Ep 191: Agents Gone Wild: Broadcom’s ASICs and the AI Frontier

0:05Hello, everyone. Welcome to another episode of The Circuit. I am Ben Beharin. Greetings, programs. I'm Jay Goldberg. Well, I wanted to kick this off talking about something pretty cool that we built. And I encourage everybody to check out. Everybody's known that our firm has gotten more into what we'll just call publishing. as a part of our research, highly encouraged to do more publishing. And so that's what the diligence stack became. And, you know, this whole sort of project started where I was like, you know, it'd be really great if just all of this research we've been doing exists in one spot so that I could be like, help me, you know, let me help me look at my data model or help me, you know, triangulate these three reports that I read.

0:54And so we were like, well, let's just build that online. And then I was like, hey, maybe other people might like this, might using an agent over our full depth of research corpus that's pretty deep, deep and wide. And so we did that. We launched something we call Career Strategies Atlas. People might have seen me promote it on Twitter. It's super cool, but it's also been an opportunity to expand our coverage beyond just deep reports. So earnings notes goes there, event notes go there, conversations I'm allowed to published, right, from management go there. So there's a lot extra that our research is feeding, but the knowledge agent to me is the most cool.

1:32And I say that for two things. And the feedback we've gotten is super interesting because people were initially like, there's no way it just runs on your data. Like it's going to still make crap up or it's going to still infer what your opinion is. And that might not be your opinion because there's still right doubt for AI. And I was like, we'll try it. Like we built this harness so that it's purely our research and it's giving you the conclusions of what we've actually said and not said it. And it'll tell you. And so I had a few folks in the beta be like, all right, well, we're going to test this and go and know that I had not written anything about SpaceX, for example, or a handful of non-neoclouds.

2:12And the corpus comes back and it's like, there is no career strategies research on this. There is no house view on these things. So anyway, people were like, all right, I'm glad it doesn't make crap up. Like it's really just on, which I think is valuable, especially when you're, I'm saying it really is our research. It's our opinions. It's our house views on things that it doesn't make stuff up. And I, initially I thought that was going to be really hard. It wasn't super easy, but I think just the learnings that we had going through this process of, of, of building something on our data, our data only, our models only, our research only, was interesting.

2:50And I do think there's some compelling takeaways to this also, that businesses are going to go through this process in some degree as well that we learned and went through. Anyway, all that to say, we launched a new research product called Atlas as a part of our firm's publishing. It's pretty sweet. I encourage you to check it out. Cool. We're also watching people burn through tokens. and that's been fascinating to watch like our token spend make sure we've kind of control a certain way check in regular to be like who's going like let's make sure that we're not going poor on token spend but uh that's been an interesting exercise also uh but i like i want people using it people are using it we're getting good feedback that's what i hoped so token usage means people are using it and that's what i wanted and hopefully finding value so but it turns out Jay, you can't just build things.

3:45AI models are good enough that even if you have a decent understanding of software, you can build some pretty complex things. And that's super interesting. And it's a brave new world of idea to execution. It's amazing what you can build now. All right, let's jump into Broadcom. This was, I mean, you know, there was other earnings. HPE had earnings. Dell had earnings. But Broadcom was kind of the focus. They are, you know, again, still one of the largest players in compute infrastructure. They had a good quarter. We can talk a little bit about kind of why people were slightly disappointed with those earnings.

4:37But I'll let you jump into the top line and then let's talk about kind of the overall narrative, what management said, custom ASIC market as a whole, which has gotten a lot more interesting with so many people in it. But let's start with top read on Broadcom. So, yeah, Broadcom reported earnings and they basically came in line with expectations both for the current quarter and for the future for guidance. and big numbers overall. I think they also talked more about the scope of their ASIC wins. They noted that Broadcom has had Google as a big customer for TPU all along. TPU is obviously massive now, but now Anthropic and OpenAI are poised to overtake Google as Broadcom's largest ASIC customers.

5:36And so I think management was pointing to that as saying, hey, look at how big this business is going. We've had this massive customer, Google, and now we have these two new customers come online in less than a year, gotten even bigger than Google. And the stock is down 4 % this week. It was down on the earnings. And I'm taking that as just the street feeling like this is all sort of expected. and there was no upside surprise. And so people were hoping for more, didn't get it, and the stock has been down this week.

