Exclusive Coverage: theCUBE at IBM Think | Holger Mueller | Constellation Research, Inc.

theCUBE.net

Clips
More from Exclusive Coverage: theCUBE at IBM Think

Holger Mueller

Principal Analyst & VP

Constellation Research

play_circle_outline The need for larger language models and more data

play_circle_outline Importance of data gravity and data availability in the cloud

play_circle_outline The significance of end-to-end workload understanding and optimization

play_circle_outline Overcoming On-Premise AI Infrastructure Challenges: Embracing Cloud Capabilities for Elasticity and Pay-as-You-Go

play_circle_outline Realizing the Significance: Data Latency, Centralization, and Foundation Model Architecture in AI

play_circle_outline The importance of solving for structured data in enterprise AI

play_circle_outline The role of partners and software licensing for IBM's success

Info

Holger Mueller | Constellation Research, Inc.

Holger Mueller

Principal Analyst & VP Constellation Research

search

>> Hello, welcome back to theCUBE's Coverage here in Boston Massachusetts. I'm John Furrier, host of theCUBE. IBM Think is happening. We are here at the pop-up CUBE. Holger Mueller is here, is CUBE alumni, CUBE contributor. Also, principal research as a constellation and a friend of theCUBE. Holger, great to see you. You finally made it back on theCUBE.

Holger Mueller

>> Yeah, I can't believe it, huh? You gave me 18 minutes notice. Am I the fastest visitor on theCUBE? The fastest guest on theCUBE?

>> You texted me before. I tried to get you on last time, but I couldn't make it work.

Holger Mueller

>> Yeah, I know, I know, I know.

>> You were flying out of Google Next. But a lot going on. I really appreciate your perspective. You were also recently at a variety of other events. Oracle had an analyst event. Google Next. The industry right now is going through a major platform re shift.

Holger Mueller

>> Yes.

>> When I say re shift, it's shifted, it's re shifting again. The tectonic plates are moving with generative AI, all layers of the stack.

Holger Mueller

>> Yes.

>> IBM's here. We had Dell has their event. What's your take on this? What are you seeing? Are you seeing the line forming between the old way and new way? Could you share your view?

Holger Mueller

>> Yeah, I think we saw this rebalancing, that on-premises was something which was getting alive again. I wrote about this six, seven years ago. I called the infinite compute platform, where you can move workloads seamlessly from the public cloud to on-premises. It has worked really well for some vendors. We had IBM, like OpenShift from Red Hat, allows you to run across multiple public clouds and on-premises even to the edge. That's exactly that. Enters with a boom, gen AI ,and all of a sudden it's all about elasticity again. Even can't get Nvidia machines for your on-premises workloads, right? So the balance always tipping back and forward. It's totally moving in towards the public cloud right now.

>> And that sounds like super cloud definition as far as I'm concerned.

Holger Mueller

>> Yeah, sorry.

>> Charles Fitzgerald would love to have a debate. Infinite cloud, what do you call it?

Holger Mueller

>> Infinite computing platforms.

>> Infinite computing platform. So we have the super cloud, it all points to the same thing. A new way to operate, which everybody wants, which is a first-party choice by the company to run their cloud operations, their infrastructure, their way, not the vendor's way. This has been a big multi-cloud problem, that you got three clouds. They all have different sets of the same services. You still need identity management, but it's different on all three. You need data, it's in all three clouds. How do you handle portability? All these things now are coming to the surface.

Holger Mueller

>> It's quick.

>> Now that we have a generative AI, the forcing function, what do you see happening right now? Obviously, do you see the gen AI market stalled a little bit on the app side? Is the action in the infrastructure? Are they retooling there? And then what happens next, after the infrastructure gets tooled up quickly, the middle layer going to get innovative. What's your-

Holger Mueller

>> What? ....

>> analysis?

Holger Mueller

>> Great question, but what gen AI does is basically collapses everything again. In my view, bigger is better when it comes to gen AI. If I know that the weather just changed from cloudy this morning in Boston to sun, I'm going to put on more ice cream in my freezers because I know people will eat ice cream, which otherwise they wouldn't. So bigger models have more awareness, are more human, will be better in business. And bigger models need more data. So data gravity, which we've taught when for a long time, is getting really, really again tangible, and that means it's going to move to certain clouds. You have to be widely, we said before, the multi cloud abilities, that made vendors across the whole stack. We can go back to 10 years ago, cloud foundry for pass. We had IBM. We can talk HashiCorp for DevOps. We can go to . It's at lots of events recently. They're user summit cross-clouds. Companies want to be cross-cloud, but if I want to do AI, and have to do a large language model, I need to get the data together.

>> And that's why the surge is on-premise, so the term AI factory, that Jensen long at Nvidia been talking about, he's promoting that, because he wants to reinvent the idea that you can create a data center on-premise, whether it's on a premise, or in a collo facility. But you can build a purpose-built set of infrastructure to service a workload.

Holger Mueller

>> Yes.

>> Do you buy that?

Holger Mueller

>> No.

>> Okay, good. Why?

Holger Mueller

>> Because Jensen was in this business, right? Five years ago, the key question of Nvidia was, can Nvidia survive, because none of the cloud vendors is using their GPUs. None of them was, five years ago. And only Microsoft's fumble of building their own custom silicon. Kevin Scott found OpenAI. And OpenAI could done for Jensen, got their first DGX machine in 2016. I don't know if you've seen the tweets on theCUBE, of , being at OpenAI at the time?

>> Yeah.

Holger Mueller

>> And the first DGX machine coming there, great seating. This dependency of OpenAI on Nvidia, has made Nvidia again a cloud company, the most valuable company in the world. So Jensen is talking about something which was very popular, out of necessity because nobody is doing this stuff. Right now, he can only sell to cloud vendors because of their premium.

>> Yeah.

