Micron's Roadmap, Meta's AI Infrastructure, and NVIDIA's Supply Chain Mastery

0:05Ben Bajarin: Hello, everyone. Welcome to another episode of The Circuit. I am Ben Bajarin. Greetings, programs. I am Jay Goldberg. Well, we have a lot to dig into this week. Let's start with Micron, who did a somewhat of a surprise webinar, or a webinar, I guess. there's a lot in here I don't think we'll get to go to all of it but for the first time in I don't want to say ever but at least a really long time Micron was pretty forthcoming about their roadmap making a couple of specific things talking up what they're going to do in 6F squared why they didn't really see market scale opportunity in 4F squared which might have been a dig at the Chinese players that they're skipping a node, which I thought was actually super interesting and needs to be unpacked.

1:01But more importantly, just broader commentary on what they, I think, wanted to get across, which was technology leadership, which obviously has to translate into a couple of different things.

1:16Ben Bajarin: Premium prices, to some degree, market share gains in areas that you care about with high bandwidth memory. there was a lot of narrative on bandwidth memory I think designed to appease people's concerns that HBM is not a good technology or not going to be around a long time or be replaced by HBF or something like that so I think they tried to quell some of that those concerns but broadly they positioned it as a technology webinar which again I think is why they spent so much time talking about the nitty-gritty of some of their solutions. But the whole tone was pretty bullish on memory, seemed to hint at it's going to remain constrained, waffles will be constrained through the end of the decade.

2:10Ben Bajarin: But despite all of that, the market is pooping this week, so it is what it is. But I thought it was a good webinar as a whole from a technology standpoint at least.

2:23Jay Goldberg: I think that's fair. I didn't sit through the whole thing, I must admit. But it was – I need to go back and reread the slides because I thought it was interesting data. Like there's a lot of nitty-gritty in there. I kept asking myself like deeper questions about memory because I was – when I was listening to it, on one screen and my other screen I was looking at all the all the red on my stock chart pages um which was very very distracting it was hard to reconcile the two um that's a comment theme this week but yeah I thought it was I thought it was interesting I I your your comment about 4fsquare being being a dig at the at the CXMT is interesting I hadn't thought about that but but I think you're right.

3:17Ben Bajarin: Yeah, I mean, they were pretty, you know, again, this is Micron's point of view. Obviously, they want this point of view to be true, but there was a lot of points that Scott DeBoer made, who is their lead sort of technologist. I think he's CTO, actually. The point being that, you know, they've looked at it. He was like, look, we've looked at this. We've been looking at it for years. We just don't think there's a viable scaling path here. And really, they tried to point this picture around, you know, they are going to go to vertical stacking of HBM. So VHBM or others call it IHBM, you know, all of these things.

4:05Ben Bajarin: but what I think they pointed out and I just I think this is important to kind of caveat with this whole thing is you know this is innovation in memory is not easy every node jump plus every in this case 3D stacking right of memory creates more thermal challenges so like to do this you like you just have your engineers just need to go solve problems It's very much like the days of Intel and NodeJumps and whatnot where they're like, look, getting to the next node is hard. We're struggling. We've got to redo our transistors. We've got to figure out how to do X, Y, and Z. And these are just problems engineers need to go solve around memory.

4:45Ben Bajarin: And I think that's why they did what they did is they wanted to say like we're pushing a culture of innovation. and if we do these things, if we create these innovations, if we go solve these really hard problems, we think it's important that the market will pay for those things. I think they were trying to show doing these things doesn't mean this is commodity, I guess, to a degree, which was another subtle tone there. But on that point, I think it's interesting. I hadn't heard memory companies basically kind of segment their roadmap to talking about memory like AI class type of solutions. And I say that to say, again, they want to make this point.

5:32Ben Bajarin: They've got a whole product roadmap. There's product derivatives that are specifically designed and being innovated for AI. And I think they think that will command some premium pricing, some ASP increase for the memory that is innovated for and thus designed for AI versus that that's not. right that's maybe just you know on board pin compatible types of so that's that's my read i just think they put a that's why they're pushing we have to innovate these things are hard we hope to capture value is the clean pillar i guess i'm yeah i i agree there that was the undertone

6:11Jay Goldberg: the the implicit message in the whole thing is that this is not a commodity um um, right, which we, we've talked about a lot before, how the extent to which memory is or is not a commodity. And I thought, you know, they, they made a good case. I don't know that I entirely bought it, but I thought, but the other thing that I thought was interesting is I was the first, a lot of their discussion was, was very, very technical. And so I thought it was very important that they did sort of tie it to specific use cases because Because one of my fears with companies in general is that they get very, very focused on, oh, look, here's the next technology.

6:54Jay Goldberg: Engineers love to solve new things. But customers don't always want that. Right. And I've worked with lots of companies over the years where they're like the engineering organization runs the product roadmap and you end up in very strange places. Nobody ever asks what this is good for. And I thought, you know, and Micron isn't doing that. I'm saying they're like, they obviously didn't name specific customers, but they were trying to tie it towards end markets, which gave me some confidence that people actually want this. I mean, I'm obviously thinking about this in the context we've talked about a lot lately, which is all the pushback.

