- Do you hate AI?

Or maybe you love it, but pretty much everyone has some kind of an opinion about it, right?

Though maybe you think we're on the cusp of a new age of discovery, where solutions to all the world's problems are just one perfect prompt away.

Or maybe you think that we're on the edge of an AI-powered apocalypse that could wreck the world.

But honestly, most people are somewhere in the middle, right?

Feeling maybe uncertain, uneasy, a whole lot of things.

Like, have you been online lately?

(ominous electronic music) AI technology has become one of the most contentious issues of our time.

How it's being built, how it's being trained, what chaos it may unleash if it's left to itself, what it's being used for, whether or not we asked for it.

Now, you'd think a question this big, well, the answers would be easy to find, but they're not.

I've spent the last couple of months obsessed with finding them for you.

But one of the hardest things to get straight answer about is what impact building all of this AI technology might have on our planet, energy, water, land, even how it's impacting people who live near these giant data centers that are popping up everywhere.

Now, it's not good enough to just trust the numbers that we hear on social media, or take a tech CEO's word for it.

I'm certainly not just gonna ask ChatGPT or something.

So I sat down with these scientists and researchers who study this stuff every day so that we could get the real answers.

- The scientific community is nowhere near being able to measure the net effect yet.

- We cannot create a new water, and it's not like electricity.

- Data centers to me are like the exhaust pipes of the AI industry.

- Of course we have a choice, and we need to choose .. - This video will hopefully help you to cut through the confusion, so that we at least can find some sort of shared truth, because no one can make a fair and informed decision about AI or really anything else without truth.

But whether you're an AI booster, or a doomer, or somewhere in-between, we all need to ask ourselves if this technology will be powering the future, what kind of a future do we want to live in?

(tense suspenseful music) Hey smart people, Joe here.

Every time someone uses AI, it comes with a price.

A lot of stuff has to happen from my question to leave my computer screen, turn into a packet of data, travel through my wifi, various networks, to end up here at a data center.

The scale of modern AI data centers is honestly hard to imagine.

I mean, these places are huge.

They can be loud, they get hot, they suck up resources, like energy and water, all to keep AI models churning.

But there have been so many claims about how big or small these impacts really are, and this industry is changing so fast that numbers or estimates from just a couple years ago are often hopelessly out-of-date.

But the loudest claims you usually see when you dive into the AI discourse are questions about how much energy and electricity AI data centers used.

There are warnings that AI will use more electricity than entire countries, and others say, "Just relax.

We can just generate more power."

(electronic dinging) (upbeat music ending) In moments of confusion like this, I think it helps to consult the history books, because it turns out, this fear about computers sucking up all of our electricity, it isn't new.

Back in the '90s, when being online was still fun, no one had heard of algorithms, and there were more than like five websites, people were already making some pretty catastrophic predictions about computers draining the US of all of its electricity.

- Back in 1999, there were a couple of guys running around talking about how the internet was gonna use half of all our electricity.

One of them would pick up a wireless palm pilot, and would say, "The networking electricity behind this wireless palm pilot is as much as a refrigerator."

And that turned out to be an overestimate by a factor of 2,000.

- The reality was that in the late '90s, the entire information technology industry accounted for more like 3% of US electricity demand.

That's significant, but it's not enough to trigger a grid meltdown.

And of course, there were already data centers back then, hosting and directing traffic.

That dates back to the first days of the internet.

Now, early data centers only consumed around one-10th of 1% of the US's total electricity.

And even as the internet grew from basically emails and websites to whatever this is we have today, it didn't crush the energy grid, right?

- Between like 2010 and maybe 2017, internet usage was growing, applications were growing, but data center power use didn't grow nearly as quickly.

- That's because even though internet use went way up, the stuff running the internet got more efficient, as it got more powerful, because engineers and scientists figured out how to do more computer stuff with less.

- Back in the early 2000s, I and some colleagues created an article that looked at the energy efficiency of computing, going back all the way to ENIAC in the mid-1940s.

And what we discovered was that energy efficiency of computing defined as computations per kilowatt hour use basically doubled every 1.6 years, starting in the mid-1940s, going through vacuum tubes, and then into the microprocessor era.

So some people call that Koomey's law, I don't do that.

- Koomey's law, I am gonna call it that, did a good job of predicting energy use and computing efficiency, until AI changed the equation.

Today feels different.