6:13I mean, what should he? I mean, I look at this two ways, right? I know there was, call it a bogey number out there, people have for Broadcom custom business in fiscal 27 of around$150 billion. I know there was hope that he would say something more than$100 billion. He did. He said$115. Most consensus notes I've seen and conversations on there with buy-side folks has been like, that's probably the floor. So then you just figure out, all right, how much more are they going to do? Is it$150 or is it less than$150? But leading into 28 was kind of what I was thinking was, would he kind of give a bigger indication for both the two years as a total, but also into 28.

7:04Them saying$230 billion in 28 is, again, good growth into a market that's probably as a whole, we'll slice this up in a bit, between$400 and$430 billion for the total of CustomASIC, including Logic, Networking, Attach, all those things. It's a good chunk of that number, but I think what's interesting is you're exactly right. They've got expanded programs with Anthropic and OpenAI. But it's becoming increasingly competitive. And this was the first time a name came out that was largely known, right? And they referenced MediaTek as, which, you know, you and I are texting about this. You know, I get the competitive dynamic that they said, right, trying to play it under that.

7:56But the reality is MediaTek is going to be getting a good chunk of additional business, but again, of a growth market, right? You can argue share taking, everybody's building out more programs, fine. But if you just look at how many gigawatts that, again, in the measure that you have in gigawatts for this, Broadcom's still doing the most. They have the largest gigawatt to custom compute program as anybody given their customer base. yeah so the the chief critiques i heard about the call aside from just not being not good enough was number one you know two two quarters ago they were basically denying competition existed within google and here they are now saying yes there's competition from media tech but we're delivering faster we're you know we're better we're getting our chips to market faster we have better ip and that's a that's a big step back um and it sort of raises the question of like, well, you said there was no MediaTek before, and now you're just better than MediaTek.

9:01Certainly, MediaTek is executing better than people expected. Inside of Google, MediaTek is no slouch in the custom ASIC business. They're doing pretty well there. So for Broadcom to now say, oh, we're still better than them, maybe, maybe. I guess, you know, you're right. It's growing very rapidly. it's a massive market. It's a massive opportunity. But still, there's more competition there. And the other critique I heard was gross margins were down. And management's defense of this was not great. What they said was, don't worry about gross margin, worry about operating margin, which is fair. I mean, I make that argument a lot too.

9:43But it doesn't matter what your percentage gross margin is. It matters how many dollars you have to pay the bank. But it's still, you know a little a little better story there would have helped them a lot i i know but you know you know what i find super entertaining and again we always keep coming back to like we we enjoy the hawk tan calls right as much as anybody is that there's always something that he said somewhere early on that he kind of made a big deal about that then he goes back and says don't worry about it right remember the whole sam thing we think this is a sam and then they start asking Sam, he said, stop worrying about Sam anymore.

10:20And I was like, hey, gross margins are going up because we're getting so much. And this was, this was always the thing. And this is why I think the gross margin point is interesting, right? The whole argument for MediaTek and others was that at some point in time, if you have enough IP to do more of your customer own tooling, why would you continue to pay Broadcom's margins? They have the highest margins of anybody in custom silicon okay great that was a known point and so he had continually hit like our margins are good you know we're strong we're not losing margin and then to be like don't worry about margin which you're right is not the right i just like my point is i like the hocktans like talk stuff up but then like oh don't worry about that anymore like it's you don't need to you don't need to worry about that anymore despite what i said earlier don't worry about yeah and in his defense like you compare him to jensen where jensen will take the question he just doesn't like he doesn't get into those arguments about sam versus tam he just takes the conversation in a whole different direction you're like you're asking about sam and he'll start telling you about robots um flying cars i i i i think it's very interesting to compare the two like i spent a lot of time thinking about this because he was talking all about operating leverage and i was out this whole spreadsheet comparing the operating leverage so revenue growth versus operating income growth and the result was a tie they're both growing revenue very strongly much faster than their growing expenses so i i think that's the that's the challenge in this sector is like i think it's a sort of challenge overall is there's a tendency to want to you know pick a winner and say, oh, this is a battle between Broadcom and NVIDIA.