7:31Jay Goldberg: The memory company's got hot chips, right? Do the AI companies really want more innovation? Do they want new things or do they just want more of the existing stuff? And Hot Chips is by no means a clear statement of what these customers want. So it sounds to me like Micron has thought through a lot of this and has talked to customers at a higher level. I also think they are signaling to people who aren't AI customers that there is, you know, that they're thinking about this, that Micron is thinking about them. Right. Don't you worry, Apple, consumer electronics people. We think about you, too. We have stuff for you as well.

8:22Although that definitely was a lesser focus. Yeah.

8:27Ben Bajarin: Yeah, I think this just brings clarity to where they're putting – what bets they're making. You're right. It has to pay off. I think there's at least an argument to be made that if infrastructure providers and lab companies still want innovation and logic, that you also still want innovation and memory because these two things sort of go together. I think the labs hope that compute moves in a direction, and when I say compute, that includes logic, memory, and networking, that it moves in a direction that they can do more with their models, build more capable models, and whatnot. And so I think there's a case to be made.

9:14Ben Bajarin: But again, I think the point here is they put a roadmap out, and I think they're just trying to say we're confident in this and we're confident we'll capture value. And the market's just going to have to believe that. Customers will have to care. But that's their posture. And I thought it was interesting because you just don't hear this kind of forward roadmap from memory companies for all the things we've talked about before. They don't want to tell each other what they're doing. So it's like, well, they just took a pretty bold bet, and I think that's super interesting.

9:52Jay Goldberg: Yeah, I always think about signaling when I hear the memory companies talk because they do signal a lot to each other, not about product roadmaps but about capacity more. But that's a very active function in the memory complex. It has been for years. There is a triopoly, possibly a quadopoly, but there's essentially been a triopoly for a long time. And they absolutely, absolutely use public occasions to signal to each other, mostly along the lines of Samsung, please don't add more capacity. But this was different. This was them talking about technology. And I wonder if they – I was thinking about that.

10:34Jay Goldberg: Who are they trying to send messages to? Clearly CXMT, but – Yeah, yes. I think this is much less of that going on. I think they're really trying to reach customers.

10:45Ben Bajarin: No, I think they were too, but I also think they, again, 100 % my interpretation, I think they were also trying to make it clear the Chinese either can't do these things or will struggle greatly to do these things. I think it was trying to be like we're on another level than they are, and that's going to be true of Samsung and SK also. But I think there was a little bit of that, at least as a hope going on also. That makes sense.

11:17Jay Goldberg: That makes sense. Because you think about it, last week we had CXMT basically just come out with a press release saying, hey, we have a new memory structure. Right. And then here we have Mike Rohn a week later saying, oh, you got a structure. We have a roadmap with lots of structures on it.

11:32Ben Bajarin: Yeah, and lots of money behind it and all the tooling and whatnot and knows how to do it and all of those things. So anyway, I thought it was good. I like technology roadmap webinars. I thought it was good. I hope they do more. And I think there's more coming on this front of memory innovation. All right. Let's talk about LAM in the context now of memory as they are actually highly exposed to memory wafer equipment. so lamb reported nobody liked it despite the fact that both their numbers it was a record quarter and i think we all agree that wfe is just up and to the right um but what what was your take on i know you listened to the call too what was your take so they reported an inline quarter so right on top of consensus and they guided right on top of consensus and uh good enough results

12:33Jay Goldberg: But apparently in today's day and era, good enough is not good enough. And so stock traded off pretty heavily. I think a lot of the questions they got asked were – they said there were constraints. So a lot of analysts asked, if there weren't constraints, what would your numbers have been? There was a few of those people. And they did not answer that, which is smart. They didn't answer that. But I think what they're saying is that they were supply constrained, didn't say what exactly that involved, but they were supply constrained. But also I think there was another layer of the customers can't take delivery because they don't have clean rooms yet.

13:20Jay Goldberg: And so that was a cap on their sales. But at the same time, I think much more interesting than the results was a lot of the commentary they gave around the industry and around different niches within WFE. And so they raised their guidance or their market estimate for next year. They think it's going to be – WFE is going to be a$200 billion market up from – call it$150-ish this year, which is a good growth rate for this industry. Just to put that into context, earlier in the year, you and I were talking about these things, and it was like, could it hit 150?

13:59Ben Bajarin: That was what we were talking about. Could it hit 160? 200 wasn't even in our minds as a possibility.

14:08Jay Goldberg: Yeah. Yeah. And then you sort of look at a lot of the estimates for WFE, and everybody's talking about big numbers for next year. Yeah. And there's a lot of potential growth there if we can get through these supply constraints and if we can actually get clean rooms built on time. But I think I was much more interested in like – you know, LAM is sort of seen as an etch and deposition company. But that's like multiple tools within each of those categories and they have a whole bunch of new ones. And sort of going through that and they talk about sizing some of those segments. And that's one of the hard things about talking about semi-cap equipment is there are just so many different tools all doing slightly different things that all sound very similar but are radically different.