I mean, AI has brought with it a completely unprecedented kind of electricity vampire in the form of hyperscale AI data centers.

Well, hyperscale isn't just tech speak for really big building, though the buildings are physically huge.

These data centers are actually giant industrial facilities, whose sole purpose is to provide massive computing power 24 hours a day.

And to do that, they're filled with tens of thousands to hundreds of thousands of powerful GPU chips to process data.

- So a rack that's full of AI servers will consume five, 10, 20, maybe 30 times the power of a rack of conventional servers that we require for standard internet operations.

- Now, a single rack of GPUs is about the size of your home refrigerator, but that single rack can use enough electricity to power 65 houses.

These data centers are filled with thousands of these racks.

In order for a data center to get the label "hyperscale" here in the US, they have to have an electric power draw of more than 100 megawatts.

For comparison, 100 megawatts is enough power for about 80,000 American homes, but data centers aren't the only ones using energy at this scale, right?

I mean, aluminum smelters, steel mills, and chemical plants, these are all enormous power users too.

And we're electrifying so many things today, from EVs to heat pumps and more.

All of these raise the demand for electricity.

The difference is these hyperscale data centers seem to be just popping up all over the place.

As of this video, there are currently 68 hyperscale data centers built or under construction in the US, with 267 more approved or planned.

As a result, between 2014 and 2024, data centers tripled how much electricity they use.

Is that a lot?

We need to put that number in context.

- The best estimate that exists is that in 2024, the electricity used for compute data centers was about 4.7% of US electricity use.

So it's an important user of electricity, but is not the dominant user of electricity.

- To put it a different way, that 2024 data centers number is about one-seventh of the electricity the US uses to power our homes each year.

That's significant, but maybe it's not as much as you might think.

The thing is when it comes to AI, 2024 numbers are already out-of-date.

This is just growing too fast.

This is an industry growing so fast, even the experts struggle to get accurate numbers for how much electricity data centers are using right now.

- One of the things that really struck me when I started doing this research was just how hard it was to get information.

Data center projects are so heavily shrouded in secrecy, and non-disclosure agreements, and a lack of transparency that actually getting valid and reliable information on the impacts of data centers is incredibly hard, whether that's how much water do they use, how much electricity are they using, how many jobs have they created.

- What we do know is that the AI industry is set to invest over $1 trillion this year in building out data centers, and all of the other infrastructure.

That's more than the US will spend on housing this year, so that's a choice.

One guess says that by 2028, data centers will expand to take up somewhere between seven to 12% of the US's total electricity use.

But honestly, that's just a forecast.

It's not a prophecy, and not everyone agrees with it.

- Nobody has any idea how much electricity data centers will use even a few years from now.

There are many uncertainties that could drive the data center electricity use up, or could drive it down.

- Now, for example, if processing chips keep getting more efficient, they might use less energy to do the same job, right?

But on the other hand, if we're doing a lot more computing, because it's faster and cheaper, total electricity use can still go up.

It's the idea that making something more efficient actually leads to using more resources and not less.

The fancy name for this counterintuitive relationship is the "Jevons Paradox."

The paradox is named after this guy.

Spoiler, this won't be the last time you hear from him.

Jevons noticed something similar with coal in Victorian Britain.

Now, more efficient steam engines use less coal for each unit of work.

That's great, but that also made steam power cheaper and more useful, so people built and used more steam engines overall.

Those engines went from running locomotives and steam ships to pumping water and powering factories.

That meant Britain used more coal overall, not less, even as things got more efficient.

We're seeing something similar happen with AI.

The Jevons Paradox is kind of a warning.

Just because AI is using a small fraction of electricity today, it doesn't mean it won't be a big honkin' problem in the future.

And that's especially true, because counting up AI's energy use at the level of like a whole country, well, that probably isn't the right way to look at it.

See, a data center doesn't plug into America.

The country isn't one big circuit where everything is connected to everything.

A data center connects to a particular electrical grid in a particular place, and that is where how much electricity it needs really matters.

- If you read the headlines, you would get the sense that there's explosive growth everywhere, and that's just not true.

Data center electricity use is a very local phenomenon.

- In some places, the local grid has enough energy for AI data centers to plug in without sucking the rest of the grid dry.

But in other places where the grid is older, or there isn't excess energy being generated, data centers can put a strain on local resources.

Here's what I mean.

Let's say a big data center wants to plug into your local grid.

