Silicon Futures for June 2026 - COMPUTEX, Nebius Inflection, Qualcomm Investor Day, and more..

Silicon Futures is a neXt Curve reThink Podcast series on AI and semiconductor tech and the industry topics that matter.

Leonard Lee (neXt Curve), Karl Freund (Cambrian AI Research), and Jim McGregor (Tirias Research), unpack a dense and drama-filled month in AI and chips.

U.S. AI labs start a whining campaign as Chinese open-weight models threaten the AI world order, NVIDIA resuscitates the AI PC, neoclouds lay out their vision for reinvention, Apple flexes its agentic AI muscle, Qualcomm brings heterogeneous computing to the AI data center game, and much more.

In this episode, Leonard, Karl and Jim talk about some of the top headlines from June of 2026.

➡️ Jim's takes from COMPUTEX 2026 (2:41)
➡️ The agentic workstation and revival of the AI PC with NVIDIA RTX Spark (3:22)
➡️ WWDC 2026: Strike 3 for Apple Intelligence? - (10:25)
➡️ Revisiting Apple's AI strategy (11:20)
➡️ Neocloud takes from Nebius Inflection 2026 (16:03)
➡️ Neocloud and their enterprise pivot play beyond GPUaaS (18:40)
➡️ Leonard's take on Cisco Live! 2026 (21:43)
➡️ The race to reduced the procurement-to-first-token for AI clusters (23:55)
➡️ The great closed AI model drama (26:49)
➡️ The great U.S. AI lab whining campaign (32:00)
➡️ Inflective reflections on Qualcomm's Investor Day 2026 (34:10)
➡️ Qualcomm's Reverse Ricky Bobby move (37:57)
➡️ Qualcomm's hybrid approach to the AI data center opportunity (43:44)
➡️ Apple's dramatic move to raise prices thanks to memory crunch (49:05)

Hit Leonard, Karl, and Jim up on LinkedIn and take part in their industry and tech insights.

Check out Jim and his research at Tirias Research at www.tiriasresearch.com.
Check out Karl and his research at Cambrian AI Research LLC at www.cambrian-ai.com. Check out Karl's Substack at: https://substack.com/@karlfreund429026

Please subscribe to our podcast which will be featured on the neXt Curve YouTube Channel. Check out the audio version on BuzzSprout or find us on your favorite Podcast platform.

Also, subscribe to the neXt Curve research portal at www.next-curve.com and our Substack (https://substack.com/@nextcurve) for the tech and industry insights that matter.

NOTE: The transcript is AI-generated and will contain errors.

DISCLAIMER: This podcast is for informational purposes only.

Transcript

Karl Freund 0:06

Next curve

Leonard Lee 0:09

Hey everyone, welcome to this episode of Next Curve's Rethink, a podcast where we, break down the latest tech and industry events and happenings in the world of semiconductors, and Carl Freund's favorite topic of all time, AI. We distill this stuff into the insights that matter on this Silicon Futures program, which we do in collaboration with Cambrium Research as well as Turius Research. And of course, everyone knows Turius Research was founded and is led by the fantabulous, incredulous... i'm probably using that term, really improperly, but- Are

Jim McGregor 0:51

you just printing out words? Are

Leonard Lee 0:53

Yeah, exactly Anything with incred, it just seems to fit you, Jim. Jim McGregor of Turius.

Jim McGregor 1:01

Sean McGregor.

Leonard Lee 1:03

Sean McGregor- of Turius Research. Gentlemen, welcome. It feels like an eternity. Hope you're well. But before we get started, hey, look, in this episode, talking about the headlines for the month of, June, right? Are we in June already?

Jim McGregor 1:22

No, we're in July, so you need to get on track.

Leonard Lee 1:24

Yeah, but we're gonna talk about the headlines from June. quick. please remember to like, share, react, and comment on the episode. Also, subscribe here on YouTube and on Buzzsprout. Listen to us on your favorite podcast platform. Opinions and statement by my wonderful guests are their own and don't reflect, mine or those of Next Curve, and we're doing this for informational purposes only, to provide an open forum for a discussion and debate on all things, semiconductor and, AI. And with that, we're gonna be talking about Computex, Nevius, Qualcomm Investor Day. The list goes on and on and on. There's so many things that happened in June. It was probably, my busiest month of the year, which typically June is pretty mellow, but this year it's absolutely insane. So let's start off with Computex. Jim, you were, the only guy there.

Jim McGregor 2:16

I know.

Leonard Lee 2:17

I was there in spirit.

Jim McGregor 2:19

No, it was, it was interesting. We've seen Computex as mostly a PC show, but it's kinda morphed over the years to focus on, more of the compute, infrastructure and everything else. I wouldn't say there were a lot of h- major announcements necessarily there, because we have a lot of companies have their own events now either earlier in the year or co- coming up, like AMD's coming up, later in July. So they're kinda saving some of their stuff. But really the one that stole the show was, really from Nvidia, you know- Yeah Nvidia coming out and talking about, getting into the PC segment with, their solution. that was pretty much, I think everyone was talking about that,

Leonard Lee 2:57

RTX Spark, right?

Jim McGregor 2:57

It's the RTX Spark- Yeah which is the same silicon as the DGX Spark. it's still using the GB10 processor, but now with Microsoft support, they're gonna be able to put in small form factor PCs and into, mobile PCs. it's gonna be interesting, and obviously that, that's, that along with the, Apple, MacBook, Neo and all that stuff, you know, all these small form factor PCs, and I think these AI mobile PCs are really kinda driving interest in the market. And that was really driving a lot of the interest at Computex. In fact, I would say that, one of the interesting things that I'm seeing is, the small form factor, either workstation or PC. I think it really becomes not just that preferred kind of productivity tool, especially for developers, but I think it eventually becomes that in-home AI appliance.

