’s changing and how much it’s growing.

The 2 things I would say are: those water statistics—here’s how much of it is actually consumed—and then, within that consumption, how much is AI? Maybe 20% to 30%. How much does that compare to towns or other things that are using water? I kind of lose sight of how clear this is to other people when I’m talking, because I’ve been so in the weeds. If you want to rephrase that for yourself, that might be useful here.

Erik Torenberg

Yeah. I do think the comparison to towns and cities is really helpful. I’ve done that for myself a little bit more on the energy side, so maybe I’ll do the energy and then you can try to do the water.

It is pretty handy that the latest chips basically run at 1,000 watts as well. The H100 was around 700 watts. There’s a little bit of extra padding on top of that for the auxiliary things that also consume power at the data center. The bigger and better chips use a little bit more. They’re getting a little bit more efficient—not necessarily a ton more efficient in terms of FLOPs per unit of energy these days, because a lot of the low-hanging fruit there has been picked—but, roughly speaking, we can still round to 1,000 watts per chip.

That’s notably the same as a home. You can then work that up to these big data center buildouts. If somebody’s going to do a 1-gigawatt data center, that’s 1 billion watts, which would be 1 million chips if each chip is at 1,000 watts. That would translate to basically 1 million homes’ worth of electricity.

What I’ve seen from the Stargate plans, for example, is that they’re trying to get to 5 gigawatts over a couple of years of this buildout. Five gigawatts would be 5 million chips, or 5 million homes, or a city of 5 million homes, which would be maybe more like 10 to 12 million people—something around the order of magnitude of New York or Los Angeles. That’s not small, but it’s only a single-digit percentage of the American population.

As you noted earlier, with electricity being roughly a quarter of our emissions footprint, if you say, “Okay, well, New York City is—I don’t know—4% or 5%, depending on whatever metro area, whatever you want to draw the line. Call it 5%, call it 4%, of the U.S. overall footprint,” and then dial that down by the 25%—only 25% of your overall emissions footprint is electricity in the first place—you’re basically saying Stargate, in its full 5-gigawatt buildout, is roughly a 1% contribution to U.S. emissions as they exist at baseline.

Andy Masley

Yeah, a ton to say about this. I actually want to flag that now that we’re talking about big data center-level stuff, I have way less confidence about how much the data center buildout affects things than I did about the individual prompts. Some people get a little bit confused by that, because isn’t there some relationship there? I can go into detail on how I think about how these individual prompts contribute to the overall picture, why this is confusing, and why it’s not.

Once we get into this, I have to say that gigawatt data centers are just mind-blowing. It’s crazy to think about buildings—just stepping back and being like, “Oh, yeah, a single building could use a percentage point of America’s household electricity.” It’s insane that that’s happening. That’s just completely mind-blowing to me.

At the individual prompt level, I’m not worried about electricity or water. At the global level, at the level of data centers, I think it’s very clear that electricity is going to be the big challenge here. At least from the estimates that I’ve seen, even at the maximum amount of projected data center growth over the next few years, the water cost seems to be about an additional 1% of America’s irrigated corn. That’s not nothing—it’s actually quite a bit—but that’s potentially adaptable. The system can handle an increase of that size.

The electricity issue is orders of magnitude larger as a proportion of America’s total grid right now. That’s going to be the huge challenge. I’m a lot more uncertain there. I could totally believe that this will be very bad for local environments.

Erik Torenberg

That's what I tend to say about that. But, yeah, gigawatt data centers are a very useful comparison. Each of these chips effectively uses as much energy as a home, or at least an apartment. A thousand watts is a really great, simple comparison to a home, and it's very convenient that it's around there because you can just do the orders-of-magnitude multiples. So, a gigawatt is something like a million homes' worth of energy; 5 gigawatts is 5 million homes—completely wild.

This is a place where I'm quite worried about the local environmental effects of the electricity specifically. There's a ton to think about there. I'm shooting off in a lot of different directions because I'm kind of a numbers person as a baseline and just want to flag a ton of stuff. That's a really great comparison between homes and data center usage.

So how do you think about this? I guess if I had to summarize the conversation so far, it's that at the very aggregate level, we're looking at something like a low-single-digit increase in emissions from AI, even with the mega-projects that have been announced all coming to fruition. From an individual-user standpoint, you almost can't get there.

Those visions are obviously predicated on a vision for the evolution of AI where it's going off and doing long-running autonomous science, coming back with discoveries, and working in the background for you nonstop. How does one get to 10,000 prompts a day, where you would potentially be using as much energy with AI as you are with a daily commute? The answer is that it has to be doing an unbelievable amount of stuff in the background for you, right? This is a very different regime of AI use.

We're accounting for all of that when we say we're getting to a low-single-digit percentage increase in emissions. Water is even less. So that's kind of, okay, great. Now, the next level of analysis that concerned citizens, including yourself, would get to would be, "Okay, well, maybe that's true."

As a brief anecdote for myself, the offset has been very real for me personally recently. I've talked about this on another episode, but my son currently has Burkitt lymphoma—Burkitt leukemia—and I've used AI nonstop to try to help me understand what that is, whether the doctors are in sync, and whether there's anything else we might be doing.

How many prompts is that across all of this? I do everything in triplicate: I'll go to ChatGPT, Gemini, and Claude, and sometimes even Grok, although it's really the main 3 that I trust the most. Even doing everything in triplicate, I'm probably at a few hundred prompts over the last month of doing this—maybe 1,000. I've still got a long way to go even to get to 1 cross-town trip.

