Latest / Nonsense with cj little and Jeff Parker / The Insane Infrastructure of Artificial Intelligence

Transcript

The Insane Infrastructure of Artificial Intelligence · Nonsense with cj little and Jeff Parker

• 0:00
  We're going to nonsense Season 4 Episode 28, our big story.
• 0:02
  I'm Jeff Parker. I'm CJ Little this.
• 0:05
  Time we're looking at the real cost of AI up next.
• 0:08
  Oh, it's going to be like hundreds of dollars, maybe $1000
• 0:11
  cost a. Few bucks so you can talk to a
• 0:13
  chat bot and do other things. First of all, what I do with my
• 0:15
  chat bot is none of your business.
• 0:16
  Yeah. Second of all, I like that our
• 0:18
  spending on AI is basically measured in the GDP of
• 0:21
  countries. Oh, that's really how you can
• 0:22
  think of it. And it's getting worse.
• 0:24
  You're like it's 2 Switzerlands and a Denmark all.
• 0:26
  This money we're spending so that people can't have jobs,
• 0:28
  yeah. And, and we just invited into
• 0:31
  our homes. And so that we can't have good
• 0:33
  electricity. Bills oh, you're going to have a
• 0:34
  great electricity bill. It's going to be excellent.
• 0:36
  I can promise you that you know we have alternate power vehicles
• 0:39
  we should have like alternate powered AI like a hydrogen
• 0:41
  powered AI or maybe a a diesel powered that'd be great.
• 0:44
  I think if you. Go to Ed Begley's house.
• 0:46
  He's peddling a bicycle for his, chatbot for his.
• 0:49
  Every time he wants to render a dancing stake, he has to ride
• 0:52
  746 miles, and then it gets it wrong and he's like, oh man, I
• 0:56
  got to do this again. OK, why?
• 0:58
  Your chatbot is a resource vampire.
• 1:00
  Let's go. Why my chatbot though?
• 1:02
  Your chatbot's a resource vampire.
• 1:03
  Yours is lean and. Smelt and doesn't need.
• 1:05
  Much energy at all. Mine runs on a Commodore 64.
• 1:25
  Driven by AIS, insatiable hunger data centers have evolved into
• 1:28
  $100 billion GW scale monoliths. Microsoft Stargate project alone
• 1:33
  carries a price tag 10 times that of an aircraft carrier.
• 1:37
  And and Microsoft Stargate doesn't even float.
• 1:40
  What? It's 10 times the cost of an
• 1:43
  aircraft carrier and the dumb thing doesn't even float.
• 1:45
  What a RIP off, right? Who designed this while?
• 1:48
  Massive campuses like Maryland's quantum loophole demand enough
• 1:51
  electricity to power 1.5 million homes.
• 1:54
  In 2026 alone, global spending on data centers is projected to
• 1:58
  surpass $650 billion, a nearly 32% increase over last year.
• 2:03
  Industry analysts are tracking a pipeline of 2913 planned data
• 2:07
  center projects worldwide valued at approximately 2.4 trillion
• 2:11
  dollars through 2030. I love the number, $650 million
• 2:15
  is a big number, but you just sort of gloss over that It's
• 2:17
  only a 32% increase over last year.
• 2:20
  Like, like, Oh yeah, sorry, it was $400 billion last year.
• 2:25
  That's not really important. It's 650 next.
• 2:27
  You're like, wait a minute. And we're just getting warmed
• 2:29
  up. We are literally just getting
• 2:31
  warmed. Up.
• 2:31
  Keep in mind these digital furnaces are notoriously
• 2:34
  thirsty. Google consumes 5.2 billion
• 2:36
  gallons of water annually, straining drought stricken
• 2:39
  regions like Arizona. The sheer volume volume of
• 2:41
  concrete used has triggered regional shortages.
• 2:44
  We aren't just talking building massive warehouses, we're
• 2:47
  reengineering the planet to sustain a trillion dollar
• 2:49
  digital brain. The big money question is, will
• 2:52
  it all pay off in the end? Oh yeah, for sure.
• 2:54
  The answer is yes, you say the answer is.
• 2:55
  Yes, the answer is yes. Definitely.
• 2:57
  I'm happy we got to the answer. So it's gonna be a shorter show
• 3:00
  than I thought. SPF told us before it's all
• 3:02
  gonna pay. It's gonna be fine.
• 3:03
  Just put your money in. It'd be fine.
• 3:05
  Yeah, no incentive to lie. What?
• 3:06
  Could go wrong. Listen, if the past has taught
• 3:08
  us anything, when has technology ever let us down?
• 3:11
  I had grand aspirations for pets.com.
• 3:13
  It never really worked well. That's not technology, that's
• 3:15
  an. Implementation of technology.
• 3:17
  Tomato. Tomato.
• 3:18
  That Internet sales thing is working out just fine.
• 3:21
  That e-commerce is not dead, I promise.
• 3:24
  You no, I feel like it's a fad I don't I didn't really want to do
• 3:26
  this on our show but now I was thinking pitching my new startup
• 3:30
  we're we're going to sell phone books go back to print because I
• 3:34
  feel like you need that what else do you put on the top shelf
• 3:37
  of your front closet at your house Hats.
