The Ultimate NVIDIA GTC 2026 Recap (with Karl Freund and Jim McGregor)

Leonard, Karl, and Jim attended the AI event of the year, NVIDIA’s GTC 2026. This year, the agentic AI theme went into overdrive with its DeepSeek moment, OpenClaw. 

Or was it a ChatGPT moment?

The trio unpacks a deluge of announcements from NVIDIA’s technology stack and across the domain-specific industry and computing solutions that are meeting the agentic AI moment.

In this episode, Leonard, Karl and Jim talk about the top headlines and share their analysis from NVIDIA GTC 2026. 

➡️ Staying ahead of the fast-rising tide of AI
➡️ Agentic AI's ChatGPT moment with OpenClaw and NVIDIA's NemoClaw
➡️ NVIDIA pivots its AI data center inference roadmap with Groq 3 LPX for agentic AI
➡️ NVIDIA is innovating at and competing with the speed of light 
➡️ NVIDIA is shifting away from GPU-centricity with Groq 3 LPX & Vera CPU systems 
➡️ NVIDIA is becoming more than a GPU company. It is heterogeneous! 
➡️ NVIDIA claims that it will become a $trillion company by end of 2027! Will it? 
➡️ Is NVIDIA GTC becoming more of a VC/Wall Street event than one for developers?
➡️ The continuous reinvention of NVIDIA and its new identity in the age of agentic AI 
➡️ NVIDIA's storage play with SDX and the AI infrastructure/factory memory tiers ➡️ The explosion of heterogeneity of computing driven by agentic AI
➡️ Does Jensen know everything, or is he learning as he goes like the rest of us?➡️ NVIDIA is becoming more application and domain specific because of agentic AI
➡️ Clarifying performance & scaling. "More for less" or "more is more"?
➡️ Jensen's Paradox: one-year cadence or extended useful life? Can you have both?
➡️ Agentic FinOps and the economics of agentic AI
➡️ Enterprises need Agentic AI strategy for ROI and for agentic risk/threats 
➡️ Did NVIDIA crack the enterprise-grade personal AI nut with NemoClaw?
➡️ Is AI Grid NVIDIA's pathway to AI-RAN or is it an edge fabric for "physical AI"? 
➡️ 2025 was a big year for AI infra networking but a light topic at GTC 2026. Why? 

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

Everybody. Welcome to, uh, this episode of Next Curve's Rethink podcast, where we break down the latest tech industry vets and happenings in the world. Semiconductors and Carl's favorite topic. Ai.

Jim McGregor 0:23

AI and gyms,

Leonard Lee 0:23

right? Yeah,

Jim McGregor 0:24

absolutely.

Leonard Lee 0:25

and we break all this stuff down into the insights that matter. I'm Leonard, the executive analyst at Next Curve, and I'm joined by. Paul Fre, who is the center of the Cambrian AI event that's happening right now. And he has a company called, or a firm called, Cambrian AI Research. And, we also have the dynamically ated, Jim McGregor. Of the fame curious research. How do you like that one, Jim? I learned that at GTC 2026.

Jim McGregor 1:01

Oh, geez.

Leonard Lee 1:02

I, I have no idea what it really means, but it is a thing. It's called dynamic ation.

Karl Freund 1:09

I'll have to look that up.

Leonard Lee 1:10

Yeah. You go to Jim should

Karl Freund 1:11

thank you, or Hitchy.

Leonard Lee 1:13

Yeah, it's really interesting com collaboration between Apple and Nvidia in the whole spatial computing space. So, really exciting stuff, but, I always associate Jim with really exciting stuff. So gentlemen, welcome.

Jim McGregor 1:28

Welcome.

Leonard Lee 1:29

And so you're probably wondering what we're gonna be talking about. Well, I already gave you a hint, GTC 2026. We're gonna do a recap and I'm gonna have the pleasure of listening to these two gentlemen who were on the analyst program I was not this year. and, I'm really curious, what your takes were since you were there, front and center with the program. And I was out in the exhibition hall running around. Trying to find a free lunch. But before we get started, remember, please, like, share, react, and comment on episode. Also subscribe here on YouTube and Buzz Sprout. Or listen to us on your favorite podcast platform. Opinions and statements by guests are their own and do not reflect those of next curve, or myself. And, we're doing this for informational purposes only to provide a open forum for discussion, debate on all things, AI and, Silicon Semiconductor and supercomputing. So, by the way, hey, look what happened. One of these hundred thousand subscribers and I wanted to, thank all of our subscribers, but also, Jim and Carl for their wonderful collaboration on the Silicon Features series. gentlemen, I attribute a good portion of this to some of the great. Insights and collaborations that we've done over the years on this program. So thank you so much and congratulations.'cause part of this. What are the bad chain?

Karl Freund 3:01

Congratulations. Congratulations to you guys as well.

Leonard Lee 3:04

That's

Karl Freund 3:04

pretty cool. That's pretty cool.

3:05

Well,

Jim McGregor 3:05

you notice he didn't send us one? No. Just kidding.

