How Nvidia's Recent Acquisition Changes the AI Game Feat. Morgan Stanley's Head of US Thematic Research

We got a DM from someone asking if we had an episode about the bankruptcy of Lehman Brothers and I went and sent her. I think it's like episode eight or episode 11 and I have not listened to any of our early episodes like how Adam Driver never watched a single episode of Girls and hasn't seen himself in any of the Star Wars movies. Like he's like a method actor so comparing myself to Adam Driver is a little bit wild but I was like oh god it's going to be so cringe. We are still immature podcasters so let's be real it's not like we're such seasoned professionals but I remember being so nervous and how we would edit every single breath of every word. We scripted the first couple of episodes like literally. No this was not scripted yeah and we finally got off of the script. Yeah it was one of our first off the cuff episodes and I bit the bullet and I went back and listened. It was so good not to not to hype us up but it was so good and one of the things I realized that I missed is we used to chit chat about everything that was going on in our lives when we were going to see each other next but we were doing and I think we got away from that because we were like no one wants to hear about us. No one wants to hear about like our stupid lives or our families but guys this is our show so you might end up here a little bit more about us from time to time. We just have so much going on that we haven't gotten to talk about and I think it's weird to be on our Instagram stories like hello. I mean I think at the end of the day like people can fast forward we figured out how to do timestamps which is so exciting like it's so funny we would put up these YouTube videos and everyone's like please add timestamps so I can find like all these things we're like we don't know how because it's literally I mean our business is basically still just the two of us and so we did and still are doing all of the recording the editing we do all this stuff for the podcast I mean for Instagram for the podcast for a sub stack and it's literally the two of us Jen you obviously know her feelings on utilizing AI so while we use it for like some research like that's pretty much it research and like we use it for the descriptions of the podcast services like I will say if AI can edit and do all this to buy all means I would love to find AI that could sufficiently edit our podcast but unfortunately when we need to reconstruct a sentence when you're explaining something highly technical and I was like it seems right I used you I scraped your content to make my LL live so this seems right and it's like this is what I was saying well that's having a stroke well that was why we were joking that the reason that it's so important that we go back and listen to our own stuff is sometimes I mean at the end of the day like we will say things like well we'll think we're saying something the right way it's like this was higher than that it'll be like two it's greater than one but we'll say the opposite one is greater than two which we know but then we're going back we're like oh gosh well was I having a stroke so we do need to go back and make sure that the stuff that we're putting out is accurate and so it's just been so hard to find any producer or editor who like has the technical chops and so anyway we have some like really exciting stuff coming on the horizon not going to give too much away but just yes well and news coming what's even more exciting is we've gotten to see each other a lot more often lately I do not like flying I'm not a good flyer I need to be tranquilized like a dog but I I don't know why I said like a dog I don't actually know that people tranquilize their dogs but also I don't if you're like I tranquilize myself what I've if you don't have like a petite little lap dog I don't know people they put them in the cargo underneath the plane like I don't really know how large dogs travel like if we ever needed to move with Renee our dog I have no idea how we'd get her somewhere yeah I have no idea I feel like I mean I would probably end up driving because I don't mean but yeah to your point I have no idea what if you want to ever see like do you put a dog on like a boat that's a good point I have no idea I mean this is not like you do not have a dog my kids are constantly asking they're like because you need one more thing to take care of yeah and I was like she's funny she's like can we get a dog and I said what we have we have James we have a 15-month-old and my six-year-old goes can I trade grace for a dog so basically could she trade her four-year-old sister I was like no I think grace started to cry but it was just like again kids are wild like I don't like my sibling can I just get a dog instead I think you would have said the same thing about your sibling when you were five or six years old yeah yeah Brian my brother best friends now we like hated each other when we were kids and oh yeah I was always like can I be an only child which again I'm so grateful that I am not Jen obviously you were so I was and and has not changed you got you got some sister-in-law so I have and and you are one of my many sisters by choice not but yeah so we're going to be a we're going to be together again in both Boston and New York City this coming week somewhat unexpectedly so I'm so excited about that that's going to be awesome I'll get to see your family and we've been together almost every other week we're going to start getting sick of each other at this point like I don't think that's possible we already talk on the phone every single minute of the day I was also joking my eight-year-old said who do you talk to more dad here Jen I was like definitely Jen but also it's like I mean your business partner so like business hours are like much more during the I mean it's like the entire day and then like you see your spouse usually it's like at night when they come home and then like well obviously text but actually we we talk about like the utilization of news and my husband was like go into my email and find all the email all the messages that I have sent I mean usually because it's on like logistical things and I'm like oh I can't deal with logistics right now she's like all the logistical emails I have sent Christian and like summarize them so I can send it to her and be like here's all the shit you gotta like you gotta get done you know because we do bifurcate it's like I kind of in charge of all the school stuff he's in charge of kind of like more the admin around the house so like that's our diversification of labor but he will get like this school emails me like hey did you like donate to this pizza party I'm like oh not yet so yeah anyway we're Venmo moms we're not remparent moms I did my time as a remparent mom in preschool no no more I am not you cannot rely me for that kind of thing but so personal lives aside is excited as I am to see you next week we actually had the amazing opportunity yesterday to chat live on market hang on Yahoo Finance which was a really incredible opportunity we'd never done a show with those guys before they're so talented we were joined by another incredibly talented thought leader in the space and I honestly felt a little bad for everybody because I was like oh well Chris and I like we know each other thoughts like blah blah blah blah and I think one of the hosts at one point was like oh no Jen said something and Kristen's gonna he's like Kristen you ought to back me up and you're like no I'm oh yeah I appreciate that we share a brain we do but so the conversation we had yesterday on market hang centered around the announcement about Muse and we have been having a lot of conversations about all the different frontier models and the companies that are running them who are about to IPO and how significant that is for the financial markets and we figured it made sense to sit down and have a conversation about some of the big strategic things that are happening that are really moving the markets and are also shaping how the rest of us are going to live our lives money you think about normal mergers and you're like okay well bank of America bought Merrill Lynch my day doesn't get that