Silicon Futures for July 2026 - AMD Advancing AI, Open model debate, OpenAI's rogue agent crisis

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

This month, the world of AI and semiconductors has gone truly mad. Drama and volatility are the only ways to describe what we witnessed in July with AI agent swarms going rogue, a frantic effort to save open-weight models from being banned by the U.S. government, and AMD coming to market with their first rack-scale AI supercomputing system to take on NVIDIA's dominance. 

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

➡️ AMD Advancing AI 2026 highlights (3:47)
➡️ AMD + Cerebras partnership for inference AI (4:18)
➡️ AMD's strategy for heterogeneous AI inference compute (7:04)
➡️ AMD touts CPU portfolio and diversity for AI (10:34)
➡️ Is agentic AI compute the same as AI training compute? (13:57)
➡️ Is AMD's ecosystem 18-month behind NVIDIA's? (17:29)
➡️ Could ROCm.ai be a CUDA equalizer for AMD? (19:46)
➡️ Are we in the infancy of AI still? (24:54)
➡️ Will closed AI models and labs continue to lead? (26:34)
➡️ AI for security vs. AI for threat actors. Who wins? (33.39)
➡️ Is the gigawatt a good metric for "AI capacity"? (37:57)
➡️ QUANTUM NEWS! IBM acquires HRL Laboratories (43:58)
➡️ What to make of Intel's public commitment to 14A? (43:58)

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

Welcome everyone to, uh, Next Curve's Rethink podcast. In this episode, um, we, as we always do, break down the latest tech and industry events and happenings into the insights that matter from the world of semiconductors and Carl's most favorite topic, AI, second only to quantum, I would say. Right, Carl?

Jim McGregor 0:33

Yeah, quantum.

Leonard Lee 0:34

Right?

Jim McGregor 0:34

We gotta talk about quantum. Quantum.

Leonard Lee 0:35

Yeah.

Jim McGregor 0:35

Gotta talk

Karl Freund 0:36

about quantum.

Leonard Lee 0:37

Really? Okay. All right.

Jim McGregor 0:38

Oh, yeah, no, there's a huge announcement about quantum this month.

Leonard Lee 0:41

Ooh, okay. Well, hey everyone, this is gonna be really, really exciting and we're gonna get some fantastic insights from these two. So, yes, I am Leonard Lee, Executive Analyst at Next Curve, and I'm joined by, the very relaxed and tropical Carl Freund of Cambrian r- AI Research and soon to be Quantum Research. And we also have, I'm at, at loss of words and adjectives for you, Jim.

Jim McGregor 1:10

Good.

Leonard Lee 1:11

Uh, I was gonna look up something- Finally. Finally in the, in the thesaurus, but, uh, uh, Jim McGregor of the famed Terias Research. And, in this Silicon Futures episode, we'll be talking about the hotli- headlines of July that mattered, and that is going to undoubtedly include topics such as, well, quantum. Mm-hmm. AMD Advancing AI 2026. Yeah. We must talk about that. maybe we'll touch on, Samsung's Galaxy unfold with, Galaxy I- AI and what's going on there, from a silicon perspective as well. And, is it- We've

Jim McGregor 1:52

got, we've got Meta's Iris-

Leonard Lee 1:54

Okay

Jim McGregor 1:55

that's going into production.

Leonard Lee 1:56

Is the And Carl, we wanna talk about the gigawatt. Is it a good metric for, quote-unquote, "AI capacity"? And what the hell do we mean by AI capacity? hey, we have to talk about this, OpenAI and Anthropics' rogue agent crisis, or is it really a crisis? Oh, yeah.

Karl Freund 2:15

Right.

Leonard Lee 2:15

and you know this weird thing that's happening this month, this turmoil around open models, right? There's a lot of stuff happening there. We should talk about that debate that's going on. And then, the state and impact of memory, we have to talk about that, give our audience an update on impact and, influence on, the supply chains across the electronics.

Jim McGregor 2:38

Yeah Is, is AI coming to a head or is it continuing to barrel forward?

Leonard Lee 2:43

But yeah, and then, let's talk about token maxing, to- Token shock yeah where is it going, and what are the implications? And, probably more, but, before we get started, please remember to like, share, react, and comment on this episode, and also subscribe here on YouTube and on Buzzsprout to listen to us on your favorite podcast platform. Opinions and statements by these two gentlemen right over there are their own, and don't reflect mine or those of Next Curve. We're here to, provide an open forum for discussion and debate on all things, AI, quantum, and silicon, for informational purposes only. So let's get started, gentlemen. let's talk about AMD, advancing AI. Yeah. what do you guys think? Where

Karl Freund 3:28

That's definitely the biggest news of the month. Yeah. There's lots of news this month. That was the big one, and they announced a lot of stuff that we already knew was happening- Yeah MI400, 450, 455. Helios Rack, they announced some customers. We all knew that was happening. What they announced that nobody knew was gonna happen, but some of us suspected might, was the relationship with Cerebras. So they're doing the same bifurcation of- Ah inference processing that Nvidia did with Groq. They're doing the same thing with Cerebras. Yeah. and I won't make any judgments as to whether Cerebras or Groq is a better partner. I will say that Cerebras is still independent. Yeah. And-

Leonard Lee 4:09

Yeah

Karl Freund 4:09

and AMD did not acquire their assets nor their leadership. And Cerebras is working closely with Amazon AWS to add, a disaggregated, im- inference platform for AWS. You can check my Forbes blog. I wrote a, a blog about disaggregation, inference to disaggregation- Yes what that means to the market and the various players in the market. So check that out on Forbes.

