Loading summary
Transcript512 lines
- [00:01]
A
Support for the show comes from l' Oreal Group, the global beauty leader defining the future of beauty through science and technology. L' Oreal Group create the beauty that moves the world.
- [00:17]
B
What does it really mean to be a neighbor?
- [00:19]
A
It's just everyday people, you know, it's just people who are retired. They have a couple hours in the afternoon, so they're gonna do patrols. And it's people who are, you know, real estate agents driving around trying to track how ice is moving and alert neighbors when things are not safe.
- [00:38]
B
The rise of mutual aid in times of crisis. That's this week on Explain it to Me new episodes Sundays, wherever you get your podcasts.
- [00:51]
A
Welcome to the Vergecast, the flagship podcast of the Daft Punk album Random Access, Memories and the Reddit thread I just found asking what would be the name of the tracks on Daft Punk's Christmas album, which I encourage everyone to go spend a lot of time reading. I'm your friend, David Pearce. Nilai Patel is here. Nilai, hello. Hello. Sean Hollister, also here. Hello, Sean.
- [01:12]
C
I am also a friend.
- [01:13]
A
This is the Vergecast Holiday Spectacular, which I would say. Nilay, is it fair to say this is just a joke that we made one time that has now gotten way away from us over time?
- [01:24]
D
Yes. To the point where I'm wearing a holiday muffin.
- [01:28]
A
Okay. I'm glad you brought this up. What is that?
- [01:32]
D
So I was gonna wear a Santa hat, but I couldn't find it. The last time I wore the Santa hat was last year on this show. So I don't know where it is. It should just be in this room. But last year, I acquired from my wife what can only be described as a holiday muff, which I'm currently wearing. It's a scarf, but it's a loop. It's very warm.
- [01:52]
C
How many stars does it have? I.
- [01:55]
D
It's a lot. I look great. I. What I. What I. What I've come to learn is that I should be a scarf guy.
- [02:01]
A
You should. I actually agree with that. I look terrible in scarves. But you're. You're kind of pulling this one off.
- [02:05]
D
I feel great about it.
- [02:06]
B
I'm.
- [02:07]
D
My neck is very warm.
- [02:08]
A
Because I would like everyone, if you're listening to this, imagine, like, the. If you've ever seen a dog wearing the cone of Shame. But it's not a cone. It's just this sort of inflatable donut they put around their necks. That's Nilai. But make it fashion. Yeah.
- [02:20]
D
You know what I mean? Like, that's no Becky's very fashionable. This looks great on her. I don't agree that it looks great on me, but I'm pulling it off right now.
- [02:31]
A
I want to say, Sean, you meanwhile, managed to make this space that you're in go from not holiday at all to extremely holiday in, like, three minutes, which I'm very impressed by.
- [02:42]
C
It turns out that my wife is quite the Christmas decorator and every other part of this house had lots of stuff to steal, so I stole all of it.
- [02:49]
A
That's great. So your room is the respite from joy and happiness and seasonal feelings. Except for right now.
- [02:56]
C
Unless those feelings are Star Fox leftist.
- [03:00]
D
I do want to say I was going to put some holiday art from the Samsung frame store on the frame TV behind me, but I was worried about copyright infringement. So this is just a winter scene painted by Van Gogh, who is dead and can't sue us. That's right, Van Gogh.
- [03:13]
A
It's a challenge calling you out. What up, Vince?
- [03:17]
D
Yeah, toe to toe, Vinny.
- [03:21]
A
So what we do every year on the Spectacular is we pick a spec or a technology or some piece of this world of tech that we live in and get like, weirdly deep into it. And we've done Nilay. Help me remember, we've done Bluetooth in the past. We've done USB C in the past. We did Matter last year. That was very fun. What else?
- [03:42]
D
We started with hdmi because the joke was that after the first Trump administration, we had so much politics on the show, we. We were going to clear the air. We were going to reset with a full hour on HDMI to do the nerdiest possible thing that was in politics. And now we have to do a spectacular every year.
- [04:00]
A
A beloved, beloved thing that we all.
- [04:03]
D
By the way, it's time to do HDMI again. I think we're. We're headed towards needing to do another HDMI episode.
- [04:08]
C
We're not going straight to 6G, so.
- [04:11]
A
Okay, Sean, I'm glad you brought this up. Here is a partial list of things that we could have done instead of the one that we've picked. We could have done 6G. It would have been a very short episode and it would have made all of us very angry. But that's an episode we could have done. We thought about doing instead of doing Matter. Last year, we almost did arm. Just the idea of ARM chips, which I think was very important. Last year continues to be very important. All things LLMs, you could spend a lot of time just digging into, like, inference as a spec would just get real weird.
- [04:43]
B
Spec.
- [04:44]
A
We could do TPUs. You want to do TPUs for an hour and a half?
- [04:47]
C
These aren't specs.
- [04:47]
D
You gotta. It's a spectacular. Not an acronym specific spectacular.
- [04:52]
C
I am very down for less. Letting the LLMs play for us.
- [04:57]
A
It's true. We could do Wi Fi. 7.
- [05:00]
D
That's a spec.
- [05:01]
A
We could do satellite Internet, which not a spec has. I'm sure there is a spec. I didn't do enough research to care, but, you know, there's specs involved. There you go. Li Fi. I thought about doing the wireless power thing where you can, like, point a charger at yourself and. And apparently won't. That won't be.
- [05:20]
D
That's. That's. That's what Sean's talking about.
- [05:21]
A
I proposed E Ink Kaleido. Uh, Travis shot that down fairly quickly. Um, Travis suggested Labubu, which may or may not be a spec. We can debate that later. But the one that we decided is ram. We're going to spend this whole episode talking about ram, because at least RAM.
- [05:38]
D
Contains within it specs.
- [05:40]
A
Ram. What is. What is everything if not a bundle of specs? You know what I mean? Think about that. We are all specs. Yeah, but ram, I think, Sean, my question to you is, would you have gone into 2025 expecting to spend as much time as you have covering specific sticks of ram?
- [06:01]
C
Absolutely not. This stuff is commodity. It's so boring until it gets fascinating. You find out how it's made, but it's just take it for granted. Right? It's everywhere, and it. You don't even think about it.
- [06:16]
E
I thought.
- [06:17]
A
Yeah, I feel like we. This. It's like a thing we write about, sort of as a clause in a sentence in a phone review about how much RAM it has. And that is most of the way that I've thought about RAM for my entire adult life. And now all of a sudden, it is like a precious rare earth mineral in the world. And we're going to talk about why. So we're going to do this show in three parts. First, we're going to just build up some knowledge of ram. Sean, you are going to just school us on what RAM is and how it works and how it became to be a thing that we all care deeply about. Then Travis, our producer, has some games that he has created, Vibe Coded, that are, I assume, going to make us all look very dumb and be deeply terrifying. And then we're going to have a conversation at the end about kind of the macroeconomic chip war of all of this. There is a big Picture question about how chips get made and who's in charge and what it means for the future of technology. And we're getting get into all of that, but that is for later and for now. Sean, we asked you to like, really give us the, the like kindergarten, starting education and how RAM works. Are you ready for this?
- [07:21]
C
Oh boy. Can't wait.
- [07:24]
A
All right, so let's, let's literally start from nothing. What, what is ram and where did it come from?
- [07:30]
C
Okay, at its very, very basic. There are now nowadays chips that are going to be on a stick or possibly patterns etched into the top of your processor that also do this memory thing. And their job is to store charge. They will store a little bit of electricity that indicates if it is a one or they will not have that amount of electricity in it, not have quite that much charge in it. It will be a zero. And these ones and zeros represent computer memory because it is code. This is the digital 01 that you hear about. This is where it is stored temporarily inside your computer, inside your phone, so that the rest of the computer can read it and say, what did I just compute? Now I can act on that previous thing I computed and do more with it. And so computers didn't always have this chip kind of memory. They had all kinds of other kinds of memory. Before this. There were switches you would have to flip or rods or gears that would be turned to a certain position in a mechanical computer that would be, this is my storage, my memory of what I just did a moment ago so that I can do more computing on top of that. But all of these forms were known as memory and the ram, the random access of that memory is the thing that the computer is doing all the time, randomly looking at various parts of it to find out what it stored temporarily in its memory that it can look back at. The kind of RAM we talk about most of the time is DRAM dynamic random access memory. This is called dynamic because the charges that are stored in it need to constantly be refreshed. The memory doesn't have. It doesn't have any. It doesn't remember what's inside of it. When it doesn't have power, it needs to constantly have power. So that's refreshing those things. But we have other kinds of memory we talk about as storage. We have SSDs which keep stuff in there longer without power. Not forever, but longer. Most of the time right now though, and the crunch that we're talking right now is mostly about DRAM memory that's going to give you that little short term for the computer to act on it.
- [09:43]
D
Again, I feel like it's important to say here that historically DRAM has been much faster than the other kinds of storage on a computer. Right? So back in the day, the hard drive was really slow, quite linear actually in how it stored information. And you would want to have a lot of RAM because it was so much faster to get information in and out. But that's sort of equalized, right? Like modern SSDs and RAM, they're closer than they were, but they're still not very close.
- [10:08]
F
Right?
- [10:09]
C
Yeah, memory is much, much faster than storage still. But, but storage is it. We're, we're at the point now where in most computers you'll have your, your, your random access memory that's, that's doing the very quick, you know, calculations back and forth between the processor and computer. Companies want to keep that mem, the cpu, the actual instruction parts of your computer as possible. So they're baking it onto the board. They're creating special rings of logic where something can pop out of the CPU and into memory and back into a CPU thread so quickly that they're getting even more performance out of it than they would if they had it on a separate stick further away from the cpu. So they prefer that higher bandwidth, that faster access if they can. But yes, the solid state drives, now they're quick enough that you could have a cache on there. Some people will talk about doing. Back in the day, you could, you could speed up your whole computer by using your DRAM as like a temporary storage for your programs. You'd like, just throw your game in there or something like that so you could load it even faster than storage. Storage is fast enough now. We don't think about that much anymore.
- [11:20]
A
Okay, so ram. Actually, I did, I did, I would say like four minutes of research. And it turns out RAM has been around much longer than I realized. Like, the way I've always thought of RAM is basically like you have a hard drive that is sort of spinning and trying to find things in a linear fashion in a row. And actually storage doesn't quite work that way. So everything is all over the place. So your drive has to constantly spin around and look for stuff, and that takes a while. And RAM just can look for it all at the same time. And that is much faster. This is basically like a thing that I learned listening to music on knockoff ipods in the early 2000s. That's the extent of my RAM education. But this is a technology that's been around since the middle of the 20th century. Right. Where did this actually come from in the first place?