Leonard Lee 3:48

Yeah.

Jim McGregor 3:49

There was also gaming there. Obviously, it's always a big gaming show. the biggest news there obviously was from kind of the Intel camp, they have their solution for the handheld games. So, One Player X, ASUS, and MSI all introduced handheld gaming consoles, with the Intel solution. Acer upgraded theirs with the AMD solution. So a lot of interest in handheld gaming, although I'm still not understanding why, because I don't see anybody use those outside of the home or office. Yeah. But, interest there and, other stuff, we didn't see as much around the data center, but still a lot- Yeah, that's interesting of interest in workstations, especially for AI developers. Yeah.

Leonard Lee 4:30

Well, you know, that was a thing last year, right? You and I were talking- Mm-hmm about how the workstation is going to be sort of this beachhead for enterprise AI, and then now we're seeing this evolution of this whole AI PC talk track, that hit like a, you know, a patch of molasses, over the last couple of years, but it's being rejuvenated in a way with, a lot of what, you know, let's be frank, what Apple has done with the, the Mac Mini- Mm-hmm w- e- especially, with the advent of open Claude, so this whole agentic movement really taking off o- on top of, ironically, Apple, which is, was supposed to be, behind on AI, all of a sudden everyone, you know, these things are sold out, right? but one of the things I thought was really interesting with RTX is that whole chip architecture is really mirroring what Apple has done, right? Yes. And what Apple and Qualcomm has done. With the unified memory, you have also, the, what do you call it? Integration of, two die, in a similar fashion as, a Fusion Ultra. And so, you know what? It, it's gonna be interesting to see how, uh, NVIDIA competes against Apple. I think Apple has a lot of game in this segment, right? This emerging segment. And, I think it's been highly underappreciated. And if Apple has game, that, I think, Qualcomm, is, a either a close second or peer, especially for, the PC segment. So, you know, the thing about NVIDIA though is that they're just that 5,000-pound gorilla, and, we'll see whether or not they can, they can make, AI PC happen, in, in the way that everyone was expecting two years ago, you know?

Jim McGregor 6:18

Well, and that's what's interesting is Apple has kinda gone down the stack to being in more cost-effective PCs- Yeah more mainstream, PCs in the, $599, $699, price range, while most of the PC industry keeps going up, even with- the ARM-based solutions. Qualcomm did introduce the Snapdragon C, which is gonna supposedly go into that, that price range. Yeah, the lower price range. But, a- and Intel introduced their Wildcat Lake that's going to that price range. But for the ma- vast majority of it, especially when, uh, at Computex, vast majority of all the PC stuff was still aiming at that, $2,000 PC.

Leonard Lee 6:52

Yeah, yeah.

Karl Freund 6:53

Yeah, I think many of us thought that the, the whole, NVIDIA platform, the DGX Spark, would take off if Microsoft added their support. Well, now they've got it. Yeah. So now they've got an operating system, they got a full stack, they got all the software NVIDIA's ever written and sold- Yeah all available on a relatively affordable, still a little bit expensive, but- relatively affordable-

Jim McGregor 7:17

Depends on how much memory you put on it.

Karl Freund 7:19

That is, that, that is absolutely the case It's absolutely the case. So it'll be very interesting to see what kind of demand they realize. If, if there is such a thing as an AI PC market- Yeah then Nvidia should lead it. i- but we're not sure there is such a market, so we'll find out.

Jim McGregor 7:39

And you have to remember that if you're putting 128 or 256 gigs, like what they, recommend for, like, the DGX Spark, if you're putting that much memory in something, you're pricing at the price of a small car. You know, you're, you're talking- Yeah $4,000 or $5,000 or more. You, especially with memory limitations today. You're not talking at a $2,000- Yeah uh, well, you know.

Leonard Lee 7:58

What year are you in? Like 1980? That's like Hyundai Excel. I can't...

Karl Freund 8:03

My car, no.

Jim McGregor 8:04

Okay, okay, okay. Uh, but I'm just saying- Yeah, definitely not it's not cheap. And, and matter of fact, that was one of the key- There you go questions I think everyone asked, wa- was concerned about with the announcement- Uh-huh was we're not gonna see these things until the latter half of the year. And- Yeah is Nvidia... And now Nvidia's the top of this food chain right now. can Nvidia get enough memory? And, what are these things gonna be priced at when they come out? Because, I don't know too many people that are willing to pay five grand for a PC, but we'll see what happens.

Leonard Lee 8:31

Yeah. Tell that to people at Apple. and speaking of Apple, WWDC, any thoughts there?

Jim McGregor 8:40

first off, they were focused obviously on AI and having an AI solution Yeah And, it's basically, the foundation of that is gonna be Google Gemini. Now- Yeah what I liked, the fact, what I really liked about it is the fact that, it works across all the Apple applications. Yeah. Now, when you pick up Gemini on your Android phone today, it doesn't work across all, even the Google applications, it doesn't work across them. It, it's a pain in the ass to get it to work across that. Now, I assume Apple's gonna have the same challenges with third-party apps, not getting- Yeah, yeah so I think that's good in making it actually a true Apple experience. Yeah. Now, my only concern, and maybe this is just a short-term strategy, and I'm thinking it probably is, that I don't think making it or, or, you know, really basing it just on Google Gemini is the best solution. I think in the long term, you wanna have all these different solutions maybe- Yeah as possible baselines, whether it's Claude or Perplexity or Gemini or ChatGPT, because let's face it, we're seeing so much innovation in those different models, it's changing rapidly. Yeah. And what... if it g- Gemini may be a good solution today, but it may not be the best solution tomorrow. So we'll have to wait- Well, yeah and see how that morphs. And quite honestly, Apple, knowing Apple, they may have their own solution within two or three years that kind of eliminates using anybody else's AI. So it's gonna be interesting to see how it evolves. Yeah.