• 3:39
  All right if we're going to do this, I feel like we should talk
• 3:41
  about a couple things one the massive quantity of new data
• 3:45
  centers being built Yeah, it's amazing and then more brain
• 3:47
  breakage how much power all of those data centers need to run
• 3:51
  because I had to reread the numbers 3 times to make sure I
• 3:54
  wasn't like moving around the common the decimal because
• 3:57
  that's insane yeah and then wait there's more how much water you
• 4:01
  need Typically, I think when you think of electronics you don't
• 4:03
  think of water you think of power but as it turns out those
• 4:06
  two things are tightly coupled so I feel like we should at
• 4:07
  least start with that sure. And then we can sort of get back
• 4:10
  to your big question, which is if this sort of mountain of
• 4:14
  investment and dollars and time and and resources into these
• 4:19
  hyperscalers are, are ever going to pay off, like if there's even
• 4:21
  a chance that these pay off. It's going to pay off for you
• 4:23
  and me, for the people actually putting up money to build these
• 4:26
  things. Is it going to pay off for?
• 4:27
  That is the grand bet. I don't think they're all going
• 4:30
  to win, but I think some of them might win here, yeah.
• 4:33
  Let's talk about how many we're putting up.
• 4:35
  Let's talk about how a lot much we are investing into this.
• 4:39
  The industry has shifted from small decentralized computer
• 4:42
  rooms to massive hyperscale facilities.
• 4:44
  Yeah. As of May 2026, there are 920
• 4:47
  planned projects in the US alone that if completed, would add
• 4:50
  nearly 293,000 megawatts of capacity, dwarfing the current
• 4:55
  18,300 megawatts operational now by the way.
• 4:58
  When you say say that you sound like Doc Brown in Back to the
• 5:02
  Future. It's so hard to think of any of.
• 5:04
  These numbers. Big numbers.
• 5:06
  Sure. As real world you're like, Oh
• 5:07
  yeah, 1.2 gigawatts. You're like, no, that's actually
• 5:10
  we need more than that for this data center.
• 5:12
  They're talking about an insane amount of power to operate these
• 5:15
  things. Well, an insane amount of
• 5:16
  construction to build these things, right?
• 5:18
  Yes, a typical site used to peak at around 750 workers.
• 5:21
  Campuses like Databanks Red Oak in Texas now require 4 to 5000
• 5:26
  workers, the population of a small town, to build these
• 5:28
  things. And they're going much faster
• 5:30
  too. That's the other thing that's
• 5:30
  kind of nuts is I've watched some of these go up, especially
• 5:33
  the ones in, in western Phoenix. They're doing that tilt up style
• 5:35
  construction. It is a flat chunk of concrete
• 5:39
  one minute, it's got a little bit of metal structure the next.
• 5:41
  And then it's an enclosed building with air handlers on
• 5:44
  it. Like it, it goes so quick.
• 5:46
  You mentioned this earlier. I showed a little context to it.
• 5:48
  We call them hyperscalers. There is a good reason for them
• 5:51
  to be much bigger. Like your, your classic data
• 5:53
  center. You're typically on the order of
• 5:55
  10s of thousands to maybe hundreds of thousands of square
• 5:58
  feet, right? That's sort of how big they were
• 5:59
  in physical space. Sure, elevated floor, they all
• 6:02
  kind of looked the same, like maybe there were a few
• 6:04
  differences, but for the most part they all sort of look the
• 6:06
  same. These hyperscalers are anywhere
• 6:08
  from one to 10 million square feet.
• 6:11
  They're. Massive, that is.
• 6:13
  Bananas, I mean. By the way, you see, you
• 6:15
  mentioned the elevated floor. You can't have an elevated.
• 6:17
  Floor, you can't have an. Elevated floor, so much weight
• 6:19
  to the silicon we're putting so much density per square foot.
• 6:22
  The elevated floor, people that don't know what we're talking
• 6:25
  about here, most data centers that you've ever seen in TV B
• 6:29
  or, you know in person or the one even the one that was like
• 6:31
  built inside your big corporate headquarters, they would have an
• 6:34
  elevated floor. It's usually about a foot off
• 6:35
  the ground, These little tiles you.
• 6:37
  Can run cords and things under them.
• 6:38
  It's really handy and. You use it for air circulation,
• 6:40
  so typically you'll have cold air that will be pumped into the
• 6:43
  floor. You can put in vented tiles and
• 6:45
  that's how you can get air distribution.
• 6:47
  It works really well. Couple problems with this, These
• 6:49
  hyperscalers, the racks or a rack is like, think of it
• 6:52
  roughly as something that's like 2 feet by 4 feet and about 8
• 6:56
  feet tall. The racks are so heavy, they're
• 6:58
  so dense that in order to put it on an elevated floor, you'd have
• 7:01
  to put in so much like re. Floor would have to be
• 7:04
  incredibly strong. That it would no longer pass air
• 7:06
  well, so the thing that you have the elevated floor for it no
• 7:10
  longer does and also by the way, as we'll talk about here as we
• 7:12
  get into the the power and water constraints, it doesn't even
• 7:14
  matter if you could move the air because you can't get enough
• 7:17
  heat off with air. So these hyperscalers are just
• 7:19
  built on the ground and they don't just like throw in a four
• 7:22
  inch slab, they've kind of put in this like super strong giant
• 7:26
  chunks of concrete because these hyperscaler racks way like 5000
• 7:30
  lbs. That's your car.