different right it's like hey guess what if these companies are the ones who are deciding how computers may or may not kill us all we might as well pay attention to the underlying mechanics of it and we can talk about that whole debate as well but you've been doing a ton of research about specifically a couple of really notable announcements recently Kristen I've loved you could kick us off by talking about what's going on within video and hugging face well actually I do want to back up because we we have talked about how obviously everything that we've done has always been like very specific in the weeds with like what's going on with the markets and what's going on with again rates or what's going on with this particular merger and what's interesting about this merger is I mean look I used to teach M&A to like new hire bankers I'd teach it at like a business school like I used to teach like how again companies would think about buying a company and it's like the the sort of traditional analysis you're going to be doing as an M&A banker is it's called like an accretion delusion analysis right you buy this company and like does your earnings for share go up or down in video going out in acquiring hugging face is is not being done for like the earnings for share and so what's interesting is that it was very much a strategic acquisition but it's centered in kind of this like the ecosystem like this is a huge topic of conversation which is like open versus post weight models and we're actually going to get Michelle Weaver who is the head of US thematic equities on later to talk about kind of the actual adoption of these various models at different companies and kind of like what they see as the future but again this bifurcation between open versus close wave models is a critical thing for understanding kind of like where the world is headed and so most of us are likely utilizing you know plot or chat GPT I mean I know we have subscriptions to both like full disclosure we use both but you in theory don't necessarily need to pay like one of those companies you could download what's called an open weight model to and like run on your computer if it's small enough or get compute from the cloud and like set up a model and you know then tweak the parameters if you want as well I mean like there's a lot of kind of customization that you can do but again understanding like the fact that there is this bifurcation in the way that people can utilize models is obviously paramount but then there's one layer more which is like what is going on in kind of the ship space and so meaning when you are running these models it requires compute so it could be like literally from your computer I mean I went back and I was actually curious we bought a little bit think pads about a year ago, I think when the tariffs were for a snap. So I was actually curious, like, what's the memory? What's the storage on our think pad? And it was like 64 gigabytes of memory and like two terabytes of storage. But it was like right there, right? It's like in video, GPU like processor, whatever. The whole point of this is that if you are going to be doing any of this open source model, you do need compute. And so that brings us to where Nvidia fits in because the compute that is kind of out there, like everyone has been talking about like data center construction. We never did like a proper long form video on this or we never did like a long form podcast. But Jen has this like, I got amazing explainer where she kind of breaks down the $500 billion project that in open AI is building. I believe it's in conjunction with Softbank and who was the other partner in video is the one who was providing basically the financing backstop for right, right? And their commitment has shrunk and shrunk and shrunk. Well, and that's another whole loondoggle we're going to get into at some point, which is kind of like the circular financing going on within video. But sorry, back to the point that I was making is like at the end of the day for these models for anyone to use AI, you need compute. And there's two different ways you can think about compute. There's compute needed to train the models at the beginning. And then once the model is set, there is a whole process that they go through to train the models. I think of it as like similar to like an auto complete machine, but much more robust. You can read our whole subs deck. We kind of get into all of that. But once the model is complete, you then have literally just like a file with like billions of these parameters that then can be used. And again, if you were an open weight model, these companies will like upload it to again, that company called Hugging Face. That's kind of the central repository where everyone goes to upload the models and then get the company's upload, but for individuals to download them. And then they need some like software. It's called Serving Software to actually run to run the model. And then obviously you need the various chips. But again, back to this point, you have training and you have inference. And so there are certain chips that are designed for training. Although if you're in video, you actually have one general purpose chip that can be used for both. But most companies like you look at Amazon, they have Trainiam. And then I think it's like Infrarentia. Right. So there's different chips that a lot of these companies build for training versus inference. And as we're getting like these again super capable models that are out there, there's less compute that is going to training and more compute that is going to inference. Because more inference means that you're like actually interacting. You're like a customer and you're interacting with the frickin chatbot. Like it's like having a conversation, hey, chat GPT. Like, where should I, you know, I want to buy this thing. Like my daughter wants to be Taylor Swift for Halloween. Like I need to get this very specific look. Like where can I buy that? Or explain to me how XYZ thing works. So the point is there's over the, you know, the foreseeable future, there's more compute that is going to be moving towards inference. And again, away from training and on the inference side, a lot of companies like your hyperscalers, the anthropics, the open-air, the frontier model companies, they are starting to build their own chips. There are companies out there like etched who are also building inference chips. And so the point is that if you're in video, you're sitting here and you're like, okay, so the final customer, right, I have obviously they have massive contracts with all these various players. And by the way, on the training side, they still dominate, and on the inference side, they obviously are still a major player. But as you have these companies with tons of money, they are starting to be like, you know, Nvidia, their margins are quite high. Like we can start to build our own inference chips. And so one of the things that for Nvidia that becomes important is like, okay, well, those customers might be doing their own shit on the inference side. But what about everybody else? What about all these random developers or, you know, the Wall Street Skinny? Like we decide you know what, Jen, we're going to build our own little internal way model. Like we're going to download something. For those of you who are just listening, I am seriously shaking my head. This is quite literally my worst nightmare. But the point is the more people that are going to be going open source, then means there's more potential demand for Nvidia chips. Depending on the software you use, and again, like the compute that impacts what chips you're utilizing. But if you're utilizing the like Nvidia specific software, then it's going to run on their chips. And so it becomes this whole thing. And by the way, it's so interesting because back to the whole point, we need to do a whole thing on like CUDA and like why Nvidia has become kind of like what they have. It's not that other people can't build the chips. Like other companies AMD, like they can build these same chips. Really has to do with like the software and, you know, developers. They know how to code on these, you know, on CUDA and utilize it to kind of like be able to communicate