133. Remembering David Floyer, From AI Models to AI Systems: The Next Technology Race

Hello, welcome to the Cubot episode 133. I'm John F with Dave Volante. Dave, summer salute. >> Hello, John. How you doing, buddy? >> Two weeks, man. We had tough calendars. We had came back from the AMD. You were flying home. We did a little rerun of our breaking analysis, but it's the dog days of summer on the cube here in the NYSC cube wired studio. You got the boss, we got the PaloAlto. The the the AI market's booming. The IPO windows looking really strong. maybe even to next year. Um the Red Sox are on a winning streak. They're looking like a playing a team that's that's like so loose. They had that big winning streak and uh everyone's glued to the TVs and my family. The tech streams are booming. >> Did you see that game last night? They came back from four deficits. They were down four zip and then they came back three other times and then they won in 13 innings. They were out of pitchers. They were going to have to bring in, you know, the right fielder to pitch. Crazy. It's it's great summer and I love the team mojo. This is not that you know it's just all team effort. Great work. Um but I mean two weeks I mean the world is turning fast. I mean I got to say it feels like a a big buzz here. Feel like I'm drinking from the the fire hose. We had AMD we came off that backend studio and events. Um Black Hat popped up. I was doing a video this week. Black Hat's coming up. It's actually started today. That was the first day of Black Hat. Cube was there. Um, again, black hats booming into a very relevant commercial show. Not that it wasn't relevant, it was always kind of the counter to RSA. RSA being the big bisdev show, marketing show, announcements, here's what we're doing, strategy, some tech speeds and feeds, maybe some deep dives here and there, great topics, sessions. Black Hat was always like the community, bring the whole team, get down and dirty, roll up the sleeves, talk about the core issues. Dave, this year with all the hacks and you had the open AI smugging face thing a couple months ago that was that was top of line. You're seeing all the AI agents essentially almost unsecure but the demand for agents is coming on so strong that it was an imperative for this industry to like step up and we saw tons of content coming from black hat. It's almost as if security is having that AI infrastructure moment that you and I were tracking like four years ago when we started to see the density of the servers. We called it largecale clusters and Charlie Kawaz at Broadcom's like you guys are on to something here. We interviewed at MWC and now of course Jensen called it AI factories that was all about density the role of the components around the GPUs whole another kind of operating model that's now rack scale. Michael Dell showing that on his in on his Instagram and Twitter that same kind of thing feels like it's happening in security but not kind of in the density way but in the importance of the reassembly of data and the role of cyber as really a built-in systematic play. It doesn't feel like best of breed tool just another platform. It feels like it's really coming together like it did in the AI infrastruure but in the security way cuz agents are playing a lot of value offense versus defense the rise of defense tech robotics I mean so much is going on to the surface area that it seems like cyber security is changed I mean what's your what's your takeaway because you're seeing it from the the the data side too what's your takeaway from black hat just in general am I'm seeing it right or feeling it right. It's more feeling >> the this I was watching a bunch of the videos that Christa was doing. Christa Casease um and and John Olsk, our team on the ground there. It it was like a your hair is on fire narrative at Black Hat. You know, the sky is falling. And um I I think the reality is well some of that's true because of the things that you mentioned these models going rogue. You know the reality is like it's the same old sort of basics but it's just happening at hypers speed like they've never seen before and it's a big distraction for a lot of organizations. But it comes down to first of all a AI is not going to take your job away if you're in cyber security. If like you know the technology and you know AI your your job is just fine. So, but the technology is advancing at such a rapid pace. I think the the the the whole people process technology thing is is crucial and you know the from a process standpoint the the word out of black hat is yeah there's some new malware that's being created out of AI but most of it is the the the attackers doing what the defenders are doing. They're trying to automate as much as possible and they're doing things at much much higher scale than ever seen before. And I think the other thing I would say is this has definitely been a distraction for practitioners. They're they're trying to fund AI. Uh they don't really have a a self-funding mechanism yet. So they're having to steal from other initiatives. And you're seeing that you're seeing a budget shift from from you know traditional legacy IT. And we've certainly seen it in in to a certain extent in SAS. Um and you're seeing some new money come in. But in general, I think what's happening is it's just going on at such an accelerated pace and it's distracting people from sort of what they have to get done, their main priorities. And we saw that in IBM's earnings and you know, I think it's it's a real challenge for practitioners. But at the end of the day, it's it comes back to basics. And those organizations that are, you know, using the best technology, following the best process, a have the best people, I think they're going to be okay. I think some of the narrative is just driven by the hype of the security vendors. They love chaos, John. Chaos means cash. >> Yeah. The I saw that Shiron, who's ex Intel Capital, quoted Andy Grove. He actually used the line. We've been saying on the cube I made a note on link security is so important because remember the conversations around the perimeter is dead and kind of went to that whole kind of zero trust endpoint protection but if you think about the the the new surface area including robotics um defense systems and the conversation around AI brings to like ethics who oversees it um they're taking shortcuts we see the hugging face open AI thing there's a lot of discussion that it's moving from a point solution to a systems game right so and that's why I was trying to make that comparison with the AI infrastructure because it moved from IT you know rack and stack to the systems and then just recently the AI infrastructure moved from the model race to the system race right so