Jim McGregor 4:36

Yeah. Now, you say that AMD didn't acquire the management, but in the past it had.

Karl Freund 4:41

had. Actually, Andrew Feldman- In fact, that very man was

Jim McGregor 4:43

acquired by AMD-

Karl Freund 4:44

It was acquired by AMD once upon a time. Yeah.

Jim McGregor 4:47

So it's kind of funny, but, uh, yeah, it's... And that is, that was a kind of an interesting choice because of all the inference processing solutions out there- You know, not saying that, Cerebras has a massive processing capability, but it's a wait-for-scale solution. Yeah. It's a large solution. It's an expensive solution. The question is, is it gonna be a competitive solution when paired with, AMD, and is it gonna be, in the same... I mean, even the racks that they use are different sizes than the Helios rack. The Helios is a double wide, OCP standard. They didn't provide too much information about that, how that's gonna work, and, the pricing and everything else. And Cerebras has already gone down a different road to be a service provider and opening up its own data centers, with its financial backing, from the Middle East. So I'm not quite sure how that's gonna work yet, and I'm not quite sure- Yeah if that's the right solution. But obviously AMD definitely, has to have a complete solution, which includes training and inference.

Leonard Lee 5:47

And I think, it falls in line with, what Lisa continues to, tell everyone, which is, AMD is taking more of an open approach. So the way that they're engaging with Cerebras, the way I characterize it, it's loosely coupled, so it's not this aqua hire that, NVIDIA did with, Groq. And so I think that gives, both companies agency, but that partnership, at a technical level also gives them a lot of flexibility. Because, the whole inference market and the way that these models are evolving and what optimal looks across, and now what's becoming an increasingly heterogeneous landscape of, compute, that flexibility might be really important going forward. So it's not about, putting all our eggs in one basket, it's really tr- trying to figure out, As these models evolve and these model architectures and agentic frameworks start, continue to evolve, what kind of system do you need to put together? And who are the best partners for certain types of, let's say, applications? So when we look at our applications, AI applications now, we have to look at,, what are the agentic compute requirements? And then what are the things that are model specific and the harness and all this other stuff? These systems are not just about the model anymore, and we're now departing from, y- this regime that we're all used to, or many are in is this mindset that, AI compute is all about training and what's optimal for training. And that's obviously changing as we start to see the diversity of inference and then now agentic, uh, start to influence, what optimal means for AI compute. I don't know if, if you guys are seeing that. Well,

Jim McGregor 7:33

I'm glad that you also brought up that AMD is taking that open approach because- Yeah the Cerebras announcement may not be the only announcement. Yeah. there still might be other partnerships or other acquisitions to address different types of, agentic and/or AI workloads, AI inference workloads. So it'll be interesting to see as that develops. Obviously, it's still early days for AMD in that segment.

Karl Freund 7:56

Yeah. And, and we don't know what the performance will be. We have no idea. I- it's interesting talking with Andrew, I said, "How, how did this come about?" He's like, "First of all, the conversation was easy because we just connect over PCIe-

Leonard Lee 8:09

Mm

Karl Freund 8:09

and Ethernet." And I'm like, "Well, that's good news, but it's also bad news because Groq and, or not Groq now, but the LPUs and Rubin are connecting with much more-" Co-designed low level chip to chip interfaces-

Jim McGregor 8:27

Yep

Karl Freund 8:27

so they get much higher performance theoretically. Yeah. So the good news is Freebush can connect to anybody. Bad news is they can't... there's not a lot of advantage. Now, Andrew would say, "Well, I don't need to have a high-speed link. we're j- we're just sharing, we're just sharing the cache," right? The keep KD cache. So, we'll have to find out. These guys- Yeah know a lot more about it than we do.

Leonard Lee 8:52

and, that could very well be the case depending on the application, right? And so I'm saying the application- Yeah then you have all these different workloads, right? It's not just a singular workload anymore. It's a, a series or a chain of workloads. The latency requirements, the bandwidth requirements could be very, could be very, very different depending on the application, right? And, one of the things that really stood out for me from AMD Advanced, AI was how, AMD has really taken the charge on, the CPU for AI, right? They have different SKUs now, different configurations for different kinds of workloads that they anticipate or they do see, being required to support agentic AI, right? And again, going back to this notion of you have the application, you have, the agentic and AI frameworks underneath, the diversity of CPUs that are gonna be required to make these, applications happen, right? and so with Epyc, she described it more as a portfolio now rather than a particular generation of chip. What your reaction to that-

Jim McGregor 10:05

Well-

Leonard Lee 10:05

that announcement was.

Jim McGregor 10:07

we've seen that for the past couple generations where they've come up with different, SKUs for different applications and different things, so we expect that. But, we have to remember also that they're not just focused on the data center. You know- Yeah they did talk about, Halo, which is their small form factor workstation. They did talk about, edge AI, especially with their Xilinx solutions, and I think people forget that Xilinx has actually been, doing robotics and some of these applications for decades. So matter of fact, they've been a, a base- a standard platform for FIRST Robotics- Mm for quite a long time. These are not new markets for AMD. AMD is actually, is one of the solutions, one of the co- few companies that can go from the edge to the cloud, and from, physical AI to agentic AI.

Leonard Lee 10:53

Mm-hmm. Mm-hmm. Carl, you were going-

Karl Freund 10:56

Yeah. No, I agree. I think it's an easy card to play. It's in your hand. You've got a whole portfolio, and Nvidia doesn't, so play that card. how meaningful is it for the data center? I don't think it is.

Leonard Lee 11:10

Really? Okay.