- [12:08]
C
I don't remember who quite did it first, but when you're looking at early computing, people will talk about the Babbage engines and they would have like rods that would be in a certain position and that would tell you it would store the previous calculation. And then some of the early computers would have punch cards or tape that would have, you know, punches through it and you would read that there would be various switches that you would flip or gears turn to a certain position. At one point they had vacuum tubes where, you know, you've got this idea that a vacuum tube, when it's, when it's on, you know, can store a little bit of energy in there.
- [12:48]
A
Not to interrupt you, but every single one of those things sounds so cool and futuristic. Yeah, I'm like, that sounds sick. Let's do that. That's probably better. And then it's like, no, these were all the bad ideas we had before. Vacuum tubes are great.
- [12:59]
C
Everybody love six Brussels of decompression and tremendous amounts of energy required to do anything. Right? At one point they had CRTs where you would store the bits on the screen of the CRT as a charge. You've got your electron gun firing your charges at the screen. That creates your phosphorescent image that you see on a CRT tv. Well, why not store the memory there too? And then there was ring memory, the core ring memory at some point where you'd have wires kind of all wrapped in patterns. And inside of each of those places you could store the charge for longer than you could in some of the previous technologies. You can store many of them in a lattice. And I've gone to the Computer History Museum in Mountain View and you can see hand woven memory as this thing. It looks like a bunch of cords knotted together. It's fascinating stuff.
- [13:47]
A
I still bring all of that back. I'm looking at my desk now being like my computer needs like several more vacuum tubes and something very cool will be happening. But as far as I can tell, RAM is one of those things that it's like there are a bunch of things in sort of the history of computing where a thing was created and then everybody kind of looked at it and went like, oh yeah, this is better. And it just immediately took off and won forever. And I feel like we're in year like 60 of RAM, which suggests that everybody just decided this was the right idea.
- [14:16]
C
The thing to know now, I think is that they are computer Chips, the RAM bits on your memory stick that goes in your computer, it is created basically the same way. It is the same material. And so when we're talking about the shortages that we're seeing now, these shortages may intersect with other places because what you're doing fundamentally is you've got a silicon wafer, you're taking a single big crystal of silicon. You are cutting it using like a diamond wire or a diamond, you know, edged rotary saw into incredibly thin layers, like a quarter of an inch thick. And these are being patterned with chemicals and ultraviolet light and then etched with acid or plasma into these little chips. And it's the same process used for processors, for computer processors, also used for memory. The same kinds of wafers we're talking about, about 13 inches wide are used for computer chips, they're used for memory chips. And so we've, we've digitized all this and made this much more efficient and small, but now it's kind of the same thing.
- [15:29]
A
So the gap between all of these different things has actually like shrunk a lot over time. So when we talk about like storage and processors and ram, like the distinction, it seems like matters a lot less than it used to. I hadn't really thought about that until just now, but it does. It is all becoming sort of a single system in the way that it was. Not like, I still think of RAM as like a giant stick of thing that I shove into my desktop computer. And that is in almost every case, not what RAM is anymore.
- [15:56]
D
Well, it's also not where the volume is, right? Like phones. There are more phones than laptops right there. There are more phones than cloud servers. And RAM is in all the phones. It's soldered into the phones. Those aren't sticks. You can have a lot of feelings about that. Maybe we should all put dim sticks into our phones. That'd be kind of cool. But you know, if you listen to the Apple earnings call or Dell's earnings call, all the analysts are always talking about fluctuations in DRAM pricing because there's only so many vendors of ram, which I'm sure we'll get into. They are kind of a commodity. It moves up and down like oil. And if the price of RAM is high, Apple's margins are lower because they, you know, the price of the iPhone does not fluctuate with the price of ram. So this is in the background of all these businesses is, oh, RAM is a commodity. And the price moves up and down in a way that, Sean, I don't think, like the price of M4 chips moves up and down. Right. In the same way that RAM is a commodity and the price moves up and down.
- [16:55]
C
Yeah. Now, if you want, if you want intel or Apple or I guess Samsung, if you want TSMC to create you a chip, you talk about like, do they have the capacity at their fabs to make as many chips as you want, and if they do, you pay, you know, this much, and if they don't, you, you try to pay more so you bump somebody else off of the fab or, or, or get capacity that's coming later in the year or whatever it is. With dram, there are so many wafers that are going to these, these, these three, basically three companies that are creating this stuff and they're, you know, they don't want to create too much of it, but they know generally how much many PCs and how many phones are going to be sold in a given year and they just produce a lot of it and everybody buys it off them. Yeah.
- [17:39]
A
RAM is one of the things. It is like so deeply unsexy, but it is increasingly everywhere. And right. I feel like, is the. We've been on this march and we talk about this a lot on this show, that like everything has become a computer. Right. That my dishwasher has a computer chip that probably resembles a smartphone chip in it, and so does basically every other piece of electronics in our world. Has RAM scaled the same way? Like, is it. Can you sort of spin around in your chair and probably find 100 things with RAM inside at the moment?
- [18:10]
D
I can tell you this frame TV doesn't have enough RAM in it. I could tell you that right now.
- [18:16]
A
Yeah. I can also tell you that having watched you go through the interface to select this photo, something I didn't know.
- [18:23]
C
Until we were doing this report the other day is that the solid state drives have RAM in them too. I figured you didn't need very much of that, but, you know, yeah, there's some cache in there. There's some cache in there for the solid state drive so it can run faster.
- [18:34]
A
RAM is now, it's just utterly ubiquitous that way. Right. Like, is it, Is it. It's just the same marches, sort of. As smartphones became this gigantic, massive industry, smartphone chips went everywhere and RAM has done the same thing to the point.
- [18:49]
C
Where it is baked into many of those chips. Your phone doesn't necessarily have an application processor soldered here and a RAM soldered there. It has the RAM packaged into the processor in many cases.
- [19:03]
D
That's what I call them socs. Right. That's you know, the chip people, the chip fluencers, get very mad when you call them CPUs. They're SOCs. They're systems on a chip. In the part of that system is the memory is the Ram. So I bet your dishwasher actually has an SoC. By the way, I've been watching you through your dishwasher for weeks now. You're loading it all wrong.
- [19:21]
A
My wife would tell you that's absolutely correct.
- [19:24]
D
But your dishwasher probably has an embedded SOC that is a complete system containing ram. That is, you know, whoever just bought off the shelf and shoved in there.
- [19:33]
A
Yeah. So, okay, so Sean, tell us about 20, 25. What the hell happened this year that I have to think about RAM in my day to day life.
- [19:41]
C
Now, I don't know exactly how it began, but I don't remember anybody forecasting that all of a sudden AI data centers would be buying up so much of the world's RAM that the rest of us would be scrounging for leftovers. I do remember them saying that DDR4, an outgoing standard, would be a little bit pricier because they were trying to end of life. It, it's like, okay, yeah, there's less of that to go around. So if you got an older computer, you might pay a little more to get your older RAM.
- [20:09]
A
Double data rate 4 synchronous dynamic random access memory, DDR4, SDRAM. Good Lord. Okay, so they're trying to phase out DDR4, presumably to phase in DDR5.
- [20:21]
C
They're trying, yeah, and we got, and we got DDR5. It's everywhere. If you're buying a new computer, you're going to buy that instead of DDR4. You stick it in, you know, your graphics card. So you stick a graphics version of it in there. There's even six for graphics cards and seven. But moving on from that, yes, the general standard, DDR5. And it goes in all the computers and it goes in the kind of computer you'd buy for yourself. And it goes in the laptop where you don't even think about it. And it also goes in the AI data center. And the AI data centers want as much of it as they can possibly get. Just like they want as many GPUs as they can possibly get.
- [20:56]
D
Can I actually ask about that? So I heard there's a data center RAM crunch. And I thought, oh, it's because Nvidia is selling a lot of GPUs. Data centers, don't the GPUs use a different RAM?
- [21:06]
C
Don't they use VRAM, they have their video RAM. It's on a slightly different standard, but fundamentally all the memory is coming from just a few companies who have a limited supply of wafers that they are cutting out of ingots of silicon and doing their chemical treatment processes on. And so everybody who's making these computer chips is going to be fighting for some of that supply. And the RAM makers in particular control. These three companies control 93% of the world's supply of RAM.
- [21:37]
A
Who are those three companies?
- [21:39]
C
This is Micron, SK Hynix and Samsung.
- [21:43]
A
Oh, sure. You know, the first. The first three companies I think of when I think of tech companies, SK.
- [21:48]
D
Hynix runs the world. You shut your mouth when you talk about SK Hynix.
- [21:51]
C
SK Hynix is the biggest of the three.
- [21:53]
B
Okay.
- [21:54]
C
And Samsung also. These three companies. 93% of the other companies that there are. There is only one other company that has even 5%. The rest have 1% or below.
- [22:06]
A
Wow. Okay.
- [22:07]
C
And so if you want RAM, you're going to one of these three of these three. One of them just said, we're done with consumer business. We're just going to focus on enterprise now. We're going to focus on data centers because that's where the money is. And of the other two, Samsung and SK Hynix, they may have contributed as much as 40% of the world's entire supply of memory to a single project at OpenAI going on right now to create a mass set of AI infrastructure there.
- [22:37]
F
Wow.
- [22:37]
D
So just to be clear, it's not the AI. Data centers need more regular RAM for the CPUs or more VRAM for the GPUs. It's just like RAM in general is needed for these buildouts.
- [22:52]
C
It's all of it. I mean, you'll want. If you're building your computer inside this data center, it's going to need a gpu. And that GPU is going to have memory. It's going to need a CPU with normal memory to talk to. And so it'll have some normal memory in it. It's going to need SSDs which are going to have a bit of memory on those. It's going to need all kinds of networking, fancy networking switches for their infin link or whatever. There's memory inside those. All of these things are consuming memory. And so much of it is going there now that companies are saying if this is the way things are going, we don't need to worry about consumers so much. We're going to have plenty of Money from where the profits lie. And we saw this already happen to some degree in GPUs. I mean, we saw AMD say we're not going to make a high end graphics card for gamers this year because we want to put our resources toward AI GPUs. Instead we saw Nvidia not say that, but then kind of do that anyway. Although they do have plenty of consumer GPUs, it's been harder to get them, they've been pricier. They know that they can charge gamers more because if gamers don't buy as much, that's not a problem for them. They're making $30,000 off of one of the AI GPUs. They don't need to make a thousand off of a gamer.