Holger Mueller

>> But so very interesting. At Oracle's OCI Summit two weeks ago, they pre-pay in cash, they shared, to get infrastructure being built. Suppliers are being paid in cash, so they can roll out their data centers across the whole stack. We're not talking about Nvidia, but basically today, you know can only get the Nvidia machine if you're a friend of Jensen and Colin.

>> And so is that because they want to Bogart the machines, and hoard them in, and store them? Or is it just pointing to the use case, that the large language models and the large cloud players have a use case, called consumer business? Amazon's got retail, Meta has Facebook. Google's got search, and then they have clouds behind it. So you have all these super clouds, these big cloud, hyper clouds have retail businesses. That's a consumer version of AI. Would you see that? What's your reaction to that?

Holger Mueller

>> There is a big need on the consumer side, and a large part is fueled by the advertising wars. But if people don't Google something anymore. They ask generative AI and want a summary, which then has advertised links in them. That's going to happen. You will see that. And you just saw Google at Google IO last week, announcing that more and more of the questions you Google will be generative AI answer. That needs massive compute. But the awareness of what's happening in the real world, is not only relevant for consumers to sell advertising, that's relevant for everything. At what temperature do you want to run this hotel lobby that way? Right? What loudness do we want to put on our microphone reception here, right?

>> Yeah.

Holger Mueller

>> The understanding of the real world, with multimodal models, which is the big innovation of 2024, that models can understand more things, is the key aspect. And this is why the larger is better. If I don't have my weather, my noise information together, I cannot build AI for it.

>> Dave Velante put a poll on his Twitter about who's going to open up the AI. Clearly, 73% of respondents, mostly analysts, agreed that the hyperscalers will mop up the AI business, and ultimately, probably, give language models away for free. Your thoughts on that? And does that validate what you're just saying, is that the service providers for AI will be the monster cloud players?

Holger Mueller

>> Yeah, there's no way around it, because we know that the pushback to public cloud from the early days was, if I know my workload, if I can predict it, and it's stable, then I go to public cloud. Quick side, a comment inside, right? One of the largest, most valuable companies in the world. Apple didn't build a cloud, uses the clouds for backup, which is super stable, right? You can't take so many pictures to make this a volatile part. But even then it's being used. So the problem is, generative AI workloads in need of an enterprise cannot be assessed by any CIO, CTO today. You cannot go to your board and say, "I need 10 million to buy 20 DGX machines and we'll be fine for the next five years." There is no way of knowing that. And that uncertainty comes back to, if you boil everything down, systems from a quality, what is the key quality of the cloud? It's elasticity. Elasticity from a technical perspective, from up and down, and more importantly, from a commercial perspective. You use more, you pay more. You use less, you pay less. So this is why the cloud wins.

>> All right, so let's take that cloud benefit and then try to map that onto the AI world. So elasticity, that's a technical, and agility, that's a mechanism-based value proposition. The cost side is, pay by the drink. That's economics. So they have both the economic value proposition, as well as the technical. I'm oversimplifying it, but agility, that's why startups went there first. The president of data center, we all know that rule. Okay, fast-forward to the AI world now, what are the problems that people have? I don't have the hard way to run my own stuff? I don't have the expertise to run AI? And how the hell does this impact my business, both cost efficiency and maybe new revenue?

Holger Mueller

>> Right.

>> So, okay, what's their answer? They? Go where

Holger Mueller

>> This is why you need to move to the cloud, because apart from machines-

>> You're saying that's the reason to move to the cloud?

Holger Mueller

>> Yes. People who are not in the cloud need to move from the cloud. I need to move my data closer, to right now the DGX machines. Maybe in two, three years, hopefully there's more competition, to the Intel, AMD, whatever machines, to create the generative AI that I need to run and power my enterprise.

>> Got it. Okay, let ask you another question. So I buy that, by the way, I think that's a good outlook.

Holger Mueller

>> We agree on something?

>> Yeah, we agree. I like that we're going to debate this. Okay, here's the next thing, you said end-to-end workload. If you know your workload, you go to the cloud. A lot of companies have workloads that are well-defined. They have a back-end, with a software model, and a front-end application. It could be a banking app or whatever.

Holger Mueller

>> Yes, yes.

>> And so, it can scope that workload end to end. Now by adding generative AI into the equation, they say, great, I know my workload end to end. I can now build purpose-built hardware, and or workload to support the workloads. Whether that's for response time or reasoning. So question, the optimized infrastructure for the IO of generative AI, question answer, whether that's a prompt under the covers or directly, that's going to be low latency response. Or go away and think of a problem. Reason, this use case is to use different maybe potentially clouds. Will you go to a cloud on that? And or would you build it on-premise, if the data's proprietary?

Holger Mueller

>> No, first of all, data proprietary is important. No question. You live a knowledge economy, data is the most important asset. But I think there's no concern to, if your data is on premises is safer than it's in the public cloud, bearing any potential accidents. I think your past that decade in that discussion. The key things though is, do you want to, do you believe in larger being better? I'm a strong believer in that camp. If you only want to do classic IT, and says, "I have my banking application and I'm going to do some intelligence," there's some people out there and saying, "Oh, AI is just intelligence and analytics." I think it's much, much more because it goes inside to action, to the automation of transaction, to the automation of bots, to the automation of agents. But if you do small AI, you can do that on-premise and try to figure that out. But I think the small AI will always lose against the big AI, and the big AI needs cloud. And here comes the key thing. Even so, I know my transaction application really well, it's predictable. I got my more cost-efficient on premises. I need the data of that in the public cloud to run my AI models on that. What if I see fraud, a classic example, right? I had 120 data scientists working my IT, at FiCo, 15 years ago, 20 years ago. 15, 20 years ago, right? Trying to figure out fraud. We had to get tapes from the banks. It took us half a year. We were so much behind the fraud. Today you have to be faster. Your transaction systems on premise, you once a day dump it in the cloud to figure out fraud, you could have lost millions.