it's no longer the model and I think you and I were talking the other day it's like Nvidia doesn't view themselves as a GPU company they're an AI infrastructure company so it feels like the AI infrastructure is bringing a systems kind of thinking to the agentic worlds. that's all the cyber security around it and all those other things have to kind of compon be componentized around it and and that's kind of my big takeaway and I'm seeing this play out in the neo clouds too because coreweave's out there Nebas is out there the hyperscalers are big customers but these specialty clouds or neoclouds Dave if if you told me in 2016 that Amazon would have a competitor I would have said no freaking way the barrier to entry is way too We saw HP die, try it. IBM tried to do a cloud, OpenStack try to do a cloud. And so with the the architectural shift, you got Coreweave, Nebius, NScale, and the list goes on. You have new players like Argentum coming in. So you s you know different approaches. Some are taking the I call it the iPhone approach vertically integrated because the workloads have to be managed across multiple clusters and scale. So you got to use compute for prefill and and handle the decode with the GPUs and vice versa. And so the Nscale acquisition from Nscale caught my attention as did Rafé systems in the interview we had in Peloto because you're starting to see those Kubernetes guys uh from the that cloud native that control plane surface areas actually being built into the networking in these rack scale systems. So, Caruso uh and scale these they're all building massive data centers that are rack scale rack scale next to rack scale and the workloads got to run in a multi-tenant environment and the only way to do that is control every piece of resource in the system. So that's one approach and you have the other approach like hey we're like an oil company here's the refinery what do you want and let the retail handle from like fireworks AI. So you start to see different approaches and they they're almost a hedge to each other, right? So if you're an investor in coreweave, you want to jump on the other side and have an investor investment in the other approach. So the capex numbers, the valuation numbers are unbelievable. So the AI infrastructure is booming. >> It's it's interesting and I don't think Wall Street's gotten it yet. I don't think the mainstream market has figured out that we're kind of going down that approach. >> Well, I I I want a couple things. is I just want to close out on security. The other big theme out of uh out of black hat that I picked up was getting identity at runtime because you know human identity and you know people have done sort of modeling human identity in the past and now they got to they got to model agent identity. Um and then to your other point about the complex of these large organizations, you've got to prioritize, right? You got to figure out, okay, which systems are driving revenue or are missionritical. And then to your other point about IoT, a lot of these large organizations, they're going to be far-flung and they're going to have exposures, you know, out at the edge and and maybe in physical infrastructure. And that is just, you know, a whole new ball game in terms of the speed at which these, you know, AI models can go after it. Um, one of the things that I want to talk to you about is, you know, just sort of the AI boom and you're saying that, you know, Wall Street's kind of confused. One of the things I'm looking at very closely, I went back, you know, kind of in in sort of in an homage to David Floyer. I pulled out he he left the he was working on a pile of stuff, just some great work that we never published. And so I've been going through that and I went back to our 2004 semiconductor forecast where we we we projected that the market would hit nearly a trillion by 2028. Well, guess what? We're we're going to hit a trillion and a half this year. Okay? So, we were so conservative. But what's really changed? A couple of things changed is is back then it was really about all all about Nvidia and they were kind of driving the whole ecosystem. And while they still are, the bottlenecks have clearly shifted. Back then there was you couldn't get your hands on, you know, enough GPUs and you still can't. But the bottlenecks have sh are shifting to not only power, but HBM is what's driving that. So semiconductor revenues uh this year from from 2025 to 2026 are projected to grow almost 90%, John. And then next year they're projected to to to drop back down to about you know mid20s percent growth rate. Why is that? It's because the big chunk of growth this year is coming from HBM high bandwidth memory which has as you know the prices have increased like exponentially like five 600% in the last year. So I was looking at Micron's earnings. Check this out. So Micron's sequential so quarterto quartarter ASPs on HBM increased 60% mid60s this last quarter. The unit volume the bit volume only increased mid- single digits like 6%. So what that says is the the increase in revenue in semiconductors in the whole value chain is all about price increases and yes there's still demand but it totally changes the way you have to look at this market and forecast the market and then you throw in their power and packaging and fab capacity you know and all the other stuff. There's there's this whole new dynamic that comes into play and the forecasting becomes very interesting because you have to separate unit demand from pricing and prices are you know even though the cost per token is coming down prices in general are going up. Usually tech is deflationary. It's a really interesting and complicated environment for people right now. But you can model out and do some scenarios as to okay when when does the bubble pop or what are the risks to the bubble popping when when does liquidity dry up and and and I tell you it's going to a lot of it is going to come down to the winners are going to figure out how that bottleneck shifts whether it's from like I say packaging bandwidth building data centers you can have an empty data center and you can't light it up because you don't have power and so you don't have the the right silicon or that you can't get HBMs can't you can't get GPUs because you can't get the the rack because not enough HBMs. It's so it's a very complicated equation and I think to to the point I'm going to make that you triggered me on triggered in a good way is the Neoclouds are so good at this stuff and Nvidia is funding them and they're going to fill a lot of those gaps. Well, the uh SKH Highex plans a $ 38 billion chipmaking expansion South Korea uh including building a 24.7 billion DRAM facility and a 13 billion NAND fabrication plant. So, they're already