12:09And in some senses, it is. But if that's the case, then it's a tie. They're both doing very well, at least for the moment. Yeah. And I think, again, I do want to talk a little bit about how the market's changing in custom. But the reality is these are very big numbers that keep going up, both in merchant silicon, GPU TAM, and custom ASIC TAM. What's interesting, though, is, you know, I was sort of going back through because Broadcom actually had to do this, right? They shared some percentages of kind of what their assumptions were between compute die and networking attached die. We had this conversation with Marvell, right, who's kind of now fully in the lane of networking attached when two years ago, right, we were talking about Marvell's upside necessitated them winning the whole custom program, right?

13:02Compute tile and other things, right? Even with MediaTek, right? MediaTek won the compute tiles. I don't know if we know they won IO, right? Or full networking. That could have been somebody else. So all my point to say is it used to be when you think about custom ASICs and you just said, it's going to be whatever, right? 200 billion in pick your year or even 80 billion, right? Or a hundred billion in pick your year. It was sort of like you could just go okay so and so won tpus we think tpus is this much volume and you just kind of gave them all right all of that now there's so much mixing and matching happening across like sure you went to compute tile that's great probably the highest margins but then there's you might use somebody else for networking you might use somebody else for io you might use somebody else for some degree of your front side or back side there's more diversity in the programs now not alone is there more chips like in you know uh google does two chips they'll probably do three or four all these customers i think at some point do more than one program so you've got a lot of programs you've got a lot of variation of mixing and matching of the menu and i and i just think that's interesting because it's in my opinion that makes it a lot harder to model there could be upside you know that that you might make assumptions on right that broadcom gets more of more of those programs.

14:25But that's not as clean as before when you could kind of just go, oh, you've got TPUs, you've probably got all that program, and I can come up with a dollar value for you. It's much harder now, even though it's a bigger market, right? Going to 400 north, probably in 2028. It's bigger, but you got to slice this in a much more complex way than you did before. Yeah, I was thinking about this a lot last week on the Marvell call, because Marvell is talking about its opportunities with Amazon and with Google. And it was clear from their language that they were being very precise in how they were talking about, how they were framing their opportunity at Google.

15:03And it occurred to me in that there was exactly this complexity you're talking about where they don't necessarily have the compute tile. They're winning with networking and some other form of attached product that is different than what a lot of people are probably picturing. and it is going to be one, like you said, very difficult to model. And two, I sense that companies are going to get very, very precise in how they word things. And we're going to have to go through the headache of trying to decode what they're talking about to sort of gauge their opportunity. And yeah, it's going to be, it's going to be, people are going to get this wrong.

15:44Like it's just going to get all confused and we're going to end up with, you know, 150 % market share kind of calculations. Right. Yeah. And I mean, even just with Marvell, I agree with you too. And I think this is true of Broadcom. You know, again, to some degree, media tech, when they start to get more into kind of IO and networking as well at, let's call it, bleeding edge tech opportunities. So optical. your if your your io let's just let's just use networking for example um your ethernet network attach is not going to be as high a margin as your copper network attach which is not going to be as high margin as your optical network attach and like how do you even figure out your assumptions of how much they're shipping of attach to one of those programs when that will actually determine to you like a better margin profile for something that's adjacent to the ASIC.

16:47You know, it's like, again, all of that to just say this is what happens in an increase of complexity that will be very hard to model and, again, catch some people off guard because maybe sometimes it could be even better, right, than they expected because their assumptions are wrong. You have to make a wider variety of assumptions now if you're doing custom basic modeling, probably than even Merchant for NVIDIA and AMD. Yeah, and I think I'm going to have – well, I'm not sure I'll ever do this, but somebody should probably go back and reread old Broadcom earnings calls because I'm sure that they were – they let some of this slip without meaning to.

17:35and very likely that a lot of the stuff that we thought was ASIC revenue going into whatever hyperscaler in the past was probably Tomahawks. Because they got asked about that on this call. And it was very, their answer was like kind of helpful, but also kind of confusing, right? And so everyone's going to start blurring this number. And, you know, not for nothing, we're in a world now where NVIDIA doesn't break out networking revenue anymore. And everyone's going to start, it's going to get very blurry. People – I don't want to say people are going to play games, but I will say that the ability to play games has increased.