Karl Freund 10:02

I agree with you. I, I think the, the unspoken hero here probably is not Apple, it's Google. Mm. Google has now extended their reach all the way down to all the Apple family devices, all the way to beyond their own data centers, beyond their own cloud. They're selling TPUs in large volumes to a select number, at least initially, a select number of, hyperscalers-

Leonard Lee 10:28

Yeah

Karl Freund 10:28

to really take on NVIDIA in a very big way. Um, so I think, I

Jim McGregor 10:33

think-

Karl Freund 10:33

And even,

Jim McGregor 10:34

and even having their own custom NPU that you can use as IP. Yeah. Yes. Mm.

Karl Freund 10:38

Yeah. I really think in five years we'll look back at this moment in time as a turning point for Google- Mm becoming a real leader in AI

Leonard Lee 10:48

Yeah. my take is this. first off, I don't think Apple is gonna come up with their own foundation model. They're gonna probably continue to do-- to execute on a similar strategy globally, because one of the things that's happening is this whole sovereign movement. And so by region, you're seeing a lot of regulations, especially around AI, that are creating these chasms, right? For especially US AI companies. And so le-let's take one of the largest markets, for AI in the world, which is China. Right. Google's not gonna be in China, and Apple knows that they cannot bring anything Gemini-based into the Chinese market. And so this is, this... And it might not be just China in the future. You're gonna probably see Europe also become... And they already are a, a, a, um, a, a, becoming a sovereign market. they're starting to wall themselves off and create barriers for US AI companies. And so I think they have a... I mean, you know, look, I think the folks at Apple are not stupid people They're very, very smart. They know where to invest, and, they're looking at what it takes to, diffuse AI in a, safe, private way across their portfolio devices and, services in a way that is truly-- is going to be global. but part of that is transversing all, all this fragmentation that's happening right now because of these sovereign movements. So, we'll see what happens. I'm actually not that bullish for, Google because of that. I think they're gonna be less and less a global company, and there's gonna be more regulations that are going to, start to infringe upon their market share and mind share, and I think Apple knows that. So, but we'll see how it goes. but one of the things that I thought was really interesting, I sat in on a couple of developer, sessions, and, there was one on MLX, which, is a, a software extension that they have in their AI stack that allows for you to create, like, clusters of Macs, for distributed inference and training, right? And, this is over a Thunderbolt 5. Really, really interesting stuff, but it takes, like, this whole idea of putting a model and running a model, locally on a cluster- a small cluster of these tiny little Mac Minis to the next level, right? I thought that was a really huge announcement I think the competitive landscape around this sort of outside of the AI super data center, landscape is getting really, really interesting. So let's talk about, um, NeoClouds really quickly. I, I was at the whole Nebius thingy. I know, Carl, you've been following NeoClouds for quite some time, and, what's happening with them, especially as it relates to their, AI infrastructure investments. Nebius had their first inflection analyst- Okay and media event. any reflections from that event?

Jim McGregor 13:57

Reflections or inflections? Are we talking about-

Leonard Lee 13:59

Inflections No,

Jim McGregor 13:59

just kidding. Well,

Leonard Lee 14:02

yeah. Well- Inflections

Jim McGregor 14:02

well, my, my colleague Kevin Hein was there, and, he was really, amazed, and I, I think that, some of the NeoClouds, especially like CoreWeave and Nebius, are kinda getting a bad rap because, now they've got a lot of capacity, and a lot of that capacity is being chewed up by the hyperscalers. So they're looking at it like it's just an extension of the hyperscalers. But as Nebius really pointed out at, Inflection, they have an entire software stack to help companies, really implement AI solutions. So they're more than just a GP as a service provider, more than just a hardware solution. They've got the software stack and the engineering resources to help companies actually use those resources and become an AI company. Yeah. So that was at least for us, was probably the biggest takeaway from Inflection.

Leonard Lee 14:47

Carl, any thoughts on what's happening- No with all these guys?

Karl Freund 14:51

not really. I think that what we're seeing is Nvidia is helping cloud, NeoCloud companies 'cause they don't develop their own competitive hardware chips.

Leonard Lee 15:04

Sure.

Karl Freund 15:04

Right? Sure. So the, at the end of the day, Nvidia would much rather sell to, to Nevius-

Jim McGregor 15:12

Mm

Karl Freund 15:13

than to somebody who's developing a competitive chip to push them out of-

Jim McGregor 15:17

Yeah

Karl Freund 15:18

their infrastructure, push Nvidia out of their infrastructure. So I think a smart move on both, the hyperscalers, the Neos cloud, the Neo cloud providers, and Nvidia to try to wall it off. Say, "No, if you wanna come to me, it's gonna be Nvidia."