gearing up. I talked to some Solidime folks which is doing some solid state uh last year at our GSA women's event we had from semiconductors. They Sandis was talking about high bandwidth flash that's released. their earnings were up dropped a little bit yesterday. But this is what I was trying to get at and I'm trying to like figure out a way to explain this but cyber security back to your point and why I'm tied in the infrastructure is because a lot of things we've been talking about agents it's the same music playing right so the discussion and I'm just reading some of my notes here the discussion that I've been tracking is this week with black hat you kind of pointed that it's not just AI assisted humans but it's completing work across enterprise systems so what's happening is like these clusters are being engineered up and and vertically integrate with the any scales and the rocky systems of the world. You're tying in hardware with software layer full control of the stack because you basically got multi-tenency and you need to remediate quickly. So you have to work across these systems. That's what why the AI factories are hot and that's why they're different because they got to support the AI. Now on the agent side with security they're working across systems, right? So autonomous workflows, multi-agent orchestration, um domain business pro process that could be deterministic, they all have the same system problem. If you don't, you can't just throw a security platform at it and get it right because these workloads uh and these agents that touching all the resource and that resource happens to be AI infrastructure by the way, services, memory, they're touching everything. So they're out in the wild. So the discussion of observability comes in. So there's a tie into this kind of systems game and I don't see a lot of other verticals that match it as well as security with infrastructure because the things that's going to power these agents will be the tokens, right? So you know how can I get the prefill and the decode done effectively? Well, if I don't have a lot of budget, I'm going to want I want to make sure that runs on the right piece on compute and then the right piece on either an XPU or GPU. Well, guess what? You got to bring back that together with software. That's what it like a orchestration operating system looks like. So these big neoclouds, some are leaning towards full integration because customers want to run their workloads on that multi-tenant but make it feel like it was a single tenant. That's cloud, right? That is what cloud is. Cloud is multi-tenency. But why doesn't cloud work for the enterprise? Because nobody wants to that. They want the single tenant security. They want to custom build a cluster and they run their workload on it. >> Wait, wait. What do you mean cloud doesn't work for the enterprise? >> If you have a workload, let's take the NYSE for they're a perfect example. They have their own data center. They have all this financial information. High security. Their entire clusters are built for workloads. So this one workload runs on that cluster. It's basically built for it. Nobody else touches it. It's a single tenant type situation. It's customuilt for the workload. The way these neoclouds and the way cloud computing works is everyone's in the cloud. You have multiple tenants and then you have to write the code to make sure your area works the way you want it to. Now cloud's a little bit easier with SAS, but when you're running complex domain specific enterprise data, you got to make sure that software stack is completely locked in, right? You can't just like use services that aren't going to be locked down from a security perspective. So you're seeing all the software abstractions from the Kubernetes world and things like slurm which is like it's just auler and resource management all these things are coming in. So these neoclouds want to get enterprise customers. They want them to just dial in like electricity. It's almost like a retail connection. But the enterprises want full security, full data observability and that's the play. >> But then okay so then how do you account for Google you know Google cloud growing at you know I had like 90%. Microsoft growing at what high 30s. Same with AW actually Microsoft like growing at like 40 plus percent. AWS grown I think 38% last quarter. So they're obviously somebody's buying this stuff. >> No, it's they're buying it because they're already in the cloud. So what's happening with the cloud players is that they're already there and they're doing the same exact thing. Now the neo clouds are more specialized with the whole GPU play. So they're back in the game. That was why I was saying you you asked me 2016 if that was if someone can be like a mini hyperscaler which they're trying to be right. >> They they are now in the cloud game. I mean, you look at Nvidia, they're supplying product to not just the big three or four, but four to 10 NeoClouds, right? So, you know, Jensen's essentially calling people up and telling them, "Buy now or you're going to be locked out." I mean, the GPUs are going like faster than hotcakes. So, there's going to be a supply chain shortage. So, you know, people are building. Again, it's crazy to think about it, but I think this is a shift, Dave. The architectural competition isn't about Nvidia and ARM. It's about what their the environments look like and you that was my takeaway from ARMs uh advancing AI event because >> they win too because comput >> inference >> is so it's not a it's not a war against Nvidia and ARM they're just the AI side of it. So if that's >> AMD, you keep saying ARM, >> I mean AMD, sorry, AMD, >> that's going to continue to build up, but the battle is going to be at the agentic layer when enterprises need to actually turn up and stand up workflows. What's their choice going to be? Buying a capex solution. >> Yeah. The market share, the debate about market share for AMD versus Nvidia. First of all, Nvidia is going to have the dominant market share, but it's to your point, it's irrelevant. It's all about the rest of the bottlenecks in the supply chain and the value chain. I mean, half of that if if the if the for the forecast says it's a billion and a trillion and a half this year, >> half of that is HBM is memory. It's like, you know, logic's only about 400 billion of that. So, yeah. So, it's just the market's so huge and there's so many other opportunities. The market share, everybody, of course, is going to stress about market share, but that's not the issue right now. The issue is how you build this stuff and the and the you know Jensen talks about this the supply chain is just so complicated and the bottlenecks are just everywhere and you know he