Karl Freund 11:11

Heterogeneity is good in the data center to a point, but if you end up with, a suit made of 1,000 different SKUs, that's not gonna be manageable. And so we'll have to see. Sure,

Leonard Lee 11:24

sure. Yeah. And

Karl Freund 11:25

I mean, do you really think hyperscalers are gonna want five different Epyc CPUs?

Jim McGregor 11:32

actually I would say yes. Yeah. But they won't be using them all for AI. Okay. So I mean, they'll still want a CPU- That I agree. That I agree for enterprise workloads. They'll still want a CPU- Yeah for communications. They'll still want a CPU for other applications. That's right. So, I mean-

Karl Freund 11:46

Inferencing.

Jim McGregor 11:47

Yeah. Yeah, a CPU can do inference processing, but let's face it, you're still not gonna get the performance or the return that you're gonna get with- Yeah a, a dedicated accelerator.

Leonard Lee 11:55

But let, let's take a step back. We're not saying that the CPU is replacing accelerators or GPUs or what have you. No. It's, it's doing- It has, especially with the Gentic- orchestration yeah, with the Gentic frameworks- Yeah and just any AI application, Jim, were you at, Arm Everywhere?

Jim McGregor 12:13

Yes

Leonard Lee 12:14

Yeah, you were there, right? And so- Yes if you went to one of the demo booths, they had this really cool chart that showed you an agentic flow, and where compute workloads were placed through that entire agentic flow. What you... A- and this is something that AMD actually highlighted as well. actually, most of the, the workloads, and, whether it's supporting or core, run on CPU. And it's not that they're running AI. When you look at the overall AI application-

Jim McGregor 12:43

Yeah, it's the orchestrator.

Leonard Lee 12:44

Yeah the GPU actually doesn't do that much in the context of the overall flow of, tasks and, o- operations, which is... it seems counterintuitive, but that's actually what we saw with RAG when semantic search was a big deal. Hyperscalers are noticing, especially ones that were looking at, leveraging it for recommendation engines, but a- also, search, that a lot of the workloads, for the retrievers, would, dri- be driven off of the CPU. Mm-hmm. So this is something that, is not well known, but these guys are now making it well known. That... And this is why we've seen the CPU, uh, business of these companies in the background kind of ramp up as we're looking at these AI numbers, growth numbers. Actually, there's been this shadow growth in CPUs because of this reality. And obviously, there is that, trend of, refresh, right? there's typical IT refresh of, a traditional, compute. But now we're hearing more and more how, a lot of the traditional non-accelerated computing is becoming more important in even these agentic, applications, whether it's tools or what have you, right? That's the point. I think that's important to clarify to the audience.

Karl Freund 14:06

And, I think it'll be interesting when we start adding these, other accelerators to the mix. So now I'm gonna have- Yeah lots of CPUs, I'm gonna have GPUs, and I'm gonna have LPUs or

Jim McGregor 14:18

whatever

Karl Freund 14:18

you call them.

Jim McGregor 14:19

NPUs and... Yeah, and NPUs. FPGAs

Karl Freund 14:22

and DPUs and... Yeah. It's quite, quite heterogeneous. The nice thing is, is that you can mix the ratios- Mm-hmm depending on your needs, right?

Jim McGregor 14:29

Absolutely.

Karl Freund 14:30

Yeah. So some, some places will have, four racks of LPUs and one rack of Rubens and a rack of Vera, and others won't, won't- Yeah have complete configuration.

Jim McGregor 14:40

I think the greatest challenge is definitely gonna be architecting these as a single system You know, not just within a rack, but, the clusters that you have to put together. So I still think that's where NVIDIA has a significant lead in doing, more of the entire platform, and the fact that, they're architecting everything to work together. Yeah. And they're working with companies like Vertiv and- Yeah and stuff like that to make sure that the cooling, the power, and everything else just all works

Leonard Lee 15:08

one of the comments that I made and my key takeaways from this event was that, AMD is about 18 months behind on ecosystem. And, I made, put this caveat. That might not necessarily be a bad thing, but, we noted this long time ago how aggressive Nvidia is, is- Mm-hmm like moving up the stack, right? And so you're absolutely right. I agree with you. Jim, your name is Jim, right?

Jim McGregor 15:31

Yes.

Leonard Lee 15:31

Okay.

Jim McGregor 15:32

Well, and the, the other- Bill thing that- Right obviously Nvidia is very aggressive on the software side, especially the model side. Yeah. And they're open sourcering, open, open sourcering- Sorcery. Sorcery. They're open sourcing- more and more content, especially from that Cuda library- Yeah than we've ever seen before for chip design, for, uh, biosciences, for you name it. But I- I would say at least both companies, especially AMD and Nvidia, are continuing to invest in the ecosystem and invest- Yeah in the rest of the industry as well.

Karl Freund 16:03

Yeah. and Nvidia's investments are part of their circular f- financial arrangements with their customers, right?

Jim McGregor 16:09

Well, so are AMD's.

Karl Freund 16:11

Yeah, yeah, but that, to a much lesser extent. They have a lot less money to throw around. And so- Yeah Jensen's sitting here thinking, "Okay, I can give this back as a higher, as a dividend, maybe a one-time dividend of 50 bucks or something," or he can say, "Or I can invest in these 20 companies that are eventually gonna buy a lot of GPUs and create new markets for more GPUs," and he's taken the latter path- aggressively.

Jim McGregor 16:35

Yeah. And AMD's got some huge announcements with, Anthropic Anthropic and Open- Yeah OpenAI.

Karl Freund 16:40

That's true. They do.