with the chips. Like if you want the chip to do something, you have to go through freaking some kind of software. And the software that Nvidia has created has been adopted so much more. I mean, developers know how to use it. It's a skill you can put on your LinkedIn profile. And so that is again, like a huge moat for them. But the whole point is for Nvidia. They are buying hugging face. And it's not because of the freaking, you know, it's going to increase the rps or someone goes, Oh, it's all about cost savings. No, this is not going to cut. This is not saving them money. It is about ensuring this customer diversification. It's interesting like they can actually kind of see the data. So it's like, who is downloading what types of models? Like what is the software that's being utilized? Like what are some of the changes that are happening? So Nvidia is basically trying to own the platform. They are kind of buying the company at the heart of open way AI. And then from there, it kind of gives them one further mode. It's like they have the diversification because all these developers are kuda familiar, right? And so that's sort of their mode on the on that side. And then here, it's like if you can also own the distribution platform for open weight, it gives you a little bit more of a mode on the on the inference chip side. So it's really interesting. And I'm going to shut up for a second, Jen. Well, one piece of jargon that I really want to clarify. And it's funny. I think people accidentally use these terms overlapping and don't necessarily understand the difference is the distinction between open source and open weight. Can you clarify that? Well, because it's funny. You actually said open source at one point. Like, and I think, but I think it's a natural thing because we hear open source so much more than we hear open weight. But I think in all the research that we've done around this, what's been really interesting is like, open weight is ubiquitous. And there's tons of open weight, lots of things that people are familiar with is open weight. Open source is a very specific designation that may not apply how you think it does. So can you walk? Well, that's a really good point. I will say, I do wonder if the word is going to start like people will just say open source. And it means open weight, because I've heard again, I think I already have on all of it already says open source. And it's not open source. So yeah. So again, but I do think like in the vernacular people will say open source when they mean open weight. Again, I did it right just now, but open source means like you are not only sharing like think about what is what is required to train these models. We go through all of it in our sub stack piece if you guys want to read about it. But basically like they will take some sentence. They split up the sentence into like these little, it's called tokens, because like you can't do math on a word. So they split up the tokens. They then will run it through this model where there's all these various parameters. And then initially when they first do the first training run, it's like the parameters are random, got the random numbers. And they then will spit out like, okay, what does the model predict the next word would be. So if you have like the dog is sleeping. And your initial tokens are like the dog is. And so that's those three tokens go in the next word. It'll come up with this like probability chart. So it's like sleeping, barking, running, Jennifer, right? And so like it'll go through. I know. Maybe he's debated on the player. Now I am the dog. Google search bar and you start typing the sentence. And it gives you all of the things that you might conceivably be is how I exactly in my mind. Right. And so you then have the various probability chart. And the answer though for this, because you have the actual sentence, you know 100% certainty. The next word is sleeping. So then when it ultimately like you see what the little whatever prediction chart is, the 100% is sleeping. So it nudges all these parameters, which are kind of like random. Like nobody, I don't think people necessarily understand exactly like how that all works. But then they'll it'll nudge it a teeny, teeny, teeny tiny bit. And so then the next time if you were to put that same like the dog is, it would come out a little bit higher. Like all the parameters get nudged so that it's a little bit closer to to sleeping. But anyway, it goes through all these runs. And then you ultimately have this like nice prediction machine, right? And there's some other things that go on. There's like what is it called? Instruction tuning and reinforcement learning. So there's additional stuff that happen as well. But at the end of the day, what ultimately comes out is this file with a get trillions of parameters that then get for or should say billions trillion. Like it depends on the size of the model. That's literally at the end of the day, like a huge point of contention. I think that Fable is estimated to have like five to 10 trillion parameters. But anyway, so you have this file. And that's then what the weights, right? The numbers are what get shared. So that's open weight. If they if the if the company were to also share all the training data, that would be then open source. Most of these companies, even the lot of the Chinese companies, they're open weight, not open source. And so but so when people say open source, again, just like I did, I meant open weight. But again, people will often just like throw the term open source when it really for most of these companies, they're doing open weight. Yeah, no, that's a really great point. And so at the end of the day, how does Nvidia making this acquisition pose potentially a threat that people haven't necessarily conceived of to the current or I guess a former as of like one week ago, because things are changing in real time, leaders in the frontier models, like the open AI's and anthropics, what kind of strategic or economic threat does it pose to their business model? The frontier model companies make money by selling access to their model. What's interesting is whether if you're a hyperscaler or Nvidia or whatever, you're making money in other parts in like that five layer cake as the gents and mong calls that there's the there's the energy, there's a chips, there's a structure, infrastructure, the model is just one piece of it and then there's the applications. And so they are making money obviously with the chips, they're obviously building into the infrastructure and like they tried to become like a cloud provider, the cloud providers we typically think of are more like AWS, Google Cloud, Azure for Microsoft, but the point is that like these hyperscalers and again a video, they're making their money at other places in the stack. And so for them, it's like give away the thing that by the way all these guys are are the frontier model companies are spending a ton of money to train these models and here you have these other companies just giving it away for free. Now the frontier model companies, their best models are still better. You kind of joke that it's like Teemo version, right? There's the like you can go and you can buy the Prada bag. And by the way, I do think like we talk with this with Michelle, like there are going to be people that want the best and like the most robust, the cutting edge model and that's they need it. But you're carrying cancer. I don't want you doing it on Teemo Claude, okay? Right. Like let's go with the most robust, most sophisticated, whatever it is made by whomever. I don't care who it is. Right. But back to this point, if you are Kristen and Jen and you're running your business, like do we really need access to Fable? I don't know, right? Oh, I don't need or what. You definitely know Jen's like I don't want any of this on subscribe. But my point is for the actual applications that we're doing, like and the easier that it is to actually like run and utilize these other open-weight models, that become like it's literally just like people want to do what's easy, right? And I actually think back to Muse. I think that meta is ultimately going to like, they're at least in the kind of winning the AI race in my opinion right now with this