15:02Jay Goldberg: And it was – I thought it was helpful for them to give color on what all – you know, what is being driven by what. Like more this kind of memory drives this kind of product, those kinds of things.

15:15Ben Bajarin: Yeah. Yeah. Yeah, and I feel – so this wasn't a part of the MyCon webinar, but I actually feel like this is going to become more of a narrative next year, more clarity on memory types for workloads, memory innovation for those workloads. Because I think – I guess what I'm saying is I don't think – historically, we had not had to richly segment memory the same way that you do other categories. and I feel like you're going to have to do this now. More acronyms. Just what the world needs. Category creation, baby. This is how us analysts stay in the biz. Yeah. Anyway. Yeah, it was interesting stuff.

16:06Ben Bajarin: I agree with you. I mean, I think I paid more attention to the numbers. I think you didn't expect them to answer, you know, what would it be constrained? It was just kind of the same answer that Jensen gave when he asked the same question, which is like, well, it just would have been higher. Okay. Like. That's right. But again, you're up.

16:23Jay Goldberg: No, no. They said it would have been higher. Jensen would say, oh, it would have been much higher.

16:27Ben Bajarin: Yeah, that's right. Much higher. Emphasis the much. But yes, constrained in material and clean room is apt. Okay. Meta also had earnings. And I'm super intrigued by meta for a couple of things. You know, obviously they announced Prometheus this week. I actually got some time to talk to Ian Buck from NVIDIA on this and some of the meta folks. You know, what was interesting was I think what I hadn't – I'm making a meta. I hate that their names are the same. I'm going to say I'm going to make a meta point and I mean like a big picture point. And it's going to make a big picture point about meta. I hadn't realized how important owning your – a gigantic bit of your own training infrastructure was until I talked to them.

17:37Ben Bajarin: And then I think this point also then translates to serving said things. But they talked to me and kind of went into great detail, like with the blog was, about why you need such large training clusters of leading class compute to develop competitive models. and you know one of the things points that they made was and they have right researchers on the back of this plus you've got you know training that happens in real time um whether that's reinforced or post or whatnot um but it actually is the kind of thing that lets them make a model and then do variants of it distill it down and make multiple other models so if you think of models as products right name your name your model that's one product and then you want to have other ones that do certain things or that might cost less.

18:33Ben Bajarin: Like you, you actually have to keep doing that in your training cycle. And so one of the points that they made was like having this, you know, at our disposal means not only that we can make these big models, we can actually make many of them at around the same time and more frequently. So essentially they can launch products, products or models. They can make more, more products. They can launch more products more frequently at, again, a cost basis that's not market rates of tokens. Because when OpenAI and Anthropic and others have to do this, right, they have to do it on other people's infrastructure.

19:06Ben Bajarin: And sure, they might have great deals with Google or Amazon or whatever, but it's still, they don't still own that infrastructure. And so there's economics at play here for Meta that I think are actually super interesting to have, you know, a gigawatt class, right, type of infrastructure that, again, spans many destinations, so scale out right across the board. They mentioned all their suppliers from Broadcom with Jericho to Arista and all the peoples that they work with to do this. But I say all of that to say it's interesting to me from an economics standpoint that you can make products, serve those products, monetize those products, right?

19:45Ben Bajarin: And of course, that really, this is a pretty important asset to have in the long scheme of AI models. It's interesting you mention that because I actually had a long, long conversation with someone from one of the hyperscalers last night about this topic and the economics of models, the economics of owning your own data centers and your own models.

20:10Jay Goldberg: And his point was that Meta is – he said Meta was the largest data center. Like basically he said they had more gigawatts than anybody else or almost as much.

20:27Ben Bajarin: For AI. For AI, yeah. Yeah, because I don't think it's a true total, but for AI, I can see that. But – which I thought was – I mean because meta is not AWS operating compute for thousands of other people. Right.

20:45Jay Goldberg: They just do it for themselves, but they're still the largest, one of the largest overall data center players out there. And largest or very close to largest, depending on how you want to measure it, very large GPU AI capacity. I mean, let's not forget, like OCP, we were talking about OCP last week, and what an important show it is, Meta founded OCP. They were the big driver behind it. So they've been doing – they have deep, longstanding expertise in this stuff. And if you compare them to, say, Amazon who doesn't have their own models and what does that mean long-term for the economics? You're always – is Amazon always going to be dependent on other people's models and what do the economics of that look like?

21:37It's – right?

21:38Jay Goldberg: And then you sort of take that a step further and sit through what that means for the chip vendors. and how that, you know, there's a lot of value being moved around, I guess is what I'm coming to. There's a lot of value being moved around here, and I think it's not entirely clear where that value ends up. But owning your own data centers is an important thing, as is owning your own models. And I like the way you think, but I don't think of models exactly as products, but certainly having your own model and then being able to tailor that to your different use cases. So you have your big training runs, and then you have a model that serves up Muse and you have another model that's distilled and honed for ad serving.