Leonard Lee 15:33

Well, last month you pointed out, and w- and we talked about how Nvidia broke out, their reporting for data center, right? Yeah. And then the NeoClouds,, these stuck on one side w- along with enterprise, and then there's a hyperscaler bucket. And, I think you're right, that it's definitely a captive market largely for Nvidia. One of the things that I- I got out of, being at Inflection is that the, the Neo-clouds, are... they're not quite there yet in terms of having their, next layer of quote-unquote platform-ish kind of services. they're trying to get there, right? I think one of the things that the Neo-cloud players, the, pure play Neo-cloud players are finding themselves in is a situation where they need to evolve beyond GPU as a service, right? The infrastructure as a service game. now, whether or not, because there, they, there is this, talk track that the hyperscalers or the cloud service providers who have their own Neo-cloud, stuff, are at a disadvantage. I don't know if that's necessarily true. Maybe they're not a- under as much pressure to, take their RPO and deploy, infrastructure as quickly as possible to tap and, convert that into revenue. But, they have the enterprise relationships, right? And I think that's the one thing that the Neo-clouds lack. And so I'm looking forward to seeing how these guys figure out how to actually create enterprise markets for themselves, it's easy to say, "Hey, we have AI stuff and we have a stack." But, everyone has a AI factory stack you know? Everybody. And so how do you differentiate? That's the big question mark.

Jim McGregor 17:20

Well, and to your point, as they expand out beyond just GPU as a service, and, um, Carl's right, pretty much they are all an Nvidia house from the GPU perspective. Yeah. Yeah. But from the CPU perspective, they're actually an AMD house. So, you know, how can they leverage that and use that, the power of the AMD processors as well as the rest of their ecosystem to actually offer additional services beyond just those traditional AI or GPU as a service, offerings?

Leonard Lee 17:48

Yeah. And I think the, the hope is that agentic, creates a new inflection point for them.

Jim McGregor 17:54

I think it's not just in agentic, 'cause I know that, they've talked also about other things such as being the backbone for physical AI and robotics and stuff like that. So I do give them credit. They are looking well beyond just, where they are today and where things are today. But, where do they need to be to meet those market inflection points going forward, hence the inflection

Leonard Lee 18:17

Ah, yeah,

Karl Freund 18:17

absolutely. There are a lot of inflections around here. Maybe it's the- Yes, we are It's gonna be- Inflection. Inflect- inflection spelled with a Q is actually a quantum computer company just down the road here.

Jim McGregor 18:28

That's true.

Leonard Lee 18:29

Ah.

Karl Freund 18:30

In another county

Leonard Lee 18:32

Okay. Okay. Yeah. I'm, wanna, make a really quick, comment here about I was at Cisco Live, that was the first event that I was at. And, from a silicon perspective, there was this, emphasis on Silicon One. To be honest, I didn't know too much about Silicon One until last year, when I hung out with, Martin Lund for the first time and he kept saying, "Our stuff is programmable. Our chips are programmable." I'm like, "Okay, what do you mean by that?" What he's really talking about is, it's called P4 programmability. It's like this layer that sits on top of, it's hard to describe. Anyways, the long and the short of it is, I think there's, like, a value proposition especially for hyperscalers who are starting to really- push the envelope of w- of networking, right? Especially for, their inference and their agentic, infrastructure. And so, one of the things that where Cisco is finding a home for their, networking technologies is with the hyperscalers, right? They're trying to branch out and diversify their, range in terms of the customer, segments that they hit. But they have a pretty nice home in hyperscaler. And I think, going forward, and something to watch, right, is that, Silicon One and programmable, network, what they call network, processing units, may have a compelling value proposition, especially for hyperscalers. And, we're starting to hear more about, like, kernel optimization and network optimization. There's all these optimizations that need to happen. And to have that flexibility with the f- programmability, I think, could be a key differentiator, and something that I think maybe the, the industry will start to look at a little bit more, especially as you see these large scale bespoke, requirements start to come down the pipe. And the other thing that I noticed, and it is something that we've probably observed over the past three years, this, requirement to just get stuff out And generating tokens as quickly as possible from the moment that something is procured, to, the time that it's, uh, received, getting that thing deployed. I don't know, what do you guys think about that trend of m- I think we're seeing that with Dell, HP, all these guys having sim- similar talk track, but any, inflections on that?

Jim McGregor 20:52

Well, I think, you're on a key point there is that the flexibility that they're trying to build in with their networking is critical because, the networking requirements are going to vary, not only by the type of workload you're running, but even the size of workload. So even if you're just running AI workloads in your data center, that size is gonna depend on how many GPUs are you trying to network together? how many do you need to be able to run one or millions of models simultaneously, especially if you're doing agentic AI and you're running them, in parallel or in series. Cisco is, making sure that those programmable solutions work with optical networking and even co-packaged optics to make sure that, they're gonna have that flexibility even as networks increase in terms of performance and, efficiency.

Leonard Lee 21:38

Again, something to keep an eye out on because these, large scale infrastructure, I mean, inference AI systems are just starting to, trend, I think, especially- Well- the agentic stuff.

Jim McGregor 21:51

Well, and you point to two key things. One, the hyperscalers definitely are taking any hardware that they can possibly get their hands on right now because the demand is insatiable. But also, hyperscalers are unique in the fact that they wanna have, different types of hardware, because let's face it, no two workloads are gonna run the same on each pers- piece of hardware. So some hardware is gonna run some workloads better than other hardware. So they're trying to make sure that they have kind of that, diversified solution depending on what, the customer needs. Uh, that does hurt them in certain ways. We've seen this over the past year where people really wanted Nvidia GPUs, and some of the hyperscalers just didn't have enough of them. So they're going to partner with like the Neoclouds. But, i- in the end, they have to think about all their customers, and it's a much more diverse set than just, AI workloads.