says he gets ahead of it he flies to Korea and convinces them to build more HBM and goes to Micron and convinces them to b more HBM by the way your point about NAND NAND is exploding the prices of NAND are going up even faster than than than DRAM it's like crazy I mean it's not as big a market. It's probably a third, but yeah. So, the Neocloud's well positioned. I mean, at some point, demand and supply are going to become an imbalance. I was talking to Neocloud the other day. They say, "We're not worried about that right now. We're just worried about how to meet demand." Okay, so that's you got to go. But at some point, there's got to be a shakeout of the Neoclouds because, you know, demand and supply come together. And then then it's like, okay, how do we differentiate? How do you compete against the hypers scales? you know, what do you do that's different from corewave that's different from some of these, you know, from the hyperscalers? And that's when it gets to me really interesting. Um, and and the markets, I don't know. What do you think? I mean, there's like 50 Neoclouds now. >> So, it reminds me of disc drives back in the day. Remember the 80s? There were like 70 disc drive companies and all. >> I think I mean I mean it's going to come down to energy, right? Energy and and money, right? So, there's two equations. I wrote a post about this. Obviously the energy is the bounding function and you know companies like NScale just was following those guys they their strategy was target non-tier one areas for the energy Norway West Virginia um and I think there's going to be a very clever growth opportunity for whoever can lock down the energy so I think if the if 50 is the number if they can get their hands on a gigawatt in the port of port portfolio or 10 gawatts you're in you're a public company basically because you got to lay you can layer on cloud scale on top of it. So you got the, you know, facility side of it, design it, AI factories, do all the cooling. I mean, I interviewed entrepreneur this week. Their business is to eliminate the noise that comes out of the building that updates that's at a low frequency that changes the neighbors complaints. I mean, people are complaining that the noise from these AI factories is so big. >> Oh god. >> And they get nose bleeds and vertigo. And so >> yeah, >> got the high frequency stuff, the low frequency noise. Some entrepreneur in Michigan was working on material science and they built this nice little thing they put over the fan and takes the low energy out. So, you know, you're going to see the buildout, right? And so, and I think the AI agents are going to take advantage of that infrastructure out of the gate. But you know when I look at the interviews you're doing I'm doing what the cube's doing in the field you separate kind of the practitioner pioneers and uh transitional you know operational side of things the enterprise and cloud then you got the alpha geeks right the nerds and so right now the hot market is obviously agents but if you look at the the people we've been interviewing the where the alpha nerds are going going back to say the race summit it's robotics autonomous systems edge computing uh defense tech and intelligent hardware. Those are the areas you're starting to see kind of like the canary in the coal mine for the next wave coming because those the tech involved in those areas require smaller footprints, embedded systems, NAND done smaller but similar way as big AI factories. So right now everyone's talking about Caruso uh Nscale the big companies right on the AI side AWS Azure but when you get to to those other areas it's a technical problem Dave it's like >> it's going to be interesting it's going to be interesting to see what the silicon stack looks like there because I think risk 5 is going to and is going to do very well there I think open source Linux you know is going to do well but I wanted to make a comment about the energy because Jensen a while ago, I was at GTC or wherever it was, he was talking, maybe it was in a podcast, talking about how overprovisioned the grid is and that the hyperscalers all want, you know, the perfect SLA and the operators, the grid operators giving them, you know, high uptime, you know, 99.999% availability. Did you see this research from Duke University? Duke University's Nicholas Institute did a study and they said the existing US power grid could free up and absorb roughly 100 gigawwatts. Okay, this would be annually of extra capacity equal to about 10% of current peak demand because you don't need to be at peak demand all the time, but they provision for peak demand. They could do this without building new power plants just by allowing large flexible loads like data centers to curtail their use for a fraction of a percent of the year which you would imagine that they could find a time where they could you know reduce the SLA by a fraction of a percent. This would be a huge win for the industry. Well, I mean, I think that study kind of highlights where the that it's not really a mainstream topic, but it's a huge kind of me you mentioned bottlenecks earlier. Connecting to the grid is really potentially a bottleneck or disruptive problem. So, grid safe energy is a big discussion. I interviewed the founder of On Energy last week and he pointed out they're selling essentially a UPS uninterruptible power supply that they've engineered that does one purpose. connects to the grid and converts it into a steady stream of energy so that when it gets into the the AI factories, it's clean and solid. And the problem that they're solving is that the grid is unpredictable uh for the energy. So now inside the AI data centers and AI factories, Dave, they have UPS next to the machines because if some something happens, they want to protect that asset. outside of the data center to the grid. This is a huge issue because you know we know what you know rolling blackouts look like. So one core little weird issue that's happening this is pretty popular for the Carus of the world is I just want the energy to be clean coming in like just not clean energy like in the sense of you know clean energy was like steady energy. So you know if you're going to have gigawatts of electrical capabilities you got to have it. Now, some people are saying, I'll go be in the energy generation business that's vertically integrated inside the meter or outside the facility. That's where the grid's terrible. But, you know, transmission infrastructure and interconnects have a lot of issues that are out of control of the data center. So, you're starting to see that issue. So, I think this grid discussion is going to be a massive conversation because they weren't built for huge data centers, right? So they were built