I would really emphasize how much it has saved me in the real world. The things we were considering that the AI helped us understand we didn't really need to do included moving our entire family to another city to take advantage of some other medical center that might have better care.

Another thing is that we had a possible issue with mold in my basement at the same time that my kid was going through this and was vulnerable to a possible mold infection. So I had this question of what to do about the fact that I might have mold in my house and my kid was extra vulnerable to it.

With the help of AI analysis, I was able to get comfortable buying a few HEPA filters and putting them around my house. That does have its own energy cost, which probably trumps the AI use by a lot, but it's still a lot cheaper than going out and renting another house and having to heat my own house and live in another one. Having 2 houses that I'm heating for the winter would be easily another order of magnitude up.

In just the last month, I've personally offset enough real-world transportation of my family, or renting and heating another house, to basically pay for your and my lifetime of conventional prompting AI use—if not necessarily the full Altman vision of tons of agents working for you around the clock in the background. I think that part is worth highlighting because it can feel hand-wavy to people, but I've definitely lived, in a very concrete way, how this has saved me from having to do dramatically more energy- and resource-intensive stuff, and obviously financially intensive stuff as well.

So the ROI in every dimension there has been outstanding, not even mentioning the impact that it's hopefully had in terms of reducing the overall risk to my son's long-term well-being. That's notable. But, okay, let's get into the local stuff.

These things are mega-projects, no doubt about that. One big building or cluster of buildings might use as much energy as a city of 1 million people, or you could even imagine that getting into a couple million people's worth of energy consumed at a particular site.

One way I started to think about this is that we can also talk about what kinds of energy are used, because that's going to change things a lot, too. I get the sense that in the short term, we're going to burn a lot more gas. In the medium term, we hope to have a lot more solar, and we hope to have a lot more nuclear as well.

I looked at solar and found that 1 gigawatt can be provided with basically 10 square miles of solar panels. That's not a small area, but then I asked, "Okay, let's say that's 1 gigawatt. You can get 1 gigawatt powered for 10 square miles. What if we go to the full 80 gigawatts that is projected as the total usage of the grand vision, Altman's $7 trillion buildout or whatever?"

Now you're going from 10 square miles to, let's say, 800 square miles. That's still less than 1% of the land area of the state of Michigan, where I live, or the state of Nevada, which is sunnier and much emptier, although there's a lot of empty space everywhere.

That's another thing that I think is dramatically underappreciated. The view from the cockpit of a plane is that so much of the time there's no sign of civilization in front of you. The cities are relatively small things over the vast landscape.

If we can power the entire grand vision of the $7 trillion buildout—80 gigawatts—with less than 1% of the land area of either Michigan or Nevada, it seems like land use per se isn't the problem. So how should we start to think about the impact in localities?

There obviously still is the question of Memphis, right by where the Colossus facility is being built. What's happening there? Is there something that's unfairly locally concentrated that we should be more concerned about, even if the aggregate picture all sounds fine?

Andy Masley

Yeah, a ton to say about that. A lot of different directions. I wanted to circle back and just say that your story has been really amazing to follow over the last few weeks, by the way. I'm really grateful that you're sharing so much, and obviously you're doing a ton.

The Memphis example actually brings up the single thing that I'm most worried about by far as an environmental impact of data centers. In my ranking, far and away, I think the biggest threat I'm worried about is air pollution, much more so than climate impacts as a whole. I can explain why climate impacts are much lower on my list, much more so than water.

Land use gets really ridiculous because America just has a lot of land, honestly. Data centers are incredibly compact for what they're doing. If you think about the processes happening inside of them, they're in some ways miracles of compactness. It's like Moore's law of the building or something like that. It's crazy.

With climate stuff, I think it's very likely that most of the impacts AI will have on the climate will be in how it's used, not in the data centers themselves. It seems very likely that there are so many opportunities to reduce emissions using AI. Even Google Maps—this isn't a great example of current AI—but if you believe that Google Maps has reduced total car emissions by 1% on net, that's already a huge portion of all global data center emissions. And that's just 1 app.

The thing about climate is that it's actually very fungible. If I emit 1 unit of carbon somewhere and reduce 10 units of carbon somewhere else, the harm is equivalent to me just removing 9 units of carbon, right? The impact of carbon is over the long term; it's not an immediate bad thing.

Air pollution is very different because it's just not fungible in the same way. If a nearby coal plant built a pipe directly into my home and said, "Oh, sorry. We did the math, and on net this will reduce air pollution as a whole, even though it will make it much worse for you," I'd say, "Well, that's not cool," because air pollution has immediate, very bad health consequences.

Andy Masley

Something that really puts this into stark perspective is that every year, it seems like millions of people are dying from the effects of indoor air pollution, especially. And this is more than the WHO expects will die from climate change every year, even into the 2040s and 2050s. Air pollution is already a much bigger disaster than climate change will be in the medium term, anyway.

In America, we have pretty clean air, but there are still a lot of bad things that can happen. It’s such a normalized part of our lives that I expect that even if AI is just using the normal grid and increasing the use of natural gas, along with some combination of renewables, air pollution actually seems like the biggest threat. As a baseline, it’s already very bad for us.

Memphis is something I’m very worried to comment on because I really don’t want to get this wrong. It does seem like the local community reported that they were literally smelling gas in the air, and there was a lot of movement around this. There was a recording of the Colossus using more gas turbines than they were permitted to use.