- [24:08]
A
So if I'm one of these three companies, basically I've looked at the market and I'm saying, okay, I can sell to like Dell who wants to make some laptops, and I can sell to like nothing who wants to make some smartphones, or I can sell essentially to like the five richest companies on earth who are all trying to build AI data centers and want to buy from me in unbelievable volume. And like any reasonable shareholder value maximizing company is going to pick the thing where OpenAI just writes you the largest check you've ever seen in your life.
- [24:37]
C
Antonio ran into Dell's COO the other day and got a whole like long, mostly empty spiel about how Dell has all these relationships with these suppliers and they go back years and yada yada. And it kind of boiled down to Dell may need to kind of tap the companies on the shoulder and be like, remember me, please give me some memory. We have these deals and we would like them to be fulfilled. He said that supply, that getting supply is their focus this year. Like they don't, they haven't already negotiated the supply. Apparently not to the degree that they need.
- [25:15]
A
So how, how wild has this gotten you? You saw and wrote about some like, I would say sincerely nuts outcomes of this RAM shortage. What have you seen out there in the world?
- [25:27]
C
I think the one that's gotten the most reach is that some stores are now selling it like lobster, in that they're selling it like the catch of the day market price. You have to walk into the store and ask, hey, how much is RAM today? Or you might say, you know, I'd like to order the ram. Can you, can you, can you tell me how much that'll be on my bill?
- [25:48]
A
Swirl your wine and I say, I'll have the ram. Thank you.
- [25:54]
C
It's. It's. It's the point where several of our Verge editors, myself, Richard Lawler, we've gone back and looked at how much we paid for RAM three months ago, six months ago, a year ago, and we've seen that the number is quadruple what it was.
- [26:09]
A
Good Lord.
- [26:11]
C
Sticks. Stick. You know, I see Reddit stories about somebody talking about how they're willing to trade their pricey GPU for a couple sticks of ram. And a year ago, that would have been unthinkable. The GPU was the item that everybody wanted. You'd pay 1000 1500, $2000 for this GPU. It was that hard to get a good GPU. And now they're saying, oh, well, maybe I should get $1,500 sticks of ram instead for my computer, so my computer has enough memory to open my hundred Chrome tabs.
- [26:43]
A
It feels like we keep doing this. It wasn't that long ago that all the bitcoin miners made it impossible to get GPUs, and then all the AI companies made it impossible to get GPUs. And then right before that, we had a chip shortage because everybody was in the pandemic buying, like, pelotons and stuff. And so, like, how is this just a thing that keeps happening? Like, at some point, these South Korean companies should just make more ram. Why is this so complicated?
- [27:10]
C
Some of it is that it takes a long time to spin up new facilities like these. There are only so many ultraviolet extreme ultraviolet lithography machines in the world. We're talking about dozens, not hundreds.
- [27:21]
A
Is that the thing that's made by that one, like, secretly powerful Dutch company that no one knows about, but it kind of runs the world?
- [27:27]
C
Yep. If you want to make the cutting edge chips, this is the only company that will sell you the machine. You can't buy them if you're in China because they think that's a security risk. They have to fly parts all over the world for this. It takes six months just to take the parts for this machine and construct them into the machine. There are only dozens of them in the world. It's wild. That's for the cutting edge stuff. So there are not. Not everything you're making is going to require the cutting edge machine.
- [27:56]
D
But still, I just have to say, I just googled how much does DRAM manufacturing need EUV Extreme Ultra Lithography? And there's answers. But Google's AI overview is like, yes, it absolutely needs EUV lithography.
- [28:09]
A
Give it to me now.
- [28:11]
D
Like, like this, this robot is like I need it, I need it right this second. It's very funny. It's on a curve up. It is not the dominant like Micron only just adopted euv.
- [28:23]
C
Wow.
- [28:23]
D
And Micron is building out more fabs. But to Sean's point the fabs take a long time to come online.
- [28:28]
A
Yeah.
- [28:29]
C
There are so few companies that control this. There's the ASML for your EVUV machines. The only company that's doing that. There's just the three memory manufacturers. When you come to the other side of the chip equation, tsmc, you know the big Taiwanese fab company they I think the New York Times said that they too single company was producing something like 93% of the world's cutting edge chips at that at, at the time that New York Times report came out like a couple years ago. So there's just it power is in the hands of very few. And those companies, they don't want to risk creating oversupply. In some cases it can be, you know, fundamentally detrimental to their bottom line. They could lose money if they do too much oversupply. I think Micron in particular massive profits now but it wasn't always there was a.12 years ago that was losing a bit of money for the other companies they're like well we'd like to maintain long term profits, thank you very much. We're going to feast on this as long as we can and not create more than we think we can sell and profit from. So they're happy to see their average selling prices go up.
- [29:39]
A
They're looking at Nvidia becoming a $4 trillion company and going well that went well. Maybe we also. Yan. Yeah. So that makes me feel like this, this thing is going to get worse and not better. That it's not like I assume there's, there's not some like next technology that is waiting to be implemented here. Like the world runs on RAM in a real way. Are, are we stuck in this sort of chaotic shortage for the foreseeable future or like I guess if the AI bubble pops and all these companies go out of business, things change. But short of that, how does this get better?
- [30:16]
C
One thought is that the AI bubble pops. Another thought is that maybe just a little thought.
- [30:21]
A
Yeah.
- [30:21]
D
The world economy crashes because the robot wants to bang you your laptop.
- [30:26]
A
It'll be easier to get a laptop though.
- [30:27]
D
That's true. But at least Kevin Roosevelt find true love. Shout out to K. Shout out to kr. My boy.
- [30:32]
A
Happy holidays Kevin.
- [30:36]
C
Another thought, another thought is that Maybe, maybe they will come up with a version of manufacturing RAM that works better for the AI companies and they'll switch to that. One theory going around right now is that today we make little chips on wafers and we, you know, we cut out those little chips and we turn them into CPUs and memory modules and so on. But maybe in the future we will make giant whole wafer chips. And because those whole wafer chips will be fundamentally much more efficient at moving data within them, maybe higher speeds for memory and so on, maybe the AI companies will switch to those. And so if the. I don't know where the bottleneck is, but if the bottleneck is not in slicing up wafers, but actually in printing out chips, maybe that will make the AI companies more happy.
- [31:30]
D
Wait, I just. The idea is that, let's just say Sam Altman. Sam Altman is going to buy 18 inch round memory chips.
- [31:36]
C
That's basically that. Yeah, basically.
- [31:38]
D
I mean, that's a perfect OpenAI idea. Like we will get so good at manufacturing memory that the yield will be perfect on an 18 inch round silicon wafer. Sure.
- [31:47]
A
This is how we get back to vacuum tubes. Do you know how cool that would look sitting on your desk right now?
- [31:53]
D
You're just like, the RAM guys are just going up the stairs with their huge waivers. Like instead of DIMMs, you're like feeding them in like giant CDs.
- [32:02]
B
Hell yeah.
- [32:03]
D
Sure, sure. I buy it.
- [32:04]
B
I love it.
- [32:04]
D
There's also like other stuff. There's like packaging stuff, there's like 3D packaging. Like all the same chip manufacturing stuff is in RAM now.
- [32:11]
F
Right.
- [32:12]
C
There's lots of thoughts about packaging things differently. But again, I don't know where the bottleneck is. If it's. There's only so many wafers to go around, then we're kind of in trouble if the AI continues to eat all the wafers. You know, if it's somewhere else, if it's in, you know, if we just package this differently, if we just manufacture it differently, and if that manufacturing can be spun out into a different supply chain that doesn't impact all the consumer stuff, that might be nice. Then again, I don't know. Right now we're talking about. The AI companies are talking a lot about HBM memory, High Bandwidth memory. It's all it stands for, hbm. And so if you're going to use this separate kind of HBM to get more bandwidth in your AI data center, we're thinking that's going to steal RAM production away from the consumer stuff, from the DDR, it takes up three times as much space according to one of the analysts that Emma talked to for her RAM report. And so if that steals away, that's not good. But maybe we all start using HBM. There was an AMD set of graphics cards in 2015. I used one in a PC that used HBM memory and it was great. It made for a very compact, efficient graphics card. Perhaps all of that'll trickle down and after AI isn't heating up so much of things, maybe everything gets better in consumer as a result. I don't know. Let's hope.
- [33:34]
D
All right, so the bubble pops, the AI buildout works out or we all transition to a new standard that's unproven. These are good outcomes.
- [33:41]
A
Perfect, we're doing great.
- [33:42]
D
And then you can finally buy a PS5. Like that's how it goes.
- [33:45]
A
Sure, yeah, sounds great. All right, we're going to come back to this thread of the story, but first we're going to take a break and then we're going to come back and we're going to play some RAM games, which I am told are a thing that exists and it's going to happen.
- [33:58]
D
We'll be right back.
- [34:00]
E
Support for the show comes from Built Rewards. Renters out there. Listen up. It's about time you earn something back on your biggest monthly expense. I mean, you can earn loyalty points on pretty much anything else these days, so why not on your rent? BILT is the loyalty program for renters that rewards you monthly with points and exclusive benefits in your neighborhood. With bilt, every rent payment earns you points that can be used towards flights, hotels, Lyft rides, Amazon.com purchases, and so much more. And starting this month, BILT members can earn points on mortgage payments for the first time. That means you'll be able to get rewarded wherever you live. Plus, using Bilt unlocks exclusive benefits on more than 45,000 restaurants, fitness studios, pharmacies and other neighborhood partners. Join the loyalty program for renters@joinbuilt.com Verge that's J O I N B I L T.com Verge make sure to use our URL so they know we sent you. Support for this show comes from wix. When you're building a website, finding a creative flow feels great. You're brainstorming aesthetics, fun functionality, and spitballing a growing list of ideas. But then you think, okay, how do I actually make this happen? If that's where you get stuck, then look no further than wix. You can make a great looking website with WIX and You can do it your way. Whether you want AI to jump in or prefer to do things yourself, wix is packed with actually useful AI features and agents built specifically for SMB so you can grow your business without burning out. Get a custom ready to use website in minutes with wix's AI website builder or choose from designer made templates and enjoy easy fuss. Free domain registration, web hosting included. 280 million businesses around the world rely on WIX for their websites. Because with wix you can own your individuality. Create freely and scale fearlessly. Ready to create your website? Go to wix.com that's wix.com.
