Silicon Futures for February 2026 - NVIDIA earnings, Qualcomm & HUMAIN, AMD & Meta and more!

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

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

➡️ NVIDIA's earnings results for Q4 FY26 - What did it mean?
➡️ Cadence acquires ChipStack to advance AI and agentic verification
➡️ The accelerating agentic AI disruption of 2026
➡️ AMD cuts a 6GW AI infrastructure deal with Meta, warrants included
➡️ Is Taalas proving that ASICs are the future of AI inference?
➡️ Meta's AI infrastructure diversification strategy
➡️ The ramp of sovereign AI, the current driver of AI infrastructure growth
➡️ Qualcomm delivers A100 racks to HUMAIN
➡️ Is there an AI bubble or not?
➡️ What's the deal between Intel and SambaNova>

➡️ Is Ericsson flipping the script on AI-RAN?

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 0:10

Hey everyone. Welcome, to this episode of Next Curves Rethink Podcast, where we break down the latest tech and industry events and happenings in the world of semiconductors and AI into the insights that matter. And this is our Silicon Futures series, I'm Leonard Lee, executive Analyst at Next Curve, that I do with the wonderful illustrious Carl Fre of Cambrian AI Research and of course, the DT Interconnected and five D package, not. Four and a half or three and a half for all this other nonsense. Five D packaged Jim McGregor of the fame, cheerious research. And I just wanna know you to know, because Jim, you probably don't even know what five D packaging is, but this is beyond advanced.

Jim McGregor 1:08

Well,

Leonard Lee 1:08

I, this is like ridiculous packaging, right?

Jim McGregor 1:11

I just spent the week out at Intel's advanced packaging facility, so I can honestly say I have no idea what you're talking about.

Leonard Lee 1:19

Neither do I. Then both of you say that all the time about me anyway. So, welcome gentlemen. it's good to see your faces again. after a really insane January, we. We had to suffer another insane February, right? in the world of ai, Carl's favorite world. But in this episode, we will be talking about the hot headlines of the month that mattered. And as usual, they're too many. this episode will undoubtedly touch on Nvidia. 'cause I know that. Carl's just gotta be chomping at the bit to talk about the, about yesterday, Qualcomm, Intel Cadence, a MD, Ericsson and Meta and more so

Jim McGregor 2:05

I'll be honest with you, Leonard, I'm more worried about March. We have Mobile World Congress, embedded world, OSC, snug, C-T-C-B-T-C. Just, it's ridiculous. Yeah. I'm worried about March. I think I'm gonna go nuts.

Leonard Lee 2:19

Yeah, I think you will. I think you will. Thankfully you'll be able to divide and conquer. But, Hey, before we get started, please remember to like, share and comment. And on, on this episode, and also subscribe here on YouTube and Buzzsprout. Listen to us on your favorite podcast platform. Opinions and statements by my guests are their own and don't reflect, mine of, or those of next curve. And we're doing this for informational purposes only. Please no comments on stock prices and valuations, gentlemen, to provide, open forum for discussion and debate on all things consumer and our semiconductor and, Carl's favorite topic, ai. And we will, just to make Carl happy, we will kick things off with Nvidia.

Jim McGregor 3:04

Now wait a minute. Did he just warn us, Carl, wasn't he the one we had to warn last time?

Karl Freund 3:09

Yeah, I think so.

Leonard Lee 3:10

Really? Okay. My bad. My bad. So. All gentlemen, Nvidia, what do you want to say? What do you want? Get off your chest. And that just happens to be the most recent working backwards.

Karl Freund 3:28

It's amazing. They had such a, such an incredible word. Stock market was notwithstanding Yeah. The, they actually had higher growth rate outside of hyperscale than hyperscale. Mm-hmm. Yeah. That kind, it's really a surprise. That kind of blows me away. That means the enterprises, but also it means the. Sovereign wealth. So sovereign, not wealth. Wealth, sovereign AI has been a big boost to their revenue. It was like three times their norm, three times growth over the last year. Not just 70%, 5%, three x growth in sovereign ai. And we've just really kicked that off

Leonard Lee 4:08

mm-hmm.

Karl Freund 4:09

So I think that portends, well, both for Nvidia. And other players in AI that are providing rack scale solutions. Yeah. They also mentioned in a call I just got off with Nvidia on it, mentioned that the sovereign AI is not just rack scale, they're actually deploying both rack scale and HT X based systems. So all around great quarter, They blew it away, not only in data center. They blew it away in professional visits.