Karl Freund 3:07

No, I didn't. I'll go check my mailbox. Yeah, I'll

Leonard Lee 3:12

you guys one, but, no. What I want everyone to do though is, recognize, you and, also reach out to, Jim and Carl for their insights. It's ab, they're absolutely valuable voices and, eyes and ears on the industry and what's happening in AI and the semiconductor industry broadly or even beyond. So yeah, definitely reach out for to them and, stay ahead of this, this, you. Very quickly rising tide of ai, that

Jim McGregor 3:42

everything, it's not possible. Everything. It's not possible. You can't stay ahead of it. It's impossible.

Karl Freund 3:47

No, no

Jim McGregor 3:49

Open claw. Just prove that. Tell yourself short.

Karl Freund 3:51

I'm trying to make you look

Leonard Lee 3:52

good, Jim. Open Claw, what's your problem? I was telling Jim, I've been working here on the new house in Tucson and I totally miss the importance of open claw and, I got to GTC and I'm like, holy cow, the world has changed While I was. Changing light bulbs here.

Jim McGregor 4:08

And he's looking at me. He is like, Jim, why didn't you tell me about Apple?

Karl Freund 4:10

Why did you call me and tell me?

Jim McGregor 4:12

I thought you knew.

Leonard Lee 4:16

So,

Karl Freund 4:17

Jim, takeoff for a few weeks in the world changes. Yeah,

Leonard Lee 4:20

yeah, yeah. Let's get, I mean, you brought up, let's go fucking Claude. So, I mean, what stood out this year?

Jim McGregor 4:25

That that was, that was probably one

Leonard Lee 4:26

big event. The

Jim McGregor 4:27

biggest things that stood out was open claw or Nemo claw, I should say, which is, NVIDIA's wrapper, kind of security wrapper along with their, open shell, the entire. Nemo Claw solution is designed to put a wrapper around open claw. And Open Claw is, very innovative in the fact that it's basically an agent to call other, LLMs or to call other tools and everything else. It's the easiest way we've had so far to build custom agents personally and actually on device. So it's a major innovation. Unfortunately, and we've played with it, it has its limitations. it can go rogue, because you don't really out there, the code when you give it rules. And those rules can be forgotten over time, especially if you're running into memory limitations. So it has some serious security limitations. Well, yeah. Uh, Nvidia recognized that it came out with Nemo Claw to put the security wrapper around it. Very innovative, along with tools like Lang Chain, the fact that you can build these agents so quickly, so effectively, in a very short period of time to do pretty much anything you want. And they can operate on device or in the cloud.

Karl Freund 5:33

Yeah. And maybe a hyperbole of it. It may be an overstatement, probably is, but to me, this is the killer app.

Jim McGregor 5:39

Yeah.

Karl Freund 5:40

We've been waiting for a killer app for AI for five, six years, and we thought it was chat GPT, but it's hard for a lot of people to make money off chat. GPT, open Claw, I believe is the kill the killer app of Agen ai and it's interesting, I can't imagine what it must have been like to work at NVIDIA for the last three months because three months ago, Nvidia did the unthinkable. They decide that GPUs weren't the right solution for something, and that something is called Gentech ai.

Leonard Lee 6:12

They

Karl Freund 6:12

wanted to get on top of that. They previously were working on the front end of inference processing the large context, prefill problem. and they came out with a product, announced a product for that called, CPX Ruben.

Leonard Lee 6:26

Yeah.

Karl Freund 6:27

And then as Ian Buck said, along the way, we found something better. Something different. It's not solving the large context prefill problem it's solving at the other end of ai, which is token generation at massive incredible speed and very low latency, and importantly, deterministic latency offered by GR language processing units, LPs in the last three months, not only have they integrated these new engineers and leadership into their team. They've come, they've taken their next generation chip called Rock three

Leonard Lee 7:01

three

Karl Freund 7:02

and Rock three LPU three, and so they built an entire. NVL 72 kind of rack for them and them as they now. So, they have now shifted strategy to focus on agen AI as opposed to what they were focused on and still are, which is the large context problem that with, let's say, using movies video as a input token. Okay. Or body of code, a million lines of code, using that as a near input, input, sequence. They're now focused on genetic ai and they're, they're gonna have probably anywhere from one to four racks of Brock Lp Xs. alongside of Ruben, right? So you're gonna have a Vera Ruben Rack, and you're gonna have like anywhere from zero to four LP X racks for each Ruben Rack. And it's like, wow, this is a major shift. This company has been so focused on GPUs. You got a problem, we'll solve it with GPUs. Right? Yeah. no, this company was amazingly, Agile to be able to say and fully believe, no, this is not the right problem. GPU is not the right solution for this problem. Rock has the right solution for this problem. It would take too long to acquire the company, so we'll just acquire its assets and leadership team and engineer. Yeah. so

Jim McGregor 8:29

well,

Karl Freund 8:30

it is pretty damn amazing.