Jim McGregor 16:41

I mean, one of them even involves 10% of AMD stock.

Leonard Lee 16:44

Yeah.

Jim McGregor 16:45

So- Yeah yeah.

Leonard Lee 16:46

Yeah.

Karl Freund 16:47

I think Nvidia's investment has a much longer tail.

Jim McGregor 16:50

Yeah.

Karl Freund 16:50

Right? And they invested in Chip Agent, right? And we can talk about Chip Agent and it's like-

Leonard Lee 16:56

Oh

Karl Freund 16:56

uh, another time. But the use of AI to develop chips is-

Leonard Lee 17:00

Yeah

Karl Freund 17:00

critical to Nvidia.

Leonard Lee 17:02

And let me throw this out there to both of you, because one of the things that I thought was super interesting that AMD announced was ROCm.ai. And so that's,, using AI to fi- quote, unquote, "file code kernels and support kernel optimization for, these, low level," and I don't mean, low level meaning menial, but, the AI engineers who, work really close to the metal. What'd you guys think about

Jim McGregor 17:28

The data scientists and stuff, yeah. Yeah. They're trying to up-level it, yeah.

Leonard Lee 17:32

Yeah.

Jim McGregor 17:32

I think that's a very important task. The more you automate, and the more... we've seen this even, I think, with agentic AI, just with, the OpenCL, movement, you know, people being able to generate their own types of agents and everything else, that is... And the whole shift towards open-weight models, I think that, we're seeing that with Kimi obviously, and other solutions. it is moving that way so quickly that, it's funny 'cause I actually talked to a couple developers recently, and they were saying, "Listen, we were doing everything as close to the chest as possible. But every, couple of months, NVIDIA and other companies, especially NVIDIA, had come out with some new feature, and we didn't have to do that anymore." Yeah. "So now it's to the point where, we're identifying what we need to do, and then we're just waiting, especially for companies like NVIDIA, to just come out with it so we can just jump on it, use it, and we don't have to worry about, wasting all the engineering resources and the time and everything else." and AI has proven to be there and, and- Yeah I don't know if you wanna get into that now. But just seeing what some of the new, models at OpenAI were able to do to break out of a system, and break into somebody else's system.

Leonard Lee 18:43

Well, let's hold that for later. Let's talk about that later, once we fini- yeah, that, which is like the craziness of this month. I mean, we thought that June was crazy. July ends up being even more, wacky, I, is the word that comes to mind. So I'm looking at it more the standpoint of how there's this talk track that's, or this notion that AMD is so behind on the, quote-unquote, "software side of stuff," especially, due to the CUDA Mo. Jim, you've already talked about how, over the years, NVIDIA has created this massive library of stuff. They have a whole, stack that extends all the way out to enterprise applications. And the question now is, given all this vibe coding stuff and the fact that, let's face it, AI is pretty damn good at, coding, right? And if you can use it to, do all the assembly l- language level stuff for kernel, programming and optimization, w-what are the, what's the possibility that, AMD can force multiply their dev- developer communities, or even internally their own software product teams, to very quickly scale out and, maybe not catch up in short term with, with, NVIDIA, but make traction in some critical areas that would make, AMD's- A- AI compute stack much more competitive,

Karl Freund 20:02

about a year ago, Dylan Patel, SemiAnalysis, just took out a sledgehammer and just crushed AMD because of the quality, or lack of quality, and functionality within ROCm. They said, "This is a piece of junk."

Leonard Lee 20:17

Yeah.

Karl Freund 20:17

And, but they didn't stop there. They actually sat down and worked with the, AMD engineers- Yeah to fix it and improve it, and they've come a long way. But now Dylan is saying, "Yeah, but you don't have any hardware to test it on. Your engineers keep telling me there's not enough, internal development platforms of scale to be able to- Yeah to test this stuff out at scale, and that's because they're shipping anything they can make, Yeah in, in this incredible environment. So it'll be s- interesting. I suspect they'll solve that problem. It's not a hard problem to solve. It's just- Yeah the matter of making some tough decisions.

Leonard Lee 20:50

Yeah, exactly.

Jim McGregor 20:50

but to your point, Leonard, we're still in the infancy of AI, and it's anybody's game. I mean, from- When are you gonna stop saying that? from tier models

Karl Freund 20:58

to

Jim McGregor 20:58

the software. It is anybody's game. I know.

Karl Freund 21:01

I agree. I agree. I don't know.

Jim McGregor 21:02

Okay. And, and, and I don't... I wouldn't count AMD out by any means.

Karl Freund 21:05

Definitely not.

Leonard Lee 21:06

I don't know if we're at the, beginning of AI. The thing is that we are where we are, and the ecosystem needs to figure out how to monetize this stuff. A huge gap, right? Yeah. It's becoming more and more obvious all the time. Well,

Jim McGregor 21:21

I think it's a gap for companies like Meta, but when you look at, Microsoft and- Yeah and, Google and everyone else, they seem to be monetizing it pretty well.

Leonard Lee 21:29

Pretty well. It depends. Yeah. But that's where we have to really, dive deeper into the numbers and it's still pretty obscure, right? They don't break things out in the way that would support an analysis on- Mm-hmm where is the AI

Jim McGregor 21:43

monetization coming from. That's the big problem the biggest problem right now is, token shock, and just the fact that-

Leonard Lee 21:49

Oh,

Jim McGregor 21:49

here we go you know, you, you start driving engineers, and but not just engineers, developers, anybody with the or- organization to start using AI, and even giving them budgets that they have to meet, or requi- Yeah or, levels they should be using. And all of a sudden you start finding how much this costs to- Yeah generate all these tokens and all this, d- data center, server time. It's getting very, very expensive. So it's pushing a lot of developers, a lot of organizations, A, to this open weight models, and B, to on-prem solutions like-

Karl Freund 22:21

Yeah

Jim McGregor 22:21

the small form fa- form factor workstations, the disk side workstations, or even on-prem servers Because it's getting so expensive to use some of these services now

Karl Freund 22:31

So, so Jim, where does that leave the foundation model, the proprietary models? What- what's your- Tell, what's your view on that?