Well, it's not like plugging in a fridge to your house and just calling it a day.

A hyperscale data center needs the right infrastructure to run.

We're talking high voltage connections, huge transformers, more transmission lines, maybe even its own power substation.

Connecting a data center into the local electrical grid safely costs millions of dollars, and takes years to get all that stuff in place.

And that's not even factoring in all the red tape around permits, and environmental impact studies, and who pays for all these upgrades.

Depending on where you live and how your local utilities negotiate their contracts, you might be the one footing most of the bill for new connections.

Not to mention that because of how electricity is priced through our super convoluted system of energy markets and wholesale prices, which we aren't gonna get into here, all of this can raise electric bills.

And when electric bills go up, no one is happy.

- Oh, man.

- To avoid all that hassle, and expense, and waiting, some data center companies have come up with a new plan, and this might be the most striking and potentially alarming change in this whole industry.

- We are seeing "Bring your own power" as a way to mitigate the potential impact of data centers on rate payers' electricity bills.

- Now, instead of asking, "How do we plug our data center into this local grid?

", they're thinking, "Why don't we just build our own power plants?"

- And there are some utilities around the country that simply don't have the generating capacity to meet that large new load, but the data center builders don't wanna wait.

They're all racing each other to have better AI models, better business models, and what they're doing is rather than waiting months or years to connect to their local utility, they're bypassing that by buying generators to run on their own property to make their own electricity.

- It's called "Behind-the-meter generation."

Now, in theory, this might sound like a great solution.

If a data center takes care of making their own electricity, well, there's no need for grid power, no higher electricity bills for normal people, everyone wins, right?

The problem is, well, they ain't exactly installing solar panels, and batteries, or nuclear energy to get most of this done.

The truth is, behind-the-meter energy relies almost exclusively on giant natural gas turbines.

These are basically giant jet engines that make electricity strapped onto these data centers.

- The cost of the power is immaterial to the investment.

The value of the compute, the perceived value of the compute is so high that they simply do not care how much the electricity costs.

And we know this, because no utility would put in a gas turbine and run it 24/7.

It's incredibly inefficient.

- Now, behind-the-meter power is another one of those issues around data centers, where it is hard to get good, clear numbers, and there's no one in the government or in finance who is keeping track of all of this stuff, it's just sort of out there, happening.

But here's what we know.

About 25% of announced data center builds are planning to build their own power plants, but as of now, only about two to 3% of that has actually been built.

But don't start celebrating just yet.

Satellite analysis from the International Energy Agency and others shows that about a quarter of those planned private power plants have broken ground.

So this is all likely to change pretty quick.

To put that in perspective, this one data center site under construction in West Texas called "GW Ranch," or "Gigawatt Ranch," I assume, that plans to install 35 giant gas turbine generators, which at over seven gigawatts of peak power will make it the largest power plant in the United States, not just the largest data center power plant, the largest, period, bigger than the fricking Grand Coulee Dam, y'all.

And it'll likely be the largest fossil fuel power plant on earth, burning more than 1% of all the natural gas that the US produces, for one single data center.

Now, there are some data centers, like this one in Nevada, that are using renewable energy.

It is totally possible to run a data center on renewable energy.

With enough solar panels and enough batteries, it can absolutely be done.

We have the technology.

It just by and large isn't being done.

- The argument has been, "Well, we don't have time to do that," but why do you then have time to wait seven years to get a gas turbine?

And ultimately, you know, I can build a solar battery plant in 18 months.

- Now since the beginning of 2025, this industry seems to be running toward cheaper, faster fossil fuel power plants, like a kid let loose in a candy store with no adult supervision.

- These companies have urgency, 'cause they think they're gonna be first past the post, get more customers quicker, if your choice is driven solely by urgency, you will grab the thing that's familiar, and you'll grab the thing that is available.

- I think it would be disingenuous to say that the companies, you know, don't wanna wait for renewables.

I think they do.

It's just that a lot of the companies wanna build as quickly as possible, because if they can't connect to the local power grid, because the generation or transmission capacity isn't there, then really they're left with few options other than pausing the project, unless they wanna do onsite generation.

- For society as a whole, whether we develop AI in a year, or two, or three, there's no real urgency there.

- Have you ever actually seen a data center?

Because I have.

I actually went to a hyperscale data center to see with my own eyeballs what these gas turbines really look like, and what impact they're having around data centers, in a recent episode of another show I make called "O.. I'm not going to rehash all of that in this video, but it was pretty shocking, and you should check that out when you're done here, it's an awesome video.