Leonard Lee 4:39

Yeah,

Karl Freund 4:39

Workstations. Whose growth was dramatically higher than data center growth for the first time I can recall.

Leonard Lee 4:46

Mm-hmm.

Karl Freund 4:47

so they, they, the one area where they didn't hit on all cylinders, one cylinder that wasn't plugged in, I guess was, automotive, ironically enough. Yeah. automotive was a bit disappointing.

Leonard Lee 5:00

Disappointing. Oh, it's flat, right? And, but it's,

Karl Freund 5:02

that's not good.

Leonard Lee 5:04

Yeah, it is flat, but it's up from, about a year ago.

Karl Freund 5:08

Year ago. But flat, quarter to quarter.

Leonard Lee 5:10

I think their at about a $2 billion rate, annually, which is, better.

Karl Freund 5:18

They had

Leonard Lee 5:19

had 6

Karl Freund 5:19

They had 6 billion in physical ai, which is good to see some actual results in physical ai, especially outside of automotive. So, gosh, I think the, and the green team nailed it this quarter.

Jim McGregor 5:32

They did and they even gave a ROY forecast for next quarter, and they're one of the few companies that never mentioned memory. So they kind of tells you they're on the top of the priority list for memory vendors.

Karl Freund 5:44

And Stewart. Exactly. Stewart mentioned they set their prices when they launch the product, and so as memory prices increase, they don't increase their product. But the most important

Jim McGregor 5:55

part is they have the capacity, they've got the allocation.

Karl Freund 5:58

Yeah.

Leonard Lee 5:59

I think the whole supply. Equation is a little bit more complicated than, folks might assume, but yeah, that was a surprise that there was actually no, no margin impact or at least a discernible one associated with, memory prices.

Karl Freund 6:15

Yeah.

Leonard Lee 6:15

In fact, our margins went from 73% to 75. 75, which is probably not gonna make a lot of their suppliers happy. But

Karl Freund 6:27

It was the quarter of 75 billion in data center revenue. 75% gross margin.

Leonard Lee 6:32

Yeah,

Karl Freund 6:33

everybody's gotta be. Chops at that. That was amazing.

Leonard Lee 6:37

But one of the things that I thought was really curious, was, how networking, so they're reporting, they split the reporting on, data center networking is growing like crazy. And obviously, with the introduction of Spectrum X, XGS, what is it? And then. what is it? Envy, link Fusion. All this non Nvidia, walled garden business that they're tapping into really make reaching into sort of that hyperscale, scaler business. it looks like it's driving some serious growth. They went from 3 billion, 3.2 or three or billion last year in the same quarter. To 8 billion. So they're, I outstripping the growth in compute. how all that is really segregated in terms of their very now increasingly complicated business of data center. Hard to tell. And I don't think we get tons of clarity on that, which, it's understandable. but yeah, doing the comparisons is gonna be more and more difficult going forward. I think simply because their business, their data center business is shifting so much.

Jim McGregor 7:49

Well, it is, and you also have to consider the fact that, and this came out obviously with the meta announcement and the fact that Meta's basically using their entire Nvidia, entire tech te, technology stack now they're still using, they're still building their own chassis now. Their chassis are a version of the NV L 72, but Customiz, yeah. But they're using the entire stack, especially spectrum X for the networking. So the fact that, you know it and then, and that really does make it hard. whether it's a hyperscale or building their own rack or whether it's an enterprise taking an NVL 72 rack. They are selling that. It's not just one piece or the other. They're selling the entire technology stack along with it. And that doesn't even hit on all the software components that go, that are part of that.

Leonard Lee 8:33

Yeah. So, I have a question for both of you. How do you think that, because we don't really actually talk about Broadcom all that much. Uh, maybe, they should talk to us more, maybe, I don't know. But they don't, if they don't, shame on them. Uh, but it, I mean, how do you guys think this is gonna impact Broadcom? Because this has been their wheelhouse. obviously Marvell is a big player here, whether it's scale up. Scale out and now scale across, but I think in the networking front. I think you gentlemen agree. Nvidia has their game on have, have Oh

Karl Freund 9:11

yeah.

Leonard Lee 9:11

Time and where everyone's been focusing on the accelerator. When you have the system level and, rack level, data center level conversations is not just about the accelerator. In fact, the accelerators, the thing that's improving the least or scaling the least.