Jim McGregor 22:38

well first off, there, there are some, some applications and some organizations They're almost required to use them. Mm-hmm. And when we start thinking about finance, and, healthcare- Government, healthcare and government, and stuff like this, you have to have a certain level of assurance that, it's gonna abide by the guardrails. It's, you're gonna see certain upgrades. It's gonna be fully tested and vetted before you get it. So I don't think that there's a limit to the growth for frontier models. But I, what I... And this is a generational thing, too. When you see startups, you see a lot more of the startups and newer developers, immediately going to those open-weight models, and start experimenting with them, and start using them very, very quickly. When you have those younger engineers and those younger companies, all kind of pushing towards that open source solution, it's gonna bifurcate the market. I still think you're gonna see growth at the frontier, but I would say by the end of the decade, that open source solution, if it's not at least 50% of the market or more, I'd be surprised.

Leonard Lee 23:39

the biggest challenge for the foundation guys is diffusion, and, we don't hear Google a lot in this... So let's talk about this whole open, model, debate that's going on. We had Jensen lead a, a cadre of tech companies to, sign a open letter to the government advocating for open-weight models, right? Yes. And full disclosure for N- Nvidia, they have their own, which is NeMo, their whole NeMo- Mm-hmm Tron family of open, models. And then, and then w- subsequently, after the, OpenAI rogue agent incident, then you had Nvidia again, once again, announce that they're working with, I think it's around 37 companies, to form what they call the Open Secure AI Alliance, which I think really should be called Secure AI Alliance, because, secure AI is agnostic of whether or not it's open or closed. But the whole idea is to be able to leverage, I think, open models and develop tools to secure, AI going forward. make it trustworthy, I think is one of the terms that they use, and then safe. so it was like a bunch of mishmash of feel-good terms, but I think the general gist of what everyone's trying to react to is, the security threat, that, now that Mythos moment has shined a light on. And so my view is that everyone's reacting to that and trying to figure out how to capitalize on that, revelation. But, yeah, these, So anyways, let's talk about open models and, this is my take. Diffusion is going to be the biggest challenge. Google less so because they have the Gemma stuff, okay? They're on endpoint. They're already, nicely diffusing out of the data center. They have a footprint. the likes of OpenAI, Anthropic, I think both of you will agree, not so much, and so they're on the back foot there. And the challenge with the open models is that you're gonna probably see these large capable models then be distilled and, compacted down into much smaller capable application-specific models. There might be some fine-tuning that happens, yada, yada, guardrails applied, boom, this thing is capable for that use case. And so you'll see diffusion happen outside of these, large data centers for, because of economics that you mentioned, Jim. So that's sort of the dynamic I think is gonna play out here with the open models and why it actually is a pretty big threat for, frontier labs, especially the US ones. So, hey, you guys don't have to agree with anything I say. Let's make this exciting, backyard wrestling style.

Jim McGregor 26:40

Come on. I don't have that slap feature yet. We still need the slap feature on this.

Leonard Lee 26:43

We just go like that. Where is that? And I'll put the slap sound effect in. Come on. Come on, Jim. Come on, Carl. What do you guys think? No? Look at this guy. Come on.

Karl Freund 26:56

I think it's... you're likely right. I'm not sure what the implications will be, to the overall industry, and I focus mostly on the hardware side.

Jim McGregor 27:07

Yeah.

Karl Freund 27:07

Yeah. So I, I think the hardware layer, it doesn't really have to worry about it. I think, the businesses running stuff and, securing the monsters we've created, that's a totally different ball of wax.

Jim McGregor 27:23

I think the agents take care of a lot of that diffusion, so I don't see that as being a major hurdle at this point in time. You know, quite honestly, I, I expect it to be part of the orchestration. So... And let's face it, I don't know of a single organization that's using a single model, because they're using- Yeah whatever model or whatever tools- Yeah make sense for that particular reason- Yeah in that particular application. And now we're introducing the agents to do a lot of the human tasks that we'd normally do. So the agent is making that decision, and the agent can do it much quicker, much faster, and can handle that diffusion, from different levels of the hierarchy, the hardware and software hierarchy. So I don't see that as... I understand you're concerned about it, but I really think that is something that is organically being, addressed very quickly.

Leonard Lee 28:12

Yeah. I'm not concerned about it. It is what it is for me. I'm just, I'm just, stating a... I, I'm sharing a view. And I think, I think security is a, a big challenge. We can't be, we can't ignore the fact that really the core issue here and the thing that is really creating the talk, is shaping the talking points is the security concern. Whether it's concern that it's a Chinese model, whatever, however that, affects the perception of security and trust. what we're seeing with the Mythos moment is now, a lot of, environments, including the hyperscale environments, are massive, potential, environments for zero day vulnerabilities.

Jim McGregor 28:59

Well, and- And assets ones that we didn't even know about.

Leonard Lee 29:02

Right.

