Memory, Oil, Geopolitics, and Macroeconomics

Yeah. All right, hello everyone. Welcome back to Seminolysis Weekly. We're on episode number five. I'm here today with Ray Wang, the uh leader for our memory model. Uh Joey Brookhart, our uh Latest analyst joining the team, he's covering SAS. Uh and just everything in the market, let's say. And then Malcolm our our in house economist. So

Invite the crew to talk about everything from uh memory constraints to geopolitics to the impact of AI on consumers. Yeah, we're gonna chat about I don't know, a little bit of everything today. Guys, welcome to the show. Thank you. Thank you for having us. Ready to kick off. Kick off this. Yeah. Let's start actually with a with a qu quick intro. So, Ray, you are based in Seoul now, right? In Korea. Mm-hmm. Yeah. Uh Malcolm's in Oregon and Joey you are

I'm in Akron, Ohio. Awesome. Yeah. So we are I was keeping track of this on one of the other calls. We're covering three time zones on this call, uh which is It's not quite the record for semi analysis weekly, but it's uh yeah. It's it's all about semi analysis, having the uh remote first work strategy here. So um okay, let's start with memory constraints. Basically Ray is uh I don't know. You made a call a few months ago, roughly when joining uh the company, I guess, and

We've seen this like filter through to everything in the market and make everybody's prices of AI and everything go up, you know? Neo clouds, circle prices. What's the latest update on everything that's been happening in the last couple of weeks in terms of like, is this gonna get better, is it gonna get worse? Yeah, I mean, from the data point...

Uh what I'm seeing. I think the the thing is actually getting worse in terms of the pricing as well as the costs that really go into the happy skills cat pack. Uh as well as like the consumers. So the the pricing is definitely going up and the cost for the consumers and the hyperscaler spending for for kinda the cost to rebuild out is definitely going up significantly. So uh starting with uh I believe the computes, right? Um I think

This year we have the calculation for twenty twenty six CapEx. I think thirty, forty percent of the cat packs were go all going to memory. And that's just pretty insane, right? Because you all talk about like all the other things like GPUs uh all other infrastructure stuff, or including to date together and memory itself already thirty forty percent.

And if you look at the outlook for companies like Hynix and Samsung, right, they might be the top five most profitable companies this year, or at least in the semiconductor space. Uh so I think o all all these things together just tells you like how important memory is and how bad this uh uh shortage is uh in memory. Yeah. I think the thing that I've noticed or maybe that's most interesting to me is that

Um okay. The price of servers went up as a result of memory going up. And the result was not that the hyperscalers took their allocated budget and just decided to buy less compute with it. They just increase their budget.

In other words, the claims about being power constrained or data center floor space constrained seem to mean that they're not capex constrained. They're willing to raise more money or spend more money if prices go up because there's just like No more chips to buy or no more space to put them if they were to buy the chips.

Does that make sense? Yeah, I yeah, yeah, I agree. I think one key thing is like, oh, if let's say like if they are going to deploy w one million GPUs, right? And to ensure we're gonna have a one million GPU that really come in uh you know being shipped.

Uh in twenty twenty six and you require that much memory that attached to the G GPUs. So to make sure that happens you you know Even in a uh at the expense that you need to spend a lot more money, you just need to make sure you have enough memory in terms of value.

Another another complexity here is when your pricing was uh growing rapidly, you don't want to buy it like in two Q, in three Q this year, right? You want to buy it four Q last year or one Q this year and you know Trying to whether you can secure the LTA with the memory suppliers, because that's the best procurement strategy for their data center or for their com uh compute uh procurement. Okay Malcolm I don't mean to put you on the spot, but

my uh brief understanding of economics, my limited understanding makes me think about elasticity of supply and demand here. Like, do you have analogies for what this is like where the you know, an important component of the supply chain increases in price and then suddenly everybody's just willing to pay for it and figure out how that filters through later.

Well, it it's probably the closest we've got historically is something like an oil crisis and we're seeing that right now where uh the price can be anything. And I think the thing that we've seen in recent years is it it cuts both ways and it can move very, very fast. Um so I I think the fact that we have seen negative oil prices and then uh in recent terms uh big double digit uh jumps within a single day. Like if you need something to do the work of the real economy.

You will move heaven and earth to get it. It doesn't matter. And these are our absolute hurdles that we will see uh entities basically pricing whatever they need to get this up to the economic value. And I and I think everybody who's actually using these tools on a day to day basis sees that the economic value uh is is just enormous for any marginal. And then on top of it, from the the perspective of the kind of

Intermediate purchasers, this is a uh a kind of existential threat. They have to be in this game. Like you can't choose to wait six months if you are a hyperscaler. So uh this isn't behaving even at that normal, like just rational where it has to be equal to current economic value and can go up to that, because this is kind of business model threat if they don't win this now. Makes sense. Um okay, Joey, that's making me think of

Uh the SAS seats. How many jurors seats did you pick? Uh no no more SAS. Let's just show all of them. It is interesting to kind of raise point on on the increase you know, they're just deciding to spend more on memory. You know, even the hyperscalers have kind of hinted at some layoffs.

And kind of that seat debate there and and kind of like broader macro debate within SAS on you know, you look at something like Workday or some of these larger cap, you know, software names that are exposed, you know, and and are priced on a seat basis. of what that headwind kind of is. And that's even before a lot of the like the AI disruption fears.

is is just kind of broader macro seat headwinds, um, given too how some you know saturated and kind of at the top of the S curve you know things are. So maybe no direct impacts from the Iran war and you know software seats, but you know, definitely gonna see spending elsewhere that goes into semis, you know. You know, it has to come from somewhere. How long do you think it takes to filter through straight a hormones is closed and then months, you know, a quarter, a year before

you know, a a a SaaS business gets impacted by something like that. Yeah, I mean and you look at the the really large cap software players that are, you know, three years committed up front, you know, for a contract and you know, maybe maybe those are less affected. Um, versus maybe, you know, you start to experiment with more of your discretionary like software budget with, you know, what AI tools can do and and even, you know, from a head count perspective.

Um, but I think I saw an interesting tweet about Claude's or Anthropics EO Growth Marketing Department was one person. Um and then you think of all like the social media SaaS tools, you know, even a few of them are public and and you know everything like that, it's you know, it's not hard to hook up and and vibe code some APIs to to push out content.

Um so there's a lot of I think in maybe the lower cap portion of SAS there's, you know, more immediate impacts versus like large cap, you know, sing broader macro impacts there, you know, definitely takes some time to filter through. Yeah. Yeah, I mean um so our commander in chief Dylan Uh

s tweeted that this week that being in SF is like being in Wuhan right before the pandemic. Something's happening. It's gonna hit everywhere, but so few people know it. Like I talked to you guys, I've talked to everybody that works at semi analysis. And it seems like there's just a lot of people talking about AI, talking about memory. Um But then I go see other things. I talk to my neighbors. I talk to my friends. Um and people like hate AI.