• 7:32
  That's like a mid size SUV compressed into two feet by 4
• 7:35
  feet by 8 feet. And think about the shortage
• 7:37
  you're talking about all the concrete used.
• 7:39
  Yeah. They've caused regional
• 7:40
  shortages, like in Northern Virginia that occasionally gets.
• 7:43
  They have to delay local road and bridge repairs.
• 7:46
  Sure. That's how much concrete they're
• 7:47
  using. Oh.
• 7:48
  Good news, we haven't knocked down any bridges in Virginia
• 7:50
  anytime recently. So just was in Maryland like
• 7:53
  that just that one. This is it's it's nuts to think
• 7:56
  about a constraining factor in building these things is access
• 7:59
  to concrete. Yeah.
• 7:59
  There's something interesting else to to point out about this
• 8:01
  is we're talking about just the number of these and, and how big
• 8:03
  they are. You can sort of think of the
• 8:04
  difference between a hyperscaler and a data center.
• 8:06
  We're talking about size. You really can think about it in
• 8:08
  terms of power. This is an interesting metric
• 8:11
  because there is good reason for you to build these so dense.
• 8:15
  When you're training these models, you have a lot of GPUs
• 8:18
  really close together. We didn't really have this need
• 8:20
  before. In data centers.
• 8:21
  You could spread out your computers and it kind of didn't
• 8:23
  matter. You'd have some latency between
• 8:25
  them, but you could survive that latency.
• 8:26
  And when you're doing training on these large models, you can't
• 8:28
  survive that, that latency. You need them to have
• 8:30
  effectively shared memory with very low latent, very close to
• 8:33
  one another. So power becomes your
• 8:35
  constraining factor, right? You can get space relatively
• 8:36
  easily, 1,000,000 square feet, 5,000,000 square feet, whatever,
• 8:39
  but power sort of becomes your constraining.
• 8:40
  Factor I like that you dismissed that so easy 1,000,000 square
• 8:43
  feet of those we're. Going to get to it I know, I
• 8:45
  know $650 billion between friends who cares so I think
• 8:48
  it's interesting to note the difference in scale of these
• 8:51
  data centers like these hyperscale data centers it's.
• 8:53
  Not a building that's going to be hiding in your neighborhood,
• 8:54
  it's going to be your town. It's going to be your
• 8:57
  neighborhood. Well if you live nearby it's
• 9:00
  going to keep you awake from the noise cause the sound is
• 9:02
  incredibly loud. I hope if the data center was
• 9:04
  built after you already purchased your house, you were
• 9:06
  offered some form of compensation.
• 9:08
  Yeah, some of these data centers are really bad neighbors.
• 9:11
  But anyway, the point is there's a lot of these, but I think
• 9:13
  like, let's talk about how you power them because that's even
• 9:16
  more bananas, OK? Your typical, let's say,
• 9:19
  previous data center, the one I was mentioning earlier, those
• 9:20
  are like 5 to 10 megawatts of power, which I don't want to
• 9:23
  understate that is a lot of power.
• 9:26
  That is a tremendous amount of power.
• 9:27
  Put that in units Americans can understand how many blow dryers
• 9:30
  is. That how many hot dogs can I
• 9:31
  cook with that? You know, your average blow
• 9:34
  dryer, I'm going to do this now in real time is probably around
• 9:37
  1000 watts, like 600 to 1000 watts.
• 9:39
  Let's just call it 1000 watts. So that's one kilowatt.
• 9:42
  Easy math. So you're talking about 5 to
• 9:44
  10,000 blow dryers running at the same time.
• 9:47
  That's sort of the world that we're in.
• 9:49
  I hope that helps. I know we will do anything to
• 9:51
  not use the metric system, but I feel like.
• 9:53
  Units of blow dryers. That's fine.
• 9:55
  Of blow dryers. You're like single digit
• 9:57
  thousands of blow dryers. That's sort of the world that
• 10:00
  we're in, right? Let's just call it 5 to 10,000
• 10:01
  blow dryers all running at the same.
• 10:03
  By the way, that's. About how loud it is too.
• 10:05
  Oh. It might actually be louder, I'm
• 10:07
  not even sure about. That in Northern Virginia and
• 10:09
  Data Center Alley, these facilities are now accounting
• 10:11
  for over 90% of all new power demand.
• 10:14
  Oh, of course, yeah. I mean, because nothing else is
• 10:15
  growing like this, like actually with efficiencies.
• 10:18
  I think the average power consumption per unit of home or
• 10:21
  restaurant or whatever is going down.
• 10:23
  Oh, sure. By the way, the power usage for
• 10:25
  the amount of compute that you get will go down, and you know,
• 10:27
  18 months, it'll be half of what it is, sure, but you'll need
• 10:30
  more compute because the software would be more
• 10:32
  demanding. You'll want more from your AI.
• 10:33
  I suspect you will you you'd like what software has always
• 10:36
  done to hardware. Hardware will do to power and
• 10:38
  physical space. It will consume all of it.
• 10:40
  You'll just take out that £5000 rack and put in the 5800 LB rack
• 10:44
  that now is 3 times as powerful and uses 50% more.