to run, you know, electricity for the homes for us citizens. So it's going to be very interesting how that plays into sovereignty, into quality of service and effectiveness and operational readiness and resilience for the data centers because think about it, if the grid's flaky, that's unpredictable, that could cause problems. That's what that's what we're seeing. So power reliability will be a thing. I think that's going to be an SLA thing to you what you pointed out. So yeah, I mean it's not not it's not a a problem for us as with homeowners, but if you're running a factory, >> well, what do you make of a solar and energy in space, right? Data centers in spa AI factories in space. You know, there was a point in time when people were saying that's you can't do that because of latency, blah blah blah. Now it's like they're saying, "Oh, Elon will figure it out. It's a fat complete." I don't know. I'm not qualified to answer. >> I love anything to do with space. So, I say, "Yeah, go." >> Well, the the the bull argument there would be, okay, look, it even what your projections you mentioned about the trillion dollars that's going to happen this year. It seemed ludicrous when you were doing those calculations with David. So, you know, even now you say, "Okay, I can't imagine what break fix looks like, but I've seen some demos where the robotics are doing all the space station work." If there's energy in space to be had and it continues to become a constraint, engineers will figure it out. The question will be what does that look like? We can't even probably imagine the congestion issue, reliability. But if you can launch for, you know, 50k a payload into space that has robotics in it, some of the robotics advances you're seeing, certainly in China outpacing the US, they're getting better every day. And we're only in the pregame of robotics. Forget humanoids, that's a whole another discussion. Folding my laundry. >> We're pre-training. [laughter] >> We're pregaming. Um but you know autonomous systems u that's why you know defense tech and you're seeing that in the war war fighters and commanders who makes the decision drones tactical edge. So you're starting to see these systems and I I could connect the dots of my mind and saying I can see a future where things are running in space autonomously. I mean look at SpaceX. No one thought they could do what they did over the if you went back 20 years ago and said they're going to land the the rockets back down on Earth on a pad. I mean, and the stuff that they're doing now. So, I think, you know, I think it's possible. The question is, what does it look like? How do you harvest it? Um, cyber security issues in space. Who owns the space? >> All that space junk we got to clean up, you know, we're uh we're we're serial polluters. >> I mean, it's like a it's it's a it literally is a Star Trek kind of moment, you know. So, >> hey, what about um this change of subjects? this deep mind shakeup. I have a sort of maybe a contrarian take on that. I think everybody's sort of freaking out. Google stock is down. They're like, "Oh no, Gemini's not at the frontier anymore." I actually think there's a there's a silver lining here, which is that's an upside for Google Cloud. It says to me that that Thomas Currion is a big winner of that because he's going to get more more TPU capacity and more GPU capacity. you know, if they're not that of course Google's not going to give up on, you know, training models, but if they're not going to try to compete for the frontier because it's a race to the bottom, that means that Currion can get more accelerator allocation because he's monetizing it. I mean, very clearly they're their GCP is growing at 90% plus and uh they're kicking ass. So that's a that's kind of a contrarian take and a and an upside for Google despite the stock being being down and getting crushed with the deep mind shakeup. >> You know, I think deep mine is one of those things where Google always had the edge in AI. I think looking at the model race, we've said it many times on on the cube. It's like a it's like a F1 or a NASCAR race. Cars are changing positions all the time. Bite dance just announced and they you claimed that they didn't distill anything but pre-trained AI models up to 10 trillion parameters three times larger than Kimmy K3. Um and then the mythos level kind of capabilities. So you know the model game is a leaprog game. So to me I don't really look at the Google thing. The thing about Google was the leader of Deep Mind. Um and Don Klein sent me a video on this this morning. I think you you copied on is that the guy who was in charge of this that's really he didn't want to do this like he was he's wanted to has a passion for somewhere else and you got Sergey Brin over there who's you know flexing and driving it. So it's not like Google's asleep at the switch on this one. I think Google is more of like in the we're going to leaprog next. So I would wait on C on Google. I wouldn't count them out. I think your point about the cloud game is legit because you look at AWS and Google Cloud and even Azure, right? But I would put Amazon and Google Cloud as my two favorites because they got great clouds and they got great marketplaces. I think the cloud game will be very very important because they're going to have leverage on the supply chain. They're going to have the compute power. So, if Google can just kind of sharpen the saw a little bit on the models, I don't see them getting out of the game because I think there'll be parody at some level on the frontier. And then, as we pointed out, Dave, what four years ago now, the power curve is going to look with a big fat neck and torso and a long tail. And I think everyone wins on our on our power law that we published, we took a lot of heat for, but if you go back and look at the premise of why we said that was we said there'll be a mixand match capability with specialty models. And by the way, the hottest topic is specialty intelligence like fireworks doing and you got general intelligence which the frontier models are doing. So general intelligence is a is a game of getting the the scale the big three will be there. The question of Google is do they give up or not? Now in the semiconductor business the joke is that Intel could have been in all these games. They just divested all their projects at the wrong time. Right? So it's one of these things. Do you do you keep on task or do you give up and maybe bite you in the butt later? Right. So, I would say I'd say Google should definitely stay in. Sergey should lean in. And >> if their leader didn't want to lean into it, >> let him go solve some societal problems. He's he wants to work on some pretty cool projects, moonshots, too, that are