Jim McGregor 9:28

And that's a key point in the fact that, when they bought Mellanox it was, okay, well now they're the only InfiniBand provider, but it's much beyond that. Spectrum X has become an industry wide solution that a lot of people are adopting. Yeah. A lot of people are using, and let's face it, if you go with some of solutions from, Broadcom, you're basically going with a proprietary solution with Tomahawk. So, You don't gain any significant advantage by necessarily going with Broadcom 'cause you're locked into their solution.

Leonard Lee 9:57

Yeah.

Jim McGregor 9:57

So it, it, I definitely think changes the dynamics as we look at networking. and to your point, Leonard, Nvidia is a major networking vendor. Yeah. And it is going to impact Marvell and Broadcom.

Leonard Lee 10:09

It's showing, I mean, that's the thing that really stood out for me.

Karl Freund 10:14

Yeah, I can't wait to see the market share when it comes out because it's hard to imagine Nvidia is gaining significant share in the networking market.

Leonard Lee 10:21

But you know, it also puts a little bit of a. Or at least prompts me to put a little bit of a microscope on the compute side of things. So, again, you know, it's weird, like whatever we assumed last year, we come into this year and we have to think differently about that company. So, But, okay, cool. Let's move on to,

Jim McGregor 10:44

uh, well, and, and it's, it is not even just networking. Look at what they've done with, I mean, they continue to change the server architecture. They introduced Soca with Micron last year, GTC, and it was a proprietary solution when it was introduced. Now it's a genetic standard. All three of the major, memory providers are now providing it. And we're going to, so CAM two with that JET standard, which obviously we'll probably see announcements at GTC around that. So, it is, they are changing the entire, they're changing the server architecture, they're changing the data center architecture. All of it's changing. And NVIDIA's kind of helped push all of that through, through, through the industry.

Leonard Lee 11:23

Yeah. Any final. Word there, Carl, before we move on to the next thing. 'cause I think we're Nvidia to No,

Karl Freund 11:29

I think let's move on.

Leonard Lee 11:31

Carl, you wrote something about Cadence.

Karl Freund 11:33

Yeah.

Leonard Lee 11:34

Yeah. And then their acquisition to Hexagon. I think it's the design side of there that how

Karl Freund 11:40

Yeah, that's the design side. That's the design side. Design side, yeah. Yeah. My article was about the ac their acquisition of, chip Stack. Which was a company had developed Gentech AI specifically, initially, at least for design verification. an area AI hasn't really touched much, and now they've got an agentic solution. And imagine this, you've got an agent on top. It's orchestrating everything. Yeah. Okay. It's then calling agents over time for every aspect of design.

Jim McGregor 12:11

Hmm.

Karl Freund 12:11

Even the initial design in RTL eventually, the way they do this, and so agents on top of agents, they call it super agents, kind of a funny term, but it, it, it aply applies and what they're doing is creating a mental model. Mm. I think a term is, is great. It's a mental model of what the designers intend the chip to do and how it intends the chip to do it.

Jim McGregor 12:40

And

Karl Freund 12:41

so it really embodies, it creates a ground truth for all thery to check their work against this mental model and then iterate until they can get closer to achieving the design specifications and goals of, yeah, their functionality and the relationship of various functions. Of course, PPA, is pretty dramatic. They, they've gone from like 30% to. I think it was 84%.

Leonard Lee 13:08

Yeah. I have

Karl Freund 13:08

to check my math, of design coverage. Uh, wow. All through the use of this mental model. It's brilliant. 'cause you wonder, well, okay, if you tell AI very specifically want you to do, it's gonna do it. But for agents. Agents have to have some sort of mo of model, this mental model that helps them guide them to produce better results. Yeah. And so they've just begun this journey to full agent supportive, super agent supported and led. The humans are still in loop. The design engineers must stay in the loop to close the gap. Yeah. And to check the work. and so if you're a design engineer, don't worry. This is not gonna. Take your job. It's, I think it's going to make you a better engineer.

Leonard Lee 13:54

It's gonna make you be able to,, keep the pace up. Like, I, I think we this before when we were, we, we covered, or we, yeah, we covered, cadence Live last year. Right. And Honor Ruth team, this is like a big area that they we're looking at. validation or verification. Yeah.