Jim McGregor 8:32

On top of that, they were coming out with a storage rack solution. The SDX? Yeah. Yeah. And even re-architecting future generations of racks with Chiver to where they're actually gonna have blades so they can double the density of GPUs from 72 to 144 and have, and get away from using cables and actually using back plane solutions. So, I mean, the level of innovation is just phenomenal that we see going on at Nvidia. This, Jensen likes to say that Nvidia. Innovates at the speed of light, and it definitely looks that way

Karl Freund 9:03

and that speed of light. I haven't heard them talk about speed of light much lately. When it was first discussed, it was more than just speed. You're absolutely right, Jim. They're moving an amazing speed. But it is the idea that your competition is not another company. Your competition is fundamental laws of physics, like speed of light.

Jim McGregor 9:24

Yes.

Karl Freund 9:24

So speed of light is perfection. How close can you get to perfection given the technology you have today, the resources you have today? how good can you get, forget the competition. And video guys don't think of competition. They think of the speed of light.

Jim McGregor 9:40

Yeah.

Leonard Lee 9:41

Well that's, yeah, that's, a huge shift. And I know that during the keynote. he was mentioning Cuda quite a bit. In fact, he,

Jim McGregor 9:51

well, it's, it's

Leonard Lee 9:52

20 years, kind of preface the whole thing. 20 years, right?

Jim McGregor 9:54

20 years of Cuda.

Leonard Lee 9:56

Right. But like, what's really interesting about the comment that you just made, or the assessment that you've made, Carl is this huge shift away from, the GPU.

Karl Freund 10:08

yeah,

Leonard Lee 10:08

I think that was really noticeable. But also, it wasn't just the LPU, they introduced the Vera Rack, so now,

Karl Freund 10:17

yeah,

Leonard Lee 10:18

They're making a big play into CPU, and so all of a sudden we're seeing this big change in the quote unquote accelerated computing story where I think two, three years ago, the impression was everyone's gonna move everything toward these quote unquote GPU infrastructures right now, it looks like Cuda might not have that kind of scope.

Karl Freund 10:44

I, I don't think they're moving away from GPUs. I think they're adding things beyond the GPU to the GPU based solution. So if you look at infra from Nvidia a year from now mm-hmm. You'll have racks of GPUs, all with all their fancy NV link and everything. You're gonna have racks of CPUs, which is Vera, and you're gonna have racks of LP to solve The ve, the agent AI problem of lots and lots and lots of small models.

Leonard Lee 11:16

Mm-hmm.

Karl Freund 11:17

Uh, all calling each other. And so it's a much more heterogeneous environment. I don't think it's more away from GPU, I think it's move beyond GPU.

Jim McGregor 11:25

And I think if you go a step

Leonard Lee 11:26

That even that's a huge statement right there.'cause in a lot of people's minds, Nvidia, a GP company.

Karl Freund 11:34

Yeah.

Leonard Lee 11:34

And I put it quotes because not anymore. We're not just talking about the.

Jim McGregor 11:38

And I think if you take it even a step further, they continue to work, look at the workloads, which are changing rapidly and very, very quickly, and they're trying to identify the bottle X. So, I mean, last year they, along with Micron, introduced silk cam memory just to address, the KV cash issue and everything else. Yeah, they introduced a storage solution this year to also address some of those, the storage specific solution or needs that you need for ai, not long term storage, but AI type of storage that, short term or midterm type of solution. And they keep looking at that and it goes beyond just even the hard. I mean, the fact that, uh, they continue to look at, enhancing foundational or frontier models for specific types of applications, especially for the physical world and the fact that they now have over a thousand Cuda X libraries for modeling pretty much everything from, DNA genomics all the way to the planet Earth. I mean, basically really looking at is it hardware? Is it software? Is it interconnect? What is it? What are the bottlenecks and how do we get over that? And we're even seeing that with next generation platforms, with Kyber going to those blades and everything else, it's, it is amazing that at one company has this breath to look at what has to be done to address these needs.

Karl Freund 12:58

For me, the sparkling moment of GTC when Jensen went home to, to, to, and and told his wife, Hey, today I announced we're gonna do a trillion dollars in revenue and. No. So what? No big deal. Take out the

Jim McGregor 13:13

trash.

Karl Freund 13:14

Now, Jensen, I've been telling you for days, you gotta take out the trash. That was his wife's response supposedly. And if you look at what the stock market's done, it's done the same thing. Trillion dollar, eh? So what?

Leonard Lee 13:26

Yeah. Well, you know, there's a lot of big numbers being put out there, but, you know, I, I know that you guys. You guys are very bullish about all this stuff.

Jim McGregor 13:38

There's no butt.

Leonard Lee 13:38

Yeah. No, there is a big butt. There's there's a big butt. there's a big, it's money with two's either, actually. There might be, anyway, it might

Karl Freund 13:46

be one of those too.