getable. >> So, Habis, I think that's how you pronounce his name. um the head of you know the founder of DeepMind they sold by the way they sold DeepMind to to Google for like $400 million but uh but any rate he didn't want to be running in an operating role he wanted to be you know getting this he won a Nobel Prize in chemistry the book on him and by the way I'm reading a book on him right now that was recommended [snorts] to me by VJ Kana but anyway the book on him was he was a kid genius he was really strong at at physics and neuroscience and he chose chose neuroscience because he found it more interesting. He won a Nobel Prize in chemistry. He lives in London. He from London. He doesn't want to be in Silicon Valley. So now this frees him up to do kind of pet projects and it put somebody in an operating mode. And the book on him was he wasn't great at monetizing. And I think I think that's what's going on in Google. They're like, "Look, do we try to keep up with the frontier open AI anthropic and spend all our cash and race to the bottom, or do we put our capex to where we can get fast monetization, which is where I I say I think Thomas Currion is going to get more allocation and they'll have more choice and they'll sell GPUs till the or accelerators, TPUs, GPUs, you name it, till the the cows come home." >> Yeah. I mean, you know, every time people go into so damage control, Google's doing it now. Remember when Google came out, they got hammered. Everyone gets hammered on these models because they they test some use case. Oh, it did this. It's got SWAT stickers on the images. All kinds of weird things that are not politically correct and they get called out for it. Um, I think those are gone on the frontier models. think that's shifting more towards agentic because the discussion on the agentic side is um they're making shortcut decisions because AI is lazy, right? Lazy and smart. What do smart people do? As my son Alec would say, they're lazy because they don't want to work as hard because they're smart. Um that's kind of like the agent model. So I don't I I think the frontier side is going to be reach a level of everyone's good enough and then the differentiation will come from specialism and then I think the the fusion of models. That's why I like the open weight debate because like you and I weighed in years ago I weighed in and said it's fusion game. Everyone's like jumping on that now. So I think everyone's realizing that you can mix and match models. You don't need the big general intelligence if you got a domain specific application. And by the way, if you have a domain specific application that doesn't require all the overhead of the general model, you can run a compute cluster um infrastructure cost that's cost- effective. So, we're getting into the old school IT game of remember the PC days, entrylevel PC, mid-range, high-end flagship, and everyone would buy that because all they were doing was spreadsheets. So I think AI we're going to start to see that same thing with the workloads and the and the models and the paro curves that you pointed out at GTC that we discussed was like okay you want the Vera Rubin tokens that's like that's high-end performance well guess what you want those on that's the policy of that say okay send the workloads the best stuff >> send the non-critical computation or reasoning you need or the small specialty model to this cluster We'll put on one/10enth the cost per token. I think >> yeah I you know it's it's interesting there's a frontier model backlash right now. I'm not as negative on frontier models as everybody else is. I I actually think you know they they may still be chasing AGI. I don't know maybe Google is too but to me the Frontier models I think they're going to to your point they're going to segment their markets. They're going to have N minus one versions. I'm I'm sure they're going to lean into open source. Maybe they don't you know open source their models or their I mean even though they have but you know deliver openweight models but I think they are going to compete across the entire software stack and I think I I think they're going to get a fair chunk of that because I think simplicity is going to be an advantage. Everybody's looking at it as a as a as a zero sum game and it's not. The market is so large it's going to exactly what you said. you're gonna have, you know, small, medium, large, very large, very small and I think the frontier models are going to do very well. Um, I think they're leading this charge and I think they do have to, you know, focus and prioritize. I think anthropic's doing that. I think I think Open AI is going to get a crapload of compute with with the next round. And I think they've, you know, they've really optimized to get compute. Um they're they're working hopefully they're working on some of those other bottlenecks that I talked about because if it's just compute that's a problem but I I'm not I'm more sanguin on the frontier models than many people are right now. >> I am definitely bullish on frontier models. I think again the power law we talked about it's harder to affect change and compete with the frontier models on distribution. And I think and I've I've said this before I'll say it again. I think OpenAI and Czech GBT will win the consumer side. I think that's their winning swim lane. I think Anthropic wins the intelligent enterprise large scale systems and maybe there's some overlap between both and some of those use cases. But you know there's billions of people using chat GPT and they got the voice activation. So I think that's a great opportunity for them and that's the threat to Google. So if I'm Google I worry more about chat GPT than anthropic on the consumer side on the search. If I'm Google cloud I'm thinking I want to win the agent king. I got to compete with anthropic. So I think you're going to see Google bifurcate their thinking because they have to compete with both unlike AWS. They don't really have to compete with the consumer. They just got to be the enablement kind of like the success of fireworks AI we've interviewed in Paris. Um they're extremely focused on enabling people to do the specialty intelligence and let the general intelligence stay in the frontier models. But there's also other frontiers emerging besides the language models, Dave. is frontier in in computer vision. There's frontier positions in agentic. So the word frontier means forward flanking. So >> leading a cutting edge, right? You're right. It's not just models. It's it's silicon. It's fabs. It's packaging. I mean, you're absolutely right about that. >> I you think about >> you think about China. you know, China's not in the frontier of silicon, but they're they want to be. They they will be potentially