18:12You mean you can make your – you can mix and match your margin profiles or your product? But you're right though, right? Because the history of their custom business was custom Tomahawk, custom networking for the hyperscalers, which they still do. Like any hyperscaler who does a lot of custom networking does so with Broadcom. and that has nothing to do with xpu attach that's just their own switch asics uh for their custom built racks except they're going to report it all as one yes that's what i mean that's what i mean yeah that's what i mean is that it's hard to distinguish that because nobody knows exactly the skew mix being built by any one vendor uh yeah anyway i'm just going to make myself more confused trying to think about how crazy that is.

19:00I actually think this is going to work against Broadcom because Broadcom makes really good networking products. They make better networking products than anyone else out there. Tomahawk is an incredible product. And they have lasers and optical products that nobody except maybe Lumentum can compete with. And so that's going to get lost in the mix. They're really good at this. and it's going to just be, oh, that's an ASIC. People are just going to assume it's going to be on ASIC. Yeah, I don't know how it all plays out. It's going to be confusing. Yeah, agreed. But, I mean, again, like I was just doing, I know our friend Stacy at Bernstein kind of did the same gigawatt math, but I think people landed on all customers included.

19:50you're at 10 gigawatts in fiscal 27 moving to 20 gigawatts in 28 um with still good customer diversification across across those so again that's growth i i don't know again if it says clean to just go that they're a beneficiary of xyz per gigawatt right in others it's the same like we said with nvidia you can't just assume that when they say it's 40 to 50 billion uh per gigawatt that you just assume every gigawatt is 40, 50 billion to NVIDIA. But I do think they will beat those numbers of the 115 in 27 billion and 230 in 20. I do think that's the floor. That's my current conviction.

20:39Juries out on how much we think they can beat that with some potential surprises, which again necessitates more of that program, a higher percentage of the total program to what we were talking about. So two things. One is we have NVIDIA talking about$40 to$50 billion per gigawatt going to$90 to$100 billion per gigawatt. And now you have Broadcom with$10 to$15 billion per gigawatt. Sounds like it's now going to$25 to$30 billion per gigawatt.

21:17I really would love to see the spreadsheet behind both those math and how much of that is memory. And I think that is, that's a big dynamic. I think both of them are lumping a lot of things in there, but... Yeah, I'm not sure how we're going to model it. I have no idea what Broadcom is going to do in 2028. It's a big number, maybe. But to your point, I don't know if it's big enough. I think people, I think that the street is zeroing in on some numbers. I guess that's what happened this quarter. We had a number and they sort of got it and sort of didn't. And I think that's the story for the foreseeable future.

21:57And I think there's a bigger theme now, or really there's a lack of theme, right? The first half of the year was very much about finding the bottleneck. It was the bottleneck trade for the first half of the year. which company has capacity, which company is going to benefit from the bottleneck and raise prices or has surprise capacity. And I've seen a few people say this recently, it's that that's trade is fully played out. We've turned over every rock, we've searched under every cushion couch. And we know where all the bottlenecks are. So then what, what do we do now? And I think we're in a much more tactical individual company analysis.

22:37I think there are a lot of people who could have made money last year just buying anything. And now I think you're going to have to be a lot pickier and choosier in what companies you think are going to do well. Yeah. No, and it's a good transition. I think we can weave in a handful of other earnings without going deep into the earnings. You're entirely right. The consistency across at least the range of companies I tracked this week, Dell, HP, Broadcom, obviously with NVIDIA last week, is pretty much everybody is just like our growth is constrained by supply. So you just baseline growth is constrained by supply.

23:27and to your point about the bottleneck trade and i think we talked about this last week because this was the vast majority of conversations i had with investors at hot chips and still remains is exactly like everybody's basically figured out how much can be made and to some degree how many gigawatts come online that can be activated and then installed and it's basically like i'm pricing it in because i know what you can do right next year um and and i and i always keep saying like Once we started going down even ourselves, this kind of supply chain bottleneck, past substrates into testing, because I did a report on testing this week, the industry can only scale as its deepest bottleneck.

24:09So you could get more memory or you could get more GPUs, but if you can't get more substrates, it doesn't matter. Or if you can't get enough MLCCs or if you can't test your wafers, you're constrained by whatever your lowest bottleneck is. And that's just the reality of it. And literally everybody, in fact, Dell's call was my favorite because I've known folks in Dell's supply chain for 20 something years and they are amongst the best, one of the best people when it comes to companies, when it comes to supply chain management. And Michael was just like rattling off the long list of constraints there was like literally all the way into tea glass, like just listing it off.