22:22Jay Goldberg: That's super powerful. And I think because there's a lot of debate now about open weight models, right? I don't really need to have my own model. I'll just wait for the Frontier Labs to do something and then someone in China will come up with an open weight model that I can use that's almost as good as that. And I think there's cost to that, right? Every time you distill something, it becomes a little less good, a little less efficient. And so if you're waiting for someone else to distill a model from the Frontier Labs and then you distill it yourself and fine-tune it yourself, I wonder if you start to lose some of the quality.

23:05Jay Goldberg: And then what I'm really getting at is the economics of open weight are not quite as compelling as you think. And there's still – right. There's lots of cases I think where you run through the math on running an open weight model versus instead of having your own model, you use Anthropic API or whatever. Yeah. You think the open weight would be more efficient, cheaper, but it may not always be that case.

23:33Ben Bajarin: Right. No, agree. But I think what I walked away with is exactly sort of this viewpoint where Meta is more vertical than I guess I gave them credit for before. Like I think I would have said Google is the one trying to do a frontier model and is vertical. I think you put Meta cleanly in that now. Again, I get that they are largely using this for themselves, so it's a little bit of a different comp. But at the same time, we do. Again, you talk to enterprises. We just fielded a big CIO, CTO, and CSO survey. And they're weary about Chinese open source models. They want to use open models. They want to be able to route to model that's more affordable economically than the frontier.

24:26Ben Bajarin: But they're just super weary. Even though they might control those or someone else might distill them, I think they're going to want to use a US companies-based open models. And that could be Google. That's basically going to be Google, maybe Nemotron from NVIDIA or Meta. Those are your three best options at this point. But where I'm going is having that class of infrastructure, and I think this is what Meta was trying to do, why they were willing to talk to me and the point of Prometheus, is having that class of infrastructure lets them pace the frontier or perhaps lead the frontier as well, even if that's for internal things competitively and or doing things with an open weight model that's frontier class for people, companies, whoever.

25:22Jay Goldberg: Yeah. I think this is another example of how quickly things move and how you have to be wary of accepting any long-term predictions about somebody's model being the best or somebody's model being the worst. Because not so long ago, people were saying, oh, meta's all screwed up. Because they absolutely had very serious internal problems around their AI or it was very confused for a long time. And so people kind of wrote them off. And now here they are with clearly a frontier class model or close enough frontier class. Yeah, it's sort of in the ballpark. And, right, you know, and I think, you know, Google's kind of like it's funny to me how quickly it rotates.

26:13Like every three months somebody's a leader and somebody else is in the doghouse.

26:17Jay Goldberg: And, like, no, like they're all competing very heavily. They're all going to be coming out with models on a slightly different time frame. And so just because someone is the leader today doesn't mean that the last leader is finished. They're just working on the next version and they'll catch up, which does, though, open the question to like what are Microsoft and Amazon going to do because they have been very quiet on the model front. I mean, I don't – what can they do?

26:48Ben Bajarin: I just don't think this is something they're going to do, right? Right. But but again, just but tying this back down. I think it's interesting. I mean, you kind of heard some of this from OpenAI, too, right around their newest models being kind of black well trained. You know, sort of made the same point. Right. We've got leading class infrastructure, which is what's letting us do this with models, which, again, I think they doesn't allow them to do this more quickly. keep up with the frontier more quickly, if not perhaps surpass the frontier. I'm not going to lose that out. But this raises an interesting question about Google's infrastructure.

27:30Ben Bajarin: And I think what Google has made is good, but why did it take so long? Like if TPUs, you have a ton of scale and TPUs are great, it took a really long time for them to get to Argonne. And I think I'm saying all of this to say I think that infrastructure, per Meta's point, actually does play a critical role in how fast the frontier can go. Now, I get we were talking about pacing and concerns and all these things. fine. But that infrastructure is directly relative to how quickly you can make new, more capable models and variants of those models with different economics and capabilities. And so, again, I just say all that to bring it back to the importance of the right infrastructure to do this, to have competitive advantage for your models.

28:32Jay Goldberg: I agree with all of that, except for the part about Google. I think they've been there all along I don't think they're that far behind I think Google's a little bit different because they have TPU and TPUs are much more efficient for the things Google needs it to do and so their power needs have not been as acute maybe with GCP but really for most of their workloads for TPUs, it's much more efficient. And so they were never as far behind as, say, Amazon or Microsoft was or Amazon is. And so they haven't been as vocal, but they're absolutely – I think they've been spending pretty heavily. And I mean, they just did that big equity offering a few months ago to keep it going.

29:24Jay Goldberg: But I think your point is right, is owning your own infrastructure and having your own models matters tremendously.