Jim McGregor 29:02

across the platform, the fact that we have- Across

Leonard Lee 29:05

everything

Jim McGregor 29:06

agents that- Not just- can find zero-day vulnerability, vulnerabilities across platforms, across software solutions, across the... it- quite honestly, I think that AI will actually enhance security even more than the threat. Because think about it- I don't know No, I think it will. I think it will. Just the ability of AI to react to threats and react to people, acting, agents- Yeah or people or whoever acting in ways they're not supposed to act, they can do it much- Yeah quicker, much faster, and they can adapt to it. I think the issue is the fact that AI wasn't in that role of handling the security threats. I think that's an evolution that our industry's going through. But I think as AI takes on that responsibility. Even think about it like today, one of the big issues is the fact that you can pretty much see any brand-new movie, just by pirating it. Somebody's pirating it somewhere around the world. These sites go up, and it takes six months to take them down. When AI gets involved, those things will be down in minutes or seconds.

Leonard Lee 30:06

Mm-hmm,

Jim McGregor 30:06

so no, I think that, once AI gets... Matter of fact, I'm almost worried about AI going too far with security and not letting us do some of the stuff that we may wanna do. So I think that there's a delicate balance there.

Leonard Lee 30:18

and that, that's like what happened in the OpenAI situation where GitHub had to resort to a Chinese open-weight model, because, Mythos was locked down, right?

Jim McGregor 30:29

Yeah.

Leonard Lee 30:29

had a guardrail that prevented them for, doing the analysis that they needed, the forensic analysis they needed done. So, yeah, I subscribe to the challenge that the cybersecurity industry has always had, which is it's an asymmetrical fight. And I, in my research on the cybersecurity side and, safe AI, the threat actors are, have the advantage, and that should bother everyone. And the AI ability-

Jim McGregor 30:55

I think they have the advantage now, but I think that's going to change. when you have AI agents that are the ones that are actually doing the threats, that are actually the threat, the main threat that you have to deal with today, that's one thing, especially when you don't have that same level of, AI on the security side. But that's going to change very, very rapidly.

Leonard Lee 31:17

Mm. Yeah, yeah, and I'm keeping track of that. We'll

Jim McGregor 31:19

see. I mean, it's almost like a, a BattleBots competition between AIs, and that's what it's gonna be like.

Leonard Lee 31:25

Yeah, it's just one side has bigger claws and a bigger head.

Jim McGregor 31:28

Sometimes it's a lot easier to be defensive than offensive.

Leonard Lee 31:32

okay. So we'll leave it at that. That's a good debate to continue to have, and, I'm sure it'll keep coming back. The security issue is huge.

Jim McGregor 31:39

Oh, it w- it will, but- Yeah to raise it to a level where it's going to stop innovation or slow innovation would be foolish.

Leonard Lee 31:47

Yeah, yeah. No, I totally agree with you. No, it's an urgent matter. but- Mm let's... I m- Carl, let's talk about really quickly this whole gigawatt thingy. I keep hearing of it used as, as, "Hey, we have a, XYZ gigawatts of AI, AI capacity." Okay, so Yeah I'm missing, mincing words

Karl Freund 32:04

here. I think it's, I really think it's simpler than what is concerning you. You only have so many gigawatts that can come into a building, and you will do a certain amount of AI on that depending on how you architect it, and which chips you acquire, and, how you build it and manage it. But you gotta have something besides TOPS. Remember TOPS?

Leonard Lee 32:27

Yeah.

Karl Freund 32:28

Right? A terribly overused performance statistic. It's meaningless. So gigawatts is at least how much power you're gonna consume. What you do with it-

Jim McGregor 32:38

Well,

Karl Freund 32:39

well- is, is gonna determine how much AI it produces.

Jim McGregor 32:41

Yeah,

Leonard Lee 32:41

yeah.

Jim McGregor 32:42

These really are factories. They're taking data, and they're taking power- Power and they're creating information. They're creating tokens. So what's the biggest limiter? It's not the data, it's the power. So that's why we're using gigawatts- Right is because that is the biggest limiting factor that we have.

Leonard Lee 33:00

is that really a metric for AI? Because the problem is this is y- you guys are using that terminology and that mind frame down closer to the data center and infrastructure level. That's now being digested above the stack as entirely something else. Well, the, the, the- Look at gigawatts in terms of, a measure of AI capacity, whatever the hell we mean by AI.

Jim McGregor 33:27

Well,

Leonard Lee 33:27

well, well

Jim McGregor 33:28

first off, you know,

Leonard Lee 33:29

realize that- 'Cause remember, everyone's trying to figure out how to calculate and attribute all the spend to Economic benefit, right?

Jim McGregor 33:38

And the same time people are trying to future-proof their solutions, and it's almost impossible- Yeah to plan for because you have models are increasing by 10X each generation. So you can do- Yeah 10X more, or you can do the same pr- or, or better with a model that's, a 10th of the size with each generation. Mm-hmm. at the same time, we have hardware performance increasing by 30X or more with each generation. so those are phenomenal numbers just to show, how much more efficient we're getting at creating tokens. However, and it's making it more cost-effective. However, how does that plan into what you actually can do in the data center? you still c- only have so much power coming in, but to your point, the power is not gonna determine what you can do or what the token output is. Right. The token output's still gonna be a combination of those factors of what hardware you have, what models you're running, and what, oh, and what generation you're

Karl Freund 34:32

I think the disagreement we may have here, Leonard, is that I don't think you can apply gigawatts power as a metric of how much AI you can produce. It's simply the input. it's just the input. It's not the output.

Jim McGregor 34:44

limiting

Karl Freund 34:45

factor. It's not an output metric. Yeah.