24:51Like, look, guys, like we could grow a ton. It was just this industry is severely underbuilt in capacity. But that got me thinking that relief is going to come, but it's not going to come until 2028. And so you're exactly right that I think when people do this constraint trade, bottleneck trade, look into it, this is largely like 2027. So I did this kind of brief quasi report yesterday just saying that my thesis is 2027 is peak year of constraints because what was required for your supply chain, your again, your substrate manufacturers, your power and analog, like even your foundry customers' memory, what was required was them to have the confidence to go build out more capacity.

25:46for what is, as we all know, a highly conservative industry, the semiconductor industry, to build out greenfield capacity. And if you just back out your timelines, it's not until really the last, call it eight months, maybe 12 if we're being generous, but eight months, end of 2025 and early 2026, that your big masters of the supply chain, so again, once your Broadcom, your NVIDIA, your others, started doing long enough LTAs with the supply chain that they got their confidence, and you heard this over all of their calls the last couple quarters, we now have customer visibility. We are starting to increase our CapEx.

26:26We are going to go build, Intel included. But none of that, none of that comes online in 2027. Relief does not come until 2028. And so what I'm intrigued by is if I'm right, that 2027 is kind of peak constraint. And we're, again, we're just going to be vastly undersupplied in that year. When that comes on, I think there will be some upside surprises because all of a sudden you're going to have a lot more gigawatts come online that year. It's been planned. That's how long it takes. You're going to have foundry capacity come online. You're going to have memory come online. Both Samsung SK and Micron have been talking about what's going to come on in 2028.

27:10I don't think this solves the supply imbalance. I think it might bring it closer. I think we'll still be supply constrained, but we're going to have a meaningful amount more capacity in 2028 than the semiconductor industry has ever seen before. And so I don't know how to interpret that. I'm not going to try to model this. I'm just making the observation that that's the timeline that lines up from when they started building out capacity to when we might expect some relief. And I think that's an interesting exercise based on your assumptions on who might be some of those beneficiaries to the upside if a good pocket of relief starts to come on.

27:49And 2028 might still be a constrained year per year, but it might not be as bad as constrained, meaning that you have more upside opportunity because you got more supply that you're just not going to get in 27. I'm going to take the other side of this. i'm not sure this is all exactly what i believe but i'm gonna take the other side of it because people don't listen to this for us to sing kumbaya together so i'm just gonna be and disagree with you so another way to look at this is all right 2027 i think your timing is right but that means that in 2028 we're gonna start having a lot of capacity come online and And so, yes, there'll be companies who are like, oh, we built more data centers because we unlocked the supply chain.

28:40But you'll also at the same time have companies starting to see pricing negotiations go a different direction than they have been. You mean down. Down. Yeah, down, down, right? I especially think about this in the context of memory. Yep. Where remember we had Jeremy on from Micron who's building, what, seven fabs, right? and that's Micron and Samsung's building a whole bunch and SK Hynix are building a whole bunch. They all have a lot of capacity coming online. And if you look at the trends for memory this year, memory units haven't increased. They've been flat. But memory revenues have soared.

29:20So that means pricing is driven all of this, right? Big, big increase in pricing. If units start to go up in a sort of, this goes to is it a commodity or not but like it would tend to imply that prices might come down and then i think about tsmc do you know how many fabs tsmc currently has on their construction i don't let me guess uh five twenty twenty yeah i mean across nodes right so it's not it's across nodes and that probably includes some packaging but like they're building multiple fabs in Arizona, multiple fabs in Japan, multiple fabs in Taiwan. Like, so yeah, the industry is going to have more capacity than it's ever had before, but that maybe, and maybe that does something to pricing.

30:16Yeah, I think that's fair, right? I think you could balance that with a pretty significant increase in volume, right? as a part of that to variate some of that. But I also say, like you and I have said this, I'm uncomfortable with the memory price. I think there should be better margin balance. I'm not saying they need to go back to 10%, which was their lifetime average. But 80s is rough. That's the leverage you've got now. But if they could do a lot more volume at steady margins, that that still goes to their numbers right that's still annual growth in terms of bit rate growth sales growth um so i'm not i'm not opposed to what you're saying i think there there could be some bring back to imbalance but again i don't i don't think if i'm right with how much i tried to figure out comes online barring the 20 fabs from tsmc that's hard to model what's what let's just go like leading edge-ish, which is a little easier to figure out.