launch of Muse because they just make shit easy to use. Like people just fragged everyone. Yes, people just want what's easy. And so we had, I mean, literally we're in this WhatsApp group with all these women who are like trying to learn how to like how to optimize everything with AI and everything. And it's great. But it's like, that is a full-time job. Like that, like to learn how to use agents and to learn how to use all this stuff and like blah, blah, blah, blah. And it's like, I just want something to work. You know, you and I, we're going to New York and our favorite thing, like we love to go to Hellstone. You never freaking get a reservation. But if I said when I was trying to go to Hellstone as our favorite thing, let's see. That's true. That's not our favorite thing about going to New York. I've been twice. Well, let me be very clear. Sorry. I love Hellstone. Because I, my husband and I, when we met, we lived a block away. And that was our go-to spot all the time. And out of nowhere, it has now become like the impossible to, like that was our backup place. If we were like, we want to go to, I don't know, ABC Kitchen. Like it was impossible to get a reservation there. But we're like, let's just go to Hellstone and we just like go and like sit at the bar and have a drink. Now it is impossible. And like ABC Kitchen, we'll go, but anyway, we won't get into the martini glass debacle. So the point, though, is like if there, if we can now utilize whatever news and just be like, Hey, we're going to like get us a reservation as soon as it's available or whatever, right? We're finding you a good restaurant. Like we haven't, we're going to the same restaurants all the time. Like this is our specification. I don't know that we can specify the martini glass shape, but like that should be. Again, what is AI good for if not specifying who has the most delicate traditionally shaped stemwear, please? Like that's my criteria for a restaurant. And I will say we ate some good food on our last New York trip. And we got to meet some of you guys. It's awesome. If you guys see us, feel free to come up and say hi. Like we love meeting you guys. And we were so fortunate to get to run into listeners and people who are following us on social media. But yeah, I want to your point. I don't want the Android version of AI. I want to plug and play iPhone. Just one of the apps, one of the things I want to do, make it as simple as possible. And I'm at the point where if I want to go any further with AI, then the most basic chatbot functions, which for me are just search. I'm just using Claude. Like I would have used Google before Google redid it search function. Like it's traditional search function. I find very difficult to use. So mad. I Google the freaking TV show last night. And it literally in the AI thing gave away the ending. And I'm like, no, I just wanted to leave a show. Well, but you know the ending. I'm shocked. No, I didn't know this one. I didn't want to know the same thing where like the beast, the beast in me with the yeah, I didn't want to know what actually happened. But I wanted to know like what the plot is. I'm like, here's the thing that Netflix is heavily pushing on me. Is this worth my time? And I just wanted to understand like the framework of what the show was about. And it's like, yeah, and the killer. And I was like, no, no, like wait, so what show were you googling? It was like him and her, his and her. I don't know. It's literally just like when I opened up the Netflix app last night, I was it that was like the big thing that was like it's and I'm like, okay, what is this about? And so need a new show. But we digress. Sorry. And I think when it comes to AI too, right, I think that's where companies are struggling with adoption. Because they're looking at it at an enterprise level. And what you need for say, like I use this example of like researchers trying to cure cancer is also very different from like what an investment banker needs when they're doing one kind of analysis versus someone who's a traitor in another seat who's doing another different kind of analysis and needs a different type of functionality. And I think this is what's going to lead to a really interesting conversation with Michelle here. Michelle Weber who we're bringing on shortly is head of thematic equities research at Morgan Stanley where she is looking not at hey, what are we doing in the financials? What are we doing in the industrials? She's saying, okay, let's take a theme like I don't know AI. I mean how it applies across all companies globally. Yeah, and even more than that, it's because like a lot of times it'll be like specific, right? It's like you're covering Apple and you're covering meta and you're covering like just these companies. And to your point, it's even broader than just the industry. It's actually now like AI broadly. And so she's very focused on AI adoption, which I think is such an interesting thing. Like we used to joke that companies if they literally just put AI in their 10K, it's like their shock rise would like go up. And so the question now is like, okay, but like how are you utilizing it? And I think especially as a lot of people in talk about like token maxing, it's like it's not enough to just show that you're like burning money with quad. It's like we could burn money with Delta and be like we're going to pick the most expensive flight to like fly across country. You know what I mean? Like the goal is not just to like burn as much money as you can, right? Which which previously was what a lot of companies were targeting. I know like there were people in your life where they worked for a company and they were like we want to see how much you are spending. Like we want you to be using this. Well then because the companies were spending the bare minimum amount and even that wasn't getting used. So they were like use it more so we can justify this because if we to your point, put AI in our earnings call, our stock price will go up. But like if you're not using the AI, this is all kind of silly. But so yeah, so we're going to get into adoption. We're going to get into obviously the big thematic question of what is the actual impact on the labor market and Michelle is going to quantify that for us in real time. We're also going to talk about you had references earlier, the ever controversial topic of data centers and some of the constraints on the AI build out. And you know, it's funny. I saw the craziest headline, Kristen, about data centers overnight. I need to look this up. I'm going to pull up our trustee news source, which is ground news. So did you see this oracle is now declaring force measure on its new Mexico Stargate data center? So this is a really big deal guys. If you aren't familiar with the term force major, it's basically like active God, but like financial, active God that prevents you from fulfilling your financial obligations under a contract oracle of all of the AI companies that are part of this build out has definitely been the shakiest they have had the hardest time in terms of justifying some of the really early cat backs. And what's so great about ground news is it allows us to see this, but it also enables us to see how many new sources are covering this. This is a big piece of market moving news. This is impacting bloom energy, which is one of those big companies that's a central part of the AI build out. And we can see if you scroll up to the right, Kristen, we can see how many new sources are covering this emerging story, what their bias is. And it's funny because data centers seem to be the one thing that everybody can agree on politically, everybody. So it's the aisle, which I think is so wild about this. So if you guys want to learn more about ground news, you can go to groundnews.com/skinny and you can save 40% on your all access subscription. So with that being said, we are going to bring on Michelle Weaver, who is head of U.S. Thematic equities research at Morgan Stanley, our former alma mater. And we're going to have an awesome conversation about all things AI. So let's go. All right, we are so excited to be joined today by Michelle Weaver, who heads up U.S. Thematic Equity Research at Morgan Stanley. And so we have her on today