Leonard Lee 22:42

hey, what about all this like, drama around, closed models. You had Anthropic, Mythos, and Fable, you know. they became available and then they got banned. I don't know. It's like, it seems like every day the situation changes, but this is something that, was a thing in June, right, in the world of AI. And so, any re- any inf- inflections on that?

Jim McGregor 23:10

He had

Leonard Lee 23:13

to get that in, didn't he? I know I'm totally using that term wrong, but

Karl Freund 23:15

Totally wrong.

Leonard Lee 23:16

Yeah, yeah, yeah. Any, uh, reflective inflections. I,

Karl Freund 23:20

I would, kind of shift the conversation in that vein towards OpenAI.

Leonard Lee 23:25

Okay.

Karl Freund 23:26

We, we've seen OpenAI's virtual stock price. They don't have a stock price, but we've seen their stock, and their company and brand diminish somewhat this year-

Jim McGregor 23:40

Mm-hmm

Karl Freund 23:40

as Claude continues to gain traction and a lot of revenue, and open source models continue to provide really good value to- Yeah companies who wanna take something and build a specific bespoke solution for their company. I kind of wonder if OpenAI, I know they're talking about maybe they're not gonna do an IPO this year. Yeah, which is- I, I think that their best days may be behind them. I think they missed an opportunity to get out with an IPO in a timely fashion while they were king of the hill. Yeah. And I'm not sure they're gonna stay king of the hill. I, in fact, I'm pretty sure they're not. We're already seeing things like Gemini back to Google, outperforming OpenAI. We're seeing Claude take off like crazy. Yeah. Yeah. and we're seeing open solutions like NVIDIA, NeMo, being applied- for specific solutions like BioNeMo. and I'm not sure where OpenAI goes now. Now they're gonna do their own chip, right? So they can lower- Wow their costs. Maybe that's window dressing for an IPO. I don't know.

Leonard Lee 24:45

Yeah.

Jim McGregor 24:46

I think we're gonna see a lot more of that as AI matures. I think we're gonna see certain companies excel in certain areas. Yeah. certain models, certain companies, period, th- they're gonna have expertise in certain areas, whether... and it's not just saying, well, one's enterprise, one's consumer. It's really gonna be, okay, one's really good for, materials research. One's really good for scientific modeling. One's really good for this. And we're gonna continue to see that, I think, as of, at, as things evolve. One, because they have to get more efficient because they're getting so large in a lot of cases. But two, just because of the extra expertise. You've got different experts working on different problems. even a lot of the, open source, or models that we have out there today are leveraging some of the open source stuff from NVIDIA. they're probably one of the biggest sources of open source solutions today.

Karl Freund 25:34

Yeah.

Jim McGregor 25:34

Um, who would've thought we would've said that 10 years ago? But they are. So it's- Yeah it's gonna be interesting. I think we're gonna... I, we keep saying this time and time again, that we're in the early days of AI, but everything's... We're gonna keep having these DeepSeek, revolutionary moments- Yeah every six months or every year, and it's gonna keep changing the dynamics of the market- Yeah for the at least the next five years, if not 10. What do you guys think?

Karl Freund 25:59

Yeah. I think, I think you're right, and I think the next big bombshell is- Not a language model. That training, training things on language models has gotten us very far. It's done really well. But there, there's a lot of content out there that we're not training on, things like video, which reflect the real world around us, and you add the physics to, to that, which is not part of training models today. if you start talking about real world training models, that could do to the entire, LLM market today kind of what DeepSeek did by sha- by kicking it in the shin with a more cost-effective solution.

Leonard Lee 26:38

Yeah. I think, just as long as they're not training on Marvel movies, then you're okay. Otherwise, the physics are gonna be totally off. Physics would be

Karl Freund 26:51

really weird.

Leonard Lee 26:51

Yeah. So, and then Star Wars, right? Because, there is no sound in space.

Karl Freund 26:58

Boom.

Leonard Lee 26:58

Well, you know, what is really interesting for me this year is how no one talks about AGI and S- ASI anymore. In fact, I think there's no appetite for it, which is probably one of the reasons why Rene and team should probably change the name of their CPU. Uh, but anyway, um, uh, the, the thing is, I'm coining this term where I think that's starting to get traction. It's called the reverse Ricky Bobby. If you're first, you're gonna be last. And so, you know, I think everyone's learning from the mistakes of the priors, and, I think it's gonna be... it's gonna get really ugly. this is..., The future that we imagine for AI based on a singular narrative, we get married too much to it and don't look at what's happening in adjacencies, you're gonna be in, you're gonna be in big trouble because you're not gonna see really what's happening. And, your mention of The Chinese models, they're not trivial. And right now I'm not a big fan of the AI, foundation labs, running, to, Washington and whining. I call that whining. They really need to compete and figure out- Mm-hmm how do they, how, how do they restructure their cost structures for whatever economy, AI economy they are imagining so that it can be competitive. Because, the DeepSeek moment was a bigger thing than any of these guys, were, portraying at the time. And what we do see is this serial whining, and I'm not a, I'm not a big fan of that. It's like, "Hey guys, get your game on. If you don't have your game on, you will be last." It doesn't matter that you were first.

Jim McGregor 28:43

And

Leonard Lee 28:44

to your point- Whine all day long.

Jim McGregor 28:46

Yeah. You know, and to, to your point, we're seeing a lot of innovation, especially on the hardware side. I think that's gonna change- Yeah that, within memory compute, like with the, the Groq solutions- Oh or even with the, HPC, high bandwidth compute solutions, which they're gonna be offered on the new Qualcomm C1000 processors and their, you know, AI250- Speed accelerators. You know, we're seeing a lot of innovation that's really, I think, that's gonna change a lot of those cost or those TCO dynamics, just being able to run and operate these data centers- Yeah for different types of workloads, especially inference processing.