Leonard Lee 13:47

I, I'm definitely one of the overhanging, clouds at this conference is, circular investment. I, yeah, I've. Have a lot of discussions about that, so we can't discount that. Even though from a technology front, there's a lot of excitement. But from an ecosystem standpoint, one of the things that I don't think really, was well received as a comment that Wall Street is their, there's more Wall Street people or investment or finance sector people there than developers. the conference itself has changed a lot. We saw that with, the event last year. Uh, I don't know if they had a Q day this year. No, they

Jim McGregor 14:28

did not have a

Leonard Lee 14:28

quantum day this year. Quantum. Okay. Thank God. I thought that was terrible. And then CES also, less focused on technology, more on the markets, but, I, I think the. Important thing here though, is when we look at the transformation of that, the company as we're alluding to through the comments that have been made so far, Nvidia is dramatically changing. There's obviously a huge reaction to the diversity in the com, the complexity of the, inference opportunity, and this is stuff that we've actually been talking about. Quite extensively in prior episodes, whether it's related to the memory architecture and or, the role of application specific, silicon and, presenting a different envelope of optimization for some of these in, these models in runtime. You guys mentioned adaptable. The identity of a company is evolving really rapidly. Last year it was all about, hey, we're a AI infrastructure company, right? This year, AI Factory seemed to be that identity that, Nvidia wanted to portray or at least push out there? I don't know. What do you guys think?

Jim McGregor 15:49

I don't think it really changed. I think they still mention, they may not have mentioned AI infrastructure as many times as they did last year, but I still think that's the focus, they want, and more than anything else, Carl can speak to this, but we're seeing more of a message around not just hardware, but obviously software and tools and everything else. Yeah. And support for the entire ecosystem. Like the S stx. Yeah. They're not coming out with the STX, that's going to be up to partners to actually develop that rack. and the fact that all these models and all these libraries are now open source.

Leonard Lee 16:24

Yeah. Yeah. And speaking of the SDX, they're pretty stringent on the requirements and their adherence to the reference designs. But, one of the things that I think is interesting is last year it was about GPU optimized storage, right? Mm-hmm. It seems like, Jim, what you're saying is this year's, let's say. The evolution of that idea is AI optimized, right? Yes. Because when we're looking at AgTech, the workloads are more expansive. They're not singularly focused on just the AI element.

Karl Freund 16:58

that the key problem they were solving is that an agent's working on the job you gave, it has got to work continually in the background. Yeah. For days, weeks, or months. Yeah. And so you need all of its storage that agent's. Cadre of, of other agents. They all need very fast storage to be able to continue to work on that problem until completion.

Leonard Lee 17:21

Yeah.

Karl Freund 17:22

when we talk about ai, we think in terms of asking chat, GPT something, right? That's a one shot deal. And once it's done. So the storage required for that only exists for, a few milliseconds. Yeah. Whereas an age, it may need to have its storage close by for weeks or months. Yeah. And there wasn't a good solution for that until NVIDIA came out with the STX. So that's, it's a really different animal. Again, everything's being driven right now from magenta ai. I can't say that often enough.

Leonard Lee 17:52

Well, we, last year we made that observation around tiered memory, and then the very bottom of that, that tiering was the storage bit, right? Yeah. The, G-G-G-P-U optimized storage for long, long memory, right? Mm-hmm. And so, yeah, I mean, Jim, since you're like the super technical guy, what did you see changing there in terms of the thinking around the memory tiering because. That was such a big deal for the course of last year.

Jim McGregor 18:23

Well, we came out, they came out with so camm and it was a proprietary, basically architecture last year developed between Micron and Nvidia. And within a year it became an industry standard. And we're seeing not only interest in it from anybody else that's doing AI accelerated platforms, but even from like HPC applications, they don't want necessarily the height. Density, but they want the, they want the high bandwidth. So they're looking at it from, for lower density configurations. So just the fact that, that's become a new, new, level of, hierarchy within the memory hierarchy. To support KV Cash and now, we were already talking last year about GPU attached storage. Well, now essentially what we have with, STX is accelerator attached storage, and I hate to say GPU because now we have the LPU and everything else, but you have that storage solutions. It's really dedicated to those AI accelerators to be able to handle and. Perform those solutions in addition to that. Yeah, that changing memory stack that we're developing.

Karl Freund 19:27

Yeah. And at the risk of self-promotion of the idea of Cambrian explosion, we're seeing a camian explosion of the heterogeneity.

Jim McGregor 19:34

Yes.

Karl Freund 19:35

Yeah. Of AI solutions ev. Not one solution, one size will not fit all.

Jim McGregor 19:40

No.

Karl Freund 19:40

And I think Jensen realized that, and really that's what drove him to expand his portfolio to CPU Racks and LPLP racks, along with the GPU Rack and storage racks and network. it's a remarkable shift that I don't think the market fully appreciates right now. Most people don't appreciate that right now. And then Leonard, kudos to you and your platform for helping other people. More people understand what's really going on here. It's a major ground

Leonard Lee 20:09

shift. It's you guys as well. And the fact that we had these debates instead of just agreeing with each other. I think that surfaces those salient points that, hopefully, shapes a better dialogue around all this stuff, but that helps people, maybe think,, in more lockstep with what's going on. I mean, so thank both of you guys, and again, congratulations. You guys are all part of this. Achievement here, with the YouTube channel. but no, you know what? I, here's the thing though. I don't, I really don't think, Jensen knew what was. coming down the pike. I know a lot of people love to think that, but here's the thing. Let's go back to the roadmaps that he's laid out over the years.

