Data Center Water is a Distraction

[00:00:00] Speaker 1: - Data centers are part of our everyday lives. [00:00:02] Speaker 2: - Our data centers are the home base for Metis Technologies and Services. [00:00:06] Speaker 1: - Data centers use cooling systems. It pools in outside air and some water to make the air cooler. [00:00:12] Speaker 2: - Wherever we do use water, we are committed to doing so responsibly. [00:00:16] Speaker 1: - We care about saving water and building data centers that are good for communities and the planet. - What started as an announcement for one data center has evolved into plans for multiple. - A massive data center is suing over water rights to the Colorado River. - The river is also suffering from a historic drought. [00:00:33] Speaker 3: - Industry estimates show a single hyperscale data center could use up to a billion gallons of water a year. - Families in the area are starting to see their water pressure decrease. It is decimating their water quality. - Those are the things we have to fill up [00:00:46] Speaker 2: to flush the toilets. - Making sure that we have that balance, those are questions that have to be answered. - Transparency regarding water use and environmental impacts is severely limited. [00:00:56] Speaker 4: The latest reason to oppose AI isn't artistic or economic. It's physical. The robots are coming for our water. Billions of gallons of it. And that number is only going to increase in the near future. How can you think anything else when you read the headlines and see viral videos like this? [00:01:12] Speaker 3: - I have another one as well. - So this wasn't just one well, this wasn't just one family's situation. This is what the drinking water now looks like next to that data center. And I think both of us can agree that neither one of these things are drinkable. [00:01:32] Speaker 4: But are AI data centers really uniquely poisonous water guzzlers on track to drink up both jobs and H2O? Or is their water usage more of a distraction than it is damming? That's complicated. The best indicator I get that a question is really good is when I'm asked about it in the real world, outside of the algorithm and the echo chamber. I was asked to make one of my most successful videos ever by my hairdresser after she learned I was going to Japan. I have been asked about AI data center water use no less than four times in meatspace. And after looking into it, I realized why. There's a serious information gap here, similar to a topic like nuclear waste. AI water use is not the killer argument it seems to be, no matter how much even I am willing to believe it. But before we tackle the very complicated topic that is AI and water use, today's video is sponsored by the only service I currently use to help me sift through many different media outlets and see how they're covering the exact same story. Like this one. Ground News. Ground News is an app and website that collects stories from all over the world and then categorizes them in one place according to political bias, publishing practices, who owns the media outlet, and more. Giving you the all too critical transparency that we desperately need today. Honestly, Ground News across platforms has the best cross section of any one topic that I've seen. My favorite thing that Ground News does is serve me up stories from across the political spectrum in one place so that I can check my own bias and maybe even my own echo chamber. For example, here's another headline about AI and water use that's kind of spooky. Before I dig too far into it, I scroll all the way down to see how different media outlets are covering the same story. As you can see, there's a hundred different versions of this headline and many of them are falling into the same statistical pitfalls that we're about to go through in this very video. In a day and age where it seems like a lot of people don't read past a headline, framing and checking that framing is very important. I also wake up in the morning and check my Blind Spot feed where Ground News gives me stories that I wouldn't otherwise see in a very curated information landscape. Obviously, practicing good hygiene on the internet right now is very important, and Ground News has been endorsed by the Nobel Peace Center for their effect on media literacy. If you want to try Ground News like me, you can currently get 40% off the all-access Vantage plan that I use by going to this link on screen in the description or by scanning this QR code. It is extremely refreshing to not be served stories that depend on how much they are emojied to. It forces me to actually think about what I think about a topic. And that's what we're going to do right now with AI and water use. Why do data centers use water? Simply put, the CPUs and GPUs they run on are hotter than Jensen Huang in a leather jacket. No, seriously, a single data center powering GPU puts out 10 times more heat than a human body. And there can be 100,000 of these inside a center, meaning that a large facility needs to get rid of 100 megawatts of heat. 100 million joules of energy per second to make sure that their endless racks of chips operate efficiently and without damage. This is why billionaires say dumb things about putting data centers in space. The heating and cooling problem is serious enough for crazy ideas. Water is an ideal cooling medium because it can