>> It's like a video game. If you lag, you're going to lose every time. You're going to get shot on the first person shooter games. Happens all the time.

Holger Mueller

>> Yes.

>> But the lag, you bring up a good point. Data quality and data low latency address-ability. Is it in the right place at the right time? So I want to ask you a question. Define big AI, because you said that a few times, when you say you need to clap for your big AI.

Holger Mueller

>> What?

>> Define, what's the difference big AI and small AI?

Holger Mueller

>> Right. So big AI is when you take all the data available in your enterprise, and from third parties, to build your models on that, which by definition, are very large language models. They're very similar to OpenAI's Bake the Internet, to Gemini 1.5 Bake the Internet, to Llama Bake the Internet, and so on. Maybe not as big, but very, very large. Versus you do on-premise, on-device potential AI, but there's use-case for that too. But it only uses that local data. And the question is if the small AI can make the right decisions?

>> That's awesome. So okay, let's get back into the power law. So you're starting to see specialized models, enterprises are, you can see the IQ of the enterprises. It's almost a shadow IT model, shadow AI. A lot of red, a lot of retrieval augmentation generation, even came up on today's keynote at IBM, on the partner summit here. It's an easy thing to do. If you have data, do some rag on it. It's small AI, it's not a large AI model, to your definition.

Holger Mueller

>> Yes.

>> So there's a lot of that going on. Is that an indicator of appetite or are people tire-kicking?

Holger Mueller

>> No, it's the art of the possible right now, and the possible is always limited by the vendor doing that. If you look at the vendors who have large well-managed models, Microsoft, Meta, Google, they give you the architecture. And what really matters is the foundation model architecture. How you're going to do that layer cake, from your very large, boil the ocean, have-the-all-internet language model, to the one which might give you an industry aspect, might give you a vendor aspect, might give you your customer data and your customer-specific stuff to be private and be safe from anybody else looking into that? But that's layer cake. So I don't see enough discussion about foundation model architectures, which is a big thing. And by the way, IBM was one of the first winners, as we hear, to talk about foundation model architecture last year and summer, which should give them a lead on this and it'll be interesting to see what they talk about here.

>> Yeah, that's a good point. Let's unpack that nuance of foundation model architecture. Are you referring to that from a data perspective? Or are you referring from an architecture from an infrastructure standpoint? Because if you think about the horizontal scalability needs, as well as the data model to match that, are they hand in glove? Or how do you see those two distinctions?

Holger Mueller

>> I talk about it from an AI model architecture, which means that you have a cake, layered cake, like at weddings, where there's different layers. And how can you go very quickly across them? Infrastructure there is, again, it has to be in one place. You can theoretically, we'll have the same discussion, like back in the database. Put all your data on the lake, or federate the data. We know the federation guys have lost, because you can't get the level of service. AI has to be fast and efficient. If you have to wait, latency be expensive, Nvidia GPU right now, for your data be rag augmented from the on-premise data center where you're closing the books right now, and it's super slow. You're paying through the roof. That's not affordable. So you need to create the lake house, which finally everybody has agreed on, in the public cloud, in the right place, where there's cheap custom silicon on custom hardware.

>> So let's talk about the data lake. You mentioned that in the federate. It's a great example of why centralized is actually good, an argument for centralization.

Holger Mueller

>> Yes.

>> Totally agree with you on that. Now let's take it to the next level. What's next behind Snowflake and Databricks? Because they're the data cloud 1.0, that's my words and they would not agree with that, but there's a 2.0 world coming around. I mentioned a few scenarios. For example, I'm going to have this, maybe this data lake, and I want to turn it into a high velocity series of rivers to feed my devices or other subsystems, with neural net, whatever you want to call it.

Holger Mueller

>> Yep.

>> But how do you get, at an edge device or some other location, the data in less than in a millisecond, or a nanosecond to an edge device? The laws of physics cannot be changed.

Holger Mueller

>> Correct.

>> And if data's expensive, you've got a problem.

Holger Mueller

>> Yes.

>> You're saying that means the edge device can't generate. So we're back down to, it reminds me of the tactical edge conversation. Highly availability or highly available data, there's nuance on data management. So I just don't-

Holger Mueller

>> But John, we have 5G now. It's all solved.

>> But this has to change. There has to be a changing of the model of the egress.

Holger Mueller

>> Right.

>> Your either egressing, and eliminating that completely, which it will never happen, unless you see some... Or an entrepreneur will come out and create new switches, or new systems that could handle it.

Holger Mueller

>> Yep. I think we saw some motion in the egress, largely thanks to Oracle, didn't charge much to it. There's been a reflection for Google, and AWS on the egress course. So this like, California hotel model, come in and you can never leave, by bringing your data but you can't take it out, I think that's weakening up to a certain point. And enterprise will make a big stink, if that's going to be the hindrance for them to create that one single place. Where, by the way, as we know in good IT, you should always have the backup place as well. Now to your edge problem, the bigger challenge I see as a software guy, on the edges, not necessarily the connectivity, it's that the edge has to do also something that is autonomous. Your thermal start in your home cannot make a decision, or to how it could because it lost connectivity, or it's still waiting for the data to download. So it's the autonomous part and then the augmented part, when there's connectivity, and how much data can you push to the edge to make the edge do intelligent decisions again, autonomously, because it might be disconnected again? That's the interesting question in my view.

>> That's a constraint that needs to be solved for.

Holger Mueller

>> Yes. Yes.

>> So that could either be solved by real data, or synthetic data, possibly.

Holger Mueller

>> That's .

>> Not only synthetic data is emerging. Maybe there's a copy of data. So we're starting to get into a whole... Again, this notion changes the game. So the question for you is, who wins and who loses? Because again, we're starting to see the line in the industry forming between the old way and new way.

Holger Mueller

>> Yes.