Karl Freund 20:54

Mm-hmm.

Leonard Lee 20:54

This stuff was not part of the roadmaps and he was very vocal about saying that. We give you assurance that, over a four year period, you guys can depend on us. To have that, reliable roadmap that can guide your investments. What we're looking at now it's turning into a much more complicated, outline even for AI supercomputing, and I think, You guys made the comment about adaptability. Yes. They're adapting really quickly. And one of the things I'm noticing, even with some of the hyperscaler plays, whether it's Maya 300 or, Ironwood or Google's whole line of TPU systems or it's, train. Everyone is trying to. Adapt and we're seeing roadmaps as well as platforms that are transitional. Right? And the way I looked at, the Grop three LPX system, it is a weird add-on to a Vera Reen and you mentioned Carl, you're gonna have a Vera Ruben, and then maybe one to four. LP X Racks

Jim McGregor 22:09

Actually, we should specify. You're gonna have a Vera, you're gonna have Rubens, you're gonna have STXs. You're gonna have Lp Xs. Yeah. So you're gonna have, and eventually you're gonna have cpx. So yes,

Karl Freund 22:22

and CPX will be a part of that as well for some customer.

Leonard Lee 22:24

This is a lot. This is a lot for, Customers and the industry to digest. I mean, it, it we're going away from what looked like for three years ago, a very simple looking and extensible roadmap to something that, to your point, Carl has become very heterogeneous and mm-hmm. Really a diverse stack to meet, a wide range of. Domain specific requirements or even function specific requirements. And that was something that, that really resonated with me in Jensen's talk is that he actually gave to application specific stuff. Right?

Jim McGregor 23:05

Yeah.

Leonard Lee 23:05

at least that was my. Take, I don't know how, what, how you guys,

Jim McGregor 23:09

well, I, I think it fits in with that whole view that, the data, the new unit of compute is the data center. Yeah. And the fact that, we would've thought about this as a single server. How do we architect a server? Yeah. 15, 20 years ago, how much memory are we gonna put into what kind of storage? What, what type of processor do we put an accelerator in? Now we're at the point where. these functional blocks are entire racks, and we're thinking about the data center as one unit of compute, and really how do we build this system? What pieces do we need to put in there? And let's face it, if we went back 30 or 40 years, it was, what components do we put together? And then it went into an SSC. Yeah. And then it was, so, we continue, building this, I just think that it's gotten to a point where we're, yeah. And you have to remember that we're now addressing problems. When you think about simulating the human genome, when you think about simulating the earth, when you think about simulating anything physical. Yeah. We're now reaching into areas of scientific research and analysis that we only dreamed of, previously. So, we're building systems. Bigger and better than we ever have. Just look at what we're doing with, AI and multimedia and the fact that we now really have realistic AI generation capabilities for entertainment and gaming and stuff like this. It is amazing what we can now do. the roadmap just keeps taking us to where we're gonna get. We're gonna be able to do more, and we're gonna be able to do more with less. Because, from what we've seen, every year we're seeing about a 10 x improvement in efficiency. Just in terms of the model sizes. So the model sizes are coming down 10 x At the same time, the performance of these systems, or performance efficiency of these systems is going up. 10 or a hundred x. So we're seeing a thousand X improvements in capabilities and efficiency almost every year.

Karl Freund 25:02

almost every year. It's mind

Leonard Lee 25:03

boggling. I had a, yeah, I had an interesting conversation actually on my other podcast study, iot Coffee. Talk about that performance versus scaling. And this is where things like times X whatever, a hundred times or billion times there's some charts that gentle was showing. I think one of the things that, that oftentimes gets conflated is the difference between, the scaling versus performance with performance. It's really about. Do you have more efficient compute and are you able to then, have that more efficient compute in a system that can. maximize the density or push the envelope with density, right? Whether it's, innovations on the thermal front or the power or compute, right? And then you have scaling, which is the scale out. It has really been pushed forward with scale out and now scale across, right? And so some of these times a million or larger. Multiples that we're seeing are typically on the scaling of the size of the clusters that are now spanning a data center campuses. Right? And so I think it's really important to have that, proper lens as we're ingesting these large multiples that people are throwing out there.

Jim McGregor 26:35

I think of it in the same form factor. And I'll be honest with you, I do see, anywhere from a hundred to a thousand x improvement per in the form factor every year. I was at Embedded World in MWC. And I'm talking to the network guys and saying, listen, we need to make the network part of the platform, part of the workflow, because in many cases, like physical AI slash robotics, it has to be the intelligence. It has to be the backbone of the solution because A robot has a form factor, it has a power budget, it has all these things and a performance limitation. You're gonna want a lot of that intelligence in the network to do the sensor fusion Between all these different hundreds or thousands of sensors between robots and other sensors, you're gonna want it to be able to do the orchestration between all these different, units. You're gonna want it to do LLMs that you can't do on device, and you're gonna have to be able to do that in sub one millisecond. So the network, whether it's a router, whether it's a ran, whether it's whatever, has to be able to do that. And it has to be able to handle a thousand x increase in data.

