Latest / DEEP INSIGHT / Why the Iran war could expose AIs biggest weakness

Transcript

Why the Iran war could expose AIs biggest weakness · DEEP INSIGHT

• 0:00
  Speaker 1: If you've been worried about AI taking your job or
• 0:03
  Speaker 1: I don't know, taking over the world, then this really
• 0:06
  Speaker 1: awful situation in Iran is probably actually good news for
• 0:10
  Speaker 1: you because it hits the thing that makes AI possible,
• 0:16
  Speaker 1: which is the most enormous energy intensive build out of infrastructure,
• 0:23
  Speaker 1: possibly one of the most enormous and energy intensive buildouts
• 0:26
  Speaker 1: in modern history.
• 0:28
  Speaker 2: I know that might sound like a bit of an
• 0:29
  Speaker 2: over the top claim, but.
• 0:30
  Speaker 1: Honestly, I think I've got some charts here, and I
• 0:32
  Speaker 1: think once I show them to you, I think you'll
• 0:33
  Speaker 1: see what I mean. So this is the money that
• 0:37
  Speaker 1: has been spent on data centers in the last three
• 0:40
  Speaker 1: or four years. These are the companies doing the spending,
• 0:42
  Speaker 1: and the companies that you know that make AI, like
• 0:45
  Speaker 1: open Ai, they rent the data center of the data
• 0:47
  Speaker 1: centers of these guys has to look over here this bar.
• 0:51
  Speaker 1: I think it's about four hundred billion. This number is projected,
• 0:55
  Speaker 1: so you know, take it with the pinch of salt,
• 0:56
  Speaker 1: but I think it's about right, six hundred and fifty billion.
• 1:00
  Speaker 1: This is the money being spent on data centers right now.
• 1:05
  Speaker 1: And I know it can be impossible to compute what
• 1:08
  Speaker 1: six hundred and fifty billion really means. So I've tried
• 1:11
  Speaker 1: to put it in comparison with some things that we
• 1:13
  Speaker 1: can relate to a bit more so. This direy little
• 1:18
  Speaker 1: bar down there, that is UK defense spending last year.
• 1:22
  Speaker 1: This is the spending that the UK did on the
• 1:25
  Speaker 1: whole of COVID, not one year, the whole of COVID.
• 1:29
  Speaker 1: This is how much NASA's Apollo program, which sent astronauts
• 1:33
  Speaker 1: to the Moon in the ninety sixties ninety seventies cost again,
• 1:36
  Speaker 1: not one year, the whole thing. And over here as
• 1:41
  Speaker 1: data centers. But yeah, I mean, these comparisons are little random,
• 1:45
  Speaker 1: aren't they. So let's put it into more concrete historical comparison.
• 1:50
  Speaker 1: These are big infrastructure buildouts as an average percentage of
• 1:54
  Speaker 1: US GDP, US rail rights, electrification, US highways. This bar
• 1:59
  Speaker 1: here to the dot com boom and over here again
• 2:03
  Speaker 1: a projected number.
• 2:03
  Speaker 2: So take it with a pinch of salt. That's AI.
• 2:07
  Speaker 1: This buildout is huge and it all depends on energy
• 2:11
  Speaker 1: because this is what they're building here, it is this
• 2:17
  Speaker 1: is the data center, one of the biggest data in
• 2:18
  Speaker 1: the world, data centers in the world, Microsoft Fairwater. I mean,
• 2:22
  Speaker 1: these are cars here to give you some sense of
• 2:24
  Speaker 1: the scale. And actually this thing is still under construction
• 2:27
  Speaker 1: and they're still building it, and when they finished, it
• 2:30
  Speaker 1: will be maybe the most powerful AIDENTIT center in the world,
• 2:34
  Speaker 1: and it will have cost seven billion dollars.
• 2:38
  Speaker 2: And of course, like I was.
• 2:39
  Speaker 1: Saying at the beginning, this thing needs an unbelievable amount
• 2:43
  Speaker 1: of energy. How much you ask, well, this is the
• 2:47
  Speaker 1: average annual power use for the city of Los Angeles,
• 2:52
  Speaker 1: and over here is where Fairwater is projected to get
• 2:55
  Speaker 1: when they finish. This line will go up and down
• 2:57
  Speaker 1: a bit, you know, when everyone in LA switched on
• 2:59
  Speaker 1: the air condition units. But on a typical day, fair
• 3:03
  Speaker 1: Water will draw more power the entire city of LA.