- [36:16]
A
It'S time to level the up. I'm Robin Archison and I light fires. I'm an executive founder, bestselling author, ultra marathoner, mother, proud Latina, and I'm not done yet. Announcing Project Swagger, my new weekly podcast, your transformation toolkit. I'm gonna cut through the noise and give you actionable takeaways each week in under 30 minutes. Elevate your hustle with routines, strategies and mindset shifts that I have pressure tested. I. I have burnt down this Beyonce candle, like all the way to the bottom. We have been trying to manifest. Carbs are not the enemy. I probably have a piece of bread or a bagel with me at all times and I am not exaggerating. Tune in on February 24 for episode one, building the skill of Self Talk. This is the foundation. Follow Project Swagger wherever you get your podcasts. Let's go.
- [37:12]
B
Before Minnesota, Illinois basically wrote a playbook on how to fight back against Trump's ICE crackdown. Governor J.B. pritzker told everyone in the state to take action when ICE came to town. Pull out your phones, film everything.
- [37:26]
D
They're shooting moms in the face.
- [37:27]
F
Yeah.
- [37:28]
D
So peaceful protest seems like the least you could do and what we should be encouraging people to do. They've shot somebody here in Chicago five times for just observing from her car.
- [37:42]
B
Illinois created an accountability commission, took ICE agents to court, and when Trump sent.
- [37:47]
F
In the National Guard, they blocked them from the streets and they won.
- [37:50]
B
A model for Trump resistance on the state level today explained drops every weekday and now Saturdays, too. All right, we're back.
- [38:04]
A
It's time for some RAM Games here on the Holiday Spectacular.
- [38:09]
D
What is this? What is happening to me?
- [38:11]
A
Ram games. Nilhai, get ready. This is what we do here. All right, Travis Larchuk, our producer, is here. Hi, Travis.
- [38:19]
F
Hello, everybody.
- [38:21]
A
Travis, you are in charge and you have been let loose with several AI tools that make Me nervous? What do you. What? You're in charge. What do you have for us?
- [38:30]
F
I have two games for you today. The first game, Nilay, earlier you said this was a spectacular. Not an acronym, Tacular.
- [38:40]
C
You were wrong.
- [38:42]
F
Because this first game, as we all know, nothing tech people love more than jargony acronyms and initialisms. And as you know, RAM stands for Random Access Memory in this game. I will give you an initialism for one point. Tell me, does this initialism relate to RAM in any way? Or is it some random other thing that I stuck in there?
- [39:09]
D
I don't like that Sean is laughing in a way that suggests he's already won. I have to be really honest about this. Like, I don't. I don't like this confident cackle that's coming out of my boy, Sean Hollister.
- [39:19]
C
Look at.
- [39:20]
D
He's nodding.
- [39:21]
C
I gave you all of everything you need to win.
- [39:24]
A
Yeah. This is a question of whether Nilai and I were listening.
- [39:28]
D
Are you listening? The initials. All right, do we buzz in? How does this work?
- [39:32]
F
We will go in order. That's one point. Now, if the initialism turns out to indeed be related to ram, there are bonus points available if you can tell me what it stands for. I will give you one point if you get any of the words correct. I will give you two points if you get it all correct.
- [39:49]
B
All right?
- [39:50]
A
Okay.
- [39:50]
F
All right. Ready to rock?
- [39:52]
A
Let's do it.
- [39:53]
F
Okay. Let us randomly select the order of our game.
- [39:57]
C
Oh, my God.
- [39:58]
F
And Sean is up first.
- [39:59]
C
Yes.
- [40:00]
F
I vibe coded this in case anyone's watching this interface.
- [40:04]
D
I cannot wait for this thing to start hallucinating.
- [40:09]
F
Your first initialism or acronym is cas. Or maybe cas.
- [40:17]
C
This is ram.
- [40:19]
F
It is indeed RAM for one point. Do you know what it stands for?
- [40:25]
C
Shit. What's the initialism? It has to do with latency. But why can't I remember any of.
- [40:32]
D
The words in it?
- [40:35]
C
No, I don't know what CAS stands for.
- [40:38]
F
All right. CAS stands for Column Access Strobe.
- [40:43]
C
Oh, my goodness.
- [40:43]
F
As in Column Access Strobe latency.
- [40:46]
D
This is when RAM was made of, like, weird rods. Right.
- [40:51]
A
Column Access strobe is a vacuum tube.
- [40:53]
D
Yeah, you just pull the rods as fast as you can. You got strobe the rods, boy. That's what they used to say, strobe those rods.
- [41:03]
F
Sean, that is good for one point.
- [41:07]
C
Oh.
- [41:10]
F
Nilay, I'm so sorry.
- [41:12]
D
What are we doing here?
- [41:14]
F
Go ahead. I have good news for you. Your initialism is DDR.
- [41:19]
D
Oh, that's great. It's dance. Dance Revolution. No, this is ram. Double data rate. Nailed it. Two points.
- [41:25]
F
It is RAM in the lead. Take that. Double data rate. Three points. All three points.
- [41:33]
A
David, Just playing to Nilai here. I see. That's right.
- [41:36]
F
Your initialism or acronym is unr. Unr.
- [41:43]
D
I definitely paid Travis under the table.
- [41:47]
A
I'm gonna say not ram.
- [41:50]
F
You are correct. It stands for the University of Nevada, Reno.
- [41:59]
A
To be fair, I think you should have. Let me guess. To get two more points.
- [42:04]
F
All right, Sean, we are back to you.
- [42:07]
C
Ready?
- [42:07]
F
You have UMA.
- [42:10]
C
That's RAM.
- [42:11]
F
Correct for one point.
- [42:14]
C
Universal.
- [42:16]
A
It's also UMass. Amherst. Like just to.
- [42:18]
D
Just to say it's all universities. All the not rams are universities.
- [42:23]
C
I don't know the M. Crap. Universal something. Access, I think.
- [42:26]
D
How do you not know the M? Just wait, wait, wait, wait. I'm gonna think about what? Do you want to take a guess.
- [42:32]
F
What the M stands for?
- [42:33]
D
Just like you can work.
- [42:35]
C
Okay, let's. Let's try memory. I don't know.
- [42:38]
F
Okay. You are correct about the M. The M is memory. It is unified by access.
- [42:43]
B
Unified.
- [42:44]
A
You were close.
- [42:45]
D
Wait, does he get two points because he got two words? No, it's one point. It's not one point per word.
- [42:50]
A
You get two points total. Okay, yeah.
- [42:53]
F
One point if you get at least one word. Two extra points if you get all the words.
- [42:57]
D
Got it.
- [42:58]
F
This is when CPU and GPU share the same pool of RAM. Okay, Nilai. D, O, T, A, perhaps Dota.
- [43:10]
D
I'm gonna say not RAM.
- [43:11]
A
That's a game that uses a lot of RAM.
- [43:13]
D
Yeah, I'm just thinking of Vlad Silo.
- [43:14]
B
Correct.
- [43:14]
D
Dota 2.
- [43:15]
F
That is Defense of the Ancients. An esports game. David.
- [43:19]
A
All right, hit me.
- [43:20]
F
D, I, M, M Dim.
- [43:23]
A
It's a RAM thing.
- [43:24]
F
Correct. This has come up already. Do you know what it stands for?
- [43:27]
A
This has come up already. That seems bad. Data.
- [43:33]
D
David, is memory Master.
- [43:40]
A
Data. Intra machine memory.
- [43:43]
F
You got memory.
- [43:45]
A
I get a point.
- [43:46]
F
It is dual inline memory module.
- [43:49]
A
Wait, to be fair, technically, I didn't get memory.
- [43:51]
D
Yeah. Do you have to get one of the words right? In the right.
- [43:54]
A
Yeah, I had it on the wrong M. I think I don't get that point.
- [43:57]
F
The wrong M. All right.
- [43:58]
C
I'm actually willing to say, what an upstanding guy.
- [44:02]
A
No, because otherwise you just get to say memory every time and you probably get a chance.
- [44:05]
D
No, I understand how to score points, David. I'm kind of willing to say, Sean.
- [44:09]
A
You were making a face like you would have gotten this one. Would you have gotten this one?
- [44:12]
D
I would have gotten this one.
- [44:12]
C
100%. 100%.
- [44:14]
D
There was a big moment when I worked in the computer store in, like, high school where Apple switched from Sims to dims. And it was like all the talk of Macworld magazine in, like, 1998. That's why I remember. This is just PTSD.
- [44:26]
A
You were pretty cool in high school.
- [44:27]
D
I was shit.
- [44:33]
F
All right, at the end of the second round, our scores are. Sean, 3, Nilai, 4, David, 2.
- [44:40]
A
Nilai with the easiest three points anyone's ever gotten.
- [44:43]
D
I would have gotten dim.
- [44:44]
A
You would have. That's fair.
- [44:45]
F
All right, all right. This is the last round. Sean, ctu.
- [44:51]
C
Not a RAM thing.
- [44:52]
F
That is correct. That is the Counter terrorist unit from TVS 24. All right, Nilai, ECC.
- [45:06]
D
That is RAM. That is error correction circuit.
- [45:13]
F
So close. Error correcting code.
- [45:16]
D
I knew what the first two were.
- [45:17]
B
That was good.
- [45:18]
A
All right.
- [45:19]
F
This allows RAM to catch errors and is generally used for mission critical systems and not in consumer PCs. All right, David, mathematically, no way for you to win, but let's find out what happens. That's David's bm.
- [45:34]
D
Hurts.
- [45:35]
A
High bandwidth memory.
- [45:36]
F
Hey. Correct.
- [45:38]
A
Thank you, Sean Hollister. Thank you, Sam Altman, for your giant wafers.
- [45:42]
C
We are playing the games first. Next time. Next time.
- [45:44]
D
We're playing the games first.
- [45:45]
C
Explanation comes after.
- [45:47]
F
Somehow we end the game with Sean with four points, David with five, and Nilai with six. Congratulations.
- [45:53]
A
Feels right. Sean clearly the least educated about RAM.
- [45:56]
D
Of the three of us.
- [45:59]
C
I take it on the chin.