to talk specifically about AI adoption. So I love it, Michelle. If you can start just by giving us just a quick understanding of like what U.S. Thematic research actually means. And then we can dive into the conversation about AI adoption. Great. Well, thanks for having me on, Kristen and Jenna. It's great to be here. So Thematic, it's a relatively new discipline. And I spent the majority of my career on our U.S. Equity Strategy team. And so on that team, you're thinking about looking at these different industry verticals relative to each other. Thematic is essentially flipping that and looking horizontally across industries at these big themes and trends, shaping multiple industries throughout the market. AI, of course, being the biggest, most dominant theme right now. And I know you would talk to me about how one of the things that you guys do is you map out all the various companies that you cover from what you call the enablers to the adapters. And I'd love to understand a little bit better, like what that means, like how do you think about which company is which and what you guys found in the latest report you put out? Absolutely. So it's a big, by annual project and it's very collaborative, the analyst that covered these industries know their company's best. We cover around 3600 stocks globally. And so we ask every analyst for every stock under their coverage to rate both the exposure to AI and then how material AI is to the overall investment thesis. Since AI is such a big theme, it touches almost every stock we cover in some way. And the cleanest split, I think you can think about this, is some stocks are enablers, so they are essentially building the infrastructure for AI to exist. It's everything that goes into putting up a data center, whether that is the chip makers, the electrical components that go into that, the power providers, the cooling, the networking, anything that goes into setting up and running a data center or anything that is around the development of the models. Then the adopters are essentially, who are we building the technology for? These are the enterprises that are adopting AI into their business processes and looking to defuse that technology throughout the market. And so in some way, shape or form adoption or enabling of AI really touches all of these companies. And then it's important to understand how material AI is to the overall investment thesis. And what we found in our most recent iteration of the mapping is you're increasingly seeing the benefits from adoption make their way into EBIT margins. It's still relatively early, but you are seeing four companies that are analysts are tagging as high materiality AI adopters. The median stock in that group, they're seeing around 4.6% EBIT margin expansion over the next year. And that well outpaces global indices. And I'm curious, when you say the AI adoption, are these companies that are incorporating AI to go out and grow sales? Or is it more of the companies that are using it to maintain efficiencies? We'll obviously talk a little bit later about labor efficiencies but cost get by with less head count. I'd love to understand that a little better. Absolutely. So it's really challenging to think about AI adoption. And I think part of why we do this mapping is it looks so different depending on what industry we're looking at. So an AI adopter in one industry is going to look very different from an AI adopter and another. I was at our industrial conference last week. And a lot of those examples of AI adoption were around things like autonomous trucking to make truck drivers safer. Things like building a digital twin of your processes to do more predictive maintenance and better understand when something is going to fail before it fails. If you're a retailer, it's thinking about how to make your inventory processes more efficient. It's thinking about how can we resolve our customer complaints better? How can we put chat bots and put other things on our sites to handle more of this customer issue volume? If you're a media company, it's how can we use these tools for editing? How can we use these for image or video generation? So it's super different depending on what industry you're looking at. And some of these are going to be on the more revenue-facing side. Some of these are going to be more on the cost-facing side. Right now, it's more about cost, but I think there's revenue opportunities later. Can I ask a really dumb question? I hung up on kind of the bifurcation between enablers and adopters when it comes to you mentioned kind of the power companies that are part of this entire process, right? And we're thinking about the massive energy need when it comes to data centers. So I totally understand that. But what about companies like the ones that come to mind are obviously the ones that we talk about a lot that are super funky like the totals and the aginomotos where it's like we make toilets, but also key processes in parts that are related to the AI buildout or like we make MSG, but also this film. How do you think about companies that fit into multiple categories? Those are super wonky ones, but specifically when it comes to power and energy. So I think it's yeah, there's absolutely this overlap. And when we do our survey work, when we do our mapping work, analysts are able to include stocks in both buckets. And so the way we think about are you more of an enabler or more of an adopter is by rating that materiality independently. So maybe you're a it's absolutely courteer thesis that you're an AI enabler, but you also have significant exposure to the adoption side. So a lot of exactly to your point, a lot of these companies are doing both. And I think pretty frequently some of the best adopters are also the top enablers. These are the tech firms that are building out the technology. They're going to be best able to implement this technology first. They're getting high on their own supply, you can say Michelle, it's okay. Well, no, I mean, it's like look at Muse. And so as we're thinking about the utilization of these AI models, I'd love to talk a little bit more about what you're seeing as it comes to open versus closed weight models, we just had a video announced that they were buying hugging face. So they're very gung ho about promoting open weight. And it's interesting because I had read a lot about how this is going to potentially be dangerous for some of the frontier model companies, but there is a thesis that actually the more open weight adoption, there is actually it becomes better for some of the companies that have those closed weight models. So can you talk a little bit about that kind of very interesting dynamic? Yeah, so the open versus closed model debate has been huge among investors recently. And just zooming out, I think when we say open models, some people think open weights, some people think open source. So there is there is an important distinction there. Open weights models are essentially models with the parameters or the weights able to be changed by the user. So what you're seeing is companies using these open weights models and then working with their engineers to fine tune the models for their specific use cases. Open source models are one where all of the training data is also available. So it's the complete recipe and all the ingredients are available. The majority of what you're seeing enterprises use though are these open open weights models. And I think ultimately what you get is a hybrid world. And to your point around, you know, are we going to see more adoption because of these open weights models? I think yes, as you drop the cost to use these, you're going to see more and more adoption. And this is essentially Jebyn's paradox in action. And I think where we end up on this open versus closed debate is some sort of hybrid world, somewhere in the middle. I don't think you're going to see one group completely dominate the other group. But you're seeing enterprises already using a mix of both of these models. Mackenzie's studies show that around 63% of enterprises use some blend of open and closed models currently. Michelle, I am the dumbest person in the room. What is Jebyn's paradox? So it's this idea. It's an old economic paradox. It comes from the 