absorb a lot of heat before changing phase, it's non-toxic, it's relatively abundant, and it's cheap. For now. So, one of the most efficient ways to cool anything since the beginning of human industry is to simply run some free water over it or through it. But water cooling doesn't necessarily mean water use. The difference matters. In studies of data center water use, the literature makes the distinction between water withdrawal and water consumption. Water withdrawal is the removal of water from some system, like a power plant taking water in from an ocean or river. Water consumption is the fraction of that water that is used up or not returned to the system, usually because of evaporation. The difference between these two values is water returned, like the slightly heated water coming out of a power plant and back to the sea. So, for example, a cooling system that withdraws a lot of water but returns 80% of it is very different than a cooling system that withdraws water but consumes 80% of it. This is where the apparent water use problem starts. The majority of data centers in the United States currently use water for evaporative cooling. This is when GPU heat is dumped directly into water and up to 80% of it is lost to the environment as water vapor. This kind of cooling consumes a lot of water but is very efficient on electricity. Again, water can take in a lot of heat before it evaporates away, which removes that heat from the system, reducing the need for fans or other cooling systems that use more electricity. This is the same reason why you both sweat on a hot day and need to drink a lot of water to supply that sweat. You're literally a walking data center with evaporative cooling. Of course, other industries like agriculture and power production also use a lot of water. Why you see headlines like this is because many data centers are using potable water, fresh and/or treated water that humans use for drinking and washing. Why? Well, CPUs and GPUs are literally some of the most complicated, delicate things humans have ever made. They can take thousands of individual steps to make and are now at the literal edge of physics, pushing up against the barriers of quantum mechanics. The dirtier the water that cools them, the more that they can corrode or accumulate contaminating minerals or even grow bacteria. In fact, when the chips that go into these technologies are produced, they have to use water 1000 times purer than drinking water to help do so. This adds to the overall amount of water data centers consume. It should be said that data center builders do seem at least aware of this drinking water problem and of the public's problem with it. A growing percentage of data centers are either using or promising to use other forms of cooling, like chilled air, recycled wastewater, and even direct immersion of hardware. These methods can increase the electricity needed by the center, but they drastically reduce the amount of water it consumes. We will get to the electricity water tradeoff a bit later. The punchline here is that there are ways for an AI data center to consume less potable water, even no water, and companies are actively trying them. But the worst case scenario these headlines highlight does exist: a so-called hyperscale data center using municipal drinking water to evaporatively cool racks of 100,000-plus GPUs in a water-strained area like a desert. Is the water use in this worst case actually a problem compared to water use in the general case? These numbers are eye-watering, but in context do they compare to industries and applications of similar scope and scale? If they don't, then this specific argument against AI data centers doesn't hold water, and maybe we should focus on bigger issues like power, heat, and literally losing our minds. If I had to choose the single biggest failing in any kind of scientific or health-related reporting, it would be misrepresenting numbers. As Daryl Huff points out in the classic "How to Lie with Statistics," a book everyone should read, this is most often done by hiding a baseline absolute number or percentage to focus on a relative one that is more dramatic. Here's an example. You've probably seen a headline that reads something like "Forever chemicals increases the rate of schizophrenia by 50 percent." Don't worry, I'm making this up. This sounds terrifying. I wouldn't buy any product that had that chemical in it, would you? But this is a relative increase in the incidence of schizophrenia. What is the absolute incidence? Schizophrenia currently affects about 1.2 percent of U.S. adults. A 50 percent increase brings this to 1.8 percent. Suddenly, the headline doesn't sound nearly as scary, does it? At least not as much as using the relative number did. Obscuring the absolute by focusing on the relative is lying with statistics. You'll notice that all of these headlines are using relative comparisons and numbers. Reporting on the number of gallons consumed only technically matters if we know how it compares to how many gallons are available, right? Water consumption is only as dire as these stories say if the data centers are as uniquely thirsty as they imply. If not, then the real conversation we could be having is around water use in general and what it should be allocated for. It's projected that by 2028, data centers in the United States, and not just for AI, but