31:24I still think we'll be at a supply imbalance, but I think it will be better. I mean, I've tried to run these numbers. I know others have. I think our entire industry average capacity supply to demand imbalance is between 150 and 120%. There are other areas like substrates and others where it's higher than that, but let's just use it as a blended average. So let's just say it could go back to 105 % or 110%. That's going to bring pretty significant volume relief, but I don't think you all of a sudden are like, even at parity or under. I still think we're underbuilt for the next couple of years. But again, all of that to say, if you buy what I'm saying, and even in your price scenario, which is fair, it you make us it's a strong point that the next few years every year brings a little bit of relief like we're expanding capacity we're just we're like this year and next year is like a weird transition year where the demand just ballooned and and you just had no no no timing for for relief to come but then the next few years into the end of the decade if our timelines are right and everybody does keep building, every year should get more capacity.

32:45And so – but all to say I don't think this point we're making is really appreciated or modeled or priced in. Because I get it. It's an absolute very difficult variable. I don't even know how we do it. I'm just making the point that if this is right, I don't think people are considering that in their work for what 28 and 29 could look like as capacity and relief starts to come to those who have massive demand imbalances. I think that's a fair point. Although I have been thinking about this in sort of broader terms a lot lately because I've been doing a lot of work around the debt side of things.

33:28and I've talked to a lot of investors and a lot of clients and a lot of media the last two weeks, man. And we keep coming back to this idea of capital intensive economic booms, right? And somebody along the line, I apologize, I don't remember who it was. I've talked to so many people said they'd done a study and they'd looked at every past infrastructure capital intensive boom, the cell phones 3g build out railroads canals go back you know whatever whatever it was and there's always that we always overbuilt we collectively humans always overbuilt in these times but what's what is a little bit different this time is the extent to which we are limited in our ability to build as much as we want.

34:25And so in a very sort of backhanded, positive way, the constraints we have in the industry right now are probably a good thing in that they limit some of the excesses which would be possible. And things probably normalize. By the time the supply chain catches up, things will probably normalize. I don't know. then we get into the question of AI. But you're right. But you're absolutely right. I mean, to the same degree, like I've had this conversation with enterprises, is the costs of tokens are to some degree measuring their deployment. Like if it was super cheap and easy and they had access to compute, they would run much faster.

35:13And to some degree, if token costs were low, So they would deploy – they would start doing more of these productions to work out their kinks so they can move to governed production, which is what they need to be in a steady state. And we would – again, our compute demand would be vastly imbalanced. So there is a lot of truth to that, that we are sort of measured in our build out. But like an impossible scenario is like, what does parity look like? Like at what point in time do you think you could get to supply to demand balance or a ways off from it? But then what does that world look like across the vector?

36:02So – but you're right. At least this is helping measure it. But it just – it feels so – I don't even know the word I'm looking for. It feels so – like such a hairy problem that you do the constraint work and all it does is depress you more. and every layer you get deeper you're just like ah this this is a real tough one to get out of you see you see i'm the optimist here constraints are good it's a good thing you go constraints constraints good constraints equal yay yay across the board um all right i know we want to touch on this but in light of broadcom we'll weave into nvidia buying hugging faces but but nothing's really changed right broadly on the nvidia front um i think there's still optimism people still think right unfazed by broadcom's mentions i know he took he took did he take he took a jab at gpus i think right not extensive but hawk i think there was a brief sub tone for hawk for hawk it was fairly restrained yeah yeah yeah uh so a little less heated right but between the two um and i don't think he said anything i think he just said gpus is which which is funny but yeah i don't think anything's changed materially in the broadcom versus nvidia but let's talk about the let's talk about the uh hugging face um that's right that's right so nvidia is acquiring hugging face for almost 13 billion dollars hugging face is the epicenter of ai boom right this is where all all the all the free models are hosted this is sort of the place that all the model builders go to compare themselves and to share their work it's very much a sort of central piece of the infrastructure of the software side of the AI industry.