>> The fog is lifting. You start to see it. If you do things this way, you're going to be screwed. You better be on the right side of history. So we got to get on the right side of the street. So what does that look like to you? If you can say, can you start to see the formation of the line, between old way, new way? We're in a classic disruptive enablement market. It's an enablement but it's going to disrupt something. Something will change. It's either old and you're not new, and some old becomes newly transformed. What's your take on old and new? Give us your breakdown.

Holger Mueller

>> The old will be anybody who is stuck, and on premise is what I call finite computing architectures, where you would go, and go out and size your ERP system and say, "How much do we need for the three peaks of ERP?" Closing your books, running your manufacturing, and running payroll. God beware, and never run those together. The system will stop for everybody. You have to space them out. So anybody who thinks in finite computer architectures, will be in a big of problem. Anybody who also thinks small, and is trying to get wins an other pocket, as much as they might be. So I think in 2024, you will get lots of wins and enterprises, small models, right? If you're the CTO in charge, do some small models to keep everybody happy, but keep the prize in mind, which is larger is better. And anybody who's an infinite side, who's on the cloud side, who has elasticity, architecturally, commercially, who can run algorithms who are easy to use, or easy to use by business people, we will see the same democratization of AI that we saw for coding. So low code, no code of AI needs to happen this year, hopefully latest next year. Democratization of AI, those will be the winners.

>> So just to put it out there, small model and large models, they're not mutually exclusive in your mind. You can do both.

Holger Mueller

>> You should do both, but you need to keep in mind that the large model is ultimately going to win it. The sooner you are at the large model, it will make better decisions, better ultimate decisions for you.

>> It's like a brain, it's the bigger brain, but you have other synapses in there.

Holger Mueller

>> Correct.

>> I'll give you another debate here. So okay, I think they're not mutual. You can go all in on the company models, large. You have some central brain that you create, multimodal of course. And then you've got smaller models that will work almost like plugins. It's almost like connecting into the LLMs, working together, because there's small pockets of little proprietary things, but model integration will be a reality.

Holger Mueller

>> Model integration is going to be key, and that's where the incompatibilities might come out of these small models. You have the world saying it's hot in Boston now, and the small one saying, "In May, I don't put ice cream in my things." Who are we going to trust?

>> Yeah, yeah.

Holger Mueller

>> What will we going to do? So I only see small long-term, and I'm talking two, three years exceeding, when you're on the satellite on the way to Mars, because nothing is going to happen around you anymore. If it happens on this very planet, so many things happen. The Iranian President has a helicopter accident, what does it mean for my investments, for the oil price, right? How does the oil price effect my filling up? Take an Uber or do I take public transport to come here to speak on theCUBE? But it's all connected, right? It's the famous bike in China falls over and the world changes.

>> Well, first of all, a helicopter crashed. It was reported as an accident by state television from Iran. Of course, ironic, right? Probably not.

Holger Mueller

>> But the thing is-

>> ....

Holger Mueller

>> how does this influence us? Is this not important? Is it non-event? Is my large model going to run wild? Or does my large model need the data if I'm in Iran? Because there will be probably a holiday. Don't wishing anybody to crash in a helicopter.

>> Yeah.

Holger Mueller

>> But again, I'm in retail in Iran, I'm a business in Iran. There will be 10 days of morning or something like this. I'm going to be shut down. What am I going to give my cash? The cash model, the cash prediction model, CFOs like to have them under their desk as a very small model.

>> Well a lot of people want to get a win in the enterprise. So small right now is to get confidence. So if I'm the big boss, I want to show that I have the AI moves.

Holger Mueller

>> Right, yeah.

>> So I'll throw a little rag in there, I'll do a little modeling. But, ultimately, it's a talent game as well. You need the talent, you need the skills.

Holger Mueller

>> Correct.

>> . I see little wins to give confidence and budget, by the way, frankly.

Holger Mueller

>> Yes. Yes, of course.

>> So budgets are now being allocated for AI, and so you got to get some reps.

Holger Mueller

>> There's no doubt, John, little wins are key. Like successes on the way is important for any CXO adopting any new technology. That is not changing for AI. It's more important so than all the other technologies that you have the endgame in mind, which is, in my view, and I can be wrong, talk to me in a few years, or a few months. The large side is winning. Everybody who is saying do small stuff, all the people saying, "Oh, we're doing AI since 20 years." Well kind of like machine learning or predictive analytics, nobody new generative AI two years ago.

>> Well, I've been saying, "Do the small stuff." So we're debating that now.

Holger Mueller

>> Perfect, let's debate it.

>> Okay, let's go. All right, so we're going to put that-

Holger Mueller

>> Another time, another time.

>> We're going to put a pin in that. But I do want to ask you two more questions. IBM's prospects, but first before that, we compare the internet and web to this AI movement, and a much different compressed scale, obviously. But transformative, all aspects of it, the web is kind of like AI now, everything changes. Question for you, and like the web, seeing consumer lead, consumer models are out there, OpenAI. The consumer providers, like Amazon and Meta, they're buying all the chips. I thought the broad common Nvidia guys are saying, "The consumer AI is here first. Enterprise is lagging." Dave calls it chat box 2.0, not a compelling enterprise AI direction. Will the AI gap between consumer and enterprise, which has traditionally been years in adoption matching, compress, shrink? Or is it shrinking and is it closer? Because there's data involved now, and there's actually connective tissue between consumerization of IT, which we've been talking about for over a decade, is actually happened. So if consumerization of IT has happened, you'd think that the lag between consumer AI and enterprise AI would be months, not years. What's your take on this?