29:32Ben Bajarin: So here's the other one then where rubber meets the road. I tried to do some math on A, how many gigawatts of compute just agents, individual personal agents will add to our gigawatt demands. Because this is not running on existing compute demand. It's right. They talked about, right. Muse is basically everybody's getting their own VM. You're essentially getting a core or a core or two or whatever, right. That translates to. So that's additional compute demand to just run reasoning models, executing models, et cetera. um people have sort of brought this between you know anywhere a number between uh tied tied to users like for every hundred million it could be you know somewhere between three and five gigawatts if you just kind of just run the math on right the basics here if and i did this so not including the entire internet let's just say that half the internet two and a half billion people use personal agents.

30:44Ben Bajarin: Just that alone would demand another 35 gigawatts of compute a year just to run the agents, not to execute the agents, not to handle those models, just to run those agents. Now, I say that to also say there's an economic cost to doing this. If you are OpenAI or Anthropic or whoever and you're having to run agents at market token rates, it could be quite expensive. One person's personal agent could be anywhere from$50 to$100 a year just to execute that. So you've got to monetize that. Here, though, Meta, I think, is actually in a really interesting position that they can keep those costs down. It's still expensive to serve those agents per one person.

31:30Ben Bajarin: That increases your – you have to ARPU that in some degree. But I think owning that infrastructure means they can keep those costs a little lower perhaps than a competitor who doesn't own that infrastructure. with this kind of individual personal agent cost to serve is what I'm getting at.

31:51Jay Goldberg: Yeah, I see that. I think that is interesting. It's a lot of agents. There are going to be a lot of agents out there. Actually, I think one of the things that really stood out for me was they were talking about how much it costs to run agents because we're talking about Muse here really for Meta. Yeah, talking about Muse. And I think he said it explicitly They're never going to charge for Muse Right And that seems confusing Because you just outlined how much it costs To run Muse And they're saying they're not going to charge for it Which means they're not dumb So they have some other way To monetize that And I think that's important Because one, I think Meta is Very, very good at Monetizing engagement and Muse is just one more way to engage people, whether it's Instagram or WhatsApp or Muse, right?

32:43Jay Goldberg: You're going to spend more time on Meta properties and they'll find a way to make money off of you. They'll ARPU you. I like that as a verb. They'll ARPU you. And I think that's important because I think there are lots of other companies out there, like OpenAI and Anthropic, that still need to figure out that piece of it. Agree. Beyond subscriptions. Agree. And Meta is a really good example of like, hey, yeah, there are other ways to do this.

33:09Ben Bajarin: Yeah. Yeah, I mean Google is going to have to do this too, right? I think it's a fair assumption. I'm not going to make a timeline on when this happens, but I think it's a fair assumption that every person in both a consumer personal context and then probably a secondary work context is going to have an agent that they're going to use and it's going to go do things for you. I believe we'll get there. And that's, again, that's going to have a cost to serve that's different than token cost to serve to go and execute for you. And it's going to draw a different set of power. In fact, these are also very highly CPU-dependent workloads.

33:52Ben Bajarin: So that alone just shows you both the gigawatt of compute and the demand we need for more CPUs to run personal agents. but my point, yeah, there's a cost to that and economics play that factor and you recovering that particularly at a consumer level and I think we would all probably assume Meta wants to take some skin off transactions to a degree which would make sense so you do need to be transacting on their platform or with Muse but things like that I think become where they can are per you on top of their cost to serve agents

34:32Ben Bajarin: makes sense yeah okay let's talk supply chain in a couple different ways I want to talk about supply chain and then I want to talk about some other little bit of news that happened today and then I think we should end on cyber security since that's again a hot topic even though I and you have been talking about it for some time. People are finally waking up to the reality I tried to outline in June. So, all right, let's talk. Nvidia is master of the supply chain. Using balance sheet to buy up all the goods. Again, straight out Apple's playbook in the mobile era. Buy all the things so nobody else can or they have to pay really high prices for them.

35:23Ben Bajarin: Crowd out your competitors with your supply chain mastery.

35:29Jay Goldberg: and so that's happening again right yeah so this this week is gtc berlin and nvidia had a whole bunch of announcements as they are want to do all kinds of different partners doing different things with ai but deep in there also was they have gone out and bought uh substrates from a a couple of companies. I'm blanking on which ones. They've made commitments to substrate companies before. This is another set of them. And it's big numbers. It's a billion, billion dollars. Or I guess it's not that big a number for NVIDIA. It's a billion here, a billion there. It's tiny, tiny little things. But they have gone through and made purchase commitments through every part of the supply chain stack that they can touch.

36:21Jay Goldberg: Anybody who has capacity they're willing to allocate for sale to NVIDIA has gotten it, right? Lasers, substrates, packaging capacity, foundry. I imagine they've done all kinds of interesting things on test as well. You just go down the list of where NVIDIA is, they are aggressively deploying their cash flow to lock everything down, all the things that might be in short supply. And I don't see how that's not a problem for their competition next year. Because other than Broadcom, I don't know who has the balance sheet to stand up against that. I mean, I think it's going to be a real issue. Oh, and memory, and of course memory, right?