Leonard Lee 29:19

Well, let- let's jump to Investor Day really quick. Qualcomm had their Investor Day, right? Yes speaking of like novel architectures and stuff, they introduced, what is that? HBC, high bandwidth compute.

Jim McGregor 29:30

Yeah. High, high bandwidth compute

Leonard Lee 29:32

is- I still don't really get it.

Jim McGregor 29:33

I'll give you the 30-second primer. It's LPDDR memory stacked on top of a base die, a custom base die. Basically, it's a full XPU. So think of, very similar to what they're trying to do with HBM4E, where it's gonna have a customizable die underneath DRAM stacks. This is gonna be LPDDR, so kind of the mobile memory, on a full XPU die. And then they're gonna be able to take this HBC and pair it with both, their C1000, processors, which will be coming out in the 2020 timeframe- and their AI250 accelerators in that timeframe. So they are basically trying to, re-architect everything to make it, more towards that in-memory compute for very high performance inference processing.

Karl Freund 30:17

No, I had a friendly discussion with my f- friends at Qualcomm about-

Jim McGregor 30:20

It was a argument

Karl Freund 30:21

yeah, it was an argument. Were so apologetic for not being late. They said, "We're not late. We're not late to this market at all." And I'm like, "Yes, you're very late. You've sat on the, 8 Cloud AI 100 now for seven years-

Jim McGregor 30:35

Mm-hmm.

Karl Freund 30:35

On the other hand, one could argue they've learned a lot while being close to the market, not in the market, and they realize it, it's really a memory problem. And so they- Yeah came out with this, with HBC, which I think is just a brilliant move, to lower the cost of high bandwidth computing and increase capacity tremendously. the problem is they came out and said, "Well, we're not late, but, but here's our roadmap, and now we're gonna be later." now- Yeah the roadmap

Jim McGregor 31:05

is rolling

Karl Freund 31:06

out too slowly.

Jim McGregor 31:07

And not just... And not just a memory problem, but also a networking problem. Mm-hmm. So they followed a very similar timeline or strategy that AMD, Nvidia, and Intel have followed in terms of we need to go out and acquire networking resources. Yeah. So they did that with AlphaWave. But to your point, AlphaWave is their primary, foundation into that market for the next, year and a half until they start coming out with these new products. That, and their custom silicon solutions, which- We don't really... They, gave some hints with some of the companies that were there. Obviously, there was Meta, there was AWS, there was Google, there was, Microsoft Microsoft and, Humane, were all there. but and we know Humane's already a customer. Now, we don't know- Yeah who's gonna be the custom silicon customer, 'cause they didn't officially announce that, even though they teased that they would. but they're gonna have that AlphaWave networking solution, which drastically changes their perspective on the data center market, because they're going from basically zero to, billions of dollars, $2 billion, you know, by the end of 2027, I believe. And- 27, yeah 15 billion by the end of 2029. So significantly changes their perspective. To, so to your point, I don't think that they took it as seriously as they should have, but I think that acquisition of AlphaWave and that acquisition of Modular- Yeah which they just announced, from a software perspective, making it easier to run, software on their platform, I think that, those two things are- Yeah, I,

Karl Freund 32:35

well, I couldn't agree more. I think if you're gonna show up late, show up big.

Jim McGregor 32:39

Mm-hmm.

Karl Freund 32:40

And that's what they did. They showed up with a roadmap that was jaw-dropping on what they're planning to do. So they've gone from not just late, late to the party, but not even in the party,, to positioning themselves to become a major force in the industry. I was very impressed with their roadmap. Mm-hmm. I wish it wouldn't s- taken so long to develop, but that's kind of the nature of silicon.

Leonard Lee 33:02

It's reverse Ricky Bobby.

Jim McGregor 33:05

I can't believe we're going back to Ricky Bobby again.

Leonard Lee 33:09

You know, that movie is the Bible of tech. Come on. Okay. What

Jim McGregor 33:12

did you watch this weekend over the holiday weekend, Leonard?

Leonard Lee 33:17

But, you know, hey, look, yeah, you mentioned the Modular acquisition. I thought that was really interesting. Heterogeneous, kernel optimization. Everyone says software, but it's, it's a much... My view is, or understanding, is a much lower level, thing that we're talking about in the software stack. But this idea of optimization at the kernel level across heterogeneous IP, that, that's a big deal. I, I think, that's something that N- Nvidia doesn't have at the moment because their stack is so GPU centric. this could actually be a big deal for Qualcomm and a differentiator because, I think- Most folks don't understand that these guys already have a AI s- stack, and they've been doing this stuff on a, especially from an inference perspective, for years, much longer than the PC, much longer than, a data center for inference, right? Especially, o- outside of the mega data centers. But the, the odd thing is, is a lot of the technologies, and we mentioned this before on previous, episodes, mobile technology's actually making its way up, right? Jim, you mentioned, they're using mobile memory for the HBC, right? LPDDR instead of HBM. And, I think that's something to keep in mind. and so when I saw that, I went, wow, okay, the kernel level op stuff is important, and it's cool to see that, Qualcomm tuned into that. But also, we can't forget that these guys are masters at networking technologies, right? Um- Yeah for telco. And so when we start to look at, these next gen optimizations of networking and technologies, Qualcomm's not a, a slouch here at all, right?