Karl Freund 27:39

yeah,

Jim McGregor 27:40

some of'em looked at me and said, oh no, we're gonna do all that in the cloud. I'm like, no, we, you're not. And some of'em looked at me and said, well, we'll get, there we're six G. I'm like, no, you need it. Tomorrow. I don't think they get the urgency and the need and the fact that, we really are, you talk about, yeah, we are seeing some of these astronomical numbers thrown out there for a million X and everything else with these data centers. But quite honestly, when you look at the amount of data that's being processed, it is literally increasing a hundred to a thousand X every year. And it has to be done in the same form factor,

Leonard Lee 28:12

and so that's the volumes. But if there's a suggestion that generation over generation, we're seeing these large multiples in terms of performance. The other topic that was, debated pretty heavily, at least out, I was doing my own thing, so the, I wasn't part of the analyst program. I don't know how much this was talked about, but the whole obsolescence topic, right. when you look at something like that in terms of a generational improvement. It really makes the whole, Hey, you can extend a useful life of any investment, 10 years really like untenable. Because at the end of the day, it's not about how do you repurpose, you can repurpose things. People do that all the time. It doesn't mean that it, let's say that you get a hundred times improvement in on the opex and with the energy efficiency and producing a token. That alone makes the prior generation obsolete because it's a hundred times more.

Karl Freund 29:15

But that's not what's happening. The price, the cloud price of AMP peer. Mm-hmm. Remember AMP peer?

Leonard Lee 29:22

Yeah.

Karl Freund 29:23

Okay. The cloud price of AMP here is increasing, not decreasing. So that tell me, demand is demand. Demand is

Jim McGregor 29:29

outpacing it.

Karl Freund 29:30

Thank you, Jen. Right, right. Exactly right. Demand

Leonard Lee 29:32

is not pacing. There's the price. But the question with everything related to monetization is it profitable? And the question is it profitable to run that workload on the Amp P versus,

Karl Freund 29:46

hey, FOYA, Blackwell. you fully depreciated the asset. Of course it's profitable. You fully depre appreciated it. The only cost, the only cost really is power.

Leonard Lee 29:57

Yeah,

Karl Freund 29:57

right. But so, so six, seven years old.

Jim McGregor 30:02

Yeah,

Leonard Lee 30:02

So the cost, you have to factor into opex as well. But then if it cost. 10 times more to run that workload from the opex perspective on Ampire versus Blackwell.

Karl Freund 30:15

But if I can't get Blackwell, if I have more work that I have capacity, yeah, of course I'm gonna use that ampi. It's fully depreciated asset, so I I have to disagree.

Leonard Lee 30:24

Yeah.

Karl Freund 30:24

Which is what I love about these conversations. That's fine.

Leonard Lee 30:27

Yeah. Well, that continues to be what I call, Jenssen's Paradox. And I think it's gonna continue to be a question'cause I don't think people factor in the opex end of it. And then also the idea of obsolescence is,

Jim McGregor 30:41

is of

Leonard Lee 30:42

course,

Karl Freund 30:42

in the opex. I mean, these guys know how to use

Leonard Lee 30:45

spreadsheet.

Jim McGregor 30:45

I think you're bringing up the key issue, and this is, the fact that. Obviously there's an entire ecosystem, there's an entire value chain. Yeah. And everyone needs to have a positive return on investment at some point. Yeah. In that value chain. Yeah. And that's brought up the whole question, especially by the financial community of, is there an AI bubble. Well, in terms of demand, there is no AI bubble. we, as long as I can see or that we forecasted, we have wave after wave after wave after wave coming. Yeah. That is pushing just the demands that we have that are as, up astronomically. but, does that mean that we have the business models in place for all parts of that value chain to make money effectively? That's a good question. And I don't know. And I think you do a good job of bringing that up, Leonard. I don't know that all parts of that value chain are really set up to effectively do that.

Leonard Lee 31:37

And that is still the vaccine question for the industry. Yes. Right. And then it's not just about monetization as profitable monetization and in terms of. One of the things, especially for, here's another topic that came up. especially as it relates to AgTech thin ops, the economics of, AgTech is not well known. One of the things that I'm hearing quite a lot these days from the developer community is that the unpredictability of the cost of, work, when you assign it to an an agent, it's inconsistent. Mm-hmm. Because you never know, like, like say let's for the same unit of work, the cost can vary dramatically and we have to still remember these things can still confuse themselves and get into these infinite loops.