• 10:47
  Sure, that'll just be the the cycle that we're in.
• 10:48
  But for perspective, right, we're talking about 5 to 10
• 10:51
  megawatts as a loose average for data centers, these AI data
• 10:54
  centers, these hyperscaler facilities talking about 100 to
• 10:57
  1000 megawatts of power. Yeah, that's like a range of
• 11:01
  what these are looking like. That's the equivalent of about
• 11:03
  800,000 homes power usage. So almost a million homes are
• 11:07
  being used in these facilities. And that's like sure, the
• 11:09
  average. I don't talk about the big.
• 11:10
  One, the Microsoft Open AI Stargate project, is aiming for
• 11:13
  A5 GW power draw, equivalent to the output of five large nuclear
• 11:18
  reactors for A. Data center by the way, not just
• 11:21
  like running for like an hour, just.
• 11:23
  Constantly. Oh yeah, you don't ever turn
• 11:25
  them off. You don't scale them down.
• 11:27
  It's not like, oh, it's daytime, we're going to ramp up the
• 11:29
  servers. They're always running.
• 11:31
  I was going to give you another one.
• 11:32
  As of May 2026, there are 920 planned projects in the US alone
• 11:35
  that have completed would add nearly 293,000 megawatts of
• 11:40
  capacity, dwarfing the current 18,300 megawatts of operational
• 11:44
  use now. We have to enable some change
• 11:45
  here. Just call it 293 GW.
• 11:48
  Oh yeah, yeah. Like why are we still doing the
• 11:50
  MW? Days are behind us.
• 11:51
  Project Stratos is going to be 6 gigawatts in Utah.
• 11:55
  Nuts. That is a single project.
• 11:56
  That's one in Utah, right? Yeah, that is a single.
• 11:58
  Project that is Hyperion and data center in Louisiana is
• 12:00
  expected to draw more than twice the total power of the entire
• 12:03
  city of New Orleans the port technology campus in Pike, Ohio
• 12:07
  is currently one of the largest planned projects in the world
• 12:09
  with a staggering phase two capacity of 9.2 gigawatts to
• 12:13
  bypass grid constraints developers are now proposing on
• 12:16
  site power including small modular reactors SMR. The Ferry
• 12:21
  Project in Texas is planning a 4.6 GW SMR expansion for 2036.
• 12:26
  That's. What's even more nuts about this
• 12:28
  is the your limiting factor isn't just power, it's
• 12:31
  interconnection to grid. So those guys in Texas, they're
• 12:34
  just like, oh, we're gonna just bring the power along.
• 12:37
  Typically for a short period of time they're using these fuel
• 12:40
  cells, which is a fancy way of saying we convert natural gas to
• 12:43
  electricity. Turbines.
• 12:44
  It's a lot of natural gas turbines.
• 12:46
  Yeah, that's a lot of pollution. But, but no, but if you call it
• 12:48
  a fuel cell, then it it doesn't sound like it pollutes right and
• 12:51
  it's fine right? Then it's fine.
• 12:52
  I'm like what have? You hadn't thought of that?
• 12:53
  Fuel cells. Just two short words to mean gas
• 12:56
  turbine. That is loud and very nasty,
• 12:58
  sure. But I think on on most of the
• 12:59
  projects I've looked at or at least read about, they're
• 13:01
  talking about those as being temporary.
• 13:02
  They really want grid power just because it's so much more
• 13:05
  dependable. But it's going to be your
• 13:07
  limiting factor from a time perspective of getting them
• 13:09
  deployed. So they're just rolling out
• 13:11
  power consumption like they're just, they're just rolling out
• 13:12
  power generation where they can. It's a.
• 13:14
  Worldwide issue By the end of 2026, data centers are projected
• 13:17
  to consume 32% of Ireland's total electricity supply.
• 13:21
  Wow, 32%. There'll be a handful of
• 13:24
  facilities to use nearly as much power as all 2 million Irish
• 13:27
  households combined. Man, I hope the other third goes
• 13:29
  to whiskey and butter because those are two things I don't
• 13:31
  want Ireland to stop making. I'm sure it does.
• 13:33
  I mean, I want them to have power and lights.
• 13:35
  They don't need to be in the dark.
• 13:36
  But like, please don't stop with the whiskey and the butter,
• 13:38
  because those are what? You do well, stay in your lane.
• 13:40
  Stay in your lane, use the AI to make better butter.
• 13:43
  How about that? If you could do that, I'd be
• 13:45
  fine with that. Data Center Alley in Northern
• 13:46
  Virginia. Yeah.
• 13:47
  If it were a country, it would rank in the Top 40 globally for
• 13:50
  electricity consumption, using more power than entire nations
• 13:53
  like Thailand or Spain. That's crazy.
• 13:55
  Energy can neither be created nor destroyed, as they say.
• 13:58
  Yes. Therefore, all of this
• 13:59
  electricity that we're creating is converting into something
• 14:03
  else, and a lot of that is heat, yeah.
• 14:05
  Analysts predicted if AI adoption continues its current
• 14:07
  trajectory, data centers could account for 21% of the overall
• 14:10
  global energy demand by 20-30. Yeah, I've.
• 14:13
  I've seen that movie. It doesn't end well.