It's like wildly unpopular as a high level concept, but then also in Uh a lot of people's like in my experience, free tier use of it in their daily life. Do you guys see the same stuff. I mean, you probably have different perspectives sitting in different places of the world right now. Yeah, I mean I mean in Asia it's like uh like people around me, if you are to working in you know even you are working in like a large company, right? You probably don't get that much of the token usage.

That you can use, right? And then you don't you usually don't have that much money. You don't want to spend like oh two hundred dollars on the the call pro plant whatever, right? Uh so people just be m I I guess like people in Asia being cheap.

Hot take maybe, I don't know. But you know, just like you know, just don't want to spend that much money just using like three tier stuff and well even like you free tier you can still get coding here and there, right? You might use Gemini or whatever. But Uh but like I think I think the the key thing was like uh I talk to a lot of my peers in high school or like in college. I tell them, Hey dude, like this call is pretty crazy.

And it was until I tell them they start going to try, right? But if I'm tell not telling them they're just using a T G P T or Gemini. uh as a chipbox, not really not really as a coding uh tool to building anything. I think that's the biggest difference. We uh I would say I was I won't say you know people see AI negatively but uh the concern of uh being replaced by AI is very real.

Um whether it's in Korea or in Taiwan. I think the concern is very real. But people just like, oh, the concern is real, but like how can we do, right? And let's just live our day one uh day by day. Jou-jou-jou-jou-jou-jou-jou-jou-jou-jou-jou-jou-jou-jou-jou-jou-jou Yeah, so I've got a little bit of a un unusual perspective'cause I was just sitting inside a a major uh US bank bulb uh bank, um and

For safety reasons, they ha they don't have all of these cutting edge things to turned on for every employee. It's glorified uh autocorrect. Uh still internally. They just don't want to run the risk of having any of their data breached and that's still a process that they're building out and kind of coming out and stepping into uh the semi-analysis ecosystem.

I I feel like a literal time traveler. I feel like I like came from the eighteen hundreds and suddenly there are like cars everywhere and it's fast and I don't even know how to walk on the street. Like things are crazy and like Going from even just being a a kind of vibe coder at home with uh the the twenty dollar a month probe plan for what any of these?

to having basically tokens to burn and use these as as they really should be used, kind of at a full throated, just like unambiguous, how can we do the most with the least? It's another step up. Like it's it's just crazy what you can do. I feel like if I so I've now been with Semi for eight days, I think. And if you had asked me to do what I did in the last eight days

At the bank I was at previously with I would have said I need a team of a hundred in a year. Melcalk, do you want to go back to the bank? No, no, don't send me back One way straight here, yeah, yeah. So okay, but you okay, you said something interesting and I'm curious, Ray, on your take on this as well and and Joey frankly, like it it seems like

There's a big distinction between people using AI at work and people using AI in their personal lives. Seems like all of you would have the perspective of using it in your personal life because you're just like interested in adopting new stuff and trying it out. But I think that's it's it's like really hard to see in my friends or neighbors, like

them find the use cases in their personal life almost out of thin air to actually go and try this stuff out. Like they might have one example in the last week that they might think about this thing whereas at work where you're like forced to do stuff every day that is knowledge work instead of like personal life where there's a lot of leisure and stuff that is not Th the w stuff where I'm not focused on efficiency, it's a very different, you know, experience, right?

Yeah, I mean I mean a key thing is of motivation, right? Uh not not everyone is really motivated to use AI to do certain things uh in their personal life and And and the the key question is like why why I need to use AI to do something, right? I can just chill and I can just like you know watch Netflix. I I don't need to use AI beside using ChPT to asking questions to or searching sick. Uh but it's really different from people who have experience building something with cloud or with codex.

Because once you have experience building something, that I think will trigger a lot of interest and will really kinda open your eyes in terms of like how capable all this AR are? And If that's the case, you want to use it. But I feel like you know in general when you don't have that sort of uh first level experiences.

you probably don't want to just you know don't want to blend the the coding stuff into your personal life. I just haven't I just don't see it. And it typically I think things like this require quite strong motivation um uh for you to do that. Especially everyone's busy, right?

Yeah. I I mean I may be a a weirdo uh in a lot of ways, uh let's admit it. Um but I uh like While I was working at a bank, obviously I needed to know about what was going on with AI. So I I got way into using it in my personal life. And this the number of categories of use that I I mean, once the the uh The AI knows enough about your interests, I would literally just if I was bored, say, teach me about something that I don't know about that I likely will find interesting.

that's like for my ADHD like uh like love of learning just nerd brain. That is that is heroin. I can do that all day. Um like and I've been I've been adopted like I mean I I this is gonna be oversharing but like I had like high cholesterol so I got uh Claude like get on it like how do I make actually tasty food that it uh genuinely and like it's

uh like optimizing on weird bio like uh uh culinary chemistry in order to get me cream sauces that uh don't go straight into my arteries. I mean it's it's it's amazing. Yeah. Awesome, man. Joey, how about you? Are you using this in your personal life? Yeah, I mean there's definitely areas. Um I mean I have a few examples kind of being in the Midwest, like my my buddy's a defense attorney. Um he obviously gets clients, you know, that are ch you know allegedly charged with something.

And then they refer you know they use the open AI you know Chat GPT free plan and then they send them something and it's some case that doesn't even exist. You know, where he can go in and then, you know, he looks at that and he's you know, he can connect. He first time today he asked me like, what AI system should I be using?'Cause I just have cowork and you know Word essentially is all he has. Um so that's a big change, like especially just around like hearing people I talk to.

Uh both and consumer, like I think it really be like comes down to jobs to be done. And if you're really gonna get it in those like you know, stuff that's eventually too monetizable, um, it has to like actually, you know, solve some problem in a good way. Um, I think, you know, most people are very, you know, cognizant of like agenic shopping. They're just not gonna let some agent shop. They want to go read a few Amazon reviews and

You know, they care mostly about low prices. I think it's the same with like finding flights or travel itineraries. It'd be great to say, hey, go book me a five day vacation. I don't think people are there yet. Um and that transition from travel agents to like OTAs. Yeah, then Google Flights is like you know, just uh Easy thing to use and then you know maybe we'll get there. I'm sure we'll get there eventually. Um, but there's just not the like, I don't know, killer like

It'll all be verticalized eventually, but like can kill our consumer use case where like someone or you know, eighty percent of America has to have it. Um right now, like, you know, what we're seeing with cloud code and just the usage You know, those use cases we can't even really dream of or or really see yet on the consumer side.

And that's when I think, you know, maybe favorability, things like that will begin to change and um people will have some s success with that. Um Duolingo is like the pretty obvious example in public markets that's kind of like consumer software. uh where you see some really great demos of, you know, my wife's Ethiopian. When we go to Ethiopia, I don't speak you know, I speak barely any Amharic.

I'd love to be able to put an AirPod in and you know or speak to something that then speaks back and you know there's really no use case for me. Like I don't need to know the language that well to go use Duolingo or some other like language learning process.