Holger Mueller

>> Well definitely, lag is getting smaller, because everything's moving faster. My research enterprise acceleration, you have to move faster, you have to be agile. The interesting thing so is, that there's something missing in the current generative AI, to make it really work in the enterprise. Here's the example. How much of the consumer of my relevant data is in documents? Pretty much all of it. Pictures, videos, really documents, which generative AI, the transform algorithm, can make sense of. Come to the enterprise. How much of the relevant data of an enterprise and documents versus transaction databases? It's all in transactional database. If you think about as a consumer, well, I have some transaction database information too. It's always when you interact with an enterprise. It's your paycheck, your employment date, right? It's for government, right? Birthdays, tax things, and so on. You interact with an enterprise, then as a consumer you have data, structured data. Structured data needs to be understood by transformer 2.0, or it's going to be called whatever. Then we'll see all hell break loose, because right now you cannot make sense of what is happening in the enterprise for real. Once that happens, we'll see the next part, and then we're not going to talk, "Oh, it's not relevant for the consumer." For the first time, the enterprise leads here, because technology can do what the enterprise needs to do. So I think that is the interesting thing to look, the people who are researching algorithms and research developing universities at the start-ups, that's the key price to look.

>> That's a tipping point as well. That when that happens, when that transformation of database gets converted to AI, then a whole new level kicks in.

Holger Mueller

>> Correct. Because we say it's all about the data. All my data is unstructured data. Right? And how can I get structured data into this? We can see the opposite right now, right? Every database under the sun is getting vector support. That is to make the unstructured data query-able in the structured world. Let's bring it the other way around. We have to get the transaction data out, the transactional way to get it out is to put it in the lake house. There's no question about it. If you're seeing it's all listening, you have no lake house strategy, you're fumbling on your storage.

>> You're fumbling.

Holger Mueller

>> a small models-

>> You're going to sink in the mud....

Holger Mueller

>> medium models, or large, or super large models.

>> That was an incredible amount of data you just transferred here on theCUBE, putting it out there out in the open. Final question for you as we wrap up here. By the way, this is the pop-up CUBE. We will do whatever it takes to get the content. We're in Boston for IBM Think. IBM, they're back. IBM, the storied franchise, everyone knows it. They got a good comeback, then they kind of got lost a little bit. Now they're back. Watson 1.0 is not a winner but WatsonX is a winner. It's looking good. They got Red Hat, now they got HashiCorp.

Holger Mueller

>> Yep.

>> They got a gen AI gift dropped in their lap.

Holger Mueller

>> Yes.

>> Gen AI is a gift to IBM,, in my opinion, because they have an install base. They have large-scale, they've got customers that have a lot of data, and probably aren't that mature on the "lake house front". Your take on IBM and what they need to do as a company? If you had to look at IBM, their partners, their platforms, their opportunity? Arvind was on stage, the CESN.

Holger Mueller

>> Saying?

>> "The pie's big for everyone." Never heard him say that before. Can they pull a partner ecosystem together that's not a blue washing, or just IBM Blue? What's your take on IBM, their prospects? And what do they need to do to get back into that, into the winner's circle?

Holger Mueller

>> Right. So first of all, congratulations for IBM to be back, despite all the challenges. IBM is the only large old guard, as the friends at Amazon call them, IT vendor who had a million plus revenue in the '90s and is still relevant today. They're the only one who also failed at building cloud. Oracle's the other one. They have not failed at building cloud. But their main butter and water is services. And funny, you know Tony Bear who was on the call?

>> Yeah.

Holger Mueller

>> Tony Bear, that before it happened, and me, we did this on YouTube, a call like what should IBM do? They should get rid of the technical services, because they're not helping them on the modern side. They're going to get commoditized, all success code relations to Kindle right now. So I think that's the next step. And Arvind tends to decide, is IBM a software company? Or is IBM a service company with software, which they've been for a long time? And that has frankly hurt their products. Because if the product needed a little more services, that was an okay outcome for an IBM customer and for IBM. So the question is really, how much software revenue can the new IBM produce? And then they have some interesting bets on the hardware side. Not to talk about the hardware refer cycle, but this is the big year. This should, last year would've been the year of quantum, if there wouldn't have been this gene AI thing. We saw the first time quantum to scale, and the model which IBM is doing, with the eagle process of commoditization, this putting things in the rag, something well-understood, something which can have lots of problems. But we could see the quantum breakthrough coming from this very company. I'm not going to talk about this here. But that's something to keep tabs on, because that was another key disruptive thing. We talk about gen AI, we don't talk about quantum much.

>> Yeah, quantum's right around the corner, for sure.

Holger Mueller

>> Yeah.

>> So you're saying that they should be a services company first and product second? Or product first?

Holger Mueller

>> No. I think, Arvind, I'm a product guy, I love combination product companies.

>> Yeah, I do too.

Holger Mueller

>> This is why attracting the partners, or the partners plug-ins, or the partners who are making 2, 3, 4 times the revenue because they provide the services, and I'd be honest is providing the software license? That's going to be the key question for me. The HashiCorp acquisition points and the otherwise side, I'm not a big fan of the HashiCorp acquisition in general, because we see that generative AI is doing phenomenally on the IT ops, DevOps parts. Those scripts are 90, 95, 100% there, because they're all the public domain. I don't know, five minutes, Arvind said, "What were you thinking? How can you get the multiple?" It's like, HashiCorp's a great company, no question.

>> Yeah.

Holger Mueller

>> But it's a great asset, but isn't an asset for someone who's trying to become a software company? I'm not sure. Who's someone who's trying to stabilize service revenue for another decade? I think IT ops, DevOps scripts-

>> I think it's a service play....

Holger Mueller

>> are the first to be commoditized.

>> Yeah. It's definitely a service player. It's a Red Hat. HashiCorp Red is probably the closest thing to the original Red Hat. You give it away for free and you service the enterprise.

Holger Mueller

>> Yeah, but Red Hat had so much, we see that today. All the code of Red Hat, forget the Linux, , but think of open shift. All that is not going to be touched by generative AI. In contrary, it's going to be a platform for general AI across multi-cloud. So it's a genius acquisition from that perspective. The HashiCorp code, which allows me to division things in the cloud and moving things, in my view, that was a problem of HashiCorp, and all the open source companies. They started for free and they can never find a way to charge appropriately for that. So I don't see them for software. I see them for service, and I don't know how much service pipeline HashiCorp will have.