I have to admit, I just don’t know nearly as much about this as I should. There’s a lot of contested stuff that happened. It does seem like the city has run a lot of tests since then, and this is not a continuing problem, but it might have been happening at the time. Again, I want to quadruple-check this before making any big claims, but it’s illustrative of the broader point that even if data centers operate as normal parts of the grid, the normal grid is actually somewhat bad for us right now.

That cost is often shifted to very poor communities that live in undesirable areas with more air pollution as a baseline. The area that the Memphis data center was polluting already had Grade F air. It’s an obvious case of how pollution is distributed.

If you look at the forecasts for AI, it’s forecasted to use a lot of coal and natural gas at least into 2030 or so. There are all these complicating factors, though, including agreements to buy new renewable energy. There can be huge benefits to renewable energy from economies of scale there.

The climate impacts get really uncertain for me. On net, this could actually be a benefit for the climate, both in how it’s used and in building out so many new renewables that economies of scale happen and they become cheaper overall. But I really worry about air pollution as something that communities need to take extremely seriously.

Air pollution is actually the environmental impact I’m most worried about. Water just doesn’t seem to add up in most places. In most places I’ve investigated, I just can’t find a single place where it’s impacted water access at all.

A lot of people will immediately be up in arms if I say that because they’ve seen a lot of scary stories. My claim is that basically every scary story about water access doesn’t really add up when you actually look at the numbers, or it’s related to the construction of a data center rather than its normal operations.

I’m not really worried about water in the short term, but I am pretty worried about the air pollution effects of a lot of additional gas and especially new coal plants opening up. That seems really bad. Somewhere, there’s a huge gap in my level of concern: I’m moderately worried about climate impacts, not really worried about water, and extremely worried about air pollution, just because air pollution is a very normalized, very bad thing in the country right now.

Speaker 2

Yeah. I’ve learned a little about air pollution and how bad it is for us over time. It seems like it’s kind of like lead, in the sense that you want basically as little as you can possibly get.

Speaker 1

My sense about those numbers that you led off with, in terms of millions of people dying annually from indoor air pollution—

Speaker 2

Yeah, that’s different. I realize—

Speaker 1

That’s people mostly heating and cooking their homes with wood fires, right?

Speaker 2

Yeah. In America, I can’t get good numbers on how many people are dying from air pollution. But a lot of the estimates are really shocking and range from 30,000 to 100,000 people a year, which is more than guns. It’s in the same range as cars and stuff.

It seems like it’s pretty hard to actually understand the causality here. It could be more, and it could be less. I get fired up a lot about how I’m excited about Waymo because 40,000 people die in cars each year, and I’d like to at least automate that away. That would be really amazing.

But air pollution specifically does seem, by any measure, to be killing at least tens of thousands of people every year. Thinking about how these new data centers contribute to additional deaths from air pollution is something I would really want communities to be careful with. That’s the main way I’m not really gung-ho about building these really massive data centers around vulnerable communities specifically.

Speaker 1

Yeah, a ton to say about that. I’m pretty agnostic about the specifics because I haven’t done the deep dive on exactly how bad it is to have a gas turbine a mile from you versus a coal plant 10 miles away or something like that. But that’s the main thing I would be concerned about.

In some ways, I get quite frustrated by the water conversation distracting so much from what I think is a much more serious issue. How much do we know about whether that’s a sort of threshold-effect problem? Do you have any advice for everyday listeners who are like, “Geez, I hadn’t thought about air pollution”? Should they go buy—

Speaker 2

HEPA filters for their own homes? How, if you’re just living in a place where you don’t smell gas, should you still be taking action on the margin to clean your own air more? I haven’t really thought about this enough to comment.

I know I’ve read scary stories about how, back when toll booths were common, there were disproportionate and surprising negative health impacts on people in toll booths standing next to idling cars so much. I sometimes worry about just being in a city with a lot of traffic, where maybe that on its own is pretty bad.

This is actually a place I haven’t really done a deep dive on, so I don’t know the exact answer. Basically, prioritizing the general quality of the air where you live probably matters a lot for your health. Besides that, it’s not really something I know very much about, honestly, so I don’t really want to comment too much.

Speaker 1

Yeah, I do. One thing I think we’ll never go back on is upgrading to the highest-quality filters that you can put into your furnace. We have a central forced-air furnace in the house, and those things come out amazingly dirty after 8 weeks or whatever they’re in there. You can see that there’s some benefit to it, so we’ll stay top of the line on that.

On the HEPA filters, it’s interesting because they also bring a noise-pollution element. There’s increasing noise being made about the evils of noise pollution, too, so I’m not sure how to think about trading those off. At a minimum, furnace filters are a no-brainer, it seems like.

Maybe having a couple of HEPA filters around one’s home would be a good precautionary investment in one’s long-term health.

Speaker 2

Yeah, it’s super interesting, and it’s kind of fun. It’s evidence of how interesting the AI-and-environment conversation can be. It’s like, “Oh, by the way, HEPA filters.” It just bleeds into so many other things that can be pretty interesting.

I also worry a lot in these conversations that, in purely focusing on the negatives—and I’m not trying to bracket off air pollution here, because air pollution is uniquely bad—by the time we get down to noise pollution, a lot of the positives of the data center might not be being considered.

I think about a data center as a new, huge taxable industry in an area. A lot of people will talk about how so many parts of America are in decline and people are having so much trouble because industry left. A lot of those industries were also sometimes noisy. If there was a factory near you, you would probably sometimes hear noise pollution and stuff.