Um so like that's something I see, you know, for myself. Um, but there's nothing like really widespread yet where I'm seeing, you know, people are using it all day to day in like their consumer lives and you know, my mom's using it and things like that. How do you model the adoption of AI in your head? Like it seems like it started in consumer, everybody tried out ChatGPT. There's no real obvious way to monetize all those free tier users for OpenAI. So they have to go into the business world.

They start selling enterprise, you know, integrated apps, their API, all that stuff, but then they kinda are getting their lunch eaten by Anthropic with Claude Code. So okay, let's say they figure out the enterprise business really well. Do you think those same companies, like Anthropic with Claude Code or OpenAI going back to their consumer market

end up building applications for consumer companies in the or consumers in the long run. Like they're gonna go after Duolingo head on instead of going after whatever finance or accounting or legal related like enterprise apps they're going after right now.

Yeah, I mean it's it's interesting'cause we're starting to pull like here even some internal you're or some you know data on consumer subscriptions that will be really interesting to see here in a few days. Um where you are s really starting to see you know things really inflex. Um, and like consumer, you know, spending and and subscription, um, new subscriptions and just in general and spending. Um

But yeah, I mean it's it's a lot different. Obviously, like internet, like consumer internet and consumer transports ads very traditionally. Um so

You know, there's been a lot of rumors about who they'll use on like a DSP or SSP or like a DSP perspective for people that want to place ads in EO queries. And I think, you know, maybe it trends more towards that like search engine um type ad model in the interim on the consumer side, and then we go to more verticalized, whether it's the verticalized consumer companies you know adopting open AI or you know API to run those applications or open AI itself.

Um but it's it's just still so early on the consumer side. It's really easy to measure in a You know, like knowledge, you know, knowledge seats out there, how much people spend on Office three sixty five. Um, you know, all those other any task, I'm sure you can, you know, get to a rough token estimate of what it takes to complete some task in an office, you know, knowledge work job. Um, and you can, you know, roughly put that up against number of seats.

you know, number of tasks that go on in an enterprise. It's you know, you really have to break it down from like a vertical perspective and consumer to get go even deeper into that. And, you know, it's just so tough to say where that really ends up long term. But, you know, intermediate, I mean, I think ads is is the big model um that that'll turn to. And, you know, they'll kinda have to figure that out in the interim.

What do you think would be more popular, AI that costs money or AI that's free with ads? I mean there's revealed and there's revealed preferences like that we just see every day in consumer and it's it's it's obviously ads. Um but then there's, you know, probably those twenty per yes, the eighty twenty thing, twenty percent will go pay a subscription.

To just have a lot more usage, get better you know, use a better model, um, things like that. But for the other eighty percent or plus, you know, I'm sure they're fine using a lower tier model that's you know kind of subsidized and you know by I I think one area where we may see a little bit of a a surprise there is in the healthcare space, just because people do get a little bit spooked about their privacy and I think

The idea that you could buy a a more premier uh program uh version that gives you a little bit of a uh uh a privacy shield relative to just uh typing something into the internet. um may come across as a a a better value add, especially because that is biasing to where it's a much order older population. And and maybe this is just a a population I have greater visibility to.

um just because I am a bit older than you guys. But like watching my parents, they're actually using AI now because they're asking questions about health. And they're using it

really in the Google search bar. Like l let's be honest, where are they actually using it? They're they're they're finding it not because they're going to to the tool. They're they're getting fed the tool, but it's giving them something of genuine value, so they're coming back to it. So then they're kind of populating it with the increasing questions about their own uh healthcare uh situations and whether or not they need to kind of uh

um have some more substantive conversations. And I think then we're gonna see people locking in and perhaps continuing to have one instance that they actually then have like build some context with and and uh Right. Have a a durable conversation. Yeah. So let's say we've fast forward a year, there is three times more data center power online than there is today. the GPUs that are deployed and the models that are running on them are three times more performant and like, you know, cost per token.

The result is that there's like nine times more tokens or intelligence flowing on the internet than there was today. How much of that? Nine times more in a year? gets to your parents asking questions in the Google search bar versus gets to enterprise customers versus gets to

I don't know, somebody in a totally different country that's on the free tier and and doesn't even get the Google search bar at the same level or or something like that. Like Um Do you think it's gonna be an even distribution of tokens around the world? I mean, looking at

how many more tokens you can burn by doing uh um coding. Just I mean, it's it's hundreds of thousands of tokens and the IO intensity is very, very high, uh, where the the search bar is going to be really pretty lightweight I think in terms of a especially once you have uh all of these uh kind of pre popped cash uh uh things that can be pulled up so everybody who's asking the same question about their cholesterol is going to be getting something that really is compute non intensive.

Uh so I I I'm biased towards thinking that the enterprise drives tokens for the next year. Yeah. Yeah. But I guess both in terms of seats and in terms of usage per seat.

You would imagine there's more adop even more adoption in the enterprise than there already is compared to consumer. Um because in tur like uh is is the token distributed globally, do you think p on on a per dollar basis it will be fair or even in different region?

That's the question I have, right? Because you can have the same amount of to you can enjoy certain amount of tokens, but your per dollar basis can be different, right? Uh and you know also the kind of price level in different countries is different. Maybe this is like a question to Mao Tong actually, but uh I don't know. This is a question I was thinking about the other day. So like a a purchasing power parity uh adjust in token price. Wow, this is a deep cut and I'm loving it. Um

I I think we're gonna see some of that actually effectively with uh a lot of the the Chinese competitors and and when we look at like kind of off the bleeding edge, the cost per token is so much lower. And so I think the fact that Everybody in the enterprise wants bleeding edge tech and uh that's going to keep the price of that quite high. But something that's six months, eight months, twelve months out of date, uh I mean like

Don't waste my time, please. But for uh for consumers that's an amazing product. So I think we are going to be seeing that and I think we'll probably also be seeing that with uh kind of chips as they they stop being uh kind of maximally valuable. They're gonna not be taken offline. They're going to be repurposed until we get to some actual saturation point with chips, which

I can I can really not actually imagine. Um, the only way I guess I could see that is if we have a a true bottleneck on uh on electricity showing up at some point and and then it we would we would be probably seeing uh chips being geographically relocated to places that uh haven't previously had access um but have electricity that can be uh kind of picked up uh for a relative bargain. Can you make the analogy to oil again in that concept? Like if tokens are oil, right, and there's this

uh purchasing power adjusted per token price to compare globally. Like, is that the equivalent of I mean, my mind goes to the um Big Mac index, my favorite economic uh metric, right? Like how much does a Big Mac cost in a given country?'Cause they've got a price at at a certain level but they also are exposed to some sort of adjustments to input costs for in for I don't know, water and beef and whatever else that's localized. Well yeah, that's actually a really good point because um

So a lot of prices are hyper local, so obviously your rent price is uh entirely local. Your health care is P mostly local. Legal service, these are all very local prices that are going to be locally defined. Something like oil is a truly global price. I mean i uh yes there might be a little bit of a spread between uh WTI and Brent, but like Fundamentally, there's fungibility in oil in every location in the world. And to some degree tokens are going to behave a little bit like that.

uh transport the output of uh a computer pretty much anywhere in the world uh at at uh near the speed of light. So there there is gonna be a little bit of a law of one price. Uh I think where we're gonna see uh like dispersion is cutting o edge, not cutting edge. So I think if you want a cutting edge token anywhere in the world, you'll probably be paying the same price. Yeah, makes sense

Ray, you you feeling lucky to have uh Opus four point six fast access at uh semi analysis right now or what? Well uh I I I'm I I'm glad I can use thing for free. Uh that's the best. That's the best part. As long as I have a deliverable. Because I but you know if I need to pay I will I will never do that. I I don't think I will do it. Probably at well I I think the the maximum level I will pay is like twenty dollars per mount.