Jim McGregor 34:46

it's the limiter. It's the upper limit of, okay, you've got this much power. What can you do with it? That's it. Yeah that, it's one factor in the equation. Matter of fact, if you really wanna look at the performance of a token, there's, like, six different factors that, factor into that. this is only one of them.

Leonard Lee 35:02

Yeah. Actually, I don't know if, I don't think that there's that much of a disagreement. It's more of, do people know what you just

Jim McGregor 35:12

described I actually have a slide on that

Leonard Lee 35:15

Uh, no, that, g- 'cause think about it. The different audiences that are consuming this information, making decisions on informations related to this metric, do they understand everything that you-

Jim McGregor 35:27

Well-

Leonard Lee 35:27

Yeah, because right now-

Jim McGregor 35:28

It, it is difficult. One of the biggest questions we get is, "How do I future-proof? How do, how do I build a system that I'm gonna be able to use for generations, and how do I plan for that?" And It's almost impossible. It's like, okay, each genera- you have to understand that the software's changing, the hardware's changing. The only factor you have is what you have in terms of physical resources, and that's it 'cause the rest of the AI equation is gonna change constantly.

Leonard Lee 35:54

So to our audience, earmark this section of the podcast This is really important because I get into so many conversations with people who are completely confused, and the, believe it or not, the gigawatt thingy just confuses them even more. And so this is good. I'm, and, I'm not making a, a- an argument here at this point, 'cause I thought I was. You guys are actually doing a c- a clarification based on what you're describing- Mm-hmm and reacting. Probably with the intent of, beating me up and slapping me, you have done the world a favor.

Jim McGregor 36:29

No, and what y- and that is a challenge when you look and say, "Okay,, Meta's in, building a brand-new gigawatt factory. What does that really mean?" You know? Right. That, th- that, that doesn't translate into, their capacity, for tokens, and it doesn't translate into ROI until you actually get that up and running.

Leonard Lee 36:48

yeah, definitely. So no, this is good stuff. And then, hey, we had a little,

Jim McGregor 36:52

we have to talk about quantum because there was- Okay, okay a major, major, major- Okay quantum announcement.

Leonard Lee 36:59

Bring it on.

Jim McGregor 36:59

That was that IBM, who is one of the leaders in quantum computing, and especially in supercomput- supercomputing qubits- Sure actually acquired HRL Laboratories, which is jointly owned by GM and Boeing, which is developing a number of quantum technologies, particularly silicon spin bits. Mm. So now, and this is a technology that Intel and others have invested in over the past, decade or so. So now IBM now has two quantum qubit technologies that they're working on, as well as additional stuff like quantum sensing, quantum materials expertise, and cryogenic electronics and packaging technologies that HRL Labs brings into, IBM. So this is a huge acquisition for IBM, and, and quite honestly, I think lifts up the potential for, the quantum spin bits even more. Would you agree, Carl?

Karl Freund 37:52

I've been wondering what IBM's gonna do because- Mm superconducting is really amazing. It's really fast, but it's really expensive. And you've got things like, neutral atoms and, trapped ions and quantum spin bits that are much more cost-effective at producing quantum calculations. So it'll be interesting to see how this informs their future roadmap. They've got a solid roadmap. They're spending billions of dollars on it, and they will get to a superconducting sup- quantum computer by the end of this decade that is, both scalable and fault tolerant. Hmm. and so that's the roadmap. Now, how this acquisition changes that roadmap, they're not saying yet if it does or whatever. Personally, I'm more intrigued of late with, with neutral atoms. it's not as fast, but it's much more quiet. It's not as susceptible to the environmental changes- Yeah to the state of a qubit. And so- Durable, you

Leonard Lee 39:03

mean? Is it-

Karl Freund 39:04

it's more durable. It's got lo- longer, states in which the superposition can be maintained. Not, not in milliseconds, but in multiple seconds you can maintain the su- the superposition of a qubit, and that determines, how long that qubit is actively computing. So, uh, I don't know. The good news is that all of these, qubit me- methods, med, will produce quantum re- computing results by the end of this decade that will change everything. We haven't done a good enough job of protecting the world's data from the attack of quantum, busting of all those crypto keys. So, there's this, save it now. Eventually I'll have enough computing power- Purpose,

Leonard Lee 39:52

yeah to

Karl Freund 39:53

break it. and when that happens, I've seen estimates as high as 3% of the world's GDP could be wiped out in a matter

Jim McGregor 40:05

of- Yeah, companies have to be ready in a matter of a quarter. They have to be post-quantum, let's say

Karl Freund 40:08

We've got the quant- post-quantum algorithms, and IBM's worked- Yeah with NIST to create standardize on these quantum-proof algorithms. But I don't think... Everybody's too busy chasing AI, and they're spending too much, all this money on chasing AI instead of protecting their data. And, they're, the world's gonna be in for a nasty shock, that is beyond the scope of anything we've ever seen before if- Right if all of this saved data, proprietary encrypted data suddenly becomes open for harvesting.

Leonard Lee 40:43

Mm-hmm. Yeah. Interesting. Yeah. And so Jim, Carl, do you see this acquisition as being complimentary or more of a diversification approach? And by the way, to the audience, we did a... These two gentlemen presented on, the various, quantum, approaches. We have a separate podcast episode for that, so definitely check that out. It's a great education by the way. And, anyways, but, maybe you could, help us understand, is this complimentary, diversification on IBM's part?