eventually. And then, you know, maybe they're not in the frontier of models, but you you're talking about a 10 trillion parameter model, which is no surprise to us, right? Because Bite Dance, Tik Tok had the best AI algorithms in social media. It blew everybody away. So, it's no, it's absolutely no surprise that Bite Dance has great models. No surprise. don't want to count them the Chinese out on this cuz remember remember I think the big thing is who's got the distribution so if I'm bite dance and I'm approaching mythos scale already on a pre-trained model of 10 trillion parameters which means like it's like a shortcut to the top but remember they have distribution in China you got by do you got bite dance you have all these systems and billions of users touching it too so you know don't count out the Chinese the exact argument of chatbt we just made the case we made for chatbt and cloud apply to the Chinese model. So the question is who locks down their competitive differentiation on the distribution side that's integration that's usage. So I think that's going to be the big game. And then the second act of that is that as new AI native implementations come out Dave is going to be like okay what am I integrating AI into? What am I what intelligence am I injecting into my non-intelligent process business workload? And I think that's the second wave of action. And I think the agents are pointing out the fact that and the security problems with Agentic is pointing out to the fact that it's hard as hell to do that. And all by the way, you're going to need specialized intelligence. And that's why I think last time we were ripping about deterministic workloads and deterministic being a feature, not a bug in the enterprise because if you got an endto-end workload, you could go vertical stack. you can go end to end and run that deterministically and then use reasoning skills across systems. So that's an enterprise complexity problem. That's not a throw the model at it and say go at it. >> I I have a I have another take on this. So I agree with you. However, determinism and this is where the SAS vendors you workday, Service Now, Salesforce, Oracle, SAP have an advantage, right? They have that deterministic software. But determinism in some regards is illusory because you've got determinism within your own department. The supply chain, you know, and the [snorts] software that they rely on might be have determinism. The finance department has determinism. The HR department has determinism within their own world. But the tacet knowledge across the enterprise still is problematic. And this is where I think the partnerships between the LLM vendors, if they don't f it up, and the SAS vendors can be critical because you you you want to create a substrate across the organization that dissolves those stove pipes, those that fragmentation. And that's where the the cognitive layer, the intelligence comes in. And the frontier models can be that glue, but they need the determinism. So how do they get that? Do they try to reinvent it from scratch? Do they partner and steal the alpha? Do they do M&A? You know, once they do M&A, will other SAS vendors not work with them? So they have to play this very carefully. But the opportunity is a five to 10x productivity as measured by revenue per employee where you're you're dissolving those that fragmentation and you're bringing surfacing that tacet knowledge that tribal knowledge in the enterprise that democratizes it and that is the promise of AI and I think the frontier model vendors have a great opportunity to go after that. Yeah. And and by the way, I think open source complements the big frontier models. And I'd also add to your analysis there, totally agree, is that if you look at the success of the enterprise, it's the combination of the open weights and the frontiers will coexist. And you know, determinism for agents, you know, it's going to be based upon every unique situation. So if I'm an enterprise, I'm going to have certain pet peeves I'm going to need to nail down. So I just think it's a just different ballgame and the frontier models are going to either win. Now if I'm if I'm a frontier model, I'll ask you this. If you were and I were making a decision right now, I'd say what would what what would AWS do? What would Andy Jasse do if he was running a frontier model? Um he and because remember that's the same challenge AWS had. Do we compete with the ecosystem or do we let them have it and throw a competing product for the full stack of AWS? I mean, Snowflake would not exist without AWS, but they had a similar product. >> It gets to your distribution point, and that's what a AWS had, and that they they proved that you could coop compete and cooperate and still make a ton of money. It was a win-win because they had the customer. And so, it comes down to the the the distribution. Will Anthropic and Open AI have enough of a channel direct to the customer or through channels that they can affect an ecosystem where they can create a win-win where hey we can partner with you. Yeah, we're going to do some of that that functionality as well. Maybe it's not as deep but but we'll both win or andor we'll partner together. We'll bring the cognitive layer. you bring your determinism and we'll we'll monetize the outcome. That to me is the the the most likely scenario. And to your other point about routing models, unquestionably you why would you route a model to a to a or you know route a query to a model that's 10x more, you know, cost and tokens? You wouldn't. You'd go to that open- source model. So, you know, you you hear Jay Cal ranting about this, but it's again it's he talks like it's a zero- sum game. It's not. I I think I think there's places for for all of these. >> Well, I think the other thing I would add to think about remember how we used to have conversation around all that data Amazon had on all the usage. So, they also had intellectual property around how these that's why everyone's like saying, "Oh, they're going to take your alpha." Maybe not. I mean, everyone would early on in AWS ecosystem say they're going to be able to see everything. Of course, they could see everything. Now, they made that an input to their system to be better. Maybe they might have might have done one or two bad things. Bad product manager goes rogue. But you know, first principles, Amazon never ever really went after anybody, but they use the data at scale to make the product better. So I think there's that's the choice that I think the big models have to think about. It's like, okay, if I'm going to have an ecosystem play like Open AI, then you can't take someone's alpha. You got to give them more alpha back, right? So I think there's this there's a there's a mindset