38:20I don't know if infrastructure is the right word, but it's an important sort of meeting point for AI software. And NVIDIA is buying them. And this is coming at a time when we're once again worried about free Chinese models. We seem to have these every three, four months. We have a panic about Chinese models taking over the world. and NVIDIA buying Hugging Face in the middle of one of these is noteworthy because Hugging Face is where we'd normally access where we would get these models from and I think it's an interesting move from NVIDIA's point not quite clear to me what benefit it brings to them directly uh i mean i know they have a they have a good case for it but like i think the the for me the the main point of nvidia buying them is it it sort of keeps them very much at the center of the ai software conversation i don't know that they needed to convince anybody of that i think most people on the software side of things already have a pretty high opinion of nvidia but buying hugging face keeps them you know is sort of plants the flag very firmly on that side of the ring yeah i don't i again i don't know how it generates i don't think it's going to generate huge economic returns for them but i think it's a pretty important thing to own strategically yep right um yeah and i think like there's a couple things we got questions like um which nvidia clarified but like does this mean that they'll try to limit or lock in open models to nvidia hugging face others said the same like we're nothing changes in our work i think nvidia is just incentivized and and you know to be honest with you what's interesting is

40:11almost all of nvidia's competition also benefits from open models winning like it's not like that's a oh open models wins great for video okay it's actually good for everybody And AMD needs this. Others need this, right? You might argue that they could run them better, which is kind of perhaps the hat that they want to hang on because maybe you're not going to run a frontier model better when those frontier models get baked into custom ASICs. But AMD has sent the same charge, right? Open benefits them because it creates more parity and it doesn't mean that any – it's really a battle of pure TCO, right, at that point.

40:52But I don't think this gets locked, right? Just like they come away and be like, oh, you can use these models, but they're only for NVIDIA, which was a concern, which I think both companies tried to pare down. Yeah, I think that's true. Although that, again, I think it's – that goes to the question of is this actually an economic benefit for NVIDIA? And I think that's, if they maintained its value by keeping it open and free, it's hard to see them generating a ton of economic value from it. If they closed it off and tried to, I could see them trying to close it off and monetize it. That's certainly what maybe an Oracle would do.

41:32I don't see NVIDIA going down that path. I think they're going to keep it open. They're willing to spend the money because they see it's an important asset to maintain. It's important for the industry. but i think can we both agree that you know five years from now we're gonna have both closed and open models still oh of course not not a question you you and i agree on this but i'll tell you i it's one of the most common questions i've gotten lately is which one's gonna win and i'm like they're both no one's gonna win they're both gonna be there but i there's certainly a perception in some circles that it has to be one or the other.

42:08And I don't think that's quite how it's going to play out. I think we're going to have both. And we're going to have... The main thing is that the tools we use, the AI software tools we use, will route between the two. 100%. 100%, right? That's how it's going to play out. And we're going to have both. And you're going to need both. It certainly raises questions about the future of open AI and Anthropic. and their ability to differentiate. It puts a lot of pressure on them to keep the newest, best models coming out. But so far, they've been doing that, right? Certainly, Fable had a big update this week.

42:45People seem pretty happy with it. People have been very happy with Sol at OpenAI, and Astra's coming out very soon, and that certainly looks very promising. They're going to have to keep that going, but so far, that doesn't seem to be too challenging for them. Yeah. Yeah. There's, you know, I've thought about this too. There's a very strong case to make for the frontier labs in what I would just call your mission critical functions. So cybersecurity would be one of those. Engineering would be one of those. There will be areas where high value, high margins, right, will accrue to the most valuable, most important mission critical workloads.

43:34And I extend that to nation states as well, because as much as I find this to be, and I tweeted this, like literally the most exciting time to be in this industry, like I, the tools have made working so much more exciting. Every idea I have, I can go try and make it. It's also similarly exhausting because you're like, I've got so many stupid ideas and they're not all going to work, but I'm going to try them anyway. And then I'm just like, I've sat here for too long. It's also very scary. I think the more you play out the security and the cybersecurity threats to its logical conclusion, the more freaked out I get.

44:11and and i can tell you with a fair degree of certainty open models are not going to be your defenses against rogue agents rogue nations and any rogue players and to be honest with you i i mean you tell me if you disagree but like there is a degree of this that is also like an arms race if you control ai the best ais you have leverage over other nations and just look at where the place today. There's only two countries making a go at this, us and China. And if you think for any second that the government does not know that and is 100 % convinced that frontier needs to be frontier because of our need to maintain some level of sovereignty, that these are basically the equivalent to digital nuclear bombs.