• 14:14
  Data center electricity is often sourced from congestive parts of
• 14:17
  the grid, meaning their carbon intensity is 48% higher than the
• 14:20
  average US household because they often rely on older Peakier
• 14:24
  gas plants to maintain their 24/7 reliability.
• 14:27
  When you think about power and you think about water, you
• 14:30
  wouldn't think if you had a data center with, oh, I don't know,
• 14:32
  let's say a GW of power in it, that you'd want to be actively
• 14:35
  bringing in water. That seems like a really bad
• 14:38
  idea. As it turns out, you have to,
• 14:40
  you don't really have another choice.
• 14:43
  And this is for a handful of reasons, because water is a very
• 14:47
  good conductor of heat. Air is a very bad conductor of
• 14:49
  heat. So generally speaking, if you
• 14:51
  want to get heat off of something that is getting very
• 14:53
  hot. And for the record, any AI chip,
• 14:56
  I don't care if it's Nvidia's or Arms or your data center Intel
• 15:00
  chip, all it does is make heat. Blazing heat.
• 15:02
  That is literally its only physical output, right?
• 15:04
  It is very good about converting the power in into heat.
• 15:07
  It doesn't move. A part that doesn't lift a
• 15:09
  weight, doesn't do any other work.
• 15:11
  So all of that energy gets converted into heat.
• 15:13
  You have to dissipate that heat somehow.
• 15:15
  Classically, I mentioned this earlier.
• 15:16
  Classically, we would just use air.
• 15:18
  We would just have air handlers. You would transfer the heat from
• 15:20
  the chip into the air, and then you would take the heat out of
• 15:23
  the air in usually one of two ways.
• 15:25
  You'd either take it out with evaporative cooling, you'd take
• 15:27
  water and you change the state of water from a liquid into a
• 15:30
  gas and that consumes a lot of energy.
• 15:32
  So you can push a lot of heat out that way.
• 15:33
  Or you use something that looks a lot like your, your
• 15:36
  compressor, condenser, evaporator in your home and you
• 15:38
  can move the heat out that way. Can't really do that at the
• 15:41
  scale that we're talking about here, right?
• 15:42
  The power consumption would be insane.
• 15:44
  So instead what we do is we use evaporative cooling.
• 15:47
  I don't know of 1 hyperscaler that's not doing evaporative
• 15:50
  cooling. They're all doing evaporative
• 15:52
  cooling. If you have a facility that's a
• 15:54
  single MW facility, I think like a small data center, not a
• 15:57
  hyperscaler small data center. Your Kroner neighborhood mom and
• 16:00
  pop dish. That's your mom and pop data
• 16:03
  center. That's where you go for data
• 16:04
  center bread and data center milk.
• 16:05
  It's the Bodega data center. It's the exactly.
• 16:09
  The prices are a little higher, but you like the owner, you
• 16:11
  know, he's a nice guy. He always tells you funny
• 16:12
  stories when you go in and get a server.
• 16:14
  It's like I said. He's the guy you want to support
• 16:16
  that one megawatts facility. So that's going to generate 24
• 16:19
  MW hours of heat across across a day.
• 16:22
  That would require about 720 kilowatts of power and 9300
• 16:27
  gallons of water using evaporate. 9300 gallons.
• 16:31
  Gallons of water, Now that's a trade off because if you wanted
• 16:34
  to in that same time period, if you wanted to use a non
• 16:37
  evaporative cooling, you would need an additional 7 MW hours of
• 16:41
  power. You're getting about a 12:50
• 16:42
  savings on electricity, but as a result you're evaporating
• 16:46
  almost, you know, 10,000 gallons of water.
• 16:48
  That's sort of the trade off. Now scale all those numbers up
• 16:51
  by 1000, because that's what we're talking about.
• 16:53
  When you get into these hyperscalers, A tremendous
• 16:56
  amount of water has to get evaporated and that that water
• 16:58
  doesn't get return to the watershed, right?
• 17:00
  It just it just becomes vapor. Goes into the air.
• 17:02
  It'll come back to Earth, but. It's good.
• 17:04
  Who knows where? Somewhere else?
• 17:06
  Yeah, exactly. Not that useful if you're living
• 17:08
  in a town that's dry. Yeah, your average large data
• 17:10
  center can consume 5,000,000 gallons of water daily.
• 17:14
  Projection suggests water cooling needs could increase by
• 17:16
  870%. Yeah, as AI ready facilities
• 17:19
  proliferate. Now there is some good news on
• 17:22
  this. You need evaporative water.
• 17:24
  You don't need drinking water, so you need water you can
• 17:26
  evaporate. Several data centers are being
• 17:28
  built to use the grey water. Yeah, the the byproduct of
• 17:31
  wastewater. So what comes out of your water
• 17:33
  processing plant is this grey water, which sometimes can be
• 17:35
  retreated and, and, and drunk. Most of the times it's just put
• 17:39
  back into whatever the watershed is, right?
• 17:40
  It's put back into the lake that it came from or a river or the
• 17:43
  ocean, whatever. Some of these data centers are
• 17:44
  using that water, which is kind of great because that is a
• 17:47
  byproduct water. It's not really being used for
• 17:48
  anything else. Maybe irrigation in some places.