Karl Freund 14:12

I honestly think this, I've always been a, a big fan of what Synopsis been doing, and by the way, synopsis is still doing great, but I think in the application of AI for the chip, chip, workflow, chip design workflow, I think Cadence has taken the lead.

Leonard Lee 14:26

Hmm. Interesting. Jim, any reaction?

Jim McGregor 14:29

well first off, I think this is all good. I mean, and we've seen this, this was a debate a couple years ago of, okay, do you use AI within the tools or you do you use an AI agent to use the tools? And that's basically where they're going is having AI in multiple levels. So you have it in the tools and you have the agent using the tools. the real challenge still relies on the trust around it. The fact that, especially detailed stuff like place in route when you're actually doing a transit or doing a semiconductor design, not even Nvidia really trusts it yet. So yeah, it's, it's going, it's gonna take time, but it's good to see it moving that direction. and, and you mentioned synopsis. You know, both of those companies have invested heavily in ai. It's gonna be interesting to see what happens now that. Nvidia also has an investment in synopsis. Mm-hmm. So, it's, yeah.

Leonard Lee 15:18

Yeah.

Jim McGregor 15:18

it's gonna be an interesting dynamic.

Leonard Lee 15:21

Yeah, eventually NVIDIA's gonna own everybody. yeah. So if Nvidia doesn't own everybody, it looks like Meite. What is up with this? A MD? The, what is it? Six. Six gig. That gigawatt, gigawatt. Deals. I mean,

Karl Freund 15:41

to me it shows that customization's important. Yes. Because, and, and, and that's in, that's been NVIDIA's story is that we need, we chose a MD because the chip implementation allows a MD to deliver at reasonable cost. A chip that is custom for. Metas workloads.

Jim McGregor 16:00

Yeah.

Karl Freund 16:00

and by the way, NVIDIA's not sleeping through this, Ruben, CPX is also a customization, although more in the software and memory architecture layer than in the semiconductor itself. and so NVIDIA's point is, look, this is not good for margins. This is gonna hurt a MD, or this is my interpretation of their silence that, it costs a lot more to do custom silicon. You could do another tape out as well as the design work. Chips make that easier to do than trying to change a monolithic tie like Nvidia a, Blackwell. But it is gonna impact margins. Well, one of the highlights of NVIDIA's announcements last night, 'cause their margins over 75%. Mm-hmm. So I do think, however, that, that margins aside, I do think that the specialization. Of semiconductors for specific workloads and even for specific models. I wrote an article on Forbes, which I think is now available on my substack. About a company called Tali. TAA. Oh, yeah. Yeah. A s fascinated. They've actually gone to the step of, we don't see a specific silicon for a specific model. A problem is models change every 12 to 18 months. You're gonna have tape out again, another version. But what they've done is they've got the base layers all solid. And so think of that as what the software provides today. In A GPU. So above that they just have two metal layers connecting all these features functions. and so they can actually produce a new model in just two months, not two years. That's very smart. And whether that will take off or not, I don't know. Hyperscalers not gonna be happy about having to manage that many skews. But the performance is dramatic now. Right now it's only available for alma b, but they plan to have a trillion parameter model by the end of the year with a reasoning model along the way. Probably deep seek, I would imagine.

Leonard Lee 18:03

But that's interesting because that's a statement. that's a pretty profound statement about asics and what kind of headroom they actually have for that particular. path to,

Karl Freund 18:16

absolutely

Leonard Lee 18:16

AI silicon mm-hmm. Versus GPU, right? So you have general purpose, but then, when you do have these application specific, but you know, that our scale, large scale, the economics are undeniable. if Asics can deliver. At that level, right. With that.

Karl Freund 18:39

the question is, what's the tam of a specific model? Right?

Leonard Lee 18:43

Yeah. Because I hear from a lot of developers, they don't like all this change. The models aren't getting dramatically better for whatever application they actually use a particular model for. They need better economics. Mm-hmm. And so if an asic like that. Can deliver those economics they don't care about anything else. Do you know what I'm saying? Especially if they deploy it. So they might do development in a GPU environment deployed model on asic, right?

Karl Freund 19:14

Yeah, yeah, exactly. I think we're seeing the beginning of a long term trend towards silicon specialization. Yeah. In, in ai, it's expensive, but if you have a big enough market.