- [46:02]
F
Okay, your next game is on screen. Now, as we all know, RAM is the most valuable resource in the world right now. It is Mad Max times, but for ram. So we are going to play a white elephant gift swap, where the object is to end the game with the most ram. Here is how it will work. I have nine secret gifts that are wrapped in Verge wallpaper. Each is a gadget containing an amount of ram. I will not reveal how much RAM until the game is over. Before anyone writes in, I know that not all RAM is created equal, but for the purpose of this game, we only care about quantity of RAM and not quality of RAM. So.
- [46:42]
A
So DDR4 versus 5.
- [46:45]
D
We're cool.
- [46:45]
F
Exactly. Doesn't matter. It's all worth the same.
- [46:47]
A
Because I heard we're phasing out DDR4. I don't know if you're. If you're taking that into account here.
- [46:51]
F
I heard that, and I have not taken it into account. All right, so on your turn, you will choose. This is like a white elephant Yankee swap. You will choose to either open a new gift, a random gift, or you can steal a gift from another player. If a gift gets stolen, the victim of the steal can then choose to either open a new gift or steal a gift from another player. Each gift can only be stolen one time per round. And if a new gift is open, the round is over and we go to the next person to start the next round. Once everyone has three gifts, we will play one more round. Because the person who opens the first gift has no choice during this game. In that round, your choice will either be to swap one of your gifts with somebody else's gift or to end the game. Okay. All right. So we will randomize our order to find out who will open the first.
- [47:51]
D
What if I don't want ram? What if I want the other stuff?
- [47:55]
F
Well, that is up to you, but.
- [47:58]
D
Are you looking to win in the one with my wife's family? Everybody wants the lotto tickets. That's the way it goes.
- [48:05]
F
All right, Nilai, you get the first pick. So it's just one through nine, and that's all you get.
- [48:09]
D
World of no sevens. I'm going eight.
- [48:11]
F
Eight gift. Eight is a. Whoops. Palma.
- [48:17]
A
Perfect. Congratulations, Nilai.
- [48:20]
B
Nailed it.
- [48:20]
D
You. Technically, that contains ram.
- [48:24]
F
It does contain ram. David, you can either steal that from Nilai or you can open a new gift.
- [48:29]
A
This is maybe the one advantage I have over both of you, as I know off the top of my head how much RAM the books Palm might have. I'm going to open gift number four.
- [48:38]
F
Gift number four. You have a shiny new cybertruck.
- [48:43]
A
I am less aware of the amount of RAM in that device.
- [48:46]
F
We are talking about the amount of RAM in the cybertruck's entertainment system specifically.
- [48:51]
D
Oh, my God.
- [48:53]
F
So, Sean, you can either steal Nilay's books, Palma David's cybertruck, or you could open a new gift.
- [49:00]
C
Sean, I don't want to end with a cybertruck. Also, Verge wallpaper number seven is delightful, so I'm taking that one.
- [49:07]
F
I love this purpley reddish blue wallpaper. And you got meta Ray Ban display.
- [49:15]
C
Oh.
- [49:15]
F
All right. Nilai, you can steal any of the gifts that are out there, or you can open a new gift. This is our second go round here. Round four.
- [49:24]
D
I like debating whether or not the cybertruck has more RAM than the boogs. Palma, It's a real toss up. I'm gonna. Let's open number three.
- [49:33]
F
All right, you're going to open gift number three. No stealing so far. That is a iPhone 13 mini rough, huh? The favorite phone of many a Verge podcaster and listener.
- [49:47]
C
Loved it. So good.
- [49:48]
F
Okay, David, you can steal any of the gifts that have been opened or open a new gift.
- [49:54]
A
So the dilemma here is, did Travis pick a bunch of things that don't have much ram, or have we just opened the things that don't have much ram?
- [50:02]
D
I was doing the same calculation.
- [50:05]
A
I am going to make a potentially reckless decision, and I'm going to steal the books. Palma.
- [50:09]
B
Yes.
- [50:10]
F
All right.
- [50:10]
A
Which I think has a non zero chance of having the most RAM of.
- [50:13]
B
Anything on the board.
- [50:14]
F
So far, David stole the Butz Palma from Nilay. Which means, Nilay, you can now make a steal yourself or you can open a new gift.
- [50:22]
D
I'm taking the cybertruck. Straight swap.
- [50:26]
F
Nilay steals the cybertruck from David. David, it is your turn to either steal or open a new one.
- [50:32]
A
Give me number two.
- [50:33]
F
All right, gift number two is.
- [50:35]
A
Nilai, do you agree that the cybertruck either has one or a thousand gigs of RAM and, like, nothing in between?
- [50:40]
B
Yeah.
- [50:41]
D
My bet. I was doing the same calculation as you, and I was like, I know that the puk's mama doesn't have a lot. I'm gonna roll the dice on the cybertruck. Also, it is, like, only fitting that I end up with the giant wiper in the end.
- [50:52]
A
Agreed.
- [50:52]
D
The switch 2 it is. This is a good get.
- [50:55]
B
Yeah.
- [50:55]
A
I'm gonna keep the Switch two. I feel good. I'm psyched.
- [50:57]
C
Are you gonna keep the Switch 2, though? I'm not sure if you are.
- [51:00]
F
It is Sean's turn.
- [51:02]
C
I am, in fact, taking The Nintendo Switch 2.
- [51:04]
F
The Nintendo Switch 2 has been stolen.
- [51:06]
D
What's the maximum number of steals?
- [51:08]
F
Each item can be stolen once per round.
- [51:10]
A
So, like, I can't take it back.
- [51:12]
F
Right.
- [51:13]
A
But I could, in theory, take the cybertruck. I'm gonna open number five.
- [51:17]
F
Number five is the humane AI Pin.
- [51:22]
A
This is so cruel. And also the fact that I'm gonna end up with the books, Palmer, and the humane AI Pin feels like I deserve this.
- [51:28]
D
You totally do.
- [51:29]
A
My 2025 suggests that I deserve this.
- [51:32]
D
Wait, so if there's not a maximum number of steals, the end game here gets very complicated.
- [51:38]
A
Yes.
- [51:38]
F
So, yeah, each. Each round. Each item can only be stolen one time in a round, but after that round, things are open to be stolen.
- [51:46]
C
Again, but then there's a swap at the end, so we can.
- [51:49]
F
There is a swap round at the end. Just to recap for the listeners, nilai has an iPhone, 13 mini, and a Cybertruck. David has a books palma and a humane AI Pen. And Sean has Meta Ray Ban display and Nintendo Switch 2. And it is Nilai's turn.
- [52:04]
A
Can I just pause, by the way, to say that so far, we have accidentally kind of all three picked our personalities. I was gonna say is the goal.
- [52:12]
C
To wind up with the most fitting item.
- [52:13]
D
It's just stuff we like.
- [52:15]
F
All right.
- [52:17]
D
The only correct move right now is to pick a new. So I'm taking number six.
- [52:21]
A
I think that's right.
- [52:22]
F
Number six. The game theory is afoot. A pixel 9A.
- [52:27]
D
You did pick stuff with not enough RAM.
- [52:29]
B
Yeah.
- [52:30]
A
One of these is gonna have.
- [52:32]
D
This game is gonna be made or broken on, like, 16 megabytes of RAM. I can see it.
- [52:37]
A
There's. There's going to be, like, a Mac studio at the end of this that just wins. And it's. None of the rest of it's going to matter.
- [52:44]
C
Holding out for Framework Desktop.
- [52:47]
B
That's what I do have.
- [52:48]
D
Two tiny little phones. It's adorable.
- [52:51]
A
And a cyber. I didn't say that was adorable.
- [52:57]
F
His two cell phones and a Cybertruck.
- [52:59]
A
Yo, that's our boy.
- [53:02]
F
Red flags all over the place.
- [53:05]
D
Have you heard of bitcoin?
- [53:09]
A
Give me the Switch two.
- [53:11]
F
All right. David has stolen the Switch two from Sean. Sean, you can steal or open a.
- [53:15]
C
New gift, but I can't steal the Switch 2. So I think I'm going with. Have you heard of the soundtrack Number nine?
- [53:23]
D
Everyone loves it.
- [53:25]
F
He's opening a new gift, and the gift is a Galaxy Z Fold 7.
- [53:30]
C
Now we're talking.
- [53:31]
A
That feels like it has a lot of ram.
- [53:33]
C
I know exactly how much.
- [53:35]
F
Sean, it is also your turn to start the ninth round.
- [53:39]
C
Mm.
- [53:42]
F
There's one gift left.
- [53:45]
C
I'm taking the Switch two back.
- [53:47]
F
All right. Sean takes back the Switch two.
- [53:50]
D
David, I can't take it because I'm still in the same round.
- [53:52]
A
Right, Right. But you're gonna get to go next. Got it.
- [53:55]
D
I see.
- [53:56]
A
All right, give me number one. Let's do this.
- [53:58]
F
All right. David is opening number one, and number one is an Apple Vision. Probably so.
- [54:09]
A
Wait. Okay. Serious question for the two of you looking at this. Do you have confidence about which one of these nine things has the most RAM in it?
- [54:17]
B
I'm gonna.
- [54:17]
C
No.
- [54:18]
D
I think it's a cybertruck.
- [54:20]
A
So I think there's a chance it's the cybertruck. I also think it's possible that I'm just dead wrong. But. Okay. I have the vision proof.
- [54:28]
B
Great.
- [54:28]
A
I have a books Palma. An AI Pin and a Vision. This sucks. I'm sorry, dude.
- [54:33]
F
This is how no one wants to.
- [54:34]
D
Go to your house.
- [54:40]
F
I want to try out the Apple Vision Pro.
- [54:42]
A
Yeah, Travis, you can come over. Everybody else is uninvited.
- [54:47]
D
Dead, glowing eyes. Inside it is a ram starved head set.
- [54:53]
F
All right, this is the 10th and final round. Nilay gets to start this round because he went first and didn't have a choice in the beginning. On your turn, you can either swap one of your gifts with one of anybody else's gift or just end the game. So Nilai, it is your choice.
- [55:11]
D
All right. I'm going to swap the 13 mini for the Switch 2. I feel good about that decision.
- [55:16]
F
Okay, Nilay is stealing the Switch 2 from Sean and giving Sean the iPhone 13 Mini. Sean, it is now your turn. You can swap or end the game. You have Meta Ray Ban display, Galaxy Z Fold 7 and the iPhone 13 mini.
- [55:33]
C
I am going to take the Vision Pro for the 13 mini.