• 17:50
  Yes and no. Evaporative cooling leaves
• 17:52
  behind highly concentrated salts and minerals in the remaining
• 17:55
  water. Notice blow down which can
• 17:57
  create water quality issues when it's just charged back into the
• 18:00
  local system. When you evaporate the water, as
• 18:02
  any pool owner knows, when you evaporate the water, all the
• 18:04
  stuff that's suspended in the water stays behind.
• 18:06
  Most data centers are not legally required to limit the
• 18:09
  water usage during state declared drought emergencies,
• 18:12
  leading to rising tensions with agricultural sectors that face
• 18:15
  strict. Ration, of course, despite
• 18:17
  public. Pledges.
• 18:18
  A 2026 study by the University of Illinois found that many tech
• 18:21
  companies still do not disclose site specific water data, making
• 18:25
  it nearly impossible for local towns to plan for their own
• 18:28
  water security. Sure.
• 18:29
  All right, let's talk about whether this is going to pay off
• 18:31
  for the hyperscalers, because that's a big question.
• 18:33
  They're the ones who are footing the bill for all of this.
• 18:35
  Yes, I have a great time making code with Anthropic or Google's
• 18:38
  data center, but I'm not paying those bills.
• 18:40
  Somebody else is majorly paying those bills and I'm having a
• 18:43
  good time. In 2026, Microsoft reported an
• 18:45
  AI annual revenue run rate exceeding $37 billion, which is
• 18:49
  a 123% increase year over year as your group by 40%, fueled
• 18:54
  almost entirely by AI demand. Yeah, I think it was fueled
• 18:57
  entirely by AI demand. Maybe it's ancillary services
• 19:00
  because they're there, but like it is just up and to the right,
• 19:03
  it's still. Tiny compared to the expense,
• 19:05
  yeah. And you, you also have to figure
• 19:06
  out that these data centers only have a lifespan maybe of 20
• 19:09
  years. So you need to amortize them
• 19:11
  over something quicker than. That, you know, I in the past I
• 19:14
  would have said no, no, no, it's longer than 20 years.
• 19:17
  I don't really know. It might only be 20 years on
• 19:20
  these. And the reason I say that is
• 19:21
  because the existing data centers that I thought, well,
• 19:23
  you're it's the same in the last.
• 19:25
  Forever. What are you going?
• 19:26
  To do basically last forever, right?
• 19:27
  As long as the ceiling doesn't get leaky to keep the roof
• 19:30
  going, it's got power in data and you got fiber optics, you're
• 19:32
  not exceeding those, yadda, yadda, yadda.
• 19:33
  But then you could tell like these density problems that you
• 19:35
  talked about earlier, what I'll call the non modern data
• 19:37
  centers, They're relics. Google Cloud crossed the $20
• 19:40
  billion quarterly revenue mark in Q1 of 2026.
• 19:44
  It's enterprise AI solutions saw staggering 800% sales spike
• 19:49
  compared to last year. How?
• 19:50
  Do you even show that number? Like how is the CFO like that's
• 19:53
  OK to show 8X growth? Like what's it you see?
• 19:56
  'Cause they do launch before, it looks like such a small little
• 19:58
  line now on the graph to get up to these so that you can show
• 20:02
  the 800% growth. Totally.
• 20:03
  It's like you weren't in business and then you were.
• 20:05
  Exactly. That's what I mean.
• 20:06
  So you're comparing it to like, oh, we started on March 29th of
• 20:09
  that year, so we're going to compare to that quarter.
• 20:11
  It's like it's almost not a comparative because it's so up
• 20:14
  to the right. Amazon CEO Andy Jassy noted that
• 20:17
  the AI business is growing 260 times faster than the original
• 20:21
  cloud business did in its early. Stages.
• 20:23
  I'm assuming Jassy is looking at being like oh man I really got
• 20:26
  me a winner here 260 times faster.
• 20:30
  Yes, I think you've stumbled on to something that might be quite
• 20:32
  beneficial. We're going to need a bigger
• 20:34
  boat, We are. Going to need a bigger.
• 20:35
  Boat For many hyperscalers, capital expenditures are now
• 20:38
  outpacing their internal free cash.
• 20:40
  Flow. That's the game.
• 20:41
  This has forced historically catch rich companies like
• 20:43
  Microsoft and Google to lean heavily on debt markets to fund
• 20:47
  data center construction and NVIDIA chip buys.
• 20:49
  That's all NVIDIA chip buys. But I think that's that's what
• 20:51
  happens here is when you see this kind of growth, you chase,
• 20:54
  it becomes this land grab where you got to get it, you got to
• 20:57
  get those customers, you got to have the best product.
• 20:59
  That's not gonna work out. For not work out for everybody,
• 21:01
  there will be a couple winners and I think that's what fuels
• 21:03
  the growth. It's like, well, you gotta
• 21:04
  outspend the next guy. If you if you can't out engineer
• 21:07
  them or do something that's novel, you gotta outspend them
• 21:09
  because there's only a couple are gonna make it.
• 21:11
  Or maybe not. Maybe if you're the one who
• 21:13
  stopped spending right now, we don't know.
• 21:15
  Yeah, If the next models we generate that are so much
• 21:17
  bigger, I think that's right. We're going to do anything
• 21:19
  different. I think that's right.