When new data centers are being built, they often contribute so much to the local tax base and can sometimes help the economy in other ways, too. I don’t want to say they’re always good, but I do sometimes worry that people will be like, “Maybe they contribute millions of dollars a year to the town, but they make some noise.”

I’m like, “Well, this would have prevented us from building factories and stuff, too, back in the day, and the town would have really never gotten off its feet in the first place.” I’m very happy to go into the nitty-gritty of land use, noise, and these other weird aspects of data centers, but I do worry that at that point we’re starting to lose the forest for the trees and not consider that there’s a trade-off to be made here.

Andy Masley

Like, this introduces some new noise pollution, and you get a million dollars a year in tax revenue for your town or something like that. Is the trade-off worth it? I would say so in a lot of cases, obviously with the consent of the people around it, but it's definitely important to consider the positives as well.

Speaker 1

Yeah. I mean, another thing that makes noise is neighbors. You know, I've got neighbors [laughter] who run their lawnmowers during the day and have a leaf blower or whatever.

Speaker 2

And it's like, yeah, nothing's quieter than a ghost town. So that, in the limit, doesn't really work.

Speaker 1

How about just the sort of energy cost to the consumer? I think one of the big complaints we're hearing now, too, is that this is going to jack up the cost of everybody's electricity, and that's not fair. I think there's at least something to be said for that.

I guess maybe a 2-part question would be: how big of a deal is that? What have we seen so far? And then maybe you could play policy adviser. It seems like, at the macro level of policy, you'd be like, “Build—do the AI buildout.” It kind of sounds like your mainline recommendation: it'll probably have enough offset; it'll probably work.

But then there's the local policy level of, okay, what if you were advising the Memphis city council, or whatever city council there? They have more particular concerns around these water, land, air pollution, and noise pollution issues. How much of an issue have we seen with electricity bills? And how would you advise the local decision-makers who are like, “Everybody wants to get these industry investments”? The companies are often pretty good at playing jurisdictions off against one another. Where would you say, “Okay, hold the line here for your people to make sure you're getting a good deal,” versus, “What can you accept? Maybe you give some ground, because again, it might just be worth it”?

Andy Masley

Yeah. Going into the electricity thing for a second, the Lawrence Berkeley National Laboratory, which is one of my favorite sources on data center stuff—they produce a lot of really good work—produced a general report at least very strongly implying that almost all of the recent increases in electricity costs, at the average national level anyway, seemed to be due mostly to downstream effects of inflation and a few other bumps. There was a point where the war in Ukraine was raising gas prices for a little while and stuff like that. I can't really figure out exactly what's going on, but the general overview takeaway seems to be that a lot of the increase in average electricity prices at the national level in the past 5 years has been caused entirely by issues with supply rather than demand.

And so, even though we've added a lot of additional energy demand, data centers don't seem to have contributed to that rise. There's this common statistic going around right now that since 2020, the average American electricity bill has risen by 35%. From what I can tell, data centers don't seem to have contributed at all to that 35% rise. A huge amount of it is just inflation. When you factor out inflation, the rest seems to be a combination of a few weird issues with supply.

And I'm deferring here to the report itself. The report's making these claims, and I'm just trying to summarize it and do a good job there. So at the national level, it doesn't seem to be an issue. There have definitely been local places where utilities have come out and said, “We have had to raise electricity rates because of the data center.” I haven't found anywhere that they've done that with water yet. I could be wrong, but with electricity that's definitely happened.

There's a whole separate, weird debate to be had about this. Maybe we need to pressure utilities to do more buildouts of infrastructure anyway for the green energy transition. They maybe don't want to, but we have to pay higher electricity bills for the green energy transition anyway, because electrifying everything is going to require huge amounts of new green energy. But maybe data centers eat up too much of that, and they're very zero-sum altogether and don't actually add too much to that process. There are so many unknown unknowns here that it's very hard to talk about too much, and I don't feel fully equipped to diagnose it.

But basically, I think there are definitely places where data centers have so far affected electricity prices. I think those effects are pretty overblown, in that most of those price rises have come from these other things. But this could all change in the future as well, just because data centers are projected to use such a massive amount of new energy in America. One other complicating factor here is that, from what I can tell, America's total electricity use over the next 5 years will definitely rise significantly, primarily driven by data centers, plus a few other things like the electrification of cars and stuff.

But this rise will actually be slower than it has been in the past. We were in this really weird state from 2008 to the present, or until 2021 or so, where electricity consumption in America was mostly staying flat because we both optimized a lot of things and also just shifted more to a service economy. So we lost industry and, as a result, used less total electricity, and total American electricity use is rising for the first time since the financial crisis, basically. This is a weird new situation for us to be in in the short term, but in the long term, it's actually projected to rise less than it did historically, from 1950 to 2000 or something like that.

And so I think we have it in us to manage that, honestly. The big difference here is that now we really don't want to be building a ton of new fossil fuels, which, no, we didn't really know in the 1960s or whatever. And so it's going to be harder to do that as we build out green energy, but not impossible. As that rise happens, it seems manageable to keep bills low, but data centers are so much more concentrated compared to other sources of demand. So I have to throw my hands up and say I don't really know.