• 3:07
  Speaker 1: And when you look at the build out of data
• 3:09
  Speaker 1: centers in a national perspective, you see that basically they've
• 3:13
  Speaker 1: been building multiple LA's worth of power. This is the
• 3:19
  Speaker 1: blue bars here are operational data centers, and the red
• 3:23
  Speaker 1: ones here in construction. If you combine blue and red,
• 3:27
  Speaker 1: then you get about six las worth of power or
• 3:31
  Speaker 1: three London's. Or to put it into more granular terms,
• 3:34
  Speaker 1: I waxed it out fourteen million microwaves. This is the
• 3:37
  Speaker 1: power you'd need to plug in fourteen million microwaves all
• 3:41
  Speaker 1: at the same time. And you know, I know that
• 3:43
  Speaker 1: sounds a bit silly, but there's actually not that far
• 3:46
  Speaker 1: off what they're doing. Because you know how you when
• 3:48
  Speaker 1: you run your microwave in your kitchen, it get super hot. Well,
• 3:51
  Speaker 1: that's the problem that they have to do with as well.
• 3:52
  Speaker 1: That's partly why it takes so much energy just cooling
• 3:55
  Speaker 1: down all those computers, because that's what a data center
• 3:59
  Speaker 1: is really just a massive load of computers that you
• 4:02
  Speaker 1: kind of do your work on remotely. But here's the
• 4:07
  Speaker 1: funny thing. In a way, here's the craziest part of
• 4:10
  Speaker 1: this all. Despite all this build out, there still isn't
• 4:15
  Speaker 1: enough space. This is the vacancy rating primary data centers
• 4:18
  Speaker 1: in the US, which are the best ones down at
• 4:21
  Speaker 1: one percent, they are completely full.
• 4:24
  Speaker 2: Of course, laws of supply and demand.
• 4:26
  Speaker 1: What happens when you have lots of demand none the
• 4:28
  Speaker 1: supply prices go up. Price have been going up now
• 4:33
  Speaker 1: pretty much ever since the Hour of Boom started. But
• 4:36
  Speaker 1: despite these prices, the data centers are still completely full.
• 4:40
  Speaker 2: There's this incredible squeeze.
• 4:42
  Speaker 1: And now we've got the situation in the straits of
• 4:45
  Speaker 1: Horn Moves which is going to make that squeeze even tighter. Now,
• 4:52
  Speaker 1: there's something funny about data centers. You hear a lot
• 4:55
  Speaker 1: about how much electricity they use, and if I've just
• 4:58
  Speaker 1: said it, but it's not actually true overall, this is
• 5:02
  Speaker 1: quite an important point. Data centers as a percentage of
• 5:07
  Speaker 1: US electricity consumption, I think they're about four percent. And
• 5:10
  Speaker 1: this is kind of putting that same thing in terror
• 5:12
  Speaker 1: what hours, But residential air conditioning much more used much
• 5:17
  Speaker 1: more electricity. The data centers microwave is much smaller, but
• 5:20
  Speaker 1: it's still on the same chart.
• 5:22
  Speaker 2: You wouldn't expect that. And I think what I'm getting.
• 5:25
  Speaker 1: At is the nature of the squeeze here, because it's
• 5:27
  Speaker 1: not just about total energy, it's about getting energy to
• 5:30
  Speaker 1: the right place. Putting a data center is a bit
• 5:35
  Speaker 1: like if someone came into your house and plugged in
• 5:40
  Speaker 1: a microwave into every socking and started running it. I mean,
• 5:44
  Speaker 1: your street wouldn't notice that draw that much, and your
• 5:46
  Speaker 1: city definitely wouldn't, but your fusebox would, and your energy.
• 5:49
  Speaker 2: Bills definitely would. And that's more what it's like.