shift there. Plus, you think about I mean I Alex Garper, that rant was great, but okay, well, what is that that alpha? Is that what is that IP? Is it your data? Is it your process? It's it's all of it. And so my argument would be if I'm a company and I can get to market much faster working with frontier models and obviously I'm going to have a a a combination but if I can build an operating model and that operating model now becomes my alpha. So I even if the frontier models, you know, have access to some of that data and some of that that or all of it, if I'm moving faster than my competitors and I've found a foothold in a market, that alpha becomes my operating model. So they may not care if you're, you know, sucking their brain. I mean, big companies are going to care, obviously JP Morgan, but maybe that's how those big companies get disrupted. Maybe the startups say, "Hey, we can move fast enough. We don't care. We're small enough. we we can live on smaller margins for a period of time. I mean, imagine if you try to take away I mean, everybody's tried to take away Microsoft's alpha and Google's alpha in search and you know, they've tried to copy it and mimic it, but they can't because the operating model is just so much more effective. And so, I think you're going to see new forms of comp of of of competitive advantage emerge that maybe make that less concerning because they can move faster. Well, Dave, it's been great. I'd like to end this seg this podcast uh 133 as a dedication to David Floyer, our colleague, your friend for many decades in business. You wrote a great LinkedIn post. Um I have a little bit longer post. I'm trying to get out there because I wanted to get more sound bites, but I want to dedicate this podcast to David Flor, our car league that passed away two weeks ago from an illness, been with us forever. One of the best ever. Um, and uh, want to dedicate to that. And Dave, say a few words. I mean, I have some funny stories of, you know, chuckling away and also grinding away, pounding the fist on the table, uh, calling out people on their tech in front of their face. That won't work. Prove it to me. Um, say say a few words uh to David Floyer, which by the way, 133, our episode is kind of a magic number in spiritual circles. So, it's called the angel number. So, appropo for that. Dave, share a few words. Well, thank you, John, for bringing that up. I mean, you're right. David Floyer and I think met in 1993 94 and I brought him into to IDC and I knew right away that first of all, this was a wonderful human being who cared about people. Um, and he was just a a great technologist and forecaster and he became a great analyst. He wasn't an analyst when he left IBM and he made just so many great calls because he didn't get caught up in the headlines. He didn't get caught up in the hype. He just had a very methodical way of looking at markets. He loved data. He loved to ingest data and then build his own models and then reason through what was likely to happen. And he made over the years some fantastic calls. I remember I would I would get I would I would get calls when I was at IDC from folks who were saying David Floyer's absolutely wrong when he had made a call that this product will never see the light of day. It doesn't have the economics. I mean, he made one of his most famous calls was Intel. I mean, he basically said that if if Intel continues on this path, you know, that Pat Kelziger had him on, they'll go bankrupt. So, it's not going to happen. The board won't let that happen. And that was probably one of his most famous calls and many many others. Um, and he was also >> one call one call that I think was very notable was the rise of ARM. Okay. Oh, >> yeah. >> He nailed ARM. Also, hypercon converged. Okay. He was right on that one. He called rise of arm in the enterprise in 2012. I mean [laughter] that was a pretty good call. And um the other thing about David is he had he was a very deep individual. Um he was actually a an Olympian on the UK. He was on the alternate team uh for the for Bathlon. You know you cross country ski and you shoot. He was an alternate on that team. He was a junior chess champion. Um he was an excellent athlete. Uh, you know, he had many years on me, but he schooled me in squash. He'd have me running all over the place. >> He was an excellent coach. Uh, he coached Mountain View uh uh an elite Mountain View team in soccer and they won, you know, many a championship. Um, and his coaching, I'm sure, was was challenging but fair. And I can I can just hear him out out there. You're not done yet, lads. Be an option. You know, [laughter] >> I know he I can't see him yelling hard at people. He would definitely have an edge to him. >> Um, I mean, one of the things I liked about him was this he I make sense. I didn't know he was a chess player, but when I first started working with him when we started partnering, >> he had a mind that was curious almost like a puzzle like a chess player now that I think about it. But he also was a systems thinker. >> Okay. And he loved to connect the dots and put quantification to it. And he loved to guess what the next thing was going to be, but not in kind of a haymaker way. He actually had great thoughts and again he had many seinal moments. Those were the highlights. There was other little ones storage, right? He was talking about he made some great storage calls. Uh he made some great server architectural calls, the ARM thing and then Intel. He saw Nvidia. I mean he had the system game down in my opinion. One of the best ever that I've ever seen. So >> yeah, he sure did, John. He could he could see things like like you. He had an ability to see around corners. his methodology was different. You know, he would do it through, you know, building models and spreadsheets and and just thinking um you know, you have that talent as well. I think your methodology is different. You connect dots like in real time like brain synapses, but we're going to miss him. Um >> and uh and as as I shared to to many with many folks, he left a pile of unp unpublished research that I'm combing through. And uh this week's breaking analysis is a collaboration with David Floyer posumously. So we're going to miss him. God bless him and uh and his family. >> Well, Godspeed David Floyer. This episode's dedicated to you. Thanks everyone. We got big events coming. We got Crowdstrike, VMware Explorer, Dreamforce, Workea, Neo4j, Oracle, uh, OC, Octa, Octane, Core, Dell, IBM event. I mean, tech exchange, KubeCon, reinvent, supercomputing. The second half is loaded with action, of course. >> Enjoy August while it lasts. >> The firehouse. >> Oh, wow. It's crazy, >> isn't it? >> Thanks, Dave. See you later. >> Hey, thanks, John. Thanks everybody.