There are a lot of people smarter than me who have looked into this a lot more who are pretty concerned about future electricity bills rising a lot. One very last note is that there's this funny thing that I think a lot of people have a bad misconception about: if demand for electricity goes up in your region, your electricity costs stay higher forever. That's not actually the case at all, because that would imply that very large cities would always have higher electricity bills than very rural areas. What's actually happening is that there are a lot of economies of scale where, as more demand comes online, there's a temporary increase in cost, but over time that can level out.

But another weird thing that's happening is that data center demand might just keep going up and up and up and not give grids enough time to catch up. We might be facing very long periods of higher electricity bills to fund the continuous buildout to support data centers. It's really fascinating. It's one of the most interesting topics I've ever scraped up against: how the American energy grid is actually going to deal with this. I'm kind of left, first of all, believing that people actually have a surprising lack of understanding of what's causing electricity bills to rise and fall. Even experts disagree about this a lot. The future is just so wildly uncertain.

I will say that I'm very agnostic about whether we should do the AI buildout in general, because I'm excited about AI, but I want to maximize the benefits and minimize the risk and stuff, and I'm totally open to there being a huge amount of risk involved in this for the uses of AI. So I just want to flag that.

Finally, one last thing: at the community level, I actually don't feel like I have enough knowledge to give specific communities advice on this, outside of, “Do a full, thorough analysis of the pluses and minuses here.” Don't just look at the fact that the data center uses water at all. Also consider the tax revenue and the utility revenue. There are some cases where using water is actually good because it gives the utility more revenue to upgrade aging pipes and stuff like that. Really push for renewable energy if you can, build out battery stuff, and maybe pressure the utility to build out a lot of high-voltage transmission stuff anyway. But this is all very hand-wavy. I'm kind of just some guy on this, so I don't want to speak more confidently on this one thing than I am.

Erik Torenberg

So, to try to bottom-line a couple of things there: one is, if there's a concern about the macro AI buildout, it's the macro issues of controlling the AI long term, not the marginal emissions. Locally, there can be a lot of issues, but it seems like maybe what we need for the general protection of the population is some sort of federal standard. Not to invoke the impossible, but it seems like we do have a challenge where we saw this with Amazon and its second headquarters, famously, and with all these sorts of things.

Even just pro sports teams: “Build us a stadium, or we'll go to some other city that will build us a stadium.” We do have a problematic dynamic where companies have a direct incentive, and often do choose to locate themselves in the jurisdiction that will give them the most breaks and the fewest obligations.

Nathan Labenz

And so it does seem like we are probably headed for at least instances of local governments making deals with companies where they effectively race to the bottom in terms of what protections they insist on for their local populations, ultimately leading to what might not be a great deal for those populations. But the real solution seems to be that we need some sort of cooperative equilibrium to be established, or, if not an equilibrium, maybe some sort of higher-level rule to insist that certain things get paid back into the community or remediated, so that whoever is living in the area isn’t breathing bad air. There’s enough surplus to go around for it, for sure.

It’s just a question of—I don’t mean to dismiss this, because this is always the case, right? It’s like the same was said about the China shock in the first place, which left a lot of towns that are now looking for new industry in a bad place. Well, sure, aggregate welfare goes up. You can always redistribute. Is that effectively going to happen? It hasn’t always. Some might say it hasn’t in most cases. So we face that same challenge again.

But it seems like the bottom line on all these resource dimensions, and even the pollution-externality dimensions, is that the surplus is definitely there for it. If we want to take one big takeaway, it’s: let’s get our act together this time and actually make sure that we insist on reasonable standards or get the right compensation for people who are directly and locally affected. If we do that, we should be fine.

It still shouldn’t be so much that it would send the AI race to China or prevent us from being able to use all the chatbots, AI agents, and, for that matter, image- and video-generation models that we could possibly want. But the local political economy of this can be problematic if we’re not careful about it.

Andy Masley

Yeah. It’s funny. This is something else that I worry doesn’t always come through. I feel like I’m often writing in response to the most extreme takes about data centers, where people imply that it is literally always bad to use water on data centers. I’m like, well, here are all these trade-offs. In a lot of America, it is good to use more water because economies of scale happen, and the main access problem is aging infrastructure, not raw access to fresh water.

But in doing that, I often don’t also stake out that, yeah, I’d like a lot more rigorous environmental regulations on this stuff, with the goal of doing a green-energy build-out. There’s another nuance you can get into, where certain environmental regulations can actually stifle the types of builds that we need for green energy, like high-voltage, long-distance power. There’s a lot to get into there.

And the race-to-the-bottom thing is interesting, too, because it’s definitely a huge problem. You read about how people are just delaying taxes on data centers for longer and longer periods of time, among other things. People will talk about this but then, in the same breath, be like, “Oh, and there’s no benefit to communities anyway.” My question is always, well, if there’s a race to the bottom, why are all these city governments racing in the first place?

A big part of it is that there are some tangible benefits to having a data center. All else equal, I’m from an area that is not super well-off and has, in some ways, been left behind by industry. I think I would be actively excited if a medium-sized data center were built in the vicinity that the town could then tax and use for things. I think that would be net good personally, but there are a lot of other places where I would probably be much more hesitant, especially a super-large data center that was going to build a huge amount of gas turbines everywhere.

So, yeah, I’d be personally pretty excited about a lot of very specific, targeted regulation that prevented air from being polluted in significant ways but didn’t prevent the types of grid upgrades that we need for the green-energy transition. Again, I’m kind of a hobbyist here, so I’m not speaking from a position of authority. But I think my general ask for most people considering this is: even in race-to-the-bottom dynamics, it’s important to consider what the actual trade-offs are.