Jim McGregor 32:27

Well, and we have agents using agents, so you never know how many agents it's calling or how many models it's using or how much data it's used. Matter fact, you can ask. The same question twice of an agent saying, well, I want more detail. Right? Or are you sure about this? Just to make it double check itself. And that changes, everything

Karl Freund 32:44

changes, right? And Jensen had an interesting concept that I hadn't heard of before. I hadn't really thought about. when you hire an engineer, you need to give that engineer a token budget, okay? And you make sure he's, if he doesn't spend his token budget, he's outta here. Right, because that means he's not being as productive as he could be. but to Jim's point, he could also rapidly outspend his token budget and become a cost problem. So I think this uncertainty is something we'll have to live with for a while. Yeah, it will be solved.

Leonard Lee 33:17

There's the efficacy of how tokens are spent. It's good that he brought it up, but, there's a big. Question about what is the value of a token? not from a cost perspective. The cost perspective each, the cost of each token, depending on the system is somewhat predictable. But from at the application level, it's completely. Unpredictable. It's very difficult. Whereas, let's say for a deterministic system you have for transaction costs and pricing, that can be calculated. But how do you do that when you're using agent systems? it can be variable and yes, very dramatically different. So the variance can be different. How do you cost for it? So the whole finops. Topic, I think has largely been avoided because of exactly this. and then also introduces risk to the value conversation around AG agentic ai. But anyways, I, I think this is something, we'll, probably it's great topic next year. Yeah, it's next year. It's this year. Denial. I,

Jim McGregor 34:26

I, I agree.

Karl Freund 34:26

I

Jim McGregor 34:26

it's a great topic and it's great that you keep bringing it up and that, you've made it a key focus lender. It really is.

Karl Freund 34:32

Yeah. I think remember this year is the year Gentech AI became. Real impossible.

Jim McGregor 34:39

Yeah.

Karl Freund 34:39

But we, we'll worry about the economics down the road, right? yeah. Right now people are saying, oh my God, what's I, I like the way Jensen put it. He says, every company needs to have a strategy for open claw. And I more thought about that thought, wow, he's right. This is that transformational. What is your open call strategy? Everybody from IBM to Pillsbury.

Jim McGregor 35:01

we're working on it. Matter of fact, we've been working on it for a month.

Karl Freund 35:04

Yeah.

Jim McGregor 35:04

And the first thing we found out was open claw was dangerous.

Karl Freund 35:07

Yeah, yeah. Oh my God. Exactly. I love that.

Leonard Lee 35:10

Yeah.

Karl Freund 35:10

I like the way that Nvidia says when you set up your first set up, your open claw, it has zero, access. You have to grant it access, it can't inherit your ACL L right? It can't in inherit your capabilities. You've gotta give very specific, very fine grained access and very specific files and data. Although, or Jim's rice's gonna run amuck, it's gonna put you out of business.

Leonard Lee 35:37

Although I will tell you, I, I discovered some really interesting things about through. Readiness of being able to sandbox, even put together a sandbox environment. So it's not stuff that's publicized, it's stuff that's in my research and the discovery that I did at gt, GTC. So, clients, if you wanna find out, gimme a ring.

Karl Freund 36:00

Yeah.

Leonard Lee 36:00

I want to give a shout out to all the people who have been doing AgTech on device. the likes of Apple, Google, Microsoft. Lenovo Honor all these guys that have been looking at on device and trying to figure out how to do personal ai. Jim, you alluded to this earlier, that this is really about on device and doing something more localized than, your own thing. Right. this stuff is really hard and one of the things that I really struggled with when Jensen got up on stage and said, this is enterprise grade, prove it. Because these other companies have been doing, been struggling with this for two years, right? And they're being labeled as being behind and all this other stuff. now you have this very dangerous thing called open clause. You mentioned Jim. And then in two months, probably at best, now you're gonna come and say, claim that this is enterprise grade chart looks great. I'll tell you right now. Most people don't even understand this stuff. That's true. I much less know how to secure it and I would, what I want to see in 2026 is nemo. Claw to prove its claims. I'm not believing this stuff face value because we're working on, we're working on that now.

Jim McGregor 37:21

So gi give us a month and we'll see.

Leonard Lee 37:24

Yeah, see right here. And you're gonna

Jim McGregor 37:26

right here. Call. We're doing it now. We're doing it live.

Leonard Lee 37:28

exactly,

Jim McGregor 37:29

we got three software developers on staff

Leonard Lee 37:32

And to the audience if you want to know what's up with all this stuff.'cause this is probably the most important question coming outta GTC. Jim and his frigging amazing kids and the developers and the team there. Definitely. you have to talk to Jim, but, oh, one thing before we call it, wraps. Jim, I wanted to get your take on AI grid. did you have any. Points of view on that because that, for the telco industry, it's a big thing. There's a lot headlines about it. There's a lot of attempts to associate with some sort of. Game related to AI ran, what was your like impression?