One thing that's clear is because these turbines run on fossil fuels, they also pump out pollution and carbon emissions.

- Emissions can occur from the data center itself, and that would only occur if you're burning fuels on site.

- And hyperscale data centers aren't just running one of these bad boys, they're running dozens of giant turbines.

This data center in Abilene, Texas has plans to use 50 of them.

Running massive gas-powered jet engines, perhaps unsurprisingly, has massive consequences for carbon emissions.

The carbon footprint of AI data centers is measured in three ways, or scopes.

Now, the first is scope one.

That's the emissions that we just talked about.

It covers the carbon emitted by the data center itself through behind-the-meter energy generation.

Even data centers that normally plug into the grid can have scope one emissions, because many also keep backup generators around in case the grid decides to take a surprise nap, but not every data center runs entirely on behind-the-meter energy.

So that's where scope two measurements come in.

Those look at how much carbon pollution comes from the electricity that they get from the grid.

Depending on where you live, electric grids can be powered by coal, natural gas, renewable energy, nuclear energy, or a mix of any and all of the above.

When power plants use renewable energy, they emit less carbon.

If a data center plugs into one of those cleaner grids, then its carbon footprint will be much lower, even if it's still using a bunch of electricity.

But the reverse is true too.

Data centers that get electricity from coal-powered power plants are on the hook for way higher scope two carbon emissions, and emissions from running a data center are just part of its full carbon footprint.

- But there's also what we call "scope three," and that is the embedded carbon in all of the steel and concrete that was required to build the data center in the first place.

- A bunch of carbon emissions come from clearing the land, building massive buildings, manufacturing all those chips, and maintaining a data center day in and day out.

But this is all stuff that falls into the category of scope three measurements, all the CO2 created throughout a data center's entire supply chain.

So to get the real emissions total of a data center, we just add up scope one, scope two, and scope three measurements.

It seems pretty straightforward, right?

But as you might have guessed by now, actually, getting those numbers and making sure they're accurate is not easy, because AI is currently a technological arms race.

Companies can be pretty cagey about releasing their scope one and scope two data for their data centers.

Even companies that do share information don't always break it down by location, so it's hard to calculate the footprint of one data center in one specific spot.

And those scope three numbers, they're even harder to add up.

At the end of the day, these companies are self-reporting their data with very little oversight.

So their numbers are kind of just the equivalent of, "Trust me, bro."

So our best estimate is that US data centers account for about half a percent of America's total greenhouse gas emissions today.

That's a fraction of the amount of greenhouse gases emitted by transportation or even agriculture.

But remember the Jevons Paradox?

That still applies.

Even if data centers use energy more efficiently, they'll still need more electricity as total AI use goes up.

And that means they'll probably create more emissions too.

Some think that AI carbon emissions will reach 44 megatons by 2030.

That is double how much this industry emitted in 2025, which is roughly like adding 10 million gas-burning cars to US roadways.

And that is still just 1% of the US's total CO2 emissions, but we've already established that predicting what AI will do in the future is a guessing game.

We frankly just don't know if those emissions could be lower or a lot higher, but it's important to remember that no matter how much extra emissions these things create, that's still more emissions than today.

And I don't know if you've heard, but we need to stop emitting more greenhouse gases into the atmosphere.

Climate change, people, it's real, and unlike electricity, emissions are not a local problem.

- Carbon emissions are a global issue.

The atmosphere doesn't care if I emit a gram of CO2, or you do wherever we are, it all goes to the atmosphere, has the same impacts.

- Of course, carbon isn't the only thing that data centers emit.

They also create a lot of heat.

All of that power running through copper wires and GPUs physically creates heat.

Like if you've ever worked with your laptop on your lap, you've felt that.

It's just an unavoidable consequence of physics.

And if these companies don't want to melt all of their really expensive computer chips, they've gonna deal with that heat somehow.

- These AI servers use a lot of energy, and that energy turns into heat that has to be further removed away from the servers.

- Which leads us to the now infamous claim that asking AI a question uses one bottle of water.

Now we can actually trace that number to a 2023 study co-authored by Dr.

Shaolei Ren, who you met earlier.

He found that asking OpenAI's GPT-3 model 30 medium-sized questions use roughly 500 milliliters of water.