Or or just losing enough low free version. But like having having a paid version is definitely against treasure, I'll say. Do you pay for like a streaming service, like a Netflix or a Spotify or something in your personal life? No, no, bro. No, I don't no no no no no no. I'm very curious. I I see those airpods. You you listen to music?

You go like YouTube with something? Oh oh oh so I I so I pay YouTube premium. That's the only thing I pay. Yeah. Because so here's my theory. If you you want to pay like Spotify which is like thirteen dollars, why you not just just pay the

Uh so the I think the student version for YouTube premium was like twelve dollars and then for the the normal p people it's like twenty dollars, right? And uh and on YouTube you can listen music, watch video, all kinds of things, it's like twenty dollars. So so it's more

I think it's just more valuable versus like Spotify and others. Yeah, that's my thing. Joey, how about you? Do you pay for um subscription services like that? I do, yeah. I mean I um I probably pay for too many streaming services, definitely pay for Spotify. Um I probably shouldn't, but I do pay for for quite a few things.

Yeah. Do you think we're gonna get to the point where we have like a open router all in subscription plan that is the style of how Netflix started or Spotify started so that you don't have to pay every provider equally but you can use You can just like get a monthly allocation of tokens and then send some to the video models, some to the certain text model, some to a coding model. On the consumer side or enterprise? Like

Uh good question. I think I was thinking about consumer. Yeah, I mean I definitely think enterprise will go that way if you think of like some like the infrastructure software companies like kind of sat. If you think of like the old VMware or some of the Hachicore products.

We could kind of like extract it a layer above. I guess that's the way they do it with the with the cloud providers already, right? Like they have these MSA agreements within AWS near Azure and you can imagine that gets spent on tokens instead of Instance time rent. Yeah yeah. No, no, exactly. And so even just routing and I'm sure like people this happened in like the history of like OLAP databases as well. Like people just got more and you know

Even like things like Datadog, you think of like DevOps and like application infrastructure monitoring, you eventually got like you were just spending so much on your datadog bill. It made sense at a certain like company size level and spent, like look at in housing that on your own kind of you know, whatever like, you know, OLAP database you wanted to use and kind of built some

you know, visualization layer on top of that. So I'm sure something like that similar will happen in enterprise as these bills keep on going up and you know you get really smart routing on where you want to send tasks. um that you know don't need to go to the most expensive and latest model. Unlike you know, consumer probably. Yeah, not as much. And that kinda goes back to that question like

How much are people gonna be paying for this? I mean, you've got people like Ray who sounds like don't wanna spend a lot on um you know, kind of like leisurely type type services. Um and I'm sure a lot of consumers like view AI that way.

Right now and you know, who knows like riddles. And on the consumer side, what I would really wanna to highlight is that these companies are going to be w wanting to create anything that looks like a moat. Because fundamentally Like right now, switching between any of these top tier uh AI providers, Chat GPT, uh Gemini, like Anthropic.

Like it feels the same, it looks the same, ever the buttons are the same place, like the the the learning curve to switch between these for like a uh just a on your phone chat app is so low. But the dollars associated with uh actually preserving not just uh the marginal revenue. But the historic revenue. I mean like I I think I think about Google. How many Google searches are going to be run in a non-AI context?

uh ever again. I mean I think it's done. I think I think g Google search is now a AI augmented uh task and they're they realize that they fundamentally have to protect that uh moat which I mean, five years ago like there was no competition and now suddenly like I don't ever touch uh a like just because I happen to uh ch prioritize Claude, I never do a Google search in my browser. Like that that's not a thing I do anymore, and that's existential for Google. Yeah.

I'm not sure. I'm not sure I believe it's existential if they are. Still monetizing something. Like we still in our tracking at least, token production from Google is crazy. Like the numbers that they announced. I think it's Yeah, it's a r it's a risk or it's difficult to foresee how they monetize it. Um Yeah.

I mean to me they have to protect that moat. Like like When I when I kind of like uh I I I did some back of the envelope math a couple of weeks ago and I was trying to think, okay, so what is the implied GDP that we would have to see based on all of these investments on and based on normal returns uh um on capital. And it's this enormous number and it just it it I mean, everybody's like, Oh, oh my god, you're you're you're betting on the future being uh like uh

Fairyland. Well no. That enormous number makes a lot more sense if you think of the world as being Pepsi versus Coke in the the nineteen eighties and nineties, where these big companies were in a zero sum game fighting for the marginal uh consumer where a dollar spent by Pepsi was a dollar that uh Coke had to spend, absolutely had to put up.

Or they were gonna lose ground and could go away. And so I think we may see a lot of this maybe kind of companies going toe to toe and they have to spend on this uh infrastructure. They have to spend to build this out because otherwise they lose the golden goose. And and that has been really obviously a very golden goose. I mean the margins have been

great for a long time and that goes away if they don't have a seat at the table going forward. Yeah, we we saw exactly this during the uh Chinese New Year celebration over the last couple of weeks. Doug's did some great work there or he was tracking like how much those companies were spending on those promotional programs and what the returns on them were in terms of like I mean it's a different culture, so you know, it's it seems strange to me, but literally paying people in the form of credits

that which you can use to spend on like the DoorDash equivalent just to type into the AI model every day. Right? Like paying people to juice the Dow numbers. It's a really strange growth strategy, but I mean it yeah, it's like kind of the exact cynical version of what you're describing where you can just literally have people um I don't know.

spending money for the sake of it instead of spending money Well, which could come at the benefit of consumers, right? Like it could come at the benefit of free tier consumers getting a lot more access to models as these the deployments get rolled out. assuming that they maintain the same mix Uh I no I I think that's absolutely right. I I I think um

Uh Cory Doctorow has a new uh book out that has a not safe for work title, so I won't say it here, but you can Google Cory Doctoro's newest book. Um and it's basically this uh book Uh there's n I uh I think I would be giving it away. Uh um it it it's making everything poopy, so let's just say that. Um uh it it's the thesis is basically that a lot of business models are built around you give a great service, uh

and uh that service is amazing, everybody gets hooked on it, and then you make the service bad and uh you figure out how to lock it down and make money from it. And obviously, uh, we're seeing this in AI and I think uh we're going to be seeing places trying to figure out how to make money, but I I think it is fundamentally still in the the first innings where the free tier uh

uh um has to be amazing because we are still on wrapping people. And so I think it'll be interesting to see how much pushback Uh we get Um I mean and obviously the Super Bowl commu commercials by Anthropic were kind of amazing in terms of just hitting below the belt.