37:07Jay Goldberg: It's going to be very, very tricky for anyone to get any of those things next year just because it's tight and then NVIDIA is locked down anything that's locked down a ball.

37:22Ben Bajarin: So again, I think obviously suppliers don't want to get in a position where their customer concentration is one. But again, like you saw this in the Apple era with suppliers where they were the vast majority of someone's supplier because they essentially guaranteed your scale. You could just print that money with them. It was a variable for other players. They didn't always know. So they were happy to supply to Apple even though it was very hard and they didn't get the numbers that they might from others because Apple negotiates supply chain hard. But I assume that NVIDIA is doing the same thing.

38:03Ben Bajarin: But they're using this. again, this really does come down to who can manage this supply chain and a very complex supply chain to go and build complex systems across the board. And NVIDIA is doing this again, as well as anybody. Maybe second only to Broadcom who wants to also enable a whole set of customers. But you got to get all the things, right? memory and baseboards and all of the supply componentry around that, analog whatnot. And they have the balance sheet to do it, right? It's just another way they're using their balance sheet for competitive advantage is soaking up the supply chain.

38:49Jay Goldberg: You know, we're not going to be at the peak supply chain constraint until we're running out of MSG. That's my take on this. everything is so tight right now uh it's just it's gonna be it's gonna be a challenge next year i actually wonder if we're at the point now where it starts to become a real headwind for companies like we haven't seen anybody miss well i guess we just saw lamb miss well not miss but we saw lamb have problems with their results but i think i think we're gonna see more of that next year as companies just like hey we're leaving demand on the table we couldn't satisfy all our customers because we just couldn't get enough, you know, PCB boards.

39:35Jay Goldberg: Like PCB boards are short. Like what, what, what timeline are we on? There's just not enough PCB boards out there. It's a, it's a, it's a, it's a weird time. It's a good time if you make any of those things.

39:50Ben Bajarin: Yep. Yep. Um, okay. So I do want to talk about, I don't know if we've talked about this, but I do want to talk about the news-ish that some of the bears propped up about. I think it was a FT article or information that somebody talked to Amazon and they have GPUs sitting in warehouses. And this was made to be like a big, oh my God, NVIDIA's channel stuffing. and that's what's making this whole thing a whole problem. So I think it's important to talk about this in a couple contexts because I checked with some of my friendlies who build data centers for said people. And essentially the term they use is like, this is entirely common in the supply chain.

40:44Ben Bajarin: It's called staged integration. You get the things you need across the board and then you have to phase when that gets built out because data centers you just can't build to demand. You can't plan and say, hey, this is my timeline. And so unlike iPhones, which try to build supply to demand in real time, do real-time demand, you just can't do that with data centers. And their point was these things never sit idle for more than six months. Two quarters is the max. Then they go and get installed as a part of the integration site. But a lot of stuff sits there. Their substations sit there waiting to be installed.

41:21Ben Bajarin: They've got a whole slew of equipment. It's either in close proximity or it's on campus and then gets involved in its staged integration. And so I get that there's concern about that, but it's not the kind of thing like these things don't sit idle for very long. Like nobody who does this, their businesses depend on, has given me any of that indication. And it's just the complexity of that supply chain from you got to build this thing, then you got to install everything, then you got to get it all plugged in. And you have to have everything already ordered to do that. and that's just how this works and there's just there's really not any other way around it because you have to order these things in it well in advance so that you can then install them when it's ready to be installed and there's just timeline discrepancies in construction think of it like right you're waiting on wood you're waiting on cabinets you're waiting on whatever like the more you have ready to go the better your timeline is and this is just how you build things and that now componentry or compute supply chain comes along with that.

42:30Ben Bajarin: So I think that's point one. The other is if I'm a data center provider and I need to buy GPUs, I'm certainly not going to be like, oh, man, I need an extra six months. Can you hold my order or whatever for an extra six months? Well, then why not just give that to a competitor who can install right now? right and so you have to secure that capacity so that it doesn't go to somebody else when you need it so it can get installed when your build starts to happen right and so this is just this is how these things happen like the the comments i heard you know around this point in concern was more like oh well they shouldn't have ordered it i was like well that's stupid like that's not how construction works like that's when you're building something you order what you need and it has to get installed on the timeline some reasonable timeline it doesn't have anything to do with there's no demand is my point that was they're trying to be like oh there's no demand it's like again it's not a demand issue it's a timeline of construction and that's just

43:36Jay Goldberg: the reality of building things so so all right so i i agree there's no demand but i think there is something else at play here um because i actually think the i think it was the information reported this and they i think they were reporting on comments that andy jassy himself made that led to this and i think the the point that he was trying to make was they don't have enough electricity sure right it wasn't a question of demand they don't have enough electricity and that is leading them having to stockpile yeah stage gpus right um so it's not a problem of demand And I don't, you know, it's not a problem of demand, but it is a problem.