Jim McGregor 35:09

And let's face it, with data center workloads, we always get to a point where there's an abstraction layer. Yeah. Whether it's a, a VPN or, we're using containers, whatever. We get to an abstraction layer to where it's easier to run stuff across all platforms.

Leonard Lee 35:23

Yeah.

Jim McGregor 35:23

So we're not there with AI today. But Modular is probably the closest company to creating that abstraction layer at this point in time.

Leonard Lee 35:31

Yeah. And so that, that's what I thought was really interesting about that announcement. It's not just a generic software play. Mm-hmm. it, it's deeper, and it's closer to the hardware,

Karl Freund 35:42

Let me, let me ask you guys a quest- Uh-huh let me ask you guys a question about Modular. Does it give Qualcomm more flexibility in the choice of the processor technology they choose to adopt?

Jim McGregor 35:54

I believe it does

Leonard Lee 35:55

think

Jim McGregor 35:55

This is an interesting thing, nobody talked about it. Obviously there was a rumor about Tenstorrent, and obviously they already acquired, it starts with a V, they already acquired an, a RISC-V company, late last year- to focus on RISC-V. I would venture to say if we're gonna see RISC-V solutions from Qualcomm, we're gonna see them in the data center first- Mm-hmm not in the handset or the PC, because I think- That's ah, five that's gonna have the biggest impact. So to Carl's point, I think that it does give them that easier transition to RISC-V.

Leonard Lee 36:25

You did that on purpose, right?

Jim McGregor 36:28

He did. he cued it up perfectly.

Leonard Lee 36:31

That was beautiful. we're synchronized, Not all the time, thank God. I

Karl Freund 36:34

would, I would, I would, I would agree with you, Jim. Let's not forget that the Nuvia was RISC-V based, right? it was not Arm-based. It was RISC-V. No, I'm sorry. It was Arm-based, but they had roadmaps potentially-

Leonard Lee 36:49

Yeah, yeah

Karl Freund 36:50

to switch them to RISC-V if the Qualcomm, Arm litigation didn't turn out the way it did. They had that option to move. And so I think they know how to shift the business, but the business they wanna shift is the data centers. It's not the handset. There's no win in it for them on a handset o- other than just lower royalty expenses.

Jim McGregor 37:12

And they did buy the Ventana Microsystem.

Karl Freund 37:14

yeah. Ventana

Jim McGregor 37:14

Microsystem. So that in itself was actually a huge acquisition. And to your point, I think that really changes the game.

Leonard Lee 37:22

Hey, you know what? I wanted to put this out there. one of the things I'm noticing about their, Qualcomm's, quote unquote, "Ricky Bobby," not first, but, you know, play here is, they're kind of taking a, a, a dual, like a, two-prong approach here, with entering the data center market. It's, merchant and custom, right? Mm-hmm. That I think is really something to, consider. what do you guys think?

Jim McGregor 37:52

I think it's becoming a very crowded segment. Yeah. And, and I see the interest in it and the need for it. Obviously the hyperscalers are trying to optimize around specific workloads. I don't think that ever goes away. But, now you've got Broadcom, Marvell, MediaTek- Mm-hmm Qualcomm, a host of design houses, and Intel and AMD actually, although, we usually don't put them in that camp. I think that's g- that's becoming a very, very crowded segment. I think the interesting thing is what comes next, and- Yeah I'm gonna pose th- I'm gonna pose this to you. So here's my tip

Leonard Lee 38:25

Wow

Jim McGregor 38:25

for Carl. think about when we actually get to a point where AI can actually do full chip development. Then are you going to a company like, one of these companies, or are you just going to a Synopsys or Cadence or Siemens to have them do you a custom chip?

Karl Freund 38:42

And I think the answer is yes. the AI, I think, is a potential risk to Broadcom and Marvell, the custom design. So if I can do the custom design using AI, instead of spending a lot of time and money getting Broadcom or Marvell to do my custom chip for me, that's gonna be a better solution. I think we'll get there. I don't think it's tomorrow. I don't think it's, maybe not even within five years. But if you have a long enough time horizon, I think the answer to your question is yes.

Leonard Lee 39:14

Hmm.

Jim McGregor 39:15

I agree.

Leonard Lee 39:16

Interesting. Yeah. Just as long as it doesn't turn out like vibe coding, then we're okay. You know? No vibe coding.

Jim McGregor 39:25

Obviously Qualcomm already has some customers for their custom silicon. Right. And, they're working with Microsoft for their C1000, merchant market silicon solution. So they have customers, and I think that's the key thing, is they're increasing their projection for their 2029 revenue, non-handset revenue, from 22 billion to 40 billion. That's a huge, huge increase, and it takes their TAM, especially with their continued growth in automotive, It takes their TAM into the trillions. So I mean, this does definitely alter the dynamics of, Qualcomm, and it's so rapid with data center that I think it's all in at this point.

Leonard Lee 40:05

Yeah. Well, I think that their opportunities is, their ability to diversify quickly i- you know, if and when-

Karl Freund 40:14

Right the

Leonard Lee 40:14

market turns. when we get to a point where a data center all of a sudden becomes a non-thing- they have the c- chassis, if you will, if we're to borrow from Nicole DeGault, who leads the, the automotive and industrial IoT stuff. They're better positioned for the AI diffusion, motion that happens outside of these large scale-

Karl Freund 40:38

I agree. I think Josh Young has done a magnificent job of opening the company up to lots of opportunity-

Leonard Lee 40:45

Yeah

Karl Freund 40:45

a- as he saw the end of the handset wave coming, he didn't wanna be crushed by it.