And going after Chat GPT. Like I mean I mean like nobody wants that. Like and and like they w they were just uh like I mean the fact that they were doing those uh kind of signals that they're willing to uh play a little bit dirty with uh with open AI and um And basically suggest that uh OpenAI is trying to to convert to uh revenue positive maybe a little bit faster than uh the rest of the industry is going to go and and it'll be really wa fascinating and amazing to watch.

Yeah, makes sense. Um I have to run to an eight thirty meeting, everyone. So thanks so much for having me. It's great to chat. Yeah, but Okay, Ray. Um Nogum's talking about the uh okay, let me ask you, did you follow the uh you follow the Super Bowl halftime show? Uh I didn't watch this year, unfortunately. When do you think we're gonna have BTS or somebody on the Super Bowl halftime? What do you think?

I think it will happen someday. I mean th the the capa ideals this day they are everywhere, no? Yeah yeah. I heard some rumors that it was gonna do a leaper next year. For f for who though? Who who is g who is who is going for the halftime show? The halftime show is definitely for like the wives and sisters, I think. Um, as opposed to the The guys watching the typical guys watching the the football. But uh anyway.

Oh, I was gonna say I only watched the ads. I didn't watch the Super Bowl or the top time show. I was just like I gotta see the ads. Well what are they doing? Like the um but this is the uh like the economist uh who's like This is where like big dollars are being spent to try and uh reach the absolutely largest US audience imaginable. And so they this is The fact that so much money was spent by AI and so much uh

So many people were referring to AI and so much of it was so bad. Like so many of the AI adjacent things were so horrible. That I I I feel like that is a good sign that uh Uh it's not mature yet in a lot of these spaces.

Yeah, uh let me go back and read this again. So this was an NBC poll that I referred to earlier. Let me try and just bring it up on screen here for a sec. Uh so we can look at it. So an NBC poll puts AI, comma, that is artificial intelligence, as uh 26 positive, 46 negative for a net score of minus 20 for the average. American.

Um that is higher than the Democratic Party and higher than Iran, but it is lower than ICE, Gavin Newsom, Kamala Harris, the Republican Party, Donald Trump, all the way up to uh at the top, Stephen Colbert and Pope Leo the Fourteenth, number one with a Plus thirty-four, forty-two positive, eight negative out of fifty for uh this guy. Anyway, um So everybody decides to advertise AI during the Super Bowl, which by the way was less than a month ago.

Times flying here. And then I guess like two or three weeks later, Anthropic takes that heater ad and parlays it into a existential crisis with the Department of War. There's they're working on an executive order to have them pulled out of every government agency, right? Yeah, it'll be really interesting. I don't I feel like uh In terms of winning the consumer, that is like been a gold mine. Like you look at actual sign ups and downloads, it's through the roof. But uh

But yeah, how does this actually play out? Like I've talked to my uh friends who are working in uh DOD adjacent fields and they're like, wait a minute, I'm losing clawed code? No And and so I mean I think that this is gonna be a really th fascinating thing for the and we're gonna see it fast track to the Supreme Court. We've obviously seen uh the lawsuit already coming out of Anthropic and like Uh I don't know how this plays out. We've never seen

a uh a ruling like this for a domestic company before. These tools have only ever been tested without uh offshore companies and this is some pretty uncomfortable territory for your kind of um more libertarian leaning uh conservatives. Like this is government getting into the decision of who's a good company and who's a bad company amongst American companies. And that is a really uncomfortable territory for people who don't want government making this sort of decision.

So I will be fascinated. I don't think it will line up purely on partisan lines and uh we've seen some big surprises out of the Supreme Court recently with uh some major decisions. not going towards uh the White House. I mean obviously the overturning of the tariffs was a uh was the biggest, but this is not a a clear cut case of a Supreme Court that is a hundred percent in the in the bag for the White House. And uh for the Department of War. Yeah. Yeah. Yeah, makes sense. Um Okay. Let's uh...

Maybe let's talk about some of the other impacts in the broader economy. So uh another uh storyline over the last week that I was looking at was just in terms of the job revisions that have come out. Like there was initial numbers and it always seems to go in one direction where you initially report a certain amount of jobs and then they revise them later to to bring it down. Um

Do you uh what do you attribute this or like let's say bad data to? Um but then also what do you what do you imagine the Downstream effect of that is? Not the uncertainty of the data, but just literally the fact that there are less jobs or more people looking for jobs than what we currently think. Like does yeah everybody's been saying like narrative violation, AI isn't taking people's jobs, but is it? Like

Well, so I think there's a number of things that are going on right now. Um and I want to break them down separately. The first, y your your sense that the the revisions are always in one direction. That's actually a really astute point because I have always seen the correlation of the re res revisions, the direction of that as actually one of the best signals. that you're kind of moving more in that direction than the primary data signals. And that's because of the way that the

The employment data is constructed. They're actually going out and surveying a bunch of firms and some of them aren't responding. Well, some of them aren't responding because they the firm went out of business. Some of them aren't responding because times are tough. And so when they finally do get all of that data in, it corresponds with this underlying tension. The same happens in reverse when we're in really good times, in fact, because you're missing all the new startups.

Um, I do think that we have to think a lot about the fact that uh there are a lot of kind of single person uh firms being started right now, and those take a long time to actually show up. in the the major data. And so the degree to which AI allows for smaller firms and more nimble new firms, that's going to be a slow burn to get

caught up in the data. And that's where we're gonna have to be looking at uh the difference between the household survey, which hasn't really been quite as bad uh as the as the payroll survey,'cause that household survey

You don't care if uh as an individual if you're working for yourself or if you're working for somebody else. If somebody says, Are you employed, you're gonna give the answer. And so that can look past that. So the fact that that's been a little bit stronger is a little encouraging to me.

And then the dynamics that I'm getting really kind of concerned about for what might be driving some of this uh uncertainty, there's a couple things. First is CEOs are being measured right now on are they adapting? What's the easiest way to look like you're adapting? It's to cut hiring or even cut staff. If you can do that, you look like you're winning. Even if you aren't winning.

And so I think we may see a lot of tasks that we actually kind of back end have to hire for. The other thing is one of the areas where I think AI is most transformative in a terrible way. is in the hiring chain. So I was recently out of a uh job looking for a job and if you think about it Uh the quality of

A resume doesn't tell you anything now. In the past it was a a weak signal about how good somebody was. Now literally everybody can write a top-tier resume with a few clicks that is completely customized for every job posting. The people I talk to in HR departments are overwhelmed.

They're using AI to try and fight the other AI and it's all just coming to a standstill. And you get down to the situation where the only way to find a job is to know somebody in the company. And full disclosure, I knew a guy.