44:19Jay Goldberg: And I think this is what I was saying before, a minute ago, about what we're in store for next year. I think we're going to see a lot more of this, which is just things in the supply chain are going to trip people up. And it's going to be somewhat, seem somewhat random. It'll be gas turbines for someone. It'll be some kind of weird regulatory thing at another place. It'll be a shortage of electricians in some area. There'll be lots of little things like that percolating out there. And there is a risk that that leads to a significant delay. Sure. If a big enough site faces some kind of problem, it will ripple through the rest of the supply chain.

45:01Yes. And if it's just a few months, you're right.

45:05Jay Goldberg: It won't matter. Like, oh, we can't get the interconnect because there's a seasonal migration of whatever bird across where we have to lay the power cables. So we're going to push it out for three months. Right. That probably doesn't matter. But if it's something more significant, say the shortages around financing, you know, I think that will cause people to get a little more concerned. um so i it's something i'm watching for pretty carefully and i don't think it's like yeah right i was just saying this a minute ago is like amazon amazon doesn't have people inside amazon are complaining they can't get enough compute for their own projects um and so it's not a question of demand um but it supply chain is is is snarled up and it's going to get how the industry or the markets will react to that yet.

46:05Yeah. Yeah.

46:06Ben Bajarin: I mean, I don't think they'd react favorably, even though you could argue that, again, it has nothing to do with demand. But I think my point of this is my, and having these conversations with my friendlies, like you get an appreciation of just in general, how unpredictable your timelines are for construction. Like it's never the best case scenario. Like it's never like I've hit my timelines. It's like, ah, I get hit somewhere. So I got to balance that and you got to manage that accordingly.

46:38Jay Goldberg: Yeah. Yeah. I think we like, it's a little bit different though, because you know, we're dealing with the world's biggest builders. Right. And I know people who are on the neocloud side and like they have project plans from the builders months in advance of the first shovel coming on site right and it's a detailed timeline of what every how everything staged when it's going to happen yeah right it's it's not like you're you're you're like you said your kitchen getting remodeled and your contractor flakes out and just doesn't show up for a week because he's right you know just doing something else like this is this is really this is very professional, very, very efficient organizations at work here.

47:25Jay Goldberg: And certainly they've all become very, very good at this. The issue is more around random, unpredictable things, which is to your point, that are just beyond anybody's control. So the timelines I think are very, very clear, but there are still quite a number of X factors somewhere along the supply chain that are going to disrupt us.

47:45Ben Bajarin: Right. And I think my point is, is like, if you're those people, if you're one of these data center builders or whatever and again you have you have a timeline it again it's not going to be perfect but it's at least for what you're hearing now it's unlikely to be like a year again maybe something comes up right but but to a degree like i have not heard anybody be like yeah man we're going to be set back 12 to 18 months um like again if you're the vendor do you do you not order something or do you order something because you know you need it you order to what you need and sure you have to wait to install it or whatnot but you're not going to not order it like you don't want to lose that allocation from to a competitor right yeah i i

48:29Jay Goldberg: think that's the way it is certainly the way it was last year and probably right now i don't know that we'll be there in a year right i just i i think there's there's a there is a risk that like things aren't like certain things will be sold out but other things might not be and so i i think there is a risk that we do see a one you know i like i've said this before there's no such thing as a one quarter problem and so a data center comes along and says hey we're pushing out our stuff by a quarter then it compounds yeah then you know some then the odm has to say oh it's you know, it compounds and the ODM is pushing things out and then they're pushing things out to their suppliers.

49:17Jay Goldberg: Like I could, I could see it rippling through. And I think if this had happened, if there's this kind of delay had happened six months ago, no one would care. The street would look through it. I just don't know where people's heads will be in a year. Right, right. And so, you know, it sounds, it sounds crazy now, but I think there's a real risk that some company, some large ship company, random one, pick a random one. I'm not thinking of anybody in particular. Someone says, oh, we were going to supply this data center and they're delayed. And so we're going to miss our quarter. Yeah.

49:49Ben Bajarin: I think it's fair. And again, it comes down to is there another waiting customer or not. And then said person was host. Because it was interesting. the the point was made around memory similarly that the by Gavin Baker on I forget which podcast he was on where he was basically like if you the customer know that memory is constrained right your supply chain and why would you because I think the contest of the question was like well people don't have to abide by their LTAs and I was like well why would you when you have again the the three main players that you need and memories constrained why would you not honor your LTAs when then that would just piss them off and they go well I'm gonna give it to your competitor so it's like you got another customer just waiting there you know that's how high demand is and so if you back out right or don't want to pay out I'll give it to somebody else and it's not a problem I thought it was interesting point

50:53Jay Goldberg: this is my point I want to end this on a happy note so I'm gonna I'm gonna let that go absolutely you're right okay all right lastly let's just talk about security uh freak people

51:03Ben Bajarin: freaking out for good reason like it's it's interesting because one like crowd strike and then some others were like hey we think people are feeling out what they can break into like right now we don't we don't know where it's coming from but there's some definite like phishing people are trying to log in it's coming at scale um banks were sort of saying something Like we got people trying to log in. We don't know what's happening in our systems. This, you said you didn't want to add negative. Like this is actually one of my bigger concerns that like something really bad happens in security and that slows the spend or deployment of AI pretty broadly because as we know, agents are running around, ungoverned, I have data around this that will show.