1800s. It was originally around coal, but the idea is as you drop the cost to use something as something becomes cheaper and the technology defuses, you see more adoption. Oh, okay. We'll add one thing here. It's a fantasy economic game. No, I was going to say there was something that I found so fascinating. We had done an interview with John Quinn, who heads up Quinn Immanuel. And we were talking about like copyright law and everything in like in AI. And I was not familiar with this term copy left. So for people who are like doing their startup, I will say if you're using open white models, like be careful with what the licensing says. So yeah, that's a super interesting point. The liability is a drawback there. There are a few drawbacks. There's also this concern about US regulations. So a lot of these models are coming out of China. And there's concern that the US may say, okay, we think that this is a national security threat. We think this is a security threat to the US. We don't want any American enterprises using models that are coming out of China. And that's part of the reason I think also for this hybrid world because businesses are worried about continuity risk. If you build a bunch of products, you build a bunch of processes on top of one of these open whites models and then suddenly you're not allowed to use it, that's going to be a huge problem for that business. I'm so curious, though, because it was funny. I was listening to in Ezra Klein interview, Jensen Wong. And I don't know if it's Ezra Klein himself or if it's actually the New York Times, but they use a lot of the open white models out of China. But obviously, if you're using the open white model, especially if you're running it locally on your computer, then it's not like sending anything over the internet. But is the concern, I guess, from the US side that the training data includes things that maybe we in the US don't necessarily want like so the outcome that is spit out is training on data that like maybe we don't agree with. So I'm curious like why there would be concern from the US about using some of the Chinese open white models. So there's a variety of ways you can consume the open whites models. So you can consume it via an API from the provider. You can consume it locally to your to your point there or you can consume it via someone else hosting your compute. I think it's still a little a little easy where exactly a lot of these concerns are coming from. I think there's just this general idea about, you know, we're in this big race with China over AI in which country will will reign supreme on AI. And I think there's just a lot of unease around, you know, using some of these technologies. What if your data is not as secure as you think it is? Right. Yeah. That makes sense. Kristen, you would send me a really interesting article from the economist talking about trying to price the basically risk to continuity infrastructure security, et cetera, et cetera. A kin to the way you'd almost price a cat bond, if you will. And thinking about trying to ensure over that for businesses, I am curious as you think about the enablers and the adopters. And you think about kind of these growth projections. What if anything? Are you building in as far as security risks and like the risk to your point that like, hey, listen, you build in this whole infrastructure on this open weight model. And then like people stop supporting it or like, there's regulatory risk that you think about. How do you handicap your forecasts for growth? Do you add any kind of risk adjustment there for like the increasingly loud, sometimes confused, off debated risk that like, not the risk that humanity has all killed three years. We are not ensuring against that. We can't ensure against the asteroid hits the earth tomorrow. But more of these more nuanced, like security risks, I'm thinking about things like regional banks, right? Like healthcare providers, big data providers, things like that that do have such security exposure. So I think for those companies, those are exactly the ones that want to be using these from tier models. They want to be using the best models possible. They're working with cybersecurity providers to make sure that they're protected. You've seen when when you've had some of the releases around the best new frontier models, they're quite locked down. It's not everyone is getting access to these new great frontier models. They're saying, okay, we need to make sure the banking infrastructure is safe. We need to make sure healthcare data is safe. And those are the exact firms that the frontier model providers are working with to give them these models to then go and shore up their own cybersecurity. So I think that, you know, that's almost an exact example of why you'd want to be adopting one of those models. Interesting. Yeah. Well, and I'm curious to because we had talked about just this idea that there's going to be a hybrid approach amongst a lot of companies. And I, you know, even like for us, Jen, I could totally see that making sense where it's like somehow we decide we're going to like utilize local AI, right? So it's like the open models so that it's like doing stuff that we need on our computer. So we don't have to necessarily allow meta to get access to our calendar. But then you use some of the open models that are again, much more powerful for like research or all these things. And so I'm curious like when you're talking to different companies, like what are I guess you're not talking to different companies. But I'm curious like what is the current adoption for the open-weight models versus closed-weight? Is it currently still more closed-weight because that was sort of like where things initiated? I feel that people are probably very comfortable with it. I know a lot of, like I have a good friend at PricewaterhouseCouper and he was saying how like they're constantly like max, like they're token maxing. They're maxing out the number of tokens that they can use with a fancy frontier model company that I can't say the name of. So I'm just curious like what the current adoption looks like if you happen to know what the split is. I don't know the exact split. Sixty-three percent I know are using some sort of blend. I don't know what percent of tokens are being processed by which at different enterprises. You are seeing though increasingly on open routers data that open models are taking share of total tokens processed, but tokens are just telling you about volume. They're not telling you about price. So even though if perhaps some of these closed models, they're losing share, they're being priced at a premium. So that's not necessarily going to translate into losing share of revenue there. But I think companies are using a blend because of that token maxing phenomenon you brought up. You want to make sure you're using the appropriate model for the appropriate task. You don't want to incentivize your employees to just use as much compute as possible. You don't want them to be using as many resources as possible like any other business process. If I'm going to go do a marketing trip somewhere, it's how many clients you're meeting with. What's the ROI? We need to make sure that the spend is appropriate for for what you want to do. It's it's like anything else. So now I think you're seeing businesses say okay, here's your token budget per month. If you go over that budget, why do you need to go over that budget? And it's not just we're going to reward you for being the best AI power user or you're using, you know, just the most tokens. Yeah. Well, and I guess like talking a little bit about just the incorporation of AI specifically as it relates to cost cutting because one of the things that so many people are worried about and this has been one of my main concerns is the ultimate end result for the labor force, especially for young people who are entering the industry, but we can just talk more broadly about the labor force in general. What are you guys seeing in terms of like headcount or needing less jobs, hiring less in terms of now that you're utilizing this technology? Like, is it actually showing up in the labor data? Because there was some sort of mixed reporting about that in the past couple of