for everything that data centers do, internet and streaming movies, will be consuming between 144 and 276 billion liters of water a year. This is a huge number, but it's still relative. What is the baseline? The United States uses a lot of water. A lot. In 2015, the United States consumed an estimated 1.2 trillion liters of water per day. That's 443 trillion liters of water per year. By simple math, water for all data centers in the United States represents a rounding error of overall water consumption. But it's not fair to compare a few thousand data centers to everything that a country does, from cooling power plants to taking showers. So let's compare AI water use to something less consequential. This is how much water Americans use to water their golf courses. This is how much water they use to water their lawns. Adding back in the consequential stuff, this is the amount of water Americans use to generate power, and this is the water they use to make food. This is the projected total data center water consumption in 2028. If this chart is surprising, it's because most of us never really think about where our water was going and how much of it there is. An accurate conversation about how much water data centers use can't get around this fact. If we're serious about freshwater allocation, golf courses should concern us way more. And making corn grow more efficiently should be way ahead of all of these debates. We should talk more about water use generally, especially as climate change changes the equation. But singling out data centers without context is not going to help make that happen. All things being equal, AI data centers will not uniquely affect available drinking water in the United States. But of course, all things aren't equal. Water is not evenly distributed. And right now, the majority of data centers are in fact being built in areas that have strained water budgets. That's a hard to defend idea. Any significant water use in these areas will have an outsized effect on local communities. And then news stories highlighting this real problem will again skew our understanding of what the baseline is. The complaints from residents in these stories are real, just not really representative. And I worry it crowds out news on other growing problems. Data center construction sites are creating local traffic channels. Their 24/7 information processing is increasing noise pollution, which some communities are claiming affects them like infrasonic weapons. Data centers are paving over many square kilometers of land, like compute parking lots, without providing an acceptable number of long-term jobs on that land. Jamming tens of thousands of worrying GPUs into expansive, flat buildings is also creating invisible islands of heat pollution. The extent of which companies are being less than honest about, as my friend Dr. Joe Hanson recently explored. But even here, with local environments at stake, we have to be careful about lying with statistics. This now viral tweet claimed that one of the largest planned data center projects will put out the equivalent heat of 23 Hiroshima bombs. Per day. Into the surrounding Utah's salt and soil. My knee-jerk reaction to this was that it couldn't be true. But that's because, like a country's water use, I didn't know the baseline value of heating in that area. This helpful Redditor checked the math, and it is indeed exactly true. However, he notes that the comparison is misleading. Because an atomic bomb, of course, releases all its energy in milliseconds. Which is very different than dumping heat out of cooling towers over 24 hours. He then shows that it's even more misleading by noting that the total energy hitting the same square footage each day just from the sun is equal to 50 Hiroshima bombs. More than double the data center value. None of this is to say that the proposed 40,000-acre build will be thermally tolerable. One physics professor calculated that dumping this much heat into a high desert environment, and a valley no less, could raise the nighttime temperature there by 28 degrees Fahrenheit. That's insane. That's like terraforming. And apparently, the energy technology that's supposed to supply the 9 gigawatts of power for the facility is currently being hidden from the community by NDA. Will the community actually trust the developer of this project, Shark Tank investor Kevin O'Leary, to put these real environmental concerns above profit? I'm not so sure. That's an engineering side note here. Because data centers are built relatively flat and expansive instead of tall and skinny like a skyscraper, they trap heated air closer to the ground, which can then simply roll over into nearby communities instead of moving up and dissipating in the atmosphere. This exacerbates the heat island effect. If there's one problem that does have the numbers behind it though, it's power. If the absolute amount of projected data center water use is small by comparison, the amount of electricity it's going to use is not. It's estimated that by 2028, data centers will be using 17% of all U.S. electricity. That's a shocking number. That means more pressure on aging power grids, higher electricity bills for some consumers, possibly even a complete shift in how areas get basic power, and even more water use. As the electricity will likely come from thermoelectric power plants that themselves