And he helped he uh introduced me to to Doug and Dylan and that's how it worked for me because I couldn't get through any of the filters. Because this is an adversarial situation where AI is fighting AI and it's making it really hard to hire right now. And I think We'll have to figure out some major new things in the future. What do you think about the impact of

international firms. Like obviously you're talking to two people in two different countries right now. I'm in Canada, Ray's in Korea. We're all working for the same company, all contributing to similar projects. Maybe we're in a unique position, but it does seem like there was a trend in one direction of doing your work from the internet, now everybody's in the same playing field.

as oppos where, you know, employment or jobs are fungible like oil and tokens as opposed to um needing to be localized. And so Um do you think there's a significant impact on the job market or companies productivity or um individual people's payroll, like what their take home pay is going to be based on downward upward pressure from more companies hiring abroad or being created abroad or something like that? I don't know if it really changes it that much. I mean we were

For the the the kind of non face to face jobs, uh, we got really good at this. We already saw these uh data center or the the kind of call center booms uh throughout India. Like this we've gotten really good at incorporating people. And if you look at kind of the s the the salary takes of top AI researchers as being like kind of that's gonna be your your most competitive.

Uh they're really high no matter where you are in the world. It doesn't matter if you're sitting uh um in a cheaper location. Um I feel like we're going to see a lot of dynamic that Uh if you are a scarce resource, more resources are gonna be uh a aimed at you. And and it does feel a little bit like a winner take all dynamic where there's a lot of people who will not be le leveraging these tools a lot. And it it feels a little scary for uh people who who have kind of gotten by by

basically n being a knowledge worker, but fundamentally what they're doing is taking knowledge from bucket A and putting it in in bucket B um and uh always getting into the right bucket. That's an AI task now.

And so I feel like there are a lot of knowledge workers who are at threat. But kind of the uh the people who are feisty, people who are trying uh really kind of trying to keep on the cutting edge are are going to get leverage and I I suspect we're going to see uh a big divergence of wage on

output of worker because we're it's gonna be a lot easier to actually measure worker output than it ever has been before. This has always been one of the hardest things for firms to measure. Uh and and we've used really dumb metrics. I mean like hours at work like And you g you know the guy who's like there for sixteen hours, you have no idea what he's doing and then there's the the the guy who like leaves at two and you're like, Well, that guy is necessary for everything and this guy

I don't know who what he does. And like those are things that we'll actually be able to see now because AI can surface those differences so much more powerfully. So I I do think it drives what could be a socially not ideal, like actually s sort of socially disruptive, uh, greater return to uh to individual worker productivity.

that could get really, really weird, especially uh when we've got this international market where, yeah, you can have somebody who's working in Singapore and doing amazing work and it doesn't matter for your company. um and you'd rather pay that person top dollar than a domestic person who who's really uh not advertising that much. There i to me, knowledge work it's actually a benefit to have people all over the world, which I think we benefit from

quite strongly with semi analysis. Um, even if everybody's using the same GPUs in Virginia or in West Texas to do their work, you know? Um Ray, can I uh uh th'cause while Malcolm's talking there, uh I'm thinking about your dashboard that you're building for uh the memory model. Can you just like Because Malcolm's like, okay, it's gonna be hard to measure worker productivity or output or something and

Uh a lot of out for those listening, a lot of output comes from Ray in Slack, in screenshots. This guy is typing a good amount of the day, I would say. So, um what do you like Take me through what it took to build that ash board for you. maybe compare because in my mind it would have taken a long time to build it in the past before quad code existed, but likely we you just never would have tried to build it.

Yeah, yeah. I think like, you know, well the the the fact is like you know with all call code I won't even think about it, right? Uh I will only just you know making a couple graphs on like salesheets or so other tools. Yeah, I will never be able to not only build a dashboard, but also build a sort of a interface that you can interact with the dashboard.

It just is you know unfathomable unfathomable for me. Uh but you know this whole dashboard, let's say the first beta version I have, it only took me a few hours. And I literally just threw my memory model, all the Excel sheets and all the complex number in there. And then I just help you to build a dashboard. Even though there's a mistakes here and there. Uh there's a level in incorrect. But like you can really confidently say like eighty percent of the content's there and it's correct.

It only needed more prompts to get the things right, get the formats, get the layouts, get the details correctly. It you it will take time, but like only take a few hours. And let's say if I just fully owe in for twenty-four hours for the h for the whole day to focus on the dashboard, I'm pretty sure I can get the products out there and then you know get to other teams to implement them that dashboard. And that's almost impossible for me to do

Uh with all this. It's just impossible. I I'm not a software engineer to doing that. Uh and I think that's I think the biggest uh game changer here is like, oh Uh people who used to be engineer and who being tasked to doing this kinds of things. I think it will be very, very challenging for them or at least some some of their tasks or their workload there will be replaced, right? Uh because now people who like me who ha we who doesn't have engineering background can st can do this.

I'm pretty sure many man many artists can do it as long as you dedicate the time. Yeah, so okay. So so I'd add on top of that, um like we are shortening the distance from the desirer, the person who wants the tool, like in this case Ray knows what he wants.

I've been working on a dashboard uh that's going to be like a calendar function for like how the macroeconomy is interacting with AI. It's going to be surfacing all of this macroeconomic data and hopefully we'll be able to roll it out eventually. But I'm now using it instead of Bloomberg for watching releases on economics because in one week I've built a tool that I find more powerful than Bloomberg uh for seeing live React and how do I understand the economy. Um this was

Like, there's no way I could have explained it to a coder because I would have had to live with them in the room with me. Like like ignoring the fact that AI can code faster. I can get the ideas and iterate with AI in a way that I could never do with a human coder. So I am uh my use, my value is desire. Like I'm I'm basically just like

I know what I want. I worked in this industry for twenty years. I give me exactly what I want. And when it doesn't give me what I want, I can tweak it and tweak it again and tweak it again in a a really fast loop. So it's not just that it's writing code faster. Getting rid of that distance. But that okay, but that's a thing that I think is critical,'cause r both of you have kind of flippantly said, Oh, I built it in a week but

it took you how many years to get to the point where you have the knowledge in your head to know what you desire to put that dashboard on screen. And I I think it that still takes a long time. So Like Yeah. Okay. How much Net time, accumulative time do you think it c it costs? to get to the point where, Ray, you know you want DRAM supply versus demand, you want commodity DRAM X HPM supply versus demand, you want wafer capacity by vendor, you want capital expenditure.

And then you're gonna go by HBM for the tab, and then you're gonna go for a vendor share tab. Like in other words, um what does it take to build the model that this dashboard is based on? Yeah, go ahead. Go ahead, Ray.

No I mean like what you're getting at is like the the embodied human capital of Rey is an incredibly expensive thing. And So for the people who have that embodied capital and we're seeing this all over I mean the everybody's talking about how uh it's a lot harder to hire junior people and people are freaking out. about what's the the path to the future where where seniors aren't actually training up people and they're not actually going to be going through the grunt

So we're gonna have to figure out new mechanisms of training this human capital. Like it you can't we can't get the the AI that we see today uh is dependent on having somebody with Ray's embodied knowledge. Um but we're gonna need in as an economy a new path to that'cause in the past you kind of you you you spent your years uh making charts and uh and paying your dues and and that's going away. Yeah. Yeah, yeah. I guess I I get confused when there's a uh

a level at which the embodied knowledge doesn't matter anymore. Like people drag this all the way into the future and then they think Well, eventually the model's gonna be so good that it's gonna have all of that knowledge about the memory market above and beyond what Ray knows in his experience.