Leonard Lee 40:50

I mean, look at what they're doing with- So I

Karl Freund 40:51

think he's done a

Leonard Lee 40:52

Yeah, look at what they're doing at, at industrial IoT. It's a little bit, weird because we're so used to platforms not working in IoT, and they're actually going deep and broad, and that's hard to do. And looks like they're seeing some success. They're getting mind share. but obviously, making some traction. And, you know, these platform plays are really frigging tough, right? It's almost like you don't do it, because it's so hard It

Jim McGregor 41:23

You know, most people don't realize when you're talking about industrial, just talking about, like, medical Yeah That's like 1,000 different markets. It's from- Yeah all the different scan modalities to all the wearable tech to all the- Yeah in-home, in-hospital, all the different solutions. It is insane-

Leonard Lee 41:40

Insane

Jim McGregor 41:40

the number of markets you're actually targeting. It's not one, it's thousands

Leonard Lee 41:43

Thousands, Exactly. And so, any other final points we wanna make on a- jam-packed June?

Jim McGregor 41:50

Uh, I definitely... And this was huge, you know, uh, Apple conceding that they had to increase prices because of memory. that was huge. That was a shock to the market because obviously They're one of the 800-pound gorillas that has access to capacity and to memory Yeah. So to see them do such a drastic change in terms of pricing, and it was very drastic on their higher-end systems.

Leonard Lee 42:12

30%. Yeah.

Jim McGregor 42:13

That, that is huge. I mean, that's a signal. And obviously the big focus right now is on data center capacity, and that's where most of the memory's going. But everything else, and we saw this earlier in the year with, at CES where i- or since CES, I should say, we've seen handset companies, pare down the number of SKUs they're gonna offer- and even delay- when they're going to introduce it. Same thing with the PC segment. And I feel sorry for a lot of these embedded applications and startups, especially at CES. I'm like, "These guys are not gonna have any memory allocation." it put an exclamation point on the situation. Even though we've seen announcements over the past couple weeks of, a new, mega, development, between SK Hynix and Samsung in South Korea. We've seen, Microns trying to expand, both in New York, Japan, and Idaho. Even though we're seeing these announcements, memory's gonna be continued constrained- Right. Yeah for a couple years, and it's gonna be a huge impact.

Karl Freund 43:08

It'll be very interesting to see if the US government allows, Apple to acquire memory from a Chinese provider.

Leonard Lee 43:16

Yeah.

Karl Freund 43:17

CMXT. Yeah they have asked the US for permission to buy from MXT. So we'll see where that lands, because- Maybe

Leonard Lee 43:23

it,

Karl Freund 43:24

maybe they can It won't change data center. It won't change data center at all. Well- If anything, it'll free up capacity for the data center, which would

Jim McGregor 43:31

And all the IoT and embedded guys have already done that. They've already gone off and started qualifying the- Mm-hmm the Chinese vendors and the tier two- Yeah vendors from Taiwan and stuff. They've had to just because- Yeah they can't get the best anymore.

Leonard Lee 43:42

these guys are gonna get crushed, the smaller guys. And this is- Mm-hmm this, hearkens back to, the, COVID. No, was it the COVID thing? No, it was before, where they were stockpiling. What was that? Yeah, it was COVID.

Karl Freund 43:55

COVID.

Leonard Lee 43:56

Yeah. the IoT guys got creamed. smaller guys got creamed, and this is I think one of the unintended consequences of just this, frenzy, the data center build out and infrastructure build out frenzy. It really puts a lot of, pressure on the smaller guys. But then, we're also seeing, some of the IoT-ish, chip, providers doing pretty well now, really crazy, month of June. I don't know, you guys have anything, for July other than AMD? AMD, AMD,

Karl Freund 44:27

AMD,

Jim McGregor 44:28

AMD, obviously the AMD, Advancing AI event. and then shortly after that we're gonna have, the Future of Memory and Storage, conference coming up the beginning of August.

Leonard Lee 44:38

yeah, yeah, you were at

Jim McGregor 44:39

that one. Um, and then, the middle of August, we have Hot Chips. So we have a number- Oh, yeah of events that I think are going to at least give us some very good detail on- Yeah future hardware technologies. Yeah.

Leonard Lee 44:50

I could do with a boring month. I really can.

Jim McGregor 44:54

There is no boring month anymore. There's no boring months. Everyone's staking something out. Yeah. So there's no-

Karl Freund 44:58

We thought December was gonna be boring, and Nvidia acquired- Yeah the assets of Groq, so that didn't happen.

Leonard Lee 45:02

Yeah. Well, that's my message to the AI and, data center and semiconductor industry. Chill out, guys. Slow things down just a little bit. but anyways, gentlemen, thank you so much. It's great catching up. you know what, so-

Jim McGregor 45:18

Hey, hey, hey, we, we have to finish this with another Ricky Bobby quote since you started this. can't have two number ones. That just makes 11."

Leonard Lee 45:30

And there you have it. Hey, everyone, make sure to reach out to Carl and, y- find out about Cambrian AI research. If you, don't already know about, Carl and his wonderful research, you can, contact Carl at, www.cambrian-ai.com. He's also on Substack and Forbes. And also, reach out to Jim McGregor and Teras Research at www.terasresearch.com. And of course, please subscribe to our podcast, which will be featured on the Next Curve YouTube channel. check out the audio version on Buzzsprout or find us on your favorite podcast platform. Also, subscribe to the Next Curve research portal at www.next-curve.com for the tech and industry insights that matter, and we'll see you next month. Thank you so much, gentlemen. Have a great- Thank

Jim McGregor 46:23

you

Leonard Lee 46:24

month. Hopefully boring.