>> I think that brings up a good point. We'll have to put another pin in that. The idea of agents coming will solve a lot of those services inside the platforms.

Holger Mueller

>> They're already there.

>> Whether its... They're already there. You're starting to see it.

Holger Mueller

>> If I write an IT Ops DevOps script from scratch, to move my data from worked into Sales Force on Amazon, I'm stupid.

>> Yeah.

Holger Mueller

>> I'm stupid. I shouldn't be there. I'm wasting my employer's money.

>> You bullish on agents being relevant part of the future.

Holger Mueller

>> A hundred percent. A hundred percent. We see this on the tiny things of many Copilots which have created, the assistance that Google Cloud showed us a few weeks ago. There's no question, we need more help. We are running out of people. Right? Finally, the Wall Street Journal talked about how the lack of more birds in the first world is going to change dramatically our hand to machine ratio, where we need any innovation that we need. Remember the discussion around, "Oh, we're going to get all the truck drivers obsolete?"

>> Yeah.

Holger Mueller

>> And then we had the backlog in Long Beach, close to where I live, of enough containers to be stuck from Long Beach to New York because there's not enough truck drivers. Please bring the self-driving truck to get my delivery of my container.

>> We need to re haul the highway system. Holger, great to have you on theCUBE. Again, great conversation, because what's happening is, ultimately, we're living in a whole new world now, where you're starting to see aging technology. The big get bigger, and the rich get richer, and what's the opportunity for entrepreneurs?

Holger Mueller

>> I don't buy into this whole, the big get bigger, the rich get richer. If we on this podcast figure out a better algorithm for search than Google has, Microsoft, Amazon would love to host us. Google, the founders, Larry Page and , had a great idea, but it took them years to even build the infrastructure to index the internet. Today, we can index the internet with a new algorithm any given time. So I think the barriers of entry are lower. We see that with the new AI start-ups. You have a European one, with Alex Alpha. You have a French one. French one. I don't know how long they're going to be around, but the barriers of entry, everybody likes to have loads, and we have this infinite compute system which enables startups with a great and bright idea to do something.

>> Well, I think we're both in agreement, but also might debate some of the key points around big and small. But the one thing that is certain, is that there's a platform shift happening. A combination of models are going to be interacting and intersecting, sharing data through APIs, sharing data through common platforms. Maybe they look like label up blocks, but ultimately, big systems will work on big processing things. Data will be the key. The CUBE is bringing the data. Holger, thanks for coming back on theCUBE.

Holger Mueller

>> Thanks for having me.

>> We are at IMB Think. I'm John Furrier, your host of theCUBE. Dave Velante's at Dell Tech World, we have everyone all over the world, this week and next week is conference season, bringing all the action. We're on theCUBE. We'll be right back. Stay with us for more coverage from Boston.

Holger Mueller | Constellation Research, Inc.

search

>> Hello, welcome back to theCUBE's Coverage here in Boston Massachusetts. I'm John Furrier, host of theCUBE. IBM Think is happening. We are here at the pop-up CUBE. Holger Mueller is here, is CUBE alumni, CUBE contributor. Also, principal research as a constellation and a friend of theCUBE. Holger, great to see you. You finally made it back on theCUBE.

Holger Mueller

>> Yeah, I can't believe it, huh? You gave me 18 minutes notice. Am I the fastest visitor on theCUBE? The fastest guest on theCUBE?

>> You texted me before. I tried to get you on last time, but I couldn't make it work.

Holger Mueller

>> Yeah, I know, I know, I know.

>> You were flying out of Google Next. But a lot going on. I really appreciate your perspective. You were also recently at a variety of other events. Oracle had an analyst event. Google Next. The industry right now is going through a major platform re shift.

Holger Mueller

>> Yes.

>> When I say re shift, it's shifted, it's re shifting again. The tectonic plates are moving with generative AI, all layers of the stack.

Holger Mueller

>> Yes.

>> IBM's here. We had Dell has their event. What's your take on this? What are you seeing? Are you seeing the line forming between the old way and new way? Could you share your view?

Holger Mueller

>> Yeah, I think we saw this rebalancing, that on-premises was something which was getting alive again. I wrote about this six, seven years ago. I called the infinite compute platform, where you can move workloads seamlessly from the public cloud to on-premises. It has worked really well for some vendors. We had IBM, like OpenShift from Red Hat, allows you to run across multiple public clouds and on-premises even to the edge. That's exactly that. Enters with a boom, gen AI ,and all of a sudden it's all about elasticity again. Even can't get Nvidia machines for your on-premises workloads, right? So the balance always tipping back and forward. It's totally moving in towards the public cloud right now.

>> And that sounds like super cloud definition as far as I'm concerned.

Holger Mueller

>> Yeah, sorry.

>> Charles Fitzgerald would love to have a debate. Infinite cloud, what do you call it?

Holger Mueller

>> Infinite computing platforms.

>> Infinite computing platform. So we have the super cloud, it all points to the same thing. A new way to operate, which everybody wants, which is a first-party choice by the company to run their cloud operations, their infrastructure, their way, not the vendor's way. This has been a big multi-cloud problem, that you got three clouds. They all have different sets of the same services. You still need identity management, but it's different on all three. You need data, it's in all three clouds. How do you handle portability? All these things now are coming to the surface.

Holger Mueller

>> It's quick.

>> Now that we have a generative AI, the forcing function, what do you see happening right now? Obviously, do you see the gen AI market stalled a little bit on the app side? Is the action in the infrastructure? Are they retooling there? And then what happens next, after the infrastructure gets tooled up quickly, the middle layer going to get innovative. What's your-

Holger Mueller

>> What? ....

>> analysis?