withdraw a significant fraction of the country's available water. Yes, this will happen even if companies succeed in building out all of the nuclear power that's currently promised. As your local nuclear doctor, I feel obligated to tell you that nuclear power plants do withdraw a lot of water, something that we haven't really talked about. In fact, they consume a similar amount of water as coal plants. The important difference between the two technologies, of course, is nuclear's advantages in energy density, pollution, capacity factor, and carbon footprint. Many data center builders are looking at nuclear because of these advantages. The only real problem is that the data centers are being built much faster than nuclear reactors are. I'd much rather have all the new energy needs met by nuclear and renewables, but I wanted to make sure that I wasn't hiding the water use baseline. Construction, heat, power, these are all very real issues that communities and companies will have to sort out if rapid data center expansion will be tolerated by the people and the politicians that represent them. However, significant drinking water use doesn't appear to be one of those issues. My friend Hank Green recently put it this way in a video that I did not see until after I wrote this very similar script, to be clear: [00:17:10] Speaker 2: I get why people jump on AI water use. Wasting water feels immoral. In specific places that are up against their hydrological budgets, it can absolutely make sense to not at due use and to save every bit we can. Even under the maximalist goals of AI companies, the projected increase in water use is small compared to what cities and industries already use. [00:17:30] Speaker 4: And again, to be fair, despite all the bad press, and perhaps because of it, you do see companies now promising or actively practicing a much more community-centric approach. Microsoft is investing tens of millions into new wastewater treatment capacity to handle its cooling water effluent. Amazon is switching to more data centers that use recycled water rather than drinking water. Meta even says that their data centers can be water positive by switching to closed-loop recycled water cooling and investing in local environmental projects. That's not crazy. If you feel uncomfortable here, it's because we didn't know the baseline numbers. It's similar to me to the conversation around nuclear waste. It's a hot-button issue that delays construction projects. But in reality, there's not that much of it, and you'd have to eat a lot of it for it to be dangerous. Nuclear waste still scares people, and so it distracts us from having better discussions about nuclear power generally. Case in point, we come back to this viral video of Representative Alexandria Ocasio-Cortez. Here she is strongly implying that a data center did this to local water, but it didn't. Not in the way she makes it seem. What she's showing is private well water, which data centers don't use. And well water is routinely made more cloudy by construction projects that move large amounts of soil around. You can even see in the second jar she brings out that some of that sediment has already settled. In other words, the water would have looked like this regardless of whether or not there were going to be NVIDIA chips inside the nearby construction. This kind of demonstration is misleading at best, and it undermines real concerns with congressional theater. Trust me, I have a lot of forceful things to say about AI. I would love to have another sound argument against its reckless adoption. I don't want landscapes flattened by whirring warehouses as far as the eye can see. I don't. But if any of us that have legitimate concerns are going to be taken seriously, we have to accept the numbers when we have them, lest we be dismissed entirely. That's already happening in pro-AI and artificial intelligence forums. They're dismissing all of our concerns because we get the numbers wrong. So if everyone is wrong about AI water use, what should we be focusing on? What are we being distracted from? How about the fact that no one wants any of this in the first place? AI might not use much water, relatively speaking, but it could be using none of that water. What about focusing on the jobs we are not prepared to lose? What about the freedoms and privacy we're about to give up to big data? What about the existential risk that no one seems to be able to give an acceptable annihilation percentage for? What about the alignment problem? What about the ethics of interacting with or even falling in love with conscious machines? What do we do about chatbots that are so persuasive? When they tell our children to kill themselves, they do it. There's a lot at stake here. The last grand psychological experiment we ran on ourselves like this was social media. A tool that ended up eroding democracy, unweaving the fabric of civil society, and creating both the most anxious and most depressed generation of children since we started measuring. What's going to happen if AI, a technology orders of magnitude more powerful, is allowed to change our world in the same way, without regulations and safeguards from the outcomes we know are coming. AI is going to use a lot of water, sure, but I'm worried it's going to waste an even more precious resource. You. Until next time. [00:21:15] Speaker ?: Bye. Bye. Bye.