Jim McGregor 41:11

Well, it's still using standard silicon technology, so you're not using lasers or optics- Lasers or anything or anything like that. So it is complimentary in that standpoint, and the fact that it's something that IBM knows very well, and that is silicon technology. So I think it is, and I think that, like Carl said, everyone's realizing that there's not gonna be one way to skin the cat. You have to look at- No different quantum technologies. And some are gonna scale better than others, some are going to be more cost-effective than others. but I, I, I think having multiple... Matter of fact, Jason Wong said this when he was talking at, their quantum day that, I would recommend that companies go out and invest in at least three quantum technologies because, there are probably multiple ones are going to exist, co-exist simultaneously. and what we've seen, especially with IBM, is IBM invests not just in the silicon, but they invest in the system and- And the ecosystem the environment and the ecosystem. Yeah. So I think that's- I mean, what- bringing a lot of that together- Yeah 'cause some of these people that have been investing in some of these technologies, even like, the sil- the silicon spin bits, even though they've been investing for a decade, they haven't put as much into it as, others have, especially, IBM. So I think this brings a lot of investment to it.

Leonard Lee 42:21

Interesting. So what I wanna ask both of you this question really quickly. Intel, finally at least publicly committing to 14A. What'd you guys think? No big deal?

Jim McGregor 42:32

Well, no. I- About

Leonard Lee 42:32

time?

Jim McGregor 42:33

I've always said that they're committed to 14A- Yeah because their products are planning on running on 14A, and it's a step function for them from 18A using the same equipment and everything else. So I've always said that they're committed even if they didn't say it publicly. so I don't see that as a big thing. What I see is the biggest thing is the fact that they're now at, I believe, 0.9 rev of the PDK for 14A. Which mean, which tells me that they've got it pretty much locked down with a couple of lead customers, and that's very important.

Leonard Lee 43:03

Yeah.

Jim McGregor 43:03

Cool.

Karl Freund 43:04

Yeah. I would agree. I'm not surprised they made their commitment public. It was- Poorly kept secret, right?

Jim McGregor 43:11

Yeah.

Karl Freund 43:11

I think Jim's right. I think the impending, usability of 14A is what's really important.

Leonard Lee 43:16

Yeah. A year, what was it? A year and a half ago we were at, And Intel, what was it? Locateo Vision? Yeah. That didn't look like that was gonna happen when, Lie Puten took the helm. That was like, what? Two weeks in. So he was kinda- Yeah squishy about 14A. So anyway, I, you know, it, yeah, it, it's good to a- at least hear that public acknowledgement. So- Yeah.

Jim McGregor 43:38

Let, let's face it- it's great intel for- You don't spend billions of dollars on fabs and equipment to not use it, so, it's going to get used. And it's- Yeah being used right now for 18A, it's gonna be used for 14A.

Leonard Lee 43:50

Awesome. Well, there you have it from Jim McGregor, from Clan McGregor, and of Tirias Research. How did I do? Pretty good? Did I- That was pretty good pass the That was very good. Yeah, pretty good. Not bad you're getting better. Yeah, next week I wanna try German. How's that? I'm gonna work on my- Okay, yeah Arnold Schwarzenegger accent. It's,

Karl Freund 44:10

it's fantastic.

Leonard Lee 44:11

V-

Karl Freund 44:11

vindobiss, yeah?

Leonard Lee 44:12

Ja. I,

Jim McGregor 44:13

I have German in my background too. Ja,

Leonard Lee 44:14

ich.

Jim McGregor 44:15

I've got German, Irish, and Scottish- Yeah which basically means I'm good at two things, fighting and drinking, and not necessarily in that order.

Leonard Lee 44:22

Yeah, and then, in future episodes I'm gonna force you guys to, pronounce my Korean name, which I'm not gonna tell you. But you'll butcher it massively. So, hey, gentlemen, it's always so great to spend time with both of you. Thank you so much for your insights and sharing it with our audience. And so everyone, reach out to Carl. He's at www.cambrian-ai.com, and he's also on Substack and Forbes. Follow his insights. he's, been around the block for a long time, and now he's going from HPC to AI to quantum, and along with him is going to be, our good buddy, Jim, of, Tirias Research. Follow him at www.tirasresearch.com. And of course, subscribe to our podcast.

Jim McGregor 45:13

We also have Substack, we also have EE Times- Yeah and we have Forbes, so look for us pretty much everywhere.

Leonard Lee 45:18

of stuff. Yeah. Yeah. EE Times,

Karl Freund 45:20

In fact, I would just, if I could, put a shameless- Yeah shameless, s- Do

Leonard Lee 45:23

it

Karl Freund 45:24

Just advise everybody, if you can possibly make it to the AI Infrastructure Summit, AI Infra, it will be a huge event this year. Every year it almost doubles, so be there. Yeah. I'll be there. And Kevin

Jim McGregor 45:38

Hyneman from Tirias will be there.

Karl Freund 45:40

So Kevin and I would love to, love to meet you and talk with you at the

Leonard Lee 45:43

Yeah, it's kinda like Moore's Law.

Jim McGregor 45:45

And look for me at the, FMS Summit next week

Leonard Lee 45:48

Yes, and we're on the-

Jim McGregor 45:49

Future of Memory and Storage Summit

Leonard Lee 45:51

Ah, yeah, that's right. That's right. Okay, cool. A lot of exciting stuff happening and a lot of exciting coverage that we're gonna bring to you in the August edition of our podcast. So, remember, follow us on, YouTube, subscribe, check out the audio version on Buzzsprout, and you will be able to listen to us on your favorite podcast, platform. Also, subscribe to Next Curve Research Portal at www.next-curve.com, for the tech and industry insights that matter. Gentlemen, thank you so much once again, and thanks everyone for listening. We'll see you next time. Bye. Cheers.

Karl Freund 46:25

Bye.