Holger Mueller

>> Great question, but what gen AI does is basically collapses everything again. In my view, bigger is better when it comes to gen AI. If I know that the weather just changed from cloudy this morning in Boston to sun, I'm going to put on more ice cream in my freezers because I know people will eat ice cream, which otherwise they wouldn't. So bigger models have more awareness, are more human, will be better in business. And bigger models need more data. So data gravity, which we've taught when for a long time, is getting really, really again tangible, and that means it's going to move to certain clouds. You have to be widely, we said before, the multi cloud abilities, that made vendors across the whole stack. We can go back to 10 years ago, cloud foundry for pass. We had IBM. We can talk HashiCorp for DevOps. We can go to . It's at lots of events recently. They're user summit cross-clouds. Companies want to be cross-cloud, but if I want to do AI, and have to do a large language model, I need to get the data together.

>> And that's why the surge is on-premise, so the term AI factory, that Jensen long at Nvidia been talking about, he's promoting that, because he wants to reinvent the idea that you can create a data center on-premise, whether it's on a premise, or in a collo facility. But you can build a purpose-built set of infrastructure to service a workload.

Holger Mueller

>> Yes.

>> Do you buy that?

Holger Mueller

>> No.

>> Okay, good. Why?

Holger Mueller

>> Because Jensen was in this business, right? Five years ago, the key question of Nvidia was, can Nvidia survive, because none of the cloud vendors is using their GPUs. None of them was, five years ago. And only Microsoft's fumble of building their own custom silicon. Kevin Scott found OpenAI. And OpenAI could done for Jensen, got their first DGX machine in 2016. I don't know if you've seen the tweets on theCUBE, of , being at OpenAI at the time?

>> Yeah.

Holger Mueller

>> And the first DGX machine coming there, great seating. This dependency of OpenAI on Nvidia, has made Nvidia again a cloud company, the most valuable company in the world. So Jensen is talking about something which was very popular, out of necessity because nobody is doing this stuff. Right now, he can only sell to cloud vendors because of their premium.

>> Yeah.

Holger Mueller

>> But so very interesting. At Oracle's OCI Summit two weeks ago, they pre-pay in cash, they shared, to get infrastructure being built. Suppliers are being paid in cash, so they can roll out their data centers across the whole stack. We're not talking about Nvidia, but basically today, you know can only get the Nvidia machine if you're a friend of Jensen and Colin.

>> And so is that because they want to Bogart the machines, and hoard them in, and store them? Or is it just pointing to the use case, that the large language models and the large cloud players have a use case, called consumer business? Amazon's got retail, Meta has Facebook. Google's got search, and then they have clouds behind it. So you have all these super clouds, these big cloud, hyper clouds have retail businesses. That's a consumer version of AI. Would you see that? What's your reaction to that?

Holger Mueller

>> There is a big need on the consumer side, and a large part is fueled by the advertising wars. But if people don't Google something anymore. They ask generative AI and want a summary, which then has advertised links in them. That's going to happen. You will see that. And you just saw Google at Google IO last week, announcing that more and more of the questions you Google will be generative AI answer. That needs massive compute. But the awareness of what's happening in the real world, is not only relevant for consumers to sell advertising, that's relevant for everything. At what temperature do you want to run this hotel lobby that way? Right? What loudness do we want to put on our microphone reception here, right?

>> Yeah.

Holger Mueller

>> The understanding of the real world, with multimodal models, which is the big innovation of 2024, that models can understand more things, is the key aspect. And this is why the larger is better. If I don't have my weather, my noise information together, I cannot build AI for it.

>> Dave Velante put a poll on his Twitter about who's going to open up the AI. Clearly, 73% of respondents, mostly analysts, agreed that the hyperscalers will mop up the AI business, and ultimately, probably, give language models away for free. Your thoughts on that? And does that validate what you're just saying, is that the service providers for AI will be the monster cloud players?

Holger Mueller

>> Yeah, there's no way around it, because we know that the pushback to public cloud from the early days was, if I know my workload, if I can predict it, and it's stable, then I go to public cloud. Quick side, a comment inside, right? One of the largest, most valuable companies in the world. Apple didn't build a cloud, uses the clouds for backup, which is super stable, right? You can't take so many pictures to make this a volatile part. But even then it's being used. So the problem is, generative AI workloads in need of an enterprise cannot be assessed by any CIO, CTO today. You cannot go to your board and say, "I need 10 million to buy 20 DGX machines and we'll be fine for the next five years." There is no way of knowing that. And that uncertainty comes back to, if you boil everything down, systems from a quality, what is the key quality of the cloud? It's elasticity. Elasticity from a technical perspective, from up and down, and more importantly, from a commercial perspective. You use more, you pay more. You use less, you pay less. So this is why the cloud wins.

>> All right, so let's take that cloud benefit and then try to map that onto the AI world. So elasticity, that's a technical, and agility, that's a mechanism-based value proposition. The cost side is, pay by the drink. That's economics. So they have both the economic value proposition, as well as the technical. I'm oversimplifying it, but agility, that's why startups went there first. The president of data center, we all know that rule. Okay, fast-forward to the AI world now, what are the problems that people have? I don't have the hard way to run my own stuff? I don't have the expertise to run AI? And how the hell does this impact my business, both cost efficiency and maybe new revenue?

Holger Mueller

>> Right.

>> So, okay, what's their answer? They? Go where

Holger Mueller

>> This is why you need to move to the cloud, because apart from machines-

>> You're saying that's the reason to move to the cloud?

Holger Mueller

>> Yes. People who are not in the cloud need to move from the cloud. I need to move my data closer, to right now the DGX machines. Maybe in two, three years, hopefully there's more competition, to the Intel, AMD, whatever machines, to create the generative AI that I need to run and power my enterprise.

>> Got it. Okay, let ask you another question. So I buy that, by the way, I think that's a good outlook.

Holger Mueller

>> We agree on something?