Leonard Lee 19:27

Yeah, it,

Karl Freund 19:28

it works. as long as you can turn the ships quickly, which TAs claims they will be able to do. The ta, the Team Lead Towels, two, two and a half years ago, I think they left tens torn after Jim Keller came on board and became CEO of Tens Torrent, Ibiza and, and his team of execs and designers left and formed TAUs. They thought there's a better way to do ai. We'll see. We'll see if the market, how the market reacts. They're not setting big expectations. they're saying, Hey, this is our first chip. Think of it as prototype. Not a lot of commercial use is available. Sure, Tam really is available for eight B, but as they go to reasoning models and foundation models,

Leonard Lee 20:10

yeah,

Karl Freund 20:10

it can have an impact.

Leonard Lee 20:13

So

Jim McGregor 20:15

I, I, I don't think so. Um, you know, well first off, you have to remember that you mentioned meta and, and Nvidia. You have to remember that. or a MD you have to remember that meta just last week made an announcement with Nvidia. Yeah. as well. So, I mean, they're not banking on one and they have, their own AI accelerator and, MTIA. So it's, they are, they are spreading it across different types of applications. I do agree with, Carl that I think we are gonna see more customization and I think chips and, some packaging technology like, Intel's mibt, definitely lend itself to doing that a lot easier than even what we have today with like COOs. Because of having the embedded bridges rather than having to use Interposers, I think that's gonna make it easier and easier and I don't think they're gonna have to do. A lot of D spins. I think in a lot of cases, it's a, either they want to put another D on there, or b, they want to change the memory configuration. Yeah. They wanna change the IO configuration. They want to change maybe even the compute note configuration, or tiles. So I think there's gonna be a lot of customization.

Karl Freund 21:19

So,

Jim McGregor 21:19

but

Karl Freund 21:20

Jim, do you think Nvidia will eventually adopt, templates?

Jim McGregor 21:24

Uh, well, I think that, well, technically they've already kind of gone there with the dual die solution. Mm-hmm. Although they don't like to go out a chip, so technically they're already going there and Yes. I think with the addition of engineering resources from Grok. and other talent. I think especially as we, they move into more customized solutions for inference, we're much more likely to see, chip based solutions from Nvidia a MD and in Intel, to be able to do that. I don't think doing a. Custom for these models is going to be the way to go. I think that's, I think that's a road to nowhere

Karl Freund 21:57

Maybe a bridge too far, but it'll be interesting to watch, play out

Jim McGregor 22:00

the cost and the time and everything else. I just don't believe is going to pay

Karl Freund 22:04

There's another, there's another company called et, which is doing, its cus. Silicon and asic not for a specific model, but for a family of models like transformers. So that will support all, virtually all transformers on a specific piece of silicon. Now, you're not gonna run other kinds of things on there, like, like recurrent ai.

Jim McGregor 22:26

But you also have to remember that, you know, whether it's the hyperscale that's developing their own silicon, or it's Nvidia or a MD, these guys are providing a full technology stack, just silicon. And I think that's the part that's the moat. it's really that entire stack that really becomes hard to overcome of just doing a custom chip.

Karl Freund 22:46

Well, and that's right in NVIDIA's wheelhouse, right? Yes. They've been saying that for years and years and years, and they delivered for years and years and years.

Leonard Lee 22:55

I think this will, we'll see how things pan out. I mean, obviously, there's innovation outside of the NVIDIA camp and they're also ingesting a lot of things that are not. Intrinsically, Nvidia, right? Whether it's grok, I mean, it's clear. They're looking at inference as a different, game that they have to play. And, I don't know. from my perspective, I think the jury is out whether or not they actually have the kind of moat that people think that they have. And I think when we see, the continued, growth and focus. On their portfolio. It's easy for us to settle into that trope. So, but, I think last year showed us that, there are some curve walls being thrown, in this game. But, Jim, is there anything else that, anything that popped for you this, this month that you really wanna talk about? Share with the audience?