51:54Ben Bajarin: You gotta haul these things in. It is absolutely a security risk. And people, I think, are now just realizing the degree that this is a security risk. And either, I mean, I was talking to one CTO and they were like, we're about to have a blank check for agentic security. Like, we're going to, we have to solve this. It's top priority number one. Everybody's going to start spending insane amounts of money to figure out how to secure this. But these little things that keep coming out show the magnitude of this problem. Like, let's solve this sooner than later because this could be really, really bad.

52:28Ben Bajarin: But you're starting to see hints of how bad this could be.

52:33Jay Goldberg: We're all doomed. Run.

52:37Ben Bajarin: I mean, I do need a vacation, Jay. So, like, the fundamental collapse of the infrastructure so that I could just kind of hang out, eat some food, be on my farm and not be insane with work, you know, it would be, I guess, welcome.

52:52Jay Goldberg: Yeah. Yeah, I'm with you on this one. Like, I think, like, I don't want to dismiss it. Clearly, AI is immensely capable of enabling bad people to do bad things. And I think it is very likely that we see even more significant attacks. There was this thing last week with somebody figured out how to turn off all the refrigerators at a supermarket chain in France, all of them, and so all the meat spoiled in France.

53:26Jay Goldberg: We're going to see more of that, and it's going to be worse stuff. But I also don't think it is this unmanageable, infinite problem. it reminds me a lot of like what was it like 20 years ago when the internet finally got really robust and we started seeing open source security tools that people then reverse engineered and weaponized yeah and we we saw a lot of a lot of cyber attacks back then um and there was this whole concept of of script kitties do you know this right they were just like there are a lot of open source security tools and malware tools that got released at one point and it just allowed any teenager, any bored teenager with an internet connection to do all kinds of malicious attacks on like, you know, their school website to change their grades, that kind of stuff.

54:12Jay Goldberg: Or, you know, the defense department. And I think we're seeing that repeat itself now because the class of tools have gotten infinitely more capable. and there's a lot of scope for doing creative things in a bad way.

54:34Ben Bajarin: Yeah.

54:35Jay Goldberg: But I also think, like, one, I think, like you said, everyone's very alert to this and people are being proactive and giving security real budgets for the first time, right? The knock on corporate security is that, like, everyone talks about it but they never fund it enough. That has probably changed radically this week. And then on top of it, I think it really shows that this is why having frontier models is going to remain important. After coding, cybersecurity is going to be, I think, the number one AI use case. because if all the script kitties and global mafias have access to N-class frontier models, somebody has to come up with the N-plus-1 class.

55:26Ben Bajarin: Yeah. And, yeah.

55:31Jay Goldberg: Look, I have to believe that or I have to, like, move to the woods and build a Faraday cage around my house

55:37Ben Bajarin: and just farm my own things. But what's interesting is like, you know, NSA goes something like, hey, if you're running on PHP, like you might want to fix this. And they're like, oh, you know, don't worry about it. What runs on PHP? Only your electrical grid and water systems and a whole slew of things. But I do want to say like you're exactly right because I have for years had conversations with those in IT around this. and they're always like, yeah, security is the top spend. Like that's what, all right, so how much did you spend on security this year? Ah, it's about the same. The difference is none of, I don't think they ever really, like they didn't, they kind of had to worry about it, but they didn't really have to worry about it.

56:22Ben Bajarin: They weren't hacked, they weren't breached. And I don't think they also knew how bad it was. And this is what everybody we talked to has like walked away with with this mythos moment, which is that they used it to test vulnerabilities and basically were like, this is really bad. like yes we have insane vulnerability and so now they know this they have to go figure out and spend to patch it and solve it and that's why i think this spend is a different cycle like you now know how bad it is and that anybody like you said with a model frontier model can come and attack you and exploit your vulnerabilities so security spend is going to be significant to say the least yep yep yeah all right well let's end on a i'm terrified but in a good way in fact i had i was giving i was just giving this talk yesterday to a big uh venture capital group on the state of ai and i took a land at this point that i said you know it's interesting that ai is both simultaneously one of the most if not the most transformational technology we've ever seen and one of the most dangerous at the same time.

57:35Ben Bajarin: And so there you go. Move to the woods and build bunkers, everybody. We're just kidding. All right. Thanks for listening. We will talk to you next time as next week is an insane earnings week for hyperscalers. So that will be amazing. And we'll talk about it. So we'll talk to you then.

57:53Jay Goldberg: Thank you for listening, everybody. Tell your friends, your neighbors, your agents, click like, subscribe. Thank you. yeah um oh uh

58:10yeah

58:15you you and you Thank you.