Is this worth it? Definitely say no. I don’t think that data centers should be built everywhere, but it’s important to consider the positives and the negatives, as opposed to just deciding that because they don’t see any value in AI, it shouldn’t be able to use any of the community’s water. There are a lot of instances where, even if the thing inside the data center is useless, it can actually sometimes have positive effects on the local community.

Erik Torenberg

Are there any rules of thumb that you could venture around who should not build a data center? I mean, one candidate would be: if water is already really scarce in your area, maybe you’re not a good candidate for a data center.

Andy Masley

Uh-huh. I have a really crazy take on the water stuff, actually, which is that I think data centers are quite good candidates for very water-scarce areas, like Maricopa in Arizona. A lot of new data centers are being built around the Phoenix area, and you might think, “This is ridiculous. Phoenix is a city built in a desert.” But there are a lot of other industries and commercial buildings that use a lot more water than data centers do.

The golf courses around Phoenix, from what I can tell, use something like 30 times as much water. Or the golf courses in the state of Arizona overall might use something like 30 times as much water as all data centers in Arizona. That on its own is pretty crazy to me. I don’t know if you’ve ever been to that area specifically, but it’s kind of eerie to stumble on this sickly sweet green in the middle of a desert.

And if these industries are using water, the difference between them and data centers is that data centers are often generating way more tax revenue per gallon of water used. Best I can tell right now, if you add up the total tax revenue and water used from both golf and data centers in Arizona, data centers are generating something like 50 times as much revenue per gallon used.

Basically, I wouldn’t support additional water being used in Arizona, but I would definitely support less effective industries being swapped out for data centers. If I could close all the golf courses in Arizona and replace them all with data centers in the state, that would generate billions of dollars in revenue for the state without increasing the water bills at all. Because data centers generate so much revenue per gallon used, at least right now, they actually seem like surprisingly good candidates for the desert.

There’s a lot to say about that. I definitely don’t want to increase water use there more than it already is. I don’t know. Besides that, the main rule of thumb is that it seems really preferable not to build data centers near where you have to open new coal plants or keep coal plants running to operate them.

Coal is just so bad for people’s health and for the climate that that seems good. Beyond that, I don’t have any clear rules of thumb. It’s not to say I wouldn’t object to more specific cases; it’s just kind of hard to think about on a case-by-case basis, and I haven’t really developed an overarching “Andy strategy.”

It seems a little bit goofy because all of this started with me wanting to win arguments at parties about using chatbots, and I’ve escalated to, “Here’s my complete plan for the data center build-out.” At some point I have to step back and be like, “I don’t know.” I kind of trust people who’ve been thinking about the very local ecologies of these regions and the local economics to know enough to make good decisions here. I think I’m just, in general, somewhat deferential to local experts.

Erik Torenberg

Yeah. Well, I would say the world is grappling with this in real time. One of my big themes of this overall journey of making this show is that AI is intersecting with everything, and we need people to step into these niches, whether they started in a particular niche or are jumping into it based on a motivation as pedestrian as being annoyed by stupid stuff other people are saying.

It is important that people take ownership of figuring out different corners of this whole phenomenon. So I applaud you for doing that. Obviously, more expertise and more local knowledge are always good, but we are moving very quickly into this, and I think you’ve definitely done the overall discourse a service by piping up many times, in many places, with more grounded analysis than we’re otherwise typically getting.

And then, yeah, I guess the local stuff—I’m glad we got to golf, because that was another one that I was really marveling at. I was looking at L.A.-area golf courses, and again, these are sort of giant numbers.

One thing I did want to double-click on for a second is: are we using the same meaning of “use” when we talk about water use? I’ve seen that an LA golf course uses 90 million gallons of water a year, or that LA-area golf courses use an estimated 1.6 billion gallons per year. That’s a ridiculous number, and it dwarfs what it took to train GPT-3 and things like that. But are we talking about the same kind of use, because that water sort of seeps back into the ground, right? Is that the same or different?

Andy Masley

This is where I start to scrape up against some issues. I have to admit that I’m not an expert, so I don’t want to comment too closely. I mostly try to stick to places where water is being actively chosen to be delivered by people. I don’t count—I try not to count—rainwater as an example. If crops are mainly relying on rainwater that would otherwise just be falling into the ground, you can say crops are using that, but it’s different. So, at least for agriculture, I mainly focus on irrigated agriculture specifically.

I have to admit I don’t know too much about the water dynamics of how this eventually flows back, so I’d have to circle back on this. This is an autodidact’s curse that you’re bumping up against: I know a lot about the specific water that is being delivered to these places, but whether this ultimately counts as consumption or withdrawal is a little bit beyond me, as is what the difference is and how big of a deal it is.

There was one other important difference here: the water used in data centers themselves, like I said, is potable, whereas a lot of water delivered to crops or golf courses is just fresh water. Sometimes it’s potable as well, so that’s an important difference to keep in mind. But it just doesn’t seem too costly to turn fresh water potable, based on my understanding. There’s an initial upfront infrastructure cost, like a water-treatment plant, but there are a lot of unknowns there. I wish I had a better answer; that was a good question.

I think, in general, the single really weird way that we use water that I’d like more people to know about is irrigated alfalfa farms. Again, I’m not entirely sure how this eventually flows back into the system, but they seem to use something like 1,000 times as much water as all AI used in 2023. There are a lot of irrigated alfalfa farms in Arizona and other places, too, where these are also just blooming in deserts. Most alfalfa is used to feed animals, which we then eat.