Jim McGregor 23:56

Well, I think that, one of the key things that obviously, Carl hit on with the Nvidia announcement is that the build out of sovereign ai, matter of fact, we saw the announcement also with a MD Yeah. With, TCS, with the, Tata Consultancy Services and, They're, they're subsidiary to build, sovereign AI solutions in India. And they, you know, and I think that's gonna be a key differentiator, talking about the customization. I think that's gonna be a key differentiator. And we've seen that with the US National Labs as well, to be able to provide that level of customization. So it's gonna be interesting to see. Other than that, I'll be honest with you guys, I am heads down because we have. MWC, embedded world, OFC, all this stuff coming and it is just minding least scary. I will say this though, there, there is definitely, we still see this, we've seen the shock this month of, I think people waking up in the fact that, AI's gonna change the software ecosystem and it will, matter of fact, I don't even know what a PC's going to look like when you don't, when all you have to do is tell it what you want, you know? what does the hardware and what does the software look like when the agent becomes the, so the underlying software applications, it becomes the application or it becomes the interface. It becomes everything and all you're doing is really asking it to do something. that's gonna have tremendous impact on the entire industry, hardware and software. And I think we're still trying to understand what that,

Karl Freund 25:21

and users, I think the, the humane one. I haven't heard that much about it after I can summit, but that's exactly, I trying do

Jim McGregor 25:30

he wants one person

Karl Freund 25:31

Eric hear, hear more Tarek.

Leonard Lee 25:33

I'm gonna send them video.

Karl Freund 25:34

Hey dude, it's brilliant

Leonard Lee 25:37

I

Karl Freund 25:37

strategy to make it work.

Leonard Lee 25:39

I told him, Hey man, um, oh, you know, with the Qualcomm. a 100. I thought it was a a 200 that they showed. I was like, what the hell? But, I didn't read the fine print, but hey, they're shipping racks. That's kind of cool. Shipping racks, right?

Karl Freund 25:53

Mm-hmm.

Leonard Lee 25:53

For the A 100. I told him, dude, you know? Yeah. yeah, he gave, Jim, Carl and I a tour.

Jim McGregor 26:00

the other thing I would hit on. That, you know, obviously, still seems to be lingering. The industry is this concept of an AI bubble. We have so many AI adoption waves going on right now with, with, industrial applications, with enterprise, with embedded with, and just to transition from, token generation to multimedia generation. All those are driving. I mean, our, we just published our latest forecast. And there is, I mean, the demand is still growing exponentially, and we've, we've been kind of, aggressive and we've still missed the mark for the past two years on our AI forecast. I have to be honest with you, if this is a bubble, and, and you know, we're really seeing these as major transitions. Yeah. I don't see a bubble out there. I think that we are in this, this growth mode for at least the next two to three years at least.

Leonard Lee 26:49

Yeah,

Karl Freund 26:50

getting Radox is alive and well.

Leonard Lee 26:52

Yeah. I, I, I think we are in a massive bubble. because when you look at M markets, you don't see the monetization. You still see software companies really struggle, um, to qualify. Where is the AI driven growth, that is commensurate. With the investment that's happening on infrastructure, which should be orders of magnitude larger for the end markets. And that is absolutely not visible. And then also this idea that the AI model makers, the open ais and the philanthropics of the world are going to, usurp, the, the ISVs. Well, what is the killer app so far? If there's any killer app, it's Cody. What does code generate? Code generates code. It's traditional stuff to develop software that runs on g CPUs. More than GPUs. And so this is where it doesn't make sense to me.

Jim McGregor 27:55

Yeah, I will agree with you that I think the business models to get the value back from the code generation, token generation, multi-age generation Yeah. Is still, being evaluated. And we saw this with the internet. We've seen this with a lot of different market transitions. However, the demand is there, the number of users it increases, the amount of time they're using is increasing. By our forecast by 2027, we're gonna have more users than we have people on the planet because we have agents using ai. So, yeah. Right. But would say the demand is there,

Leonard Lee 28:25

monetization has to be there. It's monetizable. Demand is what's more important than demand, where, it's pretty much given out for free.

Jim McGregor 28:36

you also have to remember that some of that monetization you don't see in terms of additional revenue. You see in terms of savings, productivity. We do see enterprises, improving productivity and everything else. So we do see value creating, being created by AI

Leonard Lee 28:53

in certain functions. But as a whole, I think that's still, a thesis that hasn't been proven out consistently.

Jim McGregor 29:01

agree to disagree.