• 5:53
  Speaker 1: And you might say, yeah, of course, but no one
• 5:54
  Speaker 1: would ever do that, But actually, you know something like
• 5:57
  Speaker 1: that did happen in Ireland. They put in all these
• 6:00
  Speaker 1: tax breaks for tech companies who started coming there, and
• 6:04
  Speaker 1: then they started building data centers. The blue line here
• 6:07
  Speaker 1: is the electricity use as a percentage of urban households
• 6:10
  Speaker 1: in islands, and the red line is the electricity use
• 6:14
  Speaker 1: of data centers. You see, it was just kept on
• 6:17
  Speaker 1: going up and up and up and up, mainly because
• 6:19
  Speaker 1: island has small has quite a small grid, and it
• 6:23
  Speaker 1: carrying going up even after they actually put a moratorium
• 6:26
  Speaker 1: in on data center building. I was in twenty twenty two.
• 6:29
  Speaker 1: I think they've they've finished it now. Although there are
• 6:31
  Speaker 1: now lots of constraints on building data centers. They basically
• 6:34
  Speaker 1: had to stop them, stop them taking over the whole grid.
• 6:39
  Speaker 2: Strange, isn't it. This is the point I'm getting at.
• 6:42
  Speaker 1: This whole system, this enormous system that has been built
• 6:46
  Speaker 1: to make AI function, and in fact the whole digital
• 6:49
  Speaker 1: world function, is under tremendous strain. And now President Trump,
• 6:54
  Speaker 1: thanks to his actions in Iran, has added.
• 6:56
  Speaker 2: That extra constraint.
• 6:58
  Speaker 1: This is going to squeeze the and already system even tighter.
• 7:03
  Speaker 1: And the eventual effect, well, I think it has to
• 7:06
  Speaker 1: be to slow down AI because although the tech companies
• 7:09
  Speaker 1: are actually getting into energy themselves. This is three Mile
• 7:12
  Speaker 1: Island in the US, the site of one of the
• 7:14
  Speaker 1: US's biggest nuclear accidents, and Microsoft is taking it over
• 7:18
  Speaker 1: to make powerful data centers. But that's going to take
• 7:21
  Speaker 1: a while. That's not going to come on board that soon.
• 7:24
  Speaker 1: So instead, I think what we're going to see is
• 7:26
  Speaker 1: more outages. I think it's going to be really hard
• 7:29
  Speaker 1: for the AI companies like Anthropic to get data to
• 7:34
  Speaker 1: us when and where.
• 7:36
  Speaker 2: We need it.
• 7:37
  Speaker 1: You can see these these red bars. This is where
• 7:40
  Speaker 1: this is poor performance, simply because there's so much demand.
• 7:44
  Speaker 1: I think also inevitably we'll see prices going up, either
• 7:47
  Speaker 1: that or kind of deals to incentivize you like this one,
• 7:49
  Speaker 1: deals to incentivize you using AI at certain times when
• 7:52
  Speaker 1: it's not so energy intensive. But lastly, and this really
• 7:57
  Speaker 1: is kind of the most important thing from a civilizational perspective,
• 8:01
  Speaker 1: if you think about it like that, I think we
• 8:03
  Speaker 1: must see some slowing of progress in AI because the
• 8:08
  Speaker 1: need for these data centers is to do the incredibly
• 8:12
  Speaker 1: energy intensive work of training AI models, and this has
• 8:15
  Speaker 1: been unbelievably effective. This chart of AI progress, I mean,
• 8:19
  Speaker 1: look at this This chart is an exponential chart.
• 8:22
  Speaker 2: I mean it's practically vertical at this point.
• 8:25
  Speaker 1: But if they can't get the energy, they can't get
• 8:28
  Speaker 1: the space in the data centers, then this chart must
• 8:31
  Speaker 1: I don't know if it's going to look much like
• 8:32
  Speaker 1: it's slowing down, but it must go a little bit
• 8:34
  Speaker 1: less vertical.
• 8:36
  Speaker 2: You think that would make so much sense?
• 8:38
  Speaker 1: And I suppose thinking about the big worry about AI,
• 8:42
  Speaker 1: well maybe this very awkward situation in Iran, which is
• 8:47
  Speaker 1: going to cause a lot of economic harm, actually might
• 8:50
  Speaker 1: be buying time for us to get ready for whatever
• 8:54
  Speaker 1: AI is going to be
• 8:57
  Speaker 2: Sure, hope we use it wisely.