I’m biased here because I’m against animal agriculture for animal-welfare reasons, but this is another place where, if we’re going to focus on one water bad guy in the country, I’d like to start with the bizarre amount of irrigated alfalfa we have and all the corn that is eventually converted into ethanol that we then use for gas, where maybe we could use something else instead.

Strong recommendation: Hank Green recently did a big video on data-center AI water stuff that I thought was masterful. I definitely disagreed with some of the ways that he hedged and some of the specific things he said, but for the most part, it was some of the best single media I’ve ever seen about it. He talks a lot about corn there, how goofy it is that we use corn, and how goofy it is that we use so much water on our lawns every year. He puts things into a lot of useful context, too. So, if people are looking for more of a deep dive on the water stuff, that’s probably the best single resource to start with. So, yeah, we should definitely start there.

Nathan Labenz

Yeah. Cool. Just to fill in a little bit of the detail on food, since you’d mentioned animal agriculture, I had ChatGPT in the background answer the question: How much energy and emissions go into the production of a hamburger? The answer comes back at a few kilowatt-hours of primary energy and a few kilograms of CO₂ equivalent.

If we just do an energy-to-energy quick conversion, let’s say it’s 2.0 kilowatt-hours. That would be running your microwave for 2 hours—3,600 seconds in an hour, so you’re talking 7,200 seconds in 2 hours—and we already previously established that one chatbot prompt is worth about a second of a microwave. So we get, again, to a comical ratio of something like 5,000 to 10,000 chatbot queries for one hamburger.

And, by the way, the exact same number came back with the question of how much energy goes into a hot shower. So, again, 5,000 to 10,000 ChatGPT queries for one hot shower. I’ve totally forgotten if I’ve already shared the story, but a friend was in a pizza place a few weeks ago and overheard 2 teenagers talking. One of them said, “Oh, this mutual friend of ours uses ChatGPT, so it’s ridiculous that she calls herself an environmentalist.” And then she immediately went and ordered a meat lover’s pizza specifically. [Laughter]

That kind of sums up a lot of the debate here for me: I worry that, in hyperfocusing so much on this, people are really losing sight of where the big environmental bad guys are. They’re not where you’d expect. A lot of them in your personal life are probably going to involve very different things, like maybe the food that you eat. But even there, I’m mostly not eating animal products for animal-welfare reasons. I have to admit that if someone’s eating a chicken sandwich around me, I actually think that it uses more resources than some of the things that I eat.

But even there, I don’t actually want to litigate the small amounts of environmental harm. Usually, if I wanted to talk to them about it, I’d say, “Oh, that was a little guy who had a really bad time in his life,” and try to talk more about the animal-welfare thing. I think saying, “Oh, and it uses this many tofu slabs or something,” would actually dilute the point that I’m making.

One other theme of a lot of my writing is that I’m quite concerned about most other aspects of AI. I love it and fear it: current chatbots are great, but I’m very worried about the consequences of even very near-term AI. I’m pretty worried that a lot of my fellow AI critics and worriers are potentially diluting many of the points they make by just adding, “Oh, this thing could make surveillance so much more effective, and every time it does that, it uses a drop of water. Isn’t that terrible?” The second point just takes away from the first, I think, in a way that a lot of people are missing.

And so a big background motivation here isn’t to say, “Everything’s fine. Data-center buildout is fine. You don’t have anything to worry about.” It’s more to say that I’d really like us to focus on these actual things and not just throw everything we have at AI. I think people who are listening to us notice very fast when you’re just reaching for any tool you can to attack something, as opposed to actually building up legitimate criticisms that you’re really worried about.

Erik Torenberg

Another way to put that, I guess, is in terms of contribution to overall p(doom): the climate-change contribution of AI is very minimal.

Andy Masley

Yeah. Yeah. And potentially even positive. Again, the International Energy Agency projects that by 2035, all AI applications together might be preventing something like 4 times as many emissions as all data centers cause. That’s wildly uncertain for a lot of different reasons. Another weird thing about that is that it will mostly rely on deep-learning models that aren’t the big chatbots being powered in these huge data centers. So there’s a lot of hand-wavy stuff to say about that.

I think the existence of AI in the short term reduces my odds of severe climate change by a little bit, but also increases the worry in a lot of other ways. Other environmental stuff, like air pollution, I do take seriously.

Nathan Labenz

Yeah, I’m really glad you mentioned that point about non-LLMs, or non-general-purpose models, driving a lot of value. I think that’s something the world would do well to keep in mind in general, but specifically on this resources point, it’s always really striking to me how small some of the models that are driving breakthrough contributions to various fields of science actually are.

I did an episode not long ago with Professor Jim Collins, who has discovered new antibiotics with models that are—I forget exactly how small they are, but I think we’re talking millions of parameters, tens of thousands of data points—super-small stuff that they can run on a couple of computers in a couple of days’ time, and you get new antibiotics out of it. It’s like, okay, this is a big deal.

Similar things probably are going to start to happen with materials science, with all sorts of optimizations. One thing we had touched on a little bit in preparing for this is optimizing energy use in buildings. There’s so much waste all over the place, right, that you could just get smarter about some of these core operations and get a ton of value from that. But when you get into materials science, then you really can move the fundamentals of energy in a significant way. Literally, one materials-science breakthrough could very plausibly offset all the inference that the world is going to use.