- [55:39]
F
Okay, giving David the 13 mini.
- [55:42]
D
Finally, David has one useful device.
- [55:46]
F
David, same choice. Swap or end the game.
- [55:49]
D
End my pain.
- [55:50]
A
Now.
- [55:52]
D
There'S at least two one move on that board that I would make.
- [55:55]
A
There's two. But I'm ending the game. I'm dancing with the ones who brought me. Let's do this.
- [56:00]
D
Humane Pin.
- [56:02]
A
Guess what's going to happen when the Humane Pin turns out to have 32 gigs of RAM? Let's go.
- [56:07]
F
David hates this game. All right, we are ending the game with Nilay having the Nintendo Switch 2, a cyber truck and a Pixel 9A.
- [56:17]
D
Yeah.
- [56:19]
F
Party. David Books Palma. He made AI pin, iPhone 13 mini. Spoiler alert. Is going to lose the game.
- [56:26]
D
And Sean.
- [56:29]
F
Meta Ray Ban Display, Galaxy Z Fold 7 and an Apple Vision Pro. All right, we all have a phone.
- [56:35]
D
We all have a phone.
- [56:37]
F
You can all call each other. This is great. Let's find out how much RAM ring rang.
- [56:43]
D
David, how's that Humane? It's okay.
- [56:46]
A
I'm looking at you on my laser projector.
- [56:49]
F
David, how much RAM does the books Palma have?
- [56:51]
A
I believe it's 6 gigs.
- [56:52]
F
It is 6.
- [56:53]
C
Wow.
- [56:54]
F
Humane AI pin has 4.
- [56:57]
A
I was gonna guess 4. I'm gonna guess the 13 mini also has 4.
- [57:03]
F
For a total of 14.
- [57:05]
A
Tough day for Dan.
- [57:06]
F
All right, Nilai, we're gonna go to you next. Nintendo Switch 2. Does anyone know how much?
- [57:12]
D
My guess was 16. That was what was in my head.
- [57:16]
F
It is 12. All right, Cybertruck.
- [57:19]
D
It's either 0 or 1,000 for David 64.
- [57:23]
F
It has 16. 8 gigabytes.
- [57:29]
D
It has More inches of wiper than RAM, brother.
- [57:34]
F
And the Pixel 9 a hat. Eight gigabytes of RAM.
- [57:38]
D
I feel good about this.
- [57:38]
A
Nilai's pink phone and his truck have the same amount of ram.
- [57:43]
D
All of those things are like influencer gadgets. That's what I'm doing.
- [57:47]
F
All right. Nilay has vaulted into the lead with 28 gigs of RAM. Sean, I've got a good feeling about this. You have met a Ray Ban display. Does anyone know four?
- [57:57]
D
Not even a debate.
- [57:58]
F
Ooh, rough start.
- [58:01]
A
Big win for the AI pin that one.
- [58:04]
F
Galaxy Z Fold 7 has 12. And this is the base Z Fold 7. 12 gigabytes of RAM. All right. Sean has 14. Nilay has 28. The Apple Vision Pro has 16 gigabytes of RAM, giving Sean the win with 30 gigs of RAM.
- [58:24]
D
Wait, David, you would have won if you held on.
- [58:28]
F
If you.
- [58:29]
D
You had held on to two things, I think you would have won.
- [58:31]
A
Right? Sean took the Vision Pro from me.
- [58:33]
C
I wasn't given that option.
- [58:35]
D
Yeah, I gotcha.
- [58:37]
A
Damn. Well played, Sean. I also really respect that you had both the most and the least. You're a full spectrum RAM guy.
- [58:46]
D
I just. My bet was that Apple always starves everything for RAM, unless you pay an additional $800. So I was out on the Vision Pro.
- [58:53]
A
Yeah, but it's $3,500. I think we paid the additional $800.
- [58:56]
D
Fair enough. Fair enough. My assumption was that Samsung overstuffed the phone with RAM. Wait, can you buy a Z Fold 7 with additional RAM?
- [59:06]
F
You can. The highest end model, I believe has 16.
- [59:09]
D
That is ridiculous. You should not be able to change the amount of RAM in your phone.
- [59:13]
C
So I have a Palm Pilot here. Anyone want to guess how much RAM is in here?
- [59:18]
D
200.
- [59:18]
A
456 megabytes.
- [59:20]
D
Oh, it's a SIM.
- [59:22]
C
I have a stick. It has two megabytes.
- [59:25]
A
Two megabytes of RAM.
- [59:28]
B
Incredible.
- [59:29]
D
That's wonderful.
- [59:30]
A
Back when a megabyte was the size of your thumb. Those are the days. All right, Travis, this was delightful. Thank you. Congratulations to Sean. Congratulations to Nilay. I lost, you know, magnanimously as the host. Just so everyone's very clear on that, we need to take one more break, and then we're gonna come back and we're gonna get back into talking about the Ship Force. We'll be right back. Residence day savings are happening now at.
- [59:55]
E
The Home Depot with up to 40% off select appliances.
- [59:57]
B
Looking to upgrade your Fridge?
- [59:59]
A
Check out LG's newest model, serving up ice in all kinds of styles. Cubed Crushed craft ice and now new mini craft ice straight from the dispenser. From cold brew to fizzy favorites, these refrigerators will have you entertaining like a pro. Shop President's day savings and get up.
- [60:13]
B
To 40% off plus free delivery on.
- [60:15]
A
Select appliances like LG at the home Depot. Free delivery on appliance purchase of $1,498 or more. Offer valid February 5th through the 25th US only. See store online for details.
- [60:23]
F
We heard you.
- [60:24]
A
Nine years of bring back the snack.
- [60:26]
D
Wrap and you've won.
- [60:27]
A
But maybe you should have asked for more. Say hello to the Hot Honey Snack wrap. Now you've really won. Go to McDonald's and get it while you can. Taxact knows filing taxes can be confusing. So we have live experts on hand who can help answer any questions you may have. Questions like can I claim my SUV is my home office? If I answer work emails in my car? If I adopted 12 dogs this year, can I list them as dependents and am I doing this right or am I doing this very, very wrong? Our experts have the answers to those questions and many others. Tax Act. Let's get them over with.
- [61:08]
C
Your little one grew three inches overnight.
- [61:11]
E
Adorable.
- [61:12]
C
Also expensive. Sell their pint sized pieces on Depop and list them in minutes with no selling fees because somewhere a dad refuses to pay full price for the clothes his kids will outgrow tomorrow and he's ready to buy your son's entire wardrobe right now. Consider your future growth bird budget secured and start selling on Depop where taste recognizes taste. Payment processing fees and boosting fees still apply. See website for details.
- [61:41]
A
All right, we're back. Nilay is gone. Sean is still here. I rearranged during the break. If you're watching, please clap. And we have replaced Nili with somebody much better. Dylan Patel is here from Semianalysis. Dylan, welcome.
- [61:51]
B
Thank you for having me.
- [61:53]
A
Real quick, I assume you know most people the brand penetration of Semianalysis is like right up there with like Coca Cola and Walmart these days. But like what, what is semi analysis?
- [62:01]
B
What do you do? Yeah, so we're a firm that does AI infrastructure research, consulting, et cetera. Started five years ago as just a blog and me consulting on the side after I quit my job and it's just sort of blown up and up and up. And now I have a very famous blog as well as companies, 50 people all up and down the stack having worked on equipment that makes chips all the way up to worked at data centers, worked on models and things like that. So we do a Lot of consulting, research, etc.
- [62:32]
A
Okay, so I want to start with a question that is sort of about the timing of launching your company, but is also just about this market in general. What I'm wondering is like, is this just what the chip market is like? Did you pick just sort of the beginning of the most bonkers time in chips history to start doing this, or is this just how chips work?
- [62:53]
B
I mean, absolutely, today is the most exciting time ever. The ability for people to make the biggest change ever. Right? Like, I think it's the craziest time ever, but chips have absolutely just been bonkers right now. Parts of it are like, annoyingly slow and parts of it are not. But you have to think about the market for. Let's just take dram, right? Memory prices are soaring. People freaking hate DRAM right now, especially general public. But you know what's interesting is this market has existed this way for years. And if I go back a decade when I was more of a forum warrior on Reddit and stuff, every three to four years, people would be oscillating between, oh, memory's so cheap, I'm going to put way more memory than I need and people in my PC or whatever, and then oscillating to like, oh, my God, memory prices are so high, they're colluding against us. They hate the American public. They hate the public in general. I think semiconductors are a very tough business, right? Memory specifically went from over 30 companies on the leading edge to just three. Right. In a couple decades, right? Every boom and bust, more companies go bankrupt. The barrier to entry is so high, despite the fact that there is an oligopoly of sorts in many industries. In fact, most of the semiconductor industry is sort of leveled out to like, hey, there's a player with like 70% share, there's a player with like 25% share, and there's like this crappy company with 5% barely hanging on, right? And obviously things are dynamic because it's technology. So shift. Stuff does shift, but at the end of the day, there's. The barrier to entry is so high, the dynamics of the market are so intense that people are just going bankrupt left and right, that it's just, it is a bonkers industry. And I've always thought it was bonkers.
- [64:33]
C
You've had these booms and busts, you've had these. This cyclical nature of the industry and get smaller and fewer players. But it's never been like this before, right? I've had, I've seen low memory prices and I've seen high memory prices, but I haven't seen there are only three players and one of them is getting out of the consumer market. And the consumers, it's not just that they can't buy it because it's high prices, it's that the supply isn't even there because there's entire new industry eating up all this dram. It's never been like this before, right?
- [65:03]
B
I definitely agree. Like, you know, if we're just talking about memory, we have seen prices spike about this fast before. You know, over the last six months, like 50%, 75% increase in contract pricing, spot pricing has risen more. But even then, right, Like I think we've seen this level of price increase. The thing that's really going to break everyone's brain is the next six months it's also going to do that again, right? It's going to double again, right? In price or something like that. And that's what's going to really make people go crazy. And I think the funny thing is, or not even funny, I think just the sad thing is the reason we're in this position is because this is actually when we talk about the boom and bust of the memory industry. 2022 PC boom, all this huge amounts of memory was being deployed because everyone's just buying a new PC. And you know, the memory companies built a lot of capacity and then all of a sudden, oh, wait a second, everyone has a new PC. I keep my PC for seven years, right? Like, why do I need a new PC as an enterprise or as a person, you know? And the PC market falls off like a cliff. And these memory companies are now losing money every quarter, right? And so from 23, 24, 25, they were not adding new wafer production capacity, right? The only thing they were doing is as AI came into vogue, they were converting the regular memory can production to hbm, right? And so this has been the longest uninterrupted period of no capacity expansion as well, right? So, so this boom and bust and then also it was the worst bust, right? One of the worst buses like sustained multiple years of losses rather than just like, hey, a year or six months because the boom of PC and Covid was so large, right? So you kind of had this like, like coalescing of like, oh my God, everyone is, we just haven't added any capacity. So like, and then on the flip side, we have the craziest boom in human history, right? Like, you know, I think AI infrastructure is undoubtedly like, you know, the biggest boom in the Chip industry on a dollar basis.