45:04And we need to have some control over that. And we also don't want it to be a monopoly. We don't want one of them to be the only one who has all of the control. You need there to be diversity there. You can make a very strong case for the frontier. Again, economics aside, how much can they make? Whatever, fine. But the frontier exists and will be valued and invested in for very, very particular strategic reasons that get very scary when you start to think about it. I agree with your broader point. I don't necessarily agree with nuclear bombs but i do agree i think it's a very important point is that there will be an incentive to invest in frontier models always because that you'll need state-of-the-art to you'll need state-of-the-art end to define defend against state-of-the-art minus one yep and i think that's important and will continue and persist and i'll tell you this this week i had a a pretty unsettling moment where I was, I was, I had Claude build me an evaluation model, right?

46:11I've just like, I was just, this is something I was just playing around with. It was never going to be published. It was just something I was just sort of, you know, mental exercise thinking through something. And I had given it a, a target valuation and it came back to me and it said, uh, I can get to your valuation, but you're not counting this very significant accounting charge that would have meaningfully altered the valuation. How would you like me to handle it? And one of the options was skip. Just ignore it. And I was like, if I were a company and I just picked skip, don't mention the$140 million accounting charge, that's fraud.

46:58Yep, absolutely. Right? So I was like, uh, again, that was never, that was just for me. It was never going to see the light of day. But like, I was like a little freaked out that like, oh, Claude's now suggesting me the morally ambiguous position. That's one of the options. Yep. Well, and I think too, there's like parts of this, you know, again, the reason the security stuff kind of freaks me out is, you know, one, I've talked to a handful of CSOs and they're like, you know, One of our biggest concerns is that everybody who tries to take down our company has access to any open model that they want and they can just – some rando in whatever country can hold our company hostage.

47:41They realize how much they have to defend against very good open models, frontier-like open models, and they will be incentivized to invest in what is better than that, regularly better than that. In fact, that's the reason why Fable and Mythos and Sol and now Astra and every model goes to enterprises and the government first to make sure that they can defend against that level of a model. And that's not changing. That's going to continue. And it, again, creates the case for them to keep reinvesting and doubling down and turning their profits back into training, better models. And eventually we'll get to much larger data sets, being able to do analog and digital like video and audio and process all this stuff and increase your data sets.

48:33So all of that to say, I firmly believe in the case for the frontier. I'm working on a follow-up to my first cybersecurity point, which again, report, which just continues to give me anxiety. but in it is also the very strong case for the frontier on a base of cybersecurity and some very actually big budgets coming for that alone in enterprise software via what ITDMs tell us. But yeah, like I said, it's an interesting time but 100%, both will win. My full conviction is you're just a normal knowledge worker at an enterprise. You do not need to go pick what model to run. You just go, look, I need a summary of this.

49:14Help me do my balance sheet. It needs to be governed. If it's going to change a file, it needs your permission so that someone can go back and rate that you actually had permission to change a file. Like, permissed and governed AI is we don't have it yet, but we need to. But you don't need to pick the model. The enterprise says these are what's approved. It routes to the right place. You just go do your work. You shouldn't have to worry about all of that. you've hit another nerve with me because when i went to bed last night i tasked claude with three three projects and i said run them overnight dangerously allow all permissions oh geez there you go dude and so i woke up this morning and by like by 7 a.m i'd run out of usage capacity yeah i actually like yeah but i like though like just thinking through it i do like that they ask you do i have permission to change this file because you and i don't have this but in an enterprise context you need you need that record right to be to be monitored that that person changed something did you have the authority to change something like that has to be recorded so i like that they are at least again like i said there's a big difference between agents running wild in an enterprise, which is what it is today, and governed production, which we're barely scratching the surface of in the enterprise.

50:35But this is all these hairy problems that just get brought up the second you are like, holy crap, agents are running wild in my enterprise. Like, what's happening? Agents gone wild. Agents gone wild, everyone. That's the title of this episode. Just kidding. Maybe maybe the second save. All right. We covered a lot of ground. next week will be fun we've got apple uh and some special guests that will come to the circuit and uh and we look forward to that and talking to you later have a great weekend well you've already had your weekend probably yeah i hope you had a good weekend thank you everybody for listening tell your friends tell your agents