But this, again, is perhaps the fastest-evolving industry in the history of technology, and those numbers, frankly, just aren't true anymore.

- So that was for a very specific model.

We should not generalize that statement to ChatGPT or to AI.

I mean, the all source of AI, you can't really say having a AI query will consume a bottle of water.

And also that number includes both direct water for cooling, and also indirect water for generating electricity.

And the indirect water is the majority of the total water usage.

We are using more computing to generate, to finish the same task than two years ago, but each token is using much, much less resource than two years ago, both are correct, and whether we're using more resources than two years ago to finish the same task or not, that's something I don't know.

- So we could use less water per question if the systems are more efficient, but probably use more water overall, because we are asking a lot more questions, more often.

There's Jevons again.

(Joe grunting) - Even if individual models.. to train and run, at the same time, you can see more deployments and more usage of these models, such that the net effect is actually larger, you know, environmental impacts of AI.

- Now, there's also a common thing people do, where they compare the amount of water AI data centers use to something like almond farms or golf courses.

So is every AI prompt secretly thirstier than me on an exercise bike, and destined to turn the world into one big desert, or are almonds the real enemy?

Or is it golf courses?

Or maybe it's cows?

It's probably cows.

I mean, look at their faces.

Very untrustworthy.

Or maybe this is a silly comparison, because guess what?

It's possible for more than one thing to be bad at once, okay?

Look, it's definitely true that data centers are using a lot of water.

In 2023, US data centers directly used about 17 billion gallons of water each year, mostly for cooling.

When you factor in the water used for generating electric.. for data centers, that number jumps up to 228 billion gallons.

That's about two-thirds of New York City's annual drinking water use.

Now researchers think, that number could skyrocket to more than 840 billion gallons of water annually.

That's more than the amount of drinking water used by New York City and Chicago each year, combined.

And shockingly, adding another New York and Chicago's worth of thirst to our water supply?

Not great.

- Because we're no.. I think for the past one million years, we're still using the same amount of water on earth.

We cannot create a new water, and it's not like electricity.

You can build more power plants, but you cannot build snowfalls, and you cannot move water from one place to another over long distances.

- Now, a lot like the whole figure out how to use less electricity thing, data centers are trying to crack the use less water code too.

Now, many data centers have turned to what's called a "closed loop cooling system" to cool servers down efficiently, and with less water loss.

Instead of cooling by evaporating water away, and then having to replace that water, these systems move around a fixed amount of water, and they remove the heat by radiation.

Now, theoretically, closed loop systems should be so efficient that the same water can flow through the pipes without being lost, for years, but filling up that loop can take a lot of water upfront.

For instance, the hyperscale data center in Abilene, Texas, that has a closed loop system.

It needed a one-time eight million gallon fill-up.

That's actually less water than my neighborhood will use in a year.

It's about three times less than the city of Abilene uses in just one day.

- For a large data center campus, I would expect this cooling water to be in the order of a million gallons.

A city uses lots of water, I mean, even for a small town with like 10,000 people, they probably need two million, four million gallons of water each day.

The water resources in term of total number or gallons is not a constraint for many communities, except for those communities that are highly water-stressed.

The constraint is really the peak water capacity infrastructure.

It turns out that many of the community water systems in the US are small water systems.

- But even though closed loop cooling system is great at grabbing heat and moving it away from a data center's precious chips, the heat still has to go somewhere.

And in the case of data centers, that's usually outside the building.

Now, there are a few ways that these data center heat sinks work.

One way is through a giant outdoor radiator, which are these big metal units with fans blowing air across hot water side of that loop to cool the water down.

This method uses less water, but moving all that heat with fans takes more electricity, and that electricity indirectly requires water.

The other option is to use evaporative cooling, where water evaporates, taking heat with it.

It's just like how sweat cools your skin.

It's a lot more energy efficient, but it uses a lot more water, especially during the summer, when it's hot inside and outside a data center, and evaporation just doesn't work that well.

That can be a big problem since the hottest times of the year are also when regions are most hard up for water.

- If we are able to pull all this water together at this national-level, even for in Arizona, we wouldn't have any water challenges.

There's no water problem here in Arizona if you can pull the water together for at the national level.

- Unfortunately, that's just not how water works.

Water is local.

Sticking with our Arizona example.

Today, data centers around Phoenix use 385 million gallons of water for cooling every year, and that number doesn't even count the water that goes into generating electricity.