- [66:59]
C
There's one way to look about this, where we're looking at like these companies, there's so few of them and they can make these massive, massive profits all of a sudden. But another way to look about this at this is a couple years ago, Micron was actually losing money. I don't mean like the revenue was, was going down. I mean like it was in the red, right? It was actually losing money.
- [67:19]
B
Yeah, for over a year too, Right. It's not even like a little bit of time, right. They had to make, they had to do layoffs. They like shut down factory lines, right. Like all of this sort of stuff, right. Samsung, Hynix and Micron, all three of them had to do this, right? So the sort of like, you know, I would say the more nuanced view is like, yes, this is crazy and the prices are soaring at the same time. They kind of had no choice to, you know, like, you know, it's like, who would have, have increased production after you just got out of the red, right. A year and a half ago, right. Like, that's just silly. But even that, no one could have predicted this.
- [67:51]
A
So that, that is my question though, because even if it's possible that there's nothing we could have done about it, even if it was inevitable, right, Which I think you just made a solid case that it is, shouldn't we all have seen this coming? Shouldn't, like all of those things that you just Described are like 3 year old trends now and if they're not that hard, at least in retrospect, to put next to each other and be like, oh, of course we were headed towards something totally disastrous here.
- [68:15]
B
The thing about this industry is that it's very easy to be bullish, but the most bullish person tends to go bankrupt.
- [68:24]
A
Right?
- [68:24]
B
And so that's the scary thing about this industry is if you overbuild the most, you end up going bankrupt. And that's how we've gone from 30 to 3. Right? And so as far as like, has anyone seen this? Like, of course I could show myself and say, look, we told all of our clients, we modeled all this production capacity, modeled all the AI production demands from a data center standpoint, from accelerator standpoint, flowed through how many wafers are required, blah, blah, blah, blah, blah. Yeah, obviously we've been writing about like, memory is going to go crazy for the last year and a half roughly. Right. And, but at the same time, it's like, you know, I'm not someone who managed a company you know, through the last 30 years where 30 competitors, where 27 competitors went bankrupt. And the scars are there, right? The industry has these scars. So there's also that front of like, is this a bubble? Well, I've seen 15 other bubbles and that's bankrupted all my competitors. Why would I, why would I go crazy this time?
- [69:13]
C
Was there anybody, though, who was like predicting this, who's like the, you know, the guy who saw the subprime mortgage crisis coming? Was there anything like that here? Because I feel like what I'd heard was DDR4 prices are going to rise because, you know, it's the old standard. They're not going to be producing as much of that anymore. You know, let's, let's get ready for DDR5. And then it felt like it was a surprise that DDR5 was in this huge crunch that everybody that these data centers wanted much of it. And even, even the companies that, that make, you know, their business on this, Dell, the COO of Dell is telling us, you know, hey, we need to. My, my focus now is to make sure we have supply, as if his company hadn't inked deals years ago to make sure that they have enough supply for their consumer products, for their commercial products.
- [69:57]
B
Yeah, yeah, so, so at the end of the day, like the memory market is like kind of a funny one in that like you can say oddr4 pricing is going to go up because it's going out of production. But the wafer production, on the wafer production level right there is sort of like the differences between DDR4.5 HBM. There are some process differences, but those process differences are not like, oh, it's a separate factory, it is the same factory. The time to change from one to the other is not that long. Right. In the case of DDR4 to 5, it is just a mask change, which can take hours. Now obviously the production timeline to make memory wafers takes months.
- [70:33]
A
Right?
- [70:33]
B
But the timelines to change stuff is quite quick. What it is is that they just didn't add wafer capacity right now whether that wafer capacity goes to DDR4, HBM or DDR5, there is some differences in sort of like what's the capabilities, right. Samsung's not that great at HBM, but they're great at DDR5. Right. Like sort of like, you know, there's these sort of mixes and discontinuities in the market, but at the end of the day the market is pretty efficient and the margins for each don't differ that much.
- [71:00]
A
Right.
- [71:01]
B
For DDR 4 versus 5 versus HBM. So what's happened is just that that memory AI is a sort of inelastic buyer, right? If you will, Nvidia buys, you know, let's take Blackwell, right? Blackwell. GPU, 192 gigabytes of memory. You know that GPU Nvidia sells for over $30,000, right? But at the same time that GPU costs them like 6, $7000 to make. And of that memory is half the cost, right? But if the memory cost goes up 20%, well they're still selling this chip for 30,000 plus, right? So you know, if it goes from 6,7000 to 8,000, their elasticity there is not crazy. You know who the elasticity is crazy for is the PC buyer, the mobile phone buyer. There the phone costs, you know, or the laptop cost, whatever it is. And if the price of memory doubles and it adds a hundred bucks to the cost of the device, then all of a sudden, you know, guess what, I might not buy, you know, I might not. Or it cost even more that, right? Like I might not, you know, that thousand dollars laptop, you know, for a little bit there were even thousand dollar laptops with 32 gigs of RAM and it was like glory glorious. Now, now like you, you know, it's kind of harder to find and I bet next year it's gonna be even tougher, right? And so there's sort of that, that interaction in the market in that like who is the elastic buyer. And unfortunately the consumer is the elastic buyer. You know, the production, you know, AI infrastructure spends going from like 10 billion to 100 billion to a trillion, right? We are scaling like crazy. And if it gets to next year, we look at the big top five hyperscalers they're going to spend or top six hyperscalers, Google, Microsoft, Amazon, Oracle, Core Weave, I'm missing one meta. You're looking at $500 billion of spend across them on AI infrastructure. Of that their spend on Nvidia is like 300 billion. And of Nvidia's 300 billion, that spend on memory, you know, after their margins, you know, Nvidia's largest supplier is SK Hynix, a memory vendor. It's not tsmc, right? So you float through and it's like, okay, actually next year the total volume that AI is consuming is ridiculous. And as far as like, hey, did anyone see this coming? There are a lot of people who are speculatively buying and things like that because the dynamics of the market are clear, right? So if you look at like there's I don't remember which PC maker it was, whether it was Dell or Lenovo or hpe. But one of them actually was super smartly about two quarters ago, secured a bunch of DRAM ahead of time and their stock, like there was a call on the earnings call where the investors were like, why did you, why did you enter all these long term supply agreements? Like, you know, you shouldn't do that. You should be just in time inventory, right? Like they were kind of like criticizing them like the Wall street person. And then now they're looking pretty smart because they secured memory at a lower price longer term. Right? But at the end of the day, right, like not all of the memory is. The other aspect of this is like, like this area for AI, the hbm, right? That manufacturing process is a little bit more complicated. The suppliers are even fewer, right. Samsung's not so good at it. So Nvidia locked in their supply for HBM totally next year. But this funniest thing is because SK Hynix is the best and because they have the best, HBM and Nvidia decided I'm going to sign a big contract with you at good margin. All of a sudden SK Hynix is going to be the least profitable of the big three memory vendors because they don't get to sell all their memory at the super inflated price that because the market is going crazy.
- [74:18]
A
So we've laid a lot of the blame for a lot of this kind of at the feet of AI data centers. But I think for people who don't follow this stuff all the time, these just seem like large buildings full of computers. And I think it's hard to put into real terms how big a project this actually is. Can you give us a sense of the scale of the build out of the AI data center right now if.
- [74:40]
B
We go back just a couple years ago, there were four or five thousand data centers in the world, but none of them were like, you know, super, super big, right? Like the biggest data center in the world three years ago is laughably small compared to what's being built now. Right. You know, the largest data center building in the world was, you know, maybe, you know, 50 megawatts, 75 megawatts, et cetera in that range. Google was building one that was like 100 megawatts. And I was like, whoa. Now when you look at like the size of like some of these campuses, right, it's like, oh, Stargate in abilene, Texas is 2 gigawatts and actually that one is kind of small relative to some of the stuff that's going up, which is five plus gigawatts. Right. You know, Meta's talked about like their Louisiana stuff. The, the, the concentration of like, value within a data center is absolutely absurd. Right? You think about it again, right? This area of Manhattan, it's 5 gigawatts, 5 gigawatts. You know, once Meta builds this entire, like, LA facility out 5 gigawatt size of Manhattan, it ends up being the entire cost of that they put into it was, was on the order of roughly $250 billion. Right? $250 billion for one site in Louisiana that they're building over, you know, a couple of years. Which, by the way, those GPUs are not going to be, you know, bleeding edge in five years, right? They're going to be, you know, they'll still be, they'll still be maybe useful in five years, but definitely in seven or eight years they won't be useful is just an insane factoid because when you think about it, it's like, like, you know, like Bill. 250, you know, it's, it's hard for humans to understand what 250 billion even means, right? That's, that's more money than the richest person in the world has. But again, like, we can't fathom what the richest person in the world even has. So it's, it's just like there's no way to contextualize this if that's where we're headed.
- [76:24]
A
A, a different market would suggest that the next thing that's going to happen is we're going to get a million chip startups who are going to come and try to make all of this stuff. And they're in the, like, your margin is my opportunity, right? That's, that's, that's this moment that we're in and every, every academic, everybody who's interested in chips, you should quit your job and go start a chip company. Like this is, this is the gold rush and we should be getting an entire new generation of folks coming out and saying, there's only three companies doing this. Let's get it back to 30, let's get it to 300, and there's going to be demand for this going forward. Is that happening? Are we headed down that road? And if not, why not?
- [77:05]
B
So there's two sort of layers in this stack, right? There is the design companies and then the manufacturing companies. As far as the design companies, there are some startups, many of them are going to fail, maybe some of them won't be successful.
- [77:17]
A
But that's true of all startups. Right. Like this is not necessarily a wildly different kind of thing to try and do.