Andy Masley

So yeah, exactly. Especially materials science for batteries and solar panels. There have already been some really exciting photovoltaic improvements enabled in part by AI, and that alone could start to dwarf a lot of what happens in data centers specifically. It’s another obvious example where computing is wildly efficient per task. It’s basically comically optimized.

I ran some goofy back-of-the-envelope calculation where I’m pretty sure that running a game of Minecraft in 1950 would have required half the energy of the United States, and now it can run on your computer. I think people really underestimate just how much energy has been optimized in computers over the last 70 years or so. Because it’s so wildly optimized, what you’re getting out of it is going to have so much more effect on the world than the actual energy that went into it.

The big hold-up to this case is that a lot of people believe that chatbots are still completely useless and are going to be the main things in data centers. If you’re still in that camp and you don’t believe that chatbots will ever have any value, then my case becomes more tenuous. But I do just have to put my chips in and be like, man, chatbots are a miracle to me. I’m sure they’re adding quite a bit of value and changing a lot of my behavior outside of the computer itself.

I think that for anyone who is very worried about the environment and AI, the whole game is going to be in the way that AI is used, not necessarily in the data centers themselves. Worrying about specific consequences of data centers, but framing the debate around the energy used in data centers, does feel disturbingly similar to seeing Amazon take off and thinking, “I should mainly worry about how much energy the website is using.”

Erik Torenberg

Yeah. Maybe one final point of comparison, and then I’ll invite your closing thoughts. It does touch on this question of how AI is going to be used. Obviously, there are some confidence bars around these estimates, but again, 80 gigawatts seems to be the sort of grand vision, with the Stargate project being 5 gigawatts of that and the full $7 trillion buildout coming in at 80 gigawatts in the analysis and estimates that I’ve run.

Compare that to total global energy use today, which is estimated at 6,000 gigawatts. That backs out from what we talked about before, with 1 gigawatt being about 1 million people. Then 1,000 gigawatts could be about 1 billion people. That’s roughly at U.S. energy levels. So 6,000 gigawatts would be about 6 billion people. There’s the electricity-to-non-electricity conversion factor there, but roughly speaking, that seems like an order-of-magnitude reasonable estimate.

We’re talking basically about a little bit more than a 1% increase. Another thing to keep in mind is that this is almost certainly going to be dominated by the increase in energy use in the developing world, through the general background process of people getting wealthier and being able to afford more energy. That’s something that I certainly do not begrudge the global poor for—their marginal energy use over the coming decades. Maybe some do, but I certainly want to go on record saying that I think it’s good when people can do more stuff and have better lives.

I don’t know if there’s any downside scenario that you think would be worth worrying about in terms of how AI is going to be used. We’ve focused on the rosy side. One other short, super-compressed version would be: if AI makes everyone richer and then we can all afford more energy, maybe there’s just a lot more energy used broadly. Is there anything more specific that you think would be the surprising, possibly negative story in the macro sense?

Andy Masley

I mean, I can get into really goofy sci-fi scenarios where it’s like, “If there’s some global war over AGI”—which I don’t actually put a super-strong probability on—but very powerful AI systems in the future might destabilize things. Nuclear war is bad for the environment. At the edge cases, it just gets really ridiculous, so I don’t want to go quite that far. I’ll bracket that for now and just say that AI can probably reduce a lot of emissions.

There do seem to be a lot of potential ways it could radically increase how much energy we use as well, which might just make the green-energy transition harder. Super-optimized shopping apps might make it much more tempting to buy more stuff. Or, economically, making manufacturing more efficient in some ways might make us use a lot more energy in total without transitioning that to green energy.

I also want to make it clear that I’m coming from a very specific perspective: it is definitely important to get the poorest people in the world a lot more access to energy, basically by any means. When I talk about cuts to emissions, I’m mainly talking about people like us and very wealthy communities globally, by comparison. I’m motivated a lot by not being a growth person, for example. There’s a lot to say about that. I’m trying to think about it. I don’t know.

Even just self-driving cars: I could see a world where that makes driving much more tempting because suddenly a 5-year-old kid can take a self-driving car to their friend’s house if the parents okay it or something like that. It opens up so much additional travel that, on net, creates a lot more energy use. It’s a funny world where the value we’re getting per unit of energy will probably be way higher because we’ll have all these nifty gadgets and stuff, but if that doesn’t come with the green-energy transition specifically, we’ll still be emitting more on net. Ultimately, the climate only cares about the total amount of CO2 in the air.

I don’t know. This all feels so speculative and weird. AI could radically reduce car trips, but maybe increase them, and reduce building energy use, but also increase it elsewhere. Materials science for green energy is good, but certain scientific discoveries for the military might make global war more likely. It starts to get really hand-wavy really fast.

Basically, there are a lot of really cool papers on this. I know the—I’m pretty sure it was the London School of Economics that recently published one that was really good. There’s a lot of cool speculative stuff about this, but it’s so uncertain that if environmentalists are really going to focus on this, I’d really want them to be soldiers in the battle to make sure that the way AI is used is good for the environment. Data centers are definitely worth worrying about, but don’t treat them as the big central part of the story.

Erik Torenberg

I think that’s probably a good place to leave it, unless you want to bottom-line it any other way for us.

Andy Masley

Yeah, basically, that was super comprehensive. Thank you. That was a blast. I just enjoyed the conversation a lot. Again, I’m just a big fan of the show. We’re very grateful to make it on.