And you actually when you're building a dashboard like that, you trust the insights of the model more than yourself where you're effectively asking the model what to build. So I'm I've had this experience a few times where I've like

in certain scenarios, effectively ask the model what to make better. And it's come up with some good ideas and some bad ones. Mm. Have you guys done that? Yeah, yeah. Uh I think it's just sort of the the feedback I'm constantly getting from with the Gen AI or just when you are doing coding. And then there's also some of the time they're like uh I'm not sure how can I make this better and they would just give you like option one to option five and like hey you can choose either one of them

to implement right and to see whether that works. If not then you go to option B or something C. Uh I think that's sort of like experience I'm getting. Uh another thing is like we are just talk to talk about that. But actually what uh I was doing over the past few weeks was like not just building the that dashboard, but also using Calcutta to build in the app. building s any kinds of scratching scratching machine, right, to to help in the research.

Uh I'm building a new telegram chat that just tracking all the lifetime semiconductors news, financial filings. uh the the some of the semiconductor news that you know I care. And then you can even to design a mechanism to wait different news to see the relevance. to your own coverage. For example, if I cover SK HANIX, right, the the weight of that or the score of that in terms of relevance is probably hundred or ninety.

But they will be the company, uh let's say the the very small semiconductor company. They are still in semiconductor space. But the relevan relevance there or the score might be ten. So like I think building those stuff just so powerful and I will never be able to do that before all the colours. And I do this without knowing anything about coding. It's a this this is that crazy.

Yeah. Yeah, yeah, yeah. How did you uh how did you come up with the idea to build a news feed of semiconductor news that's weighted based on your coverage of companies in the ecosystem? Yeah, I think I think you know it's you know, I would think like, oh, if I'm uh one of the things that you know finance people try to do is like we try to upstore the the most information we can in a short period of time so we can make an investment decision.

Uh faster than others and more accurate than others while you know uh utilizing the most information out. Uh, but there's just so much information every single day, even in the semiconductor space. So, what's the best way to do that? To extract the most valuable information that's most valuable to your own club. so that's sort of the reasoning behind it to build a whole bar

And then the and the the question coming in after this is like oh like what are the sources I should have, right? And I used to read a lot and I know where to get a source, I know where are my existing information channels. So connecting all of them together

So it becomes like an automatic process. I don't need to like, oh, every day I need to click all the things, right? I just like open this butt and then it sends things every single day as long as my map open. And then and then so I send some of the stuff to the semi-analysis Slack, right? And like, oh, this is great. Uh yeah, yeah, yeah. Okay, what's the weighting between uh geopolitics and like uh I don't know, I'm gonna come up with some example of something detailed to do with like

CXMT's uh DRAM supply for China. Like, you know, what's the What's the right mix and because you gotta cover something of both, right? To you know, y you would like to know if there's a A big issue in the the global economy or Trump like comes down with some tariff on Taiwan tomorrow or something like that, right? But Like do you have some intuition about eighty percent details, twenty percent geopolitics, macro, micro, split?

Um I think I think my one right now is fewer about uh geopolitics or macro stuff. I think that's uh just because I'm trying to like eliminate No additional noises. Just focus on purely on the industry stuff. I think that's the most the most helpful to me. Um Michael will be m very very important or to politics, but I feel like oh if I adding additional category of for information

There will be even more information for me to absorb every single day. And that's very it's that's very challenging. And even with all those information right now, like the amount of information I'm I'm getting every single day is already like So big that I kept sort of catching up. Yeah. You're gonna get so much of the macro stuff just from like living your life that it's not worth including it on the actual news feed that you're gonna use to build models.

Yeah, yeah, and if you just track tracking like Trump's uh uh true social, there's just like so much things going on. How can you catch up? Yeah, yeah. Make sense. Okay. All right, Ray, make a pick. Samsung, SK Hynex, Micron. Bam. Who's gonna have market increase or decrease this n next year? You mean the market share? Yeah, Market Share. I I don't want to choose a specific company but I I I I I will I will I I I like Samsung and Heinrich.

Yeah. Not investment advice, by the way. No investment advice, put micron. Where it is not bullish on right now. No, my point is also good. Um every company uh you know memory super you know, memory is super cycle, I will just like every c I will just hold a basket, right? But you know, if you have to I have to choose you know Sung, you know, the the best outfit I would just choose Sandsong Hynix. And I'm biased.

And I need to you know I'm in Korea so I need to uh I don't want to get killed publicly, so so yeah. I need to support Korean Korean friends. Yeah, yeah. Yeah, yeah. Yeah, I have to b have to pick the local chances. Yeah, yeah. Less of respect. All right, Malcolm, how about you? Who do you think's uh gonna win the m memory wars? The Korean guys or uh everybody else?

This is so far out of my lane that I'm like, Ooh I had to I I I had to literally uh uh build tools to to understand this space'cause this is how I I mean I'm I'm a deep macro cut kind of guy, uh, economics all the way and so I'm I'm tooling up

And I'm just thinking about how uh powerful these tools that we've been using are for getting retooled into new domains. So uh something that I've kind of surfaced internally to to people is this kind of pipeline of like kind of s uh have Claude read everything in the world about some topic and then pipe it to notebook LM and give you a podcast that you can listen to while you're working out.

And by the way, that is like a a life hack. You will you'll never sleep again, I'm sorry, but you'll learn way more than you ever did because you can you can kind of get an hour long podcast on any topic at any level of depth. Um it it's it's magical. So I I I d would say I don't yet have a vote on that, but I I

Say uh uh come back to me in a week and I'll probably have an opinion. Okay, we gotta c we gotta wrap here, but I I wanna get to one last thing which is a tweet that you sent out from the semi analysis corporate account today, which I loved. So I wanna read it out loud and then get you to expand a little bit on it. So

The the comment was basically related to feminist economics and said when an eight thousand dollar legal memo gets replaced by a twenty dollar a month subscription, Gdp drops by ninety nine point seven percent while output stays the same. Feminist Economics has built tools for this, satellite accounts, time use surveys, and replacement cost evaluation to fix this exact measurement failure. Replacement cost asks

What could this output have cost on the open market? That's how you make a twenty dollar subscription producing 150K of consulting work visible to national accounts. If we don't borrow these tools, the AI economy is gonna look like a recession while being a revolution. This is like I think a super deep insight that as a young parent I'm kind of living every day with my wife a little bit. And it's like

It's super just my personal take, it's super underappreciated how much work gets done that's not measured in GDP. So Are you anti GDP? Are you pro GDP? Like can you expand a little bit on this tweet? I think the thing we have to really understand is what is GDP, why did we make it and and what is it useful for? Because GDP was created during World War Two. in order to figure out how much of the economy we could use to make more bombs.