Leonard Lee 29:03

Yeah, no, no, that's fine. Agree, disagree. but see that, that this is why we're doing this is not because we agree. 'cause it would be freaking boring. Um, no, no. Hey, let's, let's, you know, we, we know where each other stands, but no, I can appreciate what you're saying in terms of demand. There's different qualities of demand is my point. And when we look at, we qualify if there's a, a bubble or not, it really has to do with, ultimately is there efficient allocation of capital happening? And there definitely has to be end market monetization. That's commiserate with investments that are being made, is just my point. And so that's how I'm shaping my view on, if there's an AI bubble or not. Definitely there's arguments on both sides, the aisle on this. But, hey, Jim. Samba, Nova, and Intel. Any comments there? I thought that was

Karl Freund 29:57

My opinion is back. I think Intel is taking out a very inexpensive, smart, option. Yeah. as they develop their own technology. So that's why they didn't buy San Bon Nova. Not right. Didn't buy their assets like in VI did with Rock. But they have a multi-year, agreement to work together so they can take advantage of and drive some additional revenue with Sam Nova's Tip, which looks pretty good by the way. Haven't done a full analysis yet, but, Because they don't talk to analysts either. but the, in, in video, I mean, Intel's not throwing in the towel. They plan to do their own inference specific semiconductor, which we'll hear about more later this year.

Leonard Lee 30:36

Hmm.

Jim McGregor 30:37

Completely agree. It'll be interesting if they see value there, then I would see there's an acquisition in terms of IP and or resources farther on down the road. But let's face it, they've already gone through multiple acquisitions in the form of AI that haven't necessarily paid off. So I don't, I, I think the board's gonna be very cautious about doing another one. Yep.

Leonard Lee 30:57

So, just really quickly circling back to the Qualcomm announcement with Humane and because I think we all want to get an invitation to go and take a tour, right? I wanna go with you guys. what did you guys think about that? A 100 and then s. Jim,

Jim McGregor 31:15

well, they're going in the right direction. I mean, they're targeting inference. They're leveraging their strength, which is, power efficiency and doing power efficient solutions. they realize that they have to provide a full stack, full platform solution to be able to compete in the market. So now I think they're going in the right direction. And let's face it, humane AI is, basically working with everyone. And trying to bring technology to the Middle East, through those investments. So everyone's going to benefit from it and I think this really gives, Qualcomm an opportunity to prove out its technology and its prowess and everything else. And let's face it, it has invested in AI for over a decade. Yeah. Although most of that's been towards the, edge nodes. But you know, they have that expertise in-house and they have that expertise to do, you know, very power, power efficient solutions in-house as well. So, yeah. there's no doubt that, you know, the AM market's gonna be so huge 'cause no two workloads are the same. Mm-hmm. that there's gonna be opportunity for a lot of players and, they're, they're going, they're one, let's face it, they're one of the few players that can provide a full stack solution. And that's what I think you really need to compete today.

Leonard Lee 32:20

And, Adobe is their first customer, Humane's first customer for that platform. Carl, any thoughts or reflections?

Karl Freund 32:29

No, I think I agree with Jim. I think Qualcomm's doing all the right things. I can't wait to see the AI 250 when they come out, but, meanwhile they're not sitting still. they're fighting, to gain share in a rapidly growing market. And they've got great technology, so I think they're executing well.

Leonard Lee 32:46

Yeah, hopefully we'll bump into one on one of those. A 200 a ai, two hundreds in Barcelona. Right, Jim?

Jim McGregor 32:53

I wouldn't count on that.

Leonard Lee 32:54

Oh. Bummer. but, yeah, I mean, I, I'm just surprised that they put a rec together. That was my biggest question is how quickly they would be able to do that and what would be under the hood. Okay. And then, one last thing I wanted to put out there, because we're going to MWC and there's this whole AI rant thing going on. the, I was at Erickson's, pre-event last week and, you know, Yeah, you know, they introduced a new, Ericsson silicon that has an embedded or integrated, neural, network accelerator. So they're taking a different angle on ai, ran and the silicon related to it. So it's gonna be interesting next week. To see how the whole telco and ran silicon, talk track shapes up. gentlemen, thank you and everyone, thanks for, you know, listening. make sure to reach out and follow. Carl at Cambrian AI research at www dot cambrian, hyphen ai.um.com. He's also on Substack and Forbes. And also reach out to Jim TEUs research@www.teusresearch.com. please subscribe to our podcast, it'll be featured on the next Curve YouTube channel. Check out our audio version on Buzz Brow and find us on your favorite. Podcast platform. Also subscribe to the next curve research portal@www.next curve.com for the tech and industry insights that matter. And we'll see you at the end of next month.

Karl Freund 34:24

Cheers up to GTC.

Leonard Lee 34:26

Take care. Bye. Take

Karl Freund 34:27

care. Bye-bye.