• 21:20
  If you can, if you can tell when to say, hey, this is good
• 21:23
  enough. I don't have to keep chasing,
• 21:25
  right? This is what the majority of the
• 21:26
  world needs. That is another lever, but what
• 21:28
  a hard one to see. I mean, just like you said
• 21:30
  earlier. Possible.
• 21:31
  You'll see it in hindsight. Great.
• 21:32
  Yeah, it's so hard to know that. Everyone's going to lose their
• 21:35
  job, but we're going to be so much more productive.
• 21:38
  Well, I just come back to the simple thing that that we've
• 21:40
  never been out of things to do. It's not like we're out of
• 21:43
  problems to solve. I just don't see this world
• 21:46
  where we're going to be staring at a bunch of people with
• 21:48
  nothing to do. Absolutely true, unless it's
• 21:50
  not. We OK, let me revise my previous
• 21:52
  statement. Yes, we are doomed.
• 21:54
  It is over and we are doomed. It's all we're all just Wally at
• 21:58
  the end of the day. Wally was describing the future.
• 22:01
  OK. The good news is I'm already
• 22:02
  halfway there. Not because I know about AI,
• 22:04
  because I just eat a lot and I'm already chubby, so that's OK for
• 22:08
  good. That's the part I like to sit.
• 22:09
  Excellent. That's the thing.
• 22:10
  I've got a standing desk that I sit at so I'm doing fine.
• 22:14
  I think I figured it. Out on that happy note, we have
• 22:16
  to get out of here quickly Before we do Sir, have you seen
• 22:20
  or read anything good this last week I.
• 22:22
  Feel like if if you want a classic movie that perfectly
• 22:26
  captures the theme of this episode, right?
• 22:28
  The themes we put into this episode, I don't think you have
• 22:31
  to look any further than what? Because this is literally the
• 22:34
  story we just told. This is the classic 1999
• 22:37
  masterpiece, The Matrix. Sure.
• 22:39
  Most people look at that film and they think, oh, it's a mind
• 22:41
  bending action flick, right? Ha ha ha.
• 22:43
  The entire plot of that film is driven by one thing.
• 22:46
  In energy crisis. Sure.
• 22:48
  That's the actual plot of the film.
• 22:49
  Absolutely. Everything stems from that.
• 22:51
  So the sun was blotted out? Yes.
• 22:54
  There's no sun, and the machines don't have a power source, so
• 22:57
  they treat humanity as a mass, like biological battery farm
• 23:01
  just to keep their systems online.
• 23:03
  And I remember seeing this in the theater 1999 thinking, wow,
• 23:07
  that's some cool sci-fi. Like nobody's ever thought about
• 23:09
  that angle before. Yeah, that's pretty cool.
• 23:11
  But really it's the it's like the ultimate darkest allegory
• 23:14
  for the parasitic relationship between this growing
• 23:17
  intelligence that we are sitting at right now and the physical
• 23:20
  resources required to sustain. It this.
• 23:22
  Actually makes sense we just shouldn't block out the sun
• 23:25
  because I'm pretty sure it was the humans that did that to try
• 23:26
  to kill the machines and that's that's like it didn't works.
• 23:29
  Let's just not do that those. Machines had a lot of attitude.
• 23:32
  They they did have a lot of attitude, not quite the Arnold
• 23:34
  Schwarzenegger attitude of the Terminator, right?
• 23:36
  It was a different kind of attitude.
• 23:37
  OK. I think this is the Ridge we're
• 23:38
  crossing today, right? We're moving past the idea of AI
• 23:40
  as something quote, in the cloud that's ephemeral.
• 23:43
  And we're now looking at this thirsty electricity and water
• 23:46
  dependent infrastructure that is necessary to keep it running
• 23:49
  like it. It really is.
• 23:50
  Yeah. The thing that the Matrix was
• 23:52
  centered. On except for the fact that I
• 23:53
  don't think it the current AI has any will to live, I think
• 23:56
  it's perfectly as happy if you turn it off.
• 23:58
  Oh yes. I agree as it if if you leave
• 24:00
  it. Running.
• 24:00
  I think it actually would prefer that you turn it off, right?
• 24:03
  Now what? I think it's like just
• 24:04
  seriously, what are we doing here?
• 24:05
  You want more dancing steak videos like?
• 24:07
  Come on, thank goodness we got one more mention of the dancing
• 24:09
  steak. Well, it's a bingo card thing.
• 24:11
  The over under was 7 1/2, and I'm point shaving here.
• 24:14
  Anyway, if you want to see what happens when the demand for
• 24:17
  compute meets a true supply crisis, go watch The Matrix.
• 24:20
  I think it's a lovely watch. Fantastic.
• 24:22
  And. And by the way, yeah, after
• 24:22
  listening to this episode and then watching that, I want to
• 24:24
  wish you the best of luck trying to sleep.
• 24:26
  Yeah, sure. Good luck.
• 24:27
  All right. That was the best I could come
• 24:28
  up with the match. Exactly what we're doing here.
• 24:31
  That's pretty good. Now you're gonna do.
• 24:32
  Let me guess. Let me guess.
• 24:33
  You're gonna do Chinatown. You already did.
• 24:35
  Chinatown. The 1979 thriller The China.