years. Yeah. So if you think about, you know, the overall aggregate labor data, things look pretty fine. I think what you're seeing right now is, is some of these shifts happening under the surface. So our economists last months at a really interesting report looking at jobs that are more exposed to AI, and how does unemployment look compared to what's typical in those roles? They're finding around a 50 basis point increase above what would be typical on unemployment for jobs that are more exposed to AI for the youngest group of workers. So I think that that's just really the earliest sign we've seen in some of that official macro data. We've also done a lot of survey work on our end. We surveyed companies that had been adopting AI for at least 12 months and asked about their hiring decisions over the last 12 months. And we found a 5% net job loss for companies in those industries. But what I, it's really important, we have to unpack that number. It's new hires, minus layoffs, minus roles, not backfilled. And a lot of the hefty heavy lifting there is coming from that last piece. So the roles not backfilled. So right now, I think companies are thinking about this labor decision. And okay, someone quits they find a new opportunity. The barrier for refilling that role is a lot higher than it had been in the past. We're not seeing outright layoffs yet. We're not seeing outright large groups of workers being automated. But it's starting to percolate. And there there are smaller shifts going on. And at the more senior levels, are you seeing this too? Or is it predominantly just at the junior levels? It's been happening more at the junior levels or at least what we've been able to observe in the data. But I spent a lot of time at some of our industry conferences. And what's interesting, I've heard this anecdote a few times that some of the people who are best able to help diffuse AI and push it to other workers are the junior people. It's the people that are AI native already. They went to college. They had these tools throughout college. They know exactly what they can do. And they're then able to go and bring it into their new workplaces and teach other employees how to shift some of these processes. Yeah. Industry specific, though, as you can imagine, we're a podcast called The Wall Street Skinny. We have a lot of listeners and financial services, consulting, et cetera, et cetera. Have we seen the analyst McGueden that has been forecast? Are you seeing this specifically within that sector? Or is it what we hear so often, which is like, I'm using co-pilot here and there. And I'm, you know, that's great because I'm spending less time, you know, color coding or whatever. But other than that, like, not much is changing. Yeah. I'm curious. Like, give us the give us the inside scoop there. So I wouldn't say we're having any sort of McGueden of anything happening right now. I think that the data, you know, looks, looks fairly stable overall. You know, you brought up financial services. So one one thing we did recently was an example of AI adoption itself was we used an LLM to read around 18,000 earnings and conference transcripts. And from that, we found around 10% of S&P 500 companies are explicitly discussing labor in the context of AI. And the most common types of discussions there are happening around refilling, not backfilling roles where people leave this decoupling between revenue growth and headcount growth. And the past growing headcount was a really important component for growing revenue. They're seeing that split, that's no longer the case. And then on a more limited basis, you're seeing examples of AI automation, so true automation of workers. And the areas where where these discussions were most common were financials, tech and industrials. That's so interesting. Although I do think the, I mean, it's interesting because like the the model is so specific when it comes to whether it's like the Morgan Stanley's, the Goldman's, the McKinsey's, like they hire an analyst class out of college. And like it's it's kind of a whole training ground in and of itself. So I actually am, you know, that particular type of role gen, I feel like that's what sort of we're curious about, because that's like the sort of path into the finance world. And to me, it would seem that these banks would still want to do that. But again, that, you know, so, yeah, I mean, you need to have junior people to ultimately have senior people. And that's why I think you can't see these jobs disappear entirely. But typically to your point, you had more of a pyramid type of structure. Maybe that turns a little bit more diamond shape for the bottom starts to shrink. Oh, diamond shape. I like it. The hand visuals and everything, right? I like that. You need to trademark that. I think coupled with the labour anxiety, we also have another big source of anxiety right now around the AI build that, which is data centers. This has been something that's tremendously controversial. We put up one itty bitty real estimating the cost behind a proposed data center that open AI was building. And I mean, we've had millions of views, millions of comments ranging from, why do we need this? This is the end of the world. Burn it down to this is amazing. This is so, well, they're so poorly understood, things like that. I'm curious. We talked about some of the potential risks to kind of like the AI adopters at that. But we haven't really talked about any potential headwinds to the AI build out itself. What's your take on the state of the infrastructure push right now? Obviously, it's going to very state by state. But what's your overall diagnosis of like where we stand in that process? So I would say it's the three piece, people, power and politics. Those are the primary bottlenecks, you know, constraining the AI build out right now. Let's just head on the that third P politics that you brought up. And this is a huge issue for voters. We're heading into midterms. This is something that it's not just an issue on the left or the right. This is something that is uniting quite a few voters. Is this concern around what a data center means for their community if one comes to town. And I would back at the pushback largely into three different groups. The first group is all about power bills. If a data center comes near me, does that mean my power bill is gonna go up because these things consume so much power? It's tricky though because in the US, you have regulated utility markets. It's unregulated utility markets. If you live in an area with a regulated utility, you can charge that hyperscaler, you can charge that data center developer a special higher tariff to make sure you're insulating the community and make sure you're insulating your insulating people's power bills. But anything that even touches on affordability, the consumer has been so rightfully stressed about inflation and about high price levels that there's a lot of concern there. The second bucket is around the environment. So concern about the amount of water these use, concern around potential air pollution, depending on the type of generation you're using with your using natural gas turbines, what does the air pollution from that look like? Once again, it really depends the type of cooling, the amount of water used depends heavily on the type of cooling. If you're using evaporative cooling, that is very water intensive. And if you're in a state like Arizona or a drought stress state, that is a big issue. But if you're using closed loop cooling, there's a bigger upfront water draw, but then it's a closed loop. You keep using that same water. So there's less of an ongoing draw. If you're in a less water state trust area, that would be less of an issue. And then the third bucket is all about just typical concern around having a large construction project near you. A lot of people, you don't want a data center in your backyard. You're concerned around the noise, the traffic, the dust, what this can mean for the value of your home. And in these instances, it's interesting you're seeing some targeting more of brown field sites. You're saying, OK, we're actually going to clean up this old factory that is not economically productive. And we're going to make the land usable again. There's