Jim McGregor 38:12

I'll be honest with you. I was still rather disappointed. obviously there were a lot, MWC was all about AI ran and how we're adding AI into the network. especially, with six G being, basically an AI defined. solution. however, Nvidia had a big announcement with Nokia and several, telcos over there in Europe, talking about AI ran, seeing the, I think that the telcos and the equipment companies start to see more opportunity integrate intelligence into the network, but they're still looking at it from a network management perspective. In other words, how do I more efficiently run my network? How do I do this? How do I do that? I don't think that they truly see the opportunity and they need to, I really think that the business model for the telcos has to change, and we've been telling'em this for over a decade now. Yeah. But the fact is that six G may be the last truly defined network because they're gonna have to be upgrading their networks. Yeah. Whether it's hardware, software, or both, every two to three years going forward just to keep up with this demand. Once again, like my robotics example, the network becomes a critical part. The workflow of the platform that you're looking at. In a lot of cases, you have to do the intelligence in the network because you can't afford to go to the cloud for deterministic applications. For applications that just require low latency, a 10 millisecond delay in a robot can be very hazardous. You can't afford that. You have to have sub one millisecond. Yeah. So I think that, and. I stood up in front of a whole crowd in front of Noia event and said, I don't think you guys have a sense of urgency. and I really don't. I think that, I think that they need to get urgency. So I still think that the network and it'll be interesting'cause I think this is an area of innovation and if the traditional equipment guys and telcos don't rise up to the task, somebody else will.

Leonard Lee 40:02

Yeah. And I love the fact that you were the last, last person to ask a question. I was like going, oh my God. So there he is. I haven't seen him this whole time. And you were telling, what did you tell the, The mic handler when I was up there asking

Jim McGregor 40:17

Oh, I, you were right before me. And, he was trying to hand me the mic. I says, oh, don't worry, he's gonna ramble for a while.

Leonard Lee 40:25

I love that. But my question was actually pretty short. But you know what? You keep bringing up networking. and I guess, we'll wrap up with this. I'll just make a comment. I didn't hear too much, which was really weird given, I think we all recognize how important networking is and interconnect is for the continual scaling, especially when we look at all this heterogeneous stuff. Mm-hmm. The networking has to look really different. It's not gonna be GPU or the old roadmap centric, it's gotta diversify. But the CPOs, the CPO stuff, I, I was surprised to see the semi committal, push. For the CPO, stuff, and they introduced new, rack, or a switch, right? The spectrum X.

Jim McGregor 41:12

Oh, I'll be honest with you, I'm seeing more interest, I think, more pull on the CPO stuff than I am push. So, and then instead of seeing the push, I think maybe from NVIDIA and other technology providers, I think the end market is going to pull that in even quicker. So I, I would agree. I was surprised I didn't see more from it, maybe from Nvidia, but I still think that opportunity and that demand is there because I think the hyperscalers. The neo clouds and all these guys building data centers out, realize how much that's what the value of that is from a Yeah. complexity perspective, a cost perspective, just an overall TCO. So I think we're gonna see a lot more pull on the networking side than we ever have before.

Leonard Lee 41:55

And I think in a really weird level, I'm hearing a lot of that pool that you're talking about coming from, like, the, The, Sienna, menta, that kind of level where you're talking about, this really high capacity, low latency, networking for, scale across is, I don't know, that's where I'm hearing a lot of it, but where this is real, and the requirements are being driven at that level. Not so much at, let's say the rock level, you know what I'm saying? No, I get it. Yeah. That's just what I'm noticing.

Jim McGregor 42:27

yes.

Leonard Lee 42:27

Awesome. And gentlemen, I think this is one of the best, episodes we've ever put together. we, we did some really, tasty disagreements, which I love. Isn't this why we started this Carl, right?

Karl Freund 42:39

Mm-hmm.

Leonard Lee 42:40

Yeah.

Karl Freund 42:41

It all started with our fundamental disagreement around the economics of ai, right?

Leonard Lee 42:46

Yes. People don't realize that we didn't do this to come

Karl Freund 42:50

and we're still talking about

Leonard Lee 42:51

it, like, oh, hey, I want you to just, repeat what I just said,

Karl Freund 42:54

Ah,

Leonard Lee 42:54

you guys are awesome. Love doing this with you guys, thank you so much. thanks to the audience for listening in, and please reach out to Jim McGregor. he and his team are working on one of the biggest problems, today, which is AgTech economics. taking a look at. What are the finops aspects of all of this? How do you, mm-hmm. How do you stabilize our, the ROI argument and the value propositions around, gentech applications? And of course, curious research, www.curiousresearch.com call. The mind expanding AI Cambrian expert, reach out to him to understand what is happening in the landscape of AI and the silicon, as well as software that's, changing industries and changing the semiconductor industry and pushing the AI industry forward. And also, please subscribe to our podcast. Yes, we got one of these. We're featured on YouTube, obviously because of the YouTube placard. And, check out the audio version on Bud Sprouts and, your favorite podcast platform. Also subscribe to the next research portal, at www.next-curve.com, as well as our substack. You can find some substack and of course, all this for the tech and industry insights that matter with curious research and Cambrian. AI research, LLC. Gentlemen, thank you so much and we'll see you at toward the end of the month. Right. We gotta do our March recap, so,

Jim McGregor 44:30

yep.

Karl Freund 44:31

Have a great weekend.

Jim McGregor 44:32

Cheers.

Leonard Lee 44:34

Take care. Bye.