Planes, ships, etc. It wasn't trying to understand this ivory tower idea of The economy. It was like, okay, if we make all the factories switch bases, like how much stuff could we make? And that's a really different thing than how much valuable work is being created. The economy. How much valuable work is being created in the economy involves

Every stay-at-home parent. It involves every parent who is taking care of ch kids. Every time uh y you take any sort of economic action inside of your house that somebody else could do. Mm-hmm. Uh GDP, but it is economic output. And AI is gonna behave a whole lot like that. This is this is something that's really, really weird about AI, because you can have as uh the the example I love to give is a will.

Uh ten years ago if you wanted a will, you go to a lawyer, you pay the lawyer five hundred bucks, they write up your will. I I have a nice boring life, I got a couple of kids, it's a simple will. Now I want a li will. I get my boilerplate will and it comes out of my AI that I'm already paying twenty dollars a month for. It's done. It's who cares? Like it doesn't need to be any better than that. That took a$500 thing and it went to zero.

There are so many tasks like that within the economy, and a lot of them exist in the business-to-business space. where it's particularly hard to measure and this is something I really am hoping to be able to bring out to the fore because there's a lot of activity where the only place we could hope to capture the economic value that AI is bringing in profit market.

Because that's the only place. And realistically, we may not even capture it there because it may show up as deflation, but it may vanish. Because the economic transactions are going to be going away at the same time that the price goes down. And so the data as we currently collect it may not be fit for purpose.

to understand what's going on. And we're doing this at a time where we've greatly reduced the staff size of a lot of these statistical organizations. So there's a lot of work to be done uh by

uh kind of government economists and I'm hoping to work with them. I'm hoping to talk to them. Like, how can we take these ideas? I mean, we have fifty years of amazingly rich history from feminist economics basically laying out No, you've got to measure the whole production pile. It doesn't matter if it's paid or not paid.

its economic value no matter what. It doesn't like currently GDP, if I go to the grocery store, I buy some uh groceries, uh uh I spend fifty bucks, I come home and I cook a uh six course meal for a bunch of friends and it would easily cost eight hundred bucks if I went to a restaurant. I'm a pretty good cook. I'm a I'll I'll throw that out there. Um six friends for fifty bucks. Good job, dude. Oh yeah, no no they're eating beans, but they're gonna be damn good beans. Um

Yeah, well uh yeah, my my my numbers might be presidential in quality. Like I hasn't sort walked into a grocery store in a decade. Um but uh But yeah, so GDP captures only the grocery store bill. It doesn't capture the actual value added. And it's going to be like that, just writ large for all of these activities, because in the past

Uh these are BDP transactions in a big way. Um so many things I mean Are we really going to see people paying for the same level of HR services, the same level of kind of transacting all of these elements that are being replaced? very rapidly. And that gets very strange for the economic data. It gets weird uh on the GDP side, so it could look like a contraction, and it gets even weirder on inflation. So let's go back to this example of a will.

Yeah. My will goes away. My my five hundred dollar will goes away. Bill Gates and you wanna do a will. You're not tossing it into Chat GPT. You're going to a law firm and you're going to do the best will you could possibly do. And they're gonna be using all of these tools and they're gonna they're gonna be charging you for it. So all of the cheap ones go away.

And the expensive complicated ones go up. And so we might see a situation where it looks like output goes down and the price level is going up for things like this. And that is the exact opposite of what will actually be happening. In the actual economy, the number of people who has wills goes up because now they're cheap and free, and the price has gone basically to zero for almost everybody. But The measured economy gets it entirely wrong.

Uh uh right Ralph I have a question. Uh you know, we as I analysis we always write about newsletter. What would you w what would be your first newsletter article about? Do you have uh do you uh any any topic in mind? Unsecute. Yeah, just off off the top of my head. Yeah, so this is one of the areas uh that I really want to explore. I I've been kicking this around with Doug and I want to go a lot deeper on this.

Um but I think the one that I find the the most exciting and it's gonna be a really nerdy cut. because everybody may have heard uh all of these um the the main basically there's a a weird thing going on. The market is moving like AI matters. We all know AI matters. If I go onto my Bloomberg screen and I look at what the Fed thinks is going to happen to growth in the next three years. They have twenty twenty six two point nine percent because CapEx is up from this AI stuff. That seems reasonable.

And then it goes back to uh 1-9 and 1-9, the next two years after that. I I go in and I look at what the major banks are are saying is gonna happen in GDP in the next two three years. It's the same story. It's it's kind of a little bit higher now and then it goes away.

And that's because the economics industry and this is why I basically realized I had to rebrand myself as an AI economist as opposed to a strict macroec economist, is because The the economics industry is built around what's called a dynamic stochastic general equilibrium model, and that is an awful name and you can forget it now, you've heard it and it don't D S G E. Basically, it takes a bunch of assumptions about the status quo of the economy.

And those are its priors. And it's very, very hard to get the model to move away from its priors. Everything basically tips towards stasis in the model. And so we're dealing with a situation that doesn't look static at all. And that's going to mean that these models break. Models by their very nature, whether they're Bayesian, which is my favorite category, or st or uh

uh non-Bayesian are biased towards events they've seen in the past. We are dealing with events that we haven't seen in the modern era. And and basically the I I want everybody to be able to conceptualize the magnitude of event. We have seen events this big.

It's much more like farming and the Green Revolution than it's like anything that we saw in the twentieth century. We went from ninety seven percent of the population was a farmer to now two percent of the population in the United States is a farmer. We have had big events before. We will have big events again. But if you were sitting in 1750 and said, How much do I think this synthetic fertilizer is going to change things?

Everybody would laugh you out of the room if you said, Oh, it's gonna put ninety five percent of people out of work as farmers. You would have said, Oh, so everybody's gonna starve to death. And the world is ending. Okay, that's a crazy person. Don't talk to me anymore. And it's it's that magnitude of effect that we're dealing with, and it's really going to be hard.

for any of these traditional models to capture it. So I'm thinking more in terms of how can we demonstrate when they break, as opposed to how can we model this uh out of sa uh out of period por thing. Because ultimately Saying where the economy goes with AI Um I as an economist can't answer that because that is fundamentally a question of how good can AI get?

Do I believe in a singularity or do I say, oh, if there's gonna be a bottleneck that I can't see eighteen months out where it just like that it that's as good as it gets, I guess we're done. Um that's not an economic question. That is fundamentally a technology question, but I as an economist can look at the history of things and say this is a breakpoint. It doesn't look like economics as normal over the past, say, 75, 80 years.

I think that's a great spot to end guys. We're we're a little over our allotted time anyway. Malcolm you kept branding yourself as an economist at the end. I think we gotta switch up the verbiage to tokenomemist or something like that. We've got uh hope everybody enjoyed the show. We got Ray, the memory man, uh Professor Malcolm, and uh

Yeah, myself and and Joey dropped off a little earlier. Thanks again for uh listening in and we'll catch you next week. Next time Malcolm's on we'll talk to him about UBI and uh the singularity and all the other things that he loves talking about, so That got it wrong. Yeah.