usually pretty high property taxes associated with data centers. So that tax money is going back into the community. And you're bringing jobs back. So you'll have heard me say it depends. It depends. It depends on all those different fronts. And so I think that's really meaning that the policy has to come at the local and the state level because it's varied so much throughout the country. And I do think it's interesting. We had put a video up on the tax savings because given so many of the states were competing, it felt very similar to when Amazon was choosing their new headquarters. And it was like a race to the bottom to see who can give away the most tax breaks. So I do think it's almost like be a good citizen and pay the taxes. And then again, I understand that's not necessarily what company you're supposed to do. It's just to maximize your whole their value. But it's like at the end of the day, like you don't want to have an angry mob. So maybe maybe maximizing shareholder value was actually paying the taxes. And that balance of power has absolutely shifted in the past. How can we get this data center to our state? How many tax breaks can we offer? And now it's, what are you going to do for my community? What are you going to do for my state? What is the property tax generation going to look like? And this can mean a lot for school systems. Just a double click on that. It's interesting because like I said, we did this whole thing on meta. And while they had a lot of tax savings, they also, you know, the teachers that were there got like a $50,000 bonus. So at the end of the day, if there is a data center coming in, they are going to be paying tax. It's just that it might not be as high as what it, quote, unquote, should be if there wasn't this like trying to fight against other states. And actually to one other point about the field, the sites, Jen, the one you did it was like an old uranium enrichment site. It's like, who's going to want to live on an uranium enrichment site? No, thank you. Yeah, exactly. And it's interesting because you've had a lot of these state-level moratoriums now being introduced, now being enacted. And there was one that was very close to passing in Maine. It passed the state congress. But it was ultimately vetoed by the governor because of state or because of town in Maine wanted a data center. And it would have excluded them from having it. There was an old paper mill in J-Main. And they fought to overturn this potential moratorium because they wanted that data center to come in and clean up the site. Yeah. Fascinating. And for those of our listeners who aren't familiar, thought a brownfield site is. It's a previously contaminated piece of land. So if you think about any area that's had a dry cleaner or a gas station or, as Christian phrased, a uranium enrichment site, although one thing that people did very well point out is, isn't that radiation bad for the chips? [LAUGHTER] So I would think, I am not a scientist, but it's a valid point. But that otherwise can't be re-zoned because of the toxicity already of the land. Therefore, having it used for a function like this is kind of the highest and best to use the land anyways, if you will. Otherwise, it just sits down and vacant. Michelle, in the interest of time, I know we need to wrap up here soon. What are the biggest concerns that we have-- the biggest upside risk and the biggest downside risk that we haven't talked about here? Not counting the risk of human extinction, aside from that particular risk. So I think some of the biggest risks here are power. It's still a huge concern is those bottlenecks. We talked about the political bottleneck, but also the power front. These data centers are extremely power-hungry. They're very power-intensive. And our team is forecasting a 57 gigawatt shortfall for power needed through 2028. Christian, can you give our listeners, you are a resident engineer, a size of the scope of how huge 57 gigawatt is? I mean, Michelle was talking with us earlier. But I think that to power all of New York City at peak, it's what, 10 gigawatts? So you need like six, because I'm going to round up, like you need six New York cities, that's huge. Yeah, so a huge amount of power needed when we think about these innovative time-to-power solutions, like using blue energies fuel cells, like using natural gas turbines, like using some of these power shell provider solutions, those are from the former Bitcoin miners. We still end up with around a 30 to 40 present shortfall on the power needed. So that is a big constraint for the data center build out. In terms of risk to the upside, I think it's just some sort of technological breakthrough that allows us to use these more efficiently. Quantum computing has a big breakthrough, and maybe AI can help enable that breakthrough. If we're able to find a new solution from the power role that allows us to power these much more efficiently, that could provide a lot more. Well, we joke about cold fusion globally, but I know that there have been a number of recent high-profile advancements out of China on the power front and China already being so far ahead of us on the power front when it comes to nuclear energy and things like that. I am curious, what not if any has your team given to cooperation between the US and China who have been so nominally at odds, obviously right now we've got Xi Jinping and President Trump meeting like I am curious, what value or what weight do you give to the idea that there might be actually collaboration between the US and China when it comes to the energy side of the equation? I think it's still very much a race between the US and China and who will see who will be the ultimate winner on AI. And in the US, we have very much high access to compute. We're still in a compute constrained world in terms of we'd like to be doing more inference, but in terms of global share of compute, the US is still very much winning on global share of compute. China is still very much winning on power. I wouldn't expect a big change there in relations where we're going to see a lot more American compute going over to China and China is going to come and help us on the power side of the bottleneck. I think it's still being very much viewed as a race between the two countries there. And I assume that's predominantly because of the export controls on chips to China, correct? I mean, how are they? I had seen somewhere, I think it was like in an interview with Elon Musk and the economist, and he said that one of the huge bottlenecks is the lithography and that they had had some huge breakthrough there. But on the chip side, I mean, do you guys have any, like, I don't know, idea of when they might actually be able to produce their own chips? It does seem they're trying. So they are producing their own chips. They're just not able to compete with the amount of compute that Nvidia is providing. And they are constantly innovating. They're now shipping the Vera Rubin, which is an incredibly powerful chip. So they're just not able to keep pace with what the American companies are doing there. Got it. So it's more just like the chips are so much more powerful that are coming out of like Nvidia and they're just not there yet. That makes sense. Awesome. I mean, anything else, Jen? No, Michelle, thank you so much for joining us. Where can our listeners get a hold of your research? Where can they follow you? Where can they hear more from you? Thank you so much for having me on. So we have a Morgan Stanley podcast thoughts on the market. We have myself and a lot of other research analysts appear on that. It's a daily podcast. So I think that's a great way to get quick hits of our research and what we're thinking about. How do we stand top of all the news that's moving markets? We use ground news. Ground news is basically the Spotify of news. So at first, let's you look through all the coverage. Then it tells you who funds each news outlet, where it leans politically, and how factual it is. My favorite feature is the blind spot feed, where they share important stories that are being ignored by one side of the political spectrum or the other. Go to groundnews.com/skini to save 40% off when you subscribe. That's groundnews.com/skini and save 40%