But by some estimates, that water use will go up to 3.7 billion gallons of water a year as more data centers come online in that area.

As demand for water around Phoenix goes up, the amount of available water is cratering out.

Now, Phoenix gets nearly half of its water from the Colorado River, which has lost about 25% of its flow, because of a mix of drought and overuse.

More demand plus less water, that's a bad mix.

I'm pretty sure one of the Mad Max movies started this way.

Now, the companies investing in AI and developing these technologies are not stupid.

They understand that building more fossil fuel power plants and sucking up more water from an already dry planet is not necessarily good.

But they also seem to think that one day in some uncertain future, AI will become so powerful that it will solve all of these problems.

And don't take my word for it, listen to these guys.

- When it comes to emissions, AI will pay for itself.

- Absolutely.

- If you wanna turn the corn.. lean into AI.

- AI can enable existing technology and solutions to get to a 10% reduction in global emissions by 2030.

- People talk about how much energy it takes to train an AI model, but it also takes a lot of energy to train a human.

It takes like 20 years of life.

- Now I can see how this would be an intriguing thought.

I mean, imagine if AI could make air travel more efficient, or help farmers predict their crops, or harvest times months in advance, lowering the massive amounts of food waste that contribute to climate change.

That would be awesome, but it's also just as realistic right now as me getting more subscribers than Mark Rober.

- One of the storylines that the AI industry has put forth is that climate problem.

"We should put it in as fast as possible.

People are gonna get a lot more efficient, and therefore, we should just go ahead and do it."

And that's in part to justify putting in this behind-the-meter power.

- So far it's been, you know, "Just trust AI, it's gonna solve problems.

But in the meantime, bear the brunt of the impacts."

The costs are now like, you know, we're bearing them now, and we haven't seen much evidence of the benefits yet.

It doesn't mean that they're not coming, but we need to start measuring them.

- Now it's important to remember, all of this isn't just some idea made of numbers and GPUs.

This is happening in real communities, where real people live.

xAI's Colossus data center is one of the biggest data centers in the US, and it's also located in a historically Black neighborhood.

Colossus used as many as 35 gas turbines to get the facility running, even though it's only permitted to have 15.

Naughty naughty, Mr.

Mars Man.

- If you've ever been on an airport tarmac with jet engines going, you know what it's like 100 or 200 meters from these engines, and that's what people are seeing.

- And the pollution from those is affecting the air quality in the neighborhoods around the data center.

That one analysis from 2024 estimated that pollution from data centers could contribute to 1,300 premature deaths.

That's not good.

Colossus's methane gas turbines have increased Memphis's smog by 30 to 60%, and they're releasing pollutants that are linked to lung and heart disease.

But Colossus is just one example.

With more data centers popping up in towns and cities across the US, the effects of data centers will become more widespread, and unfortunately, the people who live in these places don't always get a say in whether they get a data center at all.

- Data center development typically occurs with incredible levels of speed, and secrecy, and opacity, where community members, regardless of wealth, or race, or region, or where they live, are really left out of the conversation.

- Developers often work through shell companies to get their foot in the door.

Like in West Memphis, Arkansas, there were two shell companies named "Willowbend Capital" and "Forgelight Ventures."

Huh, those sound nice and harmless, I wonder what they do.

Well, they paved the way for a giant new Google data center, and Meta used two shell companies named "Winner" and "Wurldwide," let's not talk about the spelling, to negotiate multimillion dollar tax breaks for data centers in Fort Worth, and El Paso, Texas.

In other places, non-disclosure agreements keep officials from sharing information with citizens about potential data centers.

In Pine Island, Minnesota, city officials knew about a proposed data center for two full years before residents did.

All of that sneakery and sleight of hand can lead to some justifiably intense feelings from people who live in those communities.

- One of the major reactions that communities tend to have is a feeling of abandonment, particularly by the local government when those officials have been kind of working in secret in collaboration with the tech industry.

- But there is a reason towns say yes to these projects too.

Data centers can potentially generate serious tax revenue.

One of the world's largest data center hubs is in Loudon County, Virginia.

Here, city officials say that taxing data centers has led to a billion dollars in revenue.

- The primary benefit that data centers provide to communities is tax revenue.

Whether in the form of property taxes, or sales and use taxes, data centers do bring in tax revenue, and often that is the primary motivation for policymakers is that, "This facility will bring tax revenue to our community.