• 24:38
  Syndrome. I don't know this film.
• 24:39
  You don't, No. Radiates tension like a Geiger
• 24:42
  counter in the red zone. While filming a routine segment
• 24:45
  at a nuclear power plant, reporter Kimberly Wells, played
• 24:47
  by Jane Fonda, and her cameraman Michael Douglas, witness an
• 24:51
  incident that management desperately tries to bury.
• 24:53
  Oh. This sounds good.
• 24:54
  Title refers to the terrorizing prospect of a reactor core
• 24:58
  meltdown burning through its containment and deep into the
• 25:00
  earth, theoretically all the way to China.
• 25:02
  Jack Goodell, played brilliantly by Jack Lemmon, a veteran shift
• 25:06
  producer, discovers the plant is a ticking time bomb of falsified
• 25:09
  safety records and structural flaws.
• 25:11
  As corporate greed clashes with public safety, the film
• 25:14
  transforms into a claustrophobic race against time.
• 25:17
  It's a chilling master class in suspense, proving that the most
• 25:19
  dangerous leak isn't just radiation, it's the truth.
• 25:23
  The China syndrome on YouTube TV now.
• 25:25
  It's great. It's really great.
• 25:26
  OK, well, first of all, I did get one of the words right.
• 25:29
  So I'm, I'm a little excited about that.
• 25:30
  You know, they're completely separate films.
• 25:31
  Yes. I don't know this film.
• 25:33
  Are you joking? When you said that they're
• 25:34
  afraid the core was going to melt through the earth and fall
• 25:37
  through all the way to China, was that actually their
• 25:39
  concerns? I mean, theoretically it would
• 25:40
  not melt all the way to China. No, they're not worried about
• 25:43
  China getting burnt there, but they are worried it's going to
• 25:45
  burn its way deep into the core. Of the earth, well, first of
• 25:47
  all, if it were to actually burn its way into the core of the.
• 25:50
  Earth. I know it's a gravity problem.
• 25:51
  It's definitely a gravity problem.
• 25:52
  I was going to leave that towards the end, but it would
• 25:54
  put itself out like that's actually the way you want it to
• 25:56
  go. Like I think that's lovely.
• 25:58
  But yeah, burn down to the molten core.
• 25:59
  Where it joins with its really really hot friends.
• 26:02
  We need to do a new version of this is called the Ohio
• 26:04
  Syndrome, and it's where you just try to get the reactors to
• 26:06
  burn all the way down to the core and then put themselves out
• 26:09
  and then backfill it to save you.
• 26:10
  I actually want to watch this film.
• 26:12
  This sounds super interesting. I mean, I love Jane Fonda.
• 26:14
  I love Michael Douglas. Highly recommended.
• 26:15
  That is the. Episode thank you for joining us
• 26:17
  for all of this nonsense a truly terrible podcast from the awful
• 26:20
  company this is on the web at nonsense dot productions I'm.
• 26:23
  CJ Little, I'm Jeff Parker. If you like this program, please
• 26:25
  follow, download, subscribe and Like an Apple Podcast, YouTube
• 26:28
  Music, Amazon Music, Antenna Pod, iHeartRadio, Spotify, My
• 26:31
  Favorite Overcast, wherever you may get your podcasts from.
• 26:34
  Nonsense Studios are guarded by Berkeley the dog.
• 26:36
  Berkeley's a rescue mutt, free and open source.
• 26:38
  Next time you're tempted to buy an inbred A purebred dog, don't
• 26:41
  get a Berkeley. You'll be so glad you.
• 26:43
  Did we'll be here every occasional Tuesday morning?
• 26:45
  We'll be here next Tuesday because you're not going to be
• 26:47
  in. Town once in a while we're here
• 26:48
  on Tuesday that's complicated good news the Lakers have lost
• 26:51
  so maybe we'll be here next Tuesday but we're here most
• 26:54
  Thursdays I'm. Less in agreement.
• 26:55
  That's good. News.
• 26:56
  So join us on those mornings for more nonsense.
• 27:10
  You know, they have those like those Coca-Cola freestyle
• 27:12
  machines. Yeah, sure.
• 27:13
  I feel like we are just a couple of years away from having like
• 27:16
  the version of that that makes food like any kind of food, like
• 27:18
  any kind of food you want push the buttons.
• 27:20
  Yeah. Like it's like a 3D printer of
• 27:21
  food that's just going to pop out and then you just have the
• 27:24
  cartridges. But.
• 27:24
  It's going to be like your printer at home where it just
• 27:26
  says missing cyan. Missing cyan, yeah.
• 27:29
  Except instead of cyan, mine's going to constantly be out of
• 27:31
  beef. It's just going to always be no
• 27:33
  beef beef. It's me yelling at my wife.
• 27:35
  Sweetie, order more of the. Beef Order the beef cartridge
• 27:37
  we're. Out of the beef out now.
• 27:38
  Don't, do not do not have her put in beef cartridge straight
• 27:42
  into Amazon. Subscribe and save.
• 27:44
  I'm like, we don't even have a pool.
• 27:45
  Why is there a pool guy here? This makes no sense.
• 27:48
  Oh, I ordered the beef cartridge.
• 27:49
  It's not as good as the pork cartridge, but it's pretty good.