- [77:23]
B
Yeah. In terms of design, like the barrier to entry is there, but it's not insurmountable. But it is a high barrier to entry. Right. You have to design your chip and then tape it out at TSMC and all this cost, you know, 50 million plus dollars and if there's any issue you got to do it again. Right. So there is a barrier to entry there. That barrier to entry is really low compared to manufacturing. Right. We are not going to get many new memory companies. There is one new memory company, but that's because it's like a national level like priority from China called cxmt. But they're like, like they have national government backing. We're not going to get a new memory company in the west, we just aren't. Right. Or in Taiwan or Japan or Singapore or anywhere else in the west, there's just not going to be a new memory company. Because the physical construction of a FAB is tens of billions of dollars. Plus all of this, know how that is super trade secreted and super confidential. And it's built upon year after year after year of tens of thousands of people's R and D who are all super, super smart like PhD level engineers and researchers just building and building and building and it's incrementally building on top of each other for context. Building a chip like a leading edge chip, whether it's memory or say a phone chip, phone CPU or phone memory or laptop memory or laptop chip, that, that takes over 5,000 process steps, right? Where you take a perfectly, you know, you take a silicon wafer, it's a perfect crystalline structure, then you bombard it with ions, you put, you deposit materials, ions is for doping like you do lithography to define where things are. And you build this chip layer by layer by layer by layer by layer and any defects ruin the chip.
- [79:09]
A
So then the other thing that might happen, it seems to me is like the other thing you would hope for at the end of this is this is where we get the next breakthrough. Right. And you've written about this. I found a blog post you wrote a couple years ago basically saying we're sort of at the end of DRAM anyway, that it can't scale the way that it used to, that other parts of the process are kind of leaving its abilities behind and that maybe what we need is something beyond dram. And again we have perfect market conditions for Somebody to show up and say, actually I found something that I can make cheaper or I can make with a different set of materials that is less, you know, reliant on one multimillion dollar machine from a Dutch company that has to fly in on a Boeing 737. Is that happening? Like, is there a science breakthrough here or something?
- [79:55]
B
It takes multiple planes, not just.
- [79:57]
A
You're right, it's multiple 737. So if we can just get it down to 1 737, we will have done something. Is there any inclination of that on the horizon here?
- [80:08]
B
Yeah. So the interesting thing is when you look at. I'll zoom out to just generally semiconductors, they've sort of had Moore's Law, right? And Moore's Law was actually originally about dram, but people just sort of prescribed it to logic as well.
- [80:21]
A
Is that true? I didn't actually know that.
- [80:23]
B
Yeah, Gordon Moore was talking about dram. That's what intel made at the time. They actually lost it. They were one of the companies, they didn't go bankrupt. They pivoted to processors. But they had to leave the DRAM market because they were getting, you know, the market was too tough anyways. The number of transistors doubling every two years. Again, in an area butchering the definition. But whatever happened for decades, and Moore's Law was made before Intel's 8086, the initial coining of it was actually when they were making only dram. It happened for decades. Decades, Decades. And costs kept going down. As far as the general DRAM vendors, they do have a roadmap to continue to go 10, 15% cost reductions a year for another decade. They have some, you know, they have some innovations that they want to do over the next five years, next seven years about changing the structure of DRAM and such. But the next big one is sort of 3D dram. Right. So interestingly enough, NAND flash already went through this. The cost reductions asymptoted out the physics just didn't let you shrink it smaller and smaller. So NAND ended up having this like big revolutionary update where instead of doing planar nand, they did 3D nan. And so they can make many layers at once.
- [81:32]
A
Right.
- [81:32]
B
And that ended up resuming the cost scaling curves for nan. And that's why, you know, SSD prices have fallen so much over the last decade and even the last five years. Yes, market dislocation right now, but generally prices have fallen a lot, whereas DRAM has not really fallen much. And so if they can figure this out, that's what the DRAM vendors are all in on. Right? But that means that they're sort of. They're all in on, but so is their supply chain. Right? Applied Materials supplied, you know, LAM Research, you know asml. All these companies that are in the equipment supply chain, all of their R and D effort is also focused for memory on helping these companies figure out 3D dram because that's what they need next to keep scaling and keep the market growing and keep the industry moving forward.
- [82:12]
C
You mentioned China, you mentioned China earlier that there is the state backing there for one company in particular to start making a dent. The big companies, you know, SK Hynix and Samsung and Micron, they're making a lot of this stuff in China too, right? What keeps China from taking off?
- [82:28]
B
So as far as what they're making in China, the vast majority of SK Hynix and Samsung's capacity is in Korea. Right. The vast majority of Micron's capacity is in Singapore and Japan, Taiwan and the us Right. Micron has no capacity in China. Hynix and Samsung have some, but that's old capacity and they actually haven't upgraded it in a long time. With that said, and what prevents China from learning is, I mean it's just a ton of hard engineering, right? There's a bit of like, oh, people who worked at these companies coming back and being hired. There's all this sort of like claims of like IP theft, blah blah, blah. But at the end of the day it's like look, there's a ton of engineers who are working really hard and they have effectively unlimited money.
- [83:13]
A
Right?
- [83:14]
B
And even then China still has not been able to, you know, in 2015 they released a five year plan which was what their targets for semiconductor production in 2020 and 2025 were. And they did it again in 2020.
- [83:25]
A
Right.
- [83:26]
B
And they've missed every time despite the fact they've poured hundreds of billions of dollars. Now that's not to say they didn't make progress and that they're not going to get there. They will. I do truly believe it because Chinese engineers are really good, right? It just takes a really long time. Now I don't think this is something a venture backed business can just do, right? That amount of money is really venture nation state level stuff. Right? And that's the challenge. And as far as like China, like you know, the next breakthrough away from DRAM entirely could come from China as well. Right. There's no reason why it wouldn't. It's always easier to leapfrog than it is to sort of like incrementally catch up.
- [84:01]
C
Is this the moment though? Because with the export controls and Trump playing the games around, you know, who can and who can't get Nvidia GPUs and the national security concerns in the United States about do we want to let China have all the GPUs? Is now the moment that the nation state changes its tune and says we're going to 10x that investment in these companies to make it happen. Do you see China becoming a global power in logic and in memory the way that we already know it's a global power in all the other ways?
- [84:37]
B
Yes, I think that's absolutely going to, going to happen one day, right? China is going to catch up.
- [84:43]
A
I want to just go back to the thing you alluded to at the very beginning, which is my last question for you was going to be when does this get better and when does my computer get cheaper again? And I think the answer is a, maybe it doesn't and B, it's certainly not happening anytime soon. Is that right?
- [85:02]
B
Basically about last year, this last year, Hynix, Micron, Samsung, they've decided to start investing in building new fabs again, right? Instead of just upgrading old fabs, keeping the old fabs going, you know, changing the process, blah, blah, blah, they've decided to build new fabs. Takes 2ish years or 3 years depending on like the timescale of the construction Progress. So in 2027 we will have new wafer capacity coming online.
- [85:25]
A
And you're literally just talking about like the, the time it takes to build a building, right? Like it's not, there's not some like magic they have to do to set up these new fabs. It's just literally like they have to go do large scale construction.
- [85:35]
B
It's large scale construction in the cleanest environment humans have, which is a clean room where there's less than a particle, you know, a million, you know, like the million particles per square meter or whatever is like absurdly low, right? You know, to the point where like human like skin flakes are the most dirty thing in the like, you know, like fab, right? And that's how you wear the bunny suit. There's all the complex chemistry, right? They have like all these sulfuric acid and hydrochloric acids and they have all these different chemicals like photoresist, there's thousands of chemicals used, right? There's these tools that range all the way from these like 200, 300, $500 million tools that get flown in on multiple 7, 3, 7. The human logistics required to build these fabs is absurd. And so, yeah, it takes a couple years and then setting up the fab, making sure it's clean, running the process, et cetera. Best case, prices get cheap again in 2027 or continue to fall again in 2027. Okay, but depends on what happens with AI because as I mentioned, everyone who's building these fabs is super conservative. Right?
- [86:31]
A
Right.
- [86:32]
B
There are these 55 year old men who have seen many busts. And if you take what our predictions for AI, supply chain demand, or anyone else who's sort of like super in tune with this, what we think, number of GPUs, the types, the memory, all the volumes required, actually prices will get worse in 27 too. But that's if AI doesn't pop as a bubble. Right now I'm a maxi, I don't believe it'll pop. I think there's trillions of dollars of economic value that will be delivered through AI over the next few years. That's me. But if you do believe AI bubble will pop, then 27 will have cheaper prices. If you think AI just continues to go, go woo, then we're screwed. And actually, memory will never be cheap again and we might as well like stop using our computers.
- [87:17]
A
It is truly wild that you actually, you have to make your computer buying decision in 2026 based on macroeconomic and giant geopolitical headwinds. This is where we are in the world. Should I buy a Dell XPS 13 is actually a question about like the state of the economy in the world. This is why we do an hour on Ram Friends. This is, this is what we're doing here.
- [87:42]
F
All right.
- [87:42]
A
Dylan, Sean, thank you so much. This has been delightful. I appreciate you both doing this.
- [87:47]
B
Thank you so much.
- [87:48]
A
All right, that is it for the show. Thank you to Nilay and Sean and Dylan for being here and thank you as always for watching and listening. This is actually our last Vergecast of the year. We're all going to disappear into the holidays for a couple of weeks and we will be back in the beginning of January for ces. We're going to do an episode kind of right as all the news is happening and then we're doing a live episode on Wednesday, January 7th in Vegas at the Brooklyn Bowl. So mark it on your calendars. It's going to be in the afternoon. Come hang out, come Vergecast. We're all going to go bowling. I've never been bowling with Nilay and I have this sneaking suspicion that he's either very good or like hilariously terrible. And I'm very excited to find out which one it is. So come bowl with us. Watch the Vergecast. It's going to be very fun. Also. In the meantime, there's a bunch of decoder for you to catch up on. We have a couple more version history episodes coming up out this season. We've got TiVo left to do. We've got the Nintendo Power Glove left to do. We've got Flappy Bird left to do. They're all very fun episodes. They're all coming out the next few weeks. Make sure you go subscribe to all those shows and subscribe to the Verge if you want to get them all ad free. Until then, we're getting out of here. The Verge cast is a Verge production and part of the Vox Media Podcast Network. The show is produced by Eric Gomez, Brandon Keefer and Travis Larchuk. I hope you have a wonderful holiday. We will see you very soon. Rock and roll.