Here's an industry that wants to come .. and would provide a lot of tax revenue that we could benefit from-" - So if these data centers can bring in tons of money for communities, and these companies all keep saying that this technology is so valuable that it's going to change the course of humanity, we definitely wouldn't give them like tax breaks or something, right?

Sorry to burst your tax bubble, but right now, 38 US states provide some sort of tax incentive to lure data center operations to their neck of the woods.

- These tax breaks really represent a form of a race to the bottom, where all of these states are, you know, trying to attract development by reducing taxes, which just is a net transfer of money to these tech companies.

- They're spending $700 billion, the tech indust.. this year to put in AI data centers.

So clearly they think there's a massive value to this compute.

And so why should those you know, high-value investments be subsidized by the state?

- Good question.

But tax breaks or not, there are big questions about whether that money is worth the price.

- But you know, there are serious strings attached to that tax revenue.

You know, there are all of these costs that come along with that data center, that money doesn't come for free.

And if you're a town of 3,000 people, it's pretty hard to think about going up against, you know, billion-dollar corporations.

How much tax revenue is worth handing over so much control and agency about our community to this outside entity that does not actually care about us, they only care about, you know, using our land and our resources.

- But I mean, at least these data centers are gonna bring jobs to these communities, right?

Right?

- Once a data cent.. it doesn't have many employees, because it just doesn't need people, and one indicator of that is that I think most of these data centers are built without parking lots in front of them, because this is not a facility that's going to have many people commuting in on a daily basis to go to work, this is a facility that mostly can just be left operational and requires a few technicians and security people to, you know, keep things organized, make sure that nothing's going wrong, but it is not a workplace.

- In order for a community to know if a data center will work for them, that community needs information, like how much water it will use, how the data center will meet its energy needs, how the data center will help its surrounding community be more prosperous and resilient.

Unfortunately, that information too often just isn't easy to get, but it is out there, it just isn't being shared in the open.

- And let's be clear, they have the data.

They know how much water, power each data center uses, local utilities have that information, as well, but they're usually not allowed to release them.

- Look, let's wrap this up.

It's getting very long.

I think it's safe to say by this point that in the very short history that is AI, the truth is AI in some form or fashion is here to stay.

It isn't going anywhere.

There isn't going to be a world with no data centers.

We're not putting all that toothpaste back in the tube.

- I'll never leave.

- But it is worth asking, does it have to be done like this?

Because we've established that we could power these data centers with renewable energy if we wanted to, but the industry would have to wait longer.

And if data centers create so much heat, why not put that heat to use, heating nearby buildings, or generating power, or harvesting water from the air?

No matter what you feel about AI as a technology, and how it is being used, what is becoming very clear is that the full costs and impacts of building this technology are not being considered.

- These are incredibly valuable pieces of infrastructure, and incredibly profitable, at least that's what the companies think, so let's make sure that the people installing these facilities are following the rules and paying the full societal cost of their actions.

It's a kind of a bedrock rule of economics that people need to pay the costs of their actions.

Otherwise, it's inefficient, because they think doing the wrong thing is cheaper than it actually is, and it's also unfair, because it imposed costs on other people.

- So when somebody tells you that data centers are either an environmental catastrophe, or basically harmless, because almonds, or golf courses, or beef, or whatever use more water, they're leaving out a lot of the picture, because the truth is messier.

It always is.

But I don't think we can have honest discussions about the future of a potentially world-transformi.. if we don't have some sort of shared truth.

This video isn't to tell you how to feel about AI, but it is a step in that direction to try and at least establish that shared truth about what it means to build AI the way that we're building it today.

And unfortunately, we still don't know as much as we need to know to decide if the good can balance the bad.

- The scientific community is nowhere near being able to measure the net effects yet, it's something we all need to work on, but ideally, that's the way you wanna approach any policy, or any decision is if there are costs, you wanna weigh them against the benefits, but we don't yet have the real data to quantify that balance.

- The future of artificial intelligence will be determined by good old human intelligence, but that can't happen if we don't know what is real, and what's a hallucination, because whether you think AI is the most important technology of our lives, or you think we're spending way too much money and energy making chatbots write emails, the cost of a new AI-powered future is being decided right now.

- We need to figure out what the real benefits are of this technology, and what the costs are, and we need to answer the question as a society, "Is it worth it for this application or that application?

Is it worth it?"

Yes, we have a choice, and we need to.. - Stay curious.