>> Yeah, we agree. I like that we're going to debate this. Okay, here's the next thing, you said end-to-end workload. If you know your workload, you go to the cloud. A lot of companies have workloads that are well-defined. They have a back-end, with a software model, and a front-end application. It could be a banking app or whatever.

Holger Mueller

>> Yes, yes.

>> And so, it can scope that workload end to end. Now by adding generative AI into the equation, they say, great, I know my workload end to end. I can now build purpose-built hardware, and or workload to support the workloads. Whether that's for response time or reasoning. So question, the optimized infrastructure for the IO of generative AI, question answer, whether that's a prompt under the covers or directly, that's going to be low latency response. Or go away and think of a problem. Reason, this use case is to use different maybe potentially clouds. Will you go to a cloud on that? And or would you build it on-premise, if the data's proprietary?

Holger Mueller

>> No, first of all, data proprietary is important. No question. You live a knowledge economy, data is the most important asset. But I think there's no concern to, if your data is on premises is safer than it's in the public cloud, bearing any potential accidents. I think your past that decade in that discussion. The key things though is, do you want to, do you believe in larger being better? I'm a strong believer in that camp. If you only want to do classic IT, and says, "I have my banking application and I'm going to do some intelligence," there's some people out there and saying, "Oh, AI is just intelligence and analytics." I think it's much, much more because it goes inside to action, to the automation of transaction, to the automation of bots, to the automation of agents. But if you do small AI, you can do that on-premise and try to figure that out. But I think the small AI will always lose against the big AI, and the big AI needs cloud. And here comes the key thing. Even so, I know my transaction application really well, it's predictable. I got my more cost-efficient on premises. I need the data of that in the public cloud to run my AI models on that. What if I see fraud, a classic example, right? I had 120 data scientists working my IT, at FiCo, 15 years ago, 20 years ago. 15, 20 years ago, right? Trying to figure out fraud. We had to get tapes from the banks. It took us half a year. We were so much behind the fraud. Today you have to be faster. Your transaction systems on premise, you once a day dump it in the cloud to figure out fraud, you could have lost millions.

>> It's like a video game. If you lag, you're going to lose every time. You're going to get shot on the first person shooter games. Happens all the time.

Holger Mueller

>> Yes.

>> But the lag, you bring up a good point. Data quality and data low latency address-ability. Is it in the right place at the right time? So I want to ask you a question. Define big AI, because you said that a few times, when you say you need to clap for your big AI.

Holger Mueller

>> What?

>> Define, what's the difference big AI and small AI?

Holger Mueller

>> Right. So big AI is when you take all the data available in your enterprise, and from third parties, to build your models on that, which by definition, are very large language models. They're very similar to OpenAI's Bake the Internet, to Gemini 1.5 Bake the Internet, to Llama Bake the Internet, and so on. Maybe not as big, but very, very large. Versus you do on-premise, on-device potential AI, but there's use-case for that too. But it only uses that local data. And the question is if the small AI can make the right decisions?

>> That's awesome. So okay, let's get back into the power law. So you're starting to see specialized models, enterprises are, you can see the IQ of the enterprises. It's almost a shadow IT model, shadow AI. A lot of red, a lot of retrieval augmentation generation, even came up on today's keynote at IBM, on the partner summit here. It's an easy thing to do. If you have data, do some rag on it. It's small AI, it's not a large AI model, to your definition.

Holger Mueller

>> Yes.

>> So there's a lot of that going on. Is that an indicator of appetite or are people tire-kicking?

Holger Mueller

>> No, it's the art of the possible right now, and the possible is always limited by the vendor doing that. If you look at the vendors who have large well-managed models, Microsoft, Meta, Google, they give you the architecture. And what really matters is the foundation model architecture. How you're going to do that layer cake, from your very large, boil the ocean, have-the-all-internet language model, to the one which might give you an industry aspect, might give you a vendor aspect, might give you your customer data and your customer-specific stuff to be private and be safe from anybody else looking into that? But that's layer cake. So I don't see enough discussion about foundation model architectures, which is a big thing. And by the way, IBM was one of the first winners, as we hear, to talk about foundation model architecture last year and summer, which should give them a lead on this and it'll be interesting to see what they talk about here.

>> Yeah, that's a good point. Let's unpack that nuance of foundation model architecture. Are you referring to that from a data perspective? Or are you referring from an architecture from an infrastructure standpoint? Because if you think about the horizontal scalability needs, as well as the data model to match that, are they hand in glove? Or how do you see those two distinctions?

Holger Mueller

>> I talk about it from an AI model architecture, which means that you have a cake, layered cake, like at weddings, where there's different layers. And how can you go very quickly across them? Infrastructure there is, again, it has to be in one place. You can theoretically, we'll have the same discussion, like back in the database. Put all your data on the lake, or federate the data. We know the federation guys have lost, because you can't get the level of service. AI has to be fast and efficient. If you have to wait, latency be expensive, Nvidia GPU right now, for your data be rag augmented from the on-premise data center where you're closing the books right now, and it's super slow. You're paying through the roof. That's not affordable. So you need to create the lake house, which finally everybody has agreed on, in the public cloud, in the right place, where there's cheap custom silicon on custom hardware.

>> So let's talk about the data lake. You mentioned that in the federate. It's a great example of why centralized is actually good, an argument for centralization.

Holger Mueller

>> Yes.

>> Totally agree with you on that. Now let's take it to the next level. What's next behind Snowflake and Databricks? Because they're the data cloud 1.0, that's my words and they would not agree with that, but there's a 2.0 world coming around. I mentioned a few scenarios. For example, I'm going to have this, maybe this data lake, and I want to turn it into a high velocity series of rivers to feed my devices or other subsystems, with neural net, whatever you want to call it.

Holger Mueller

>> Yep.

>> But how do you get, at an edge device or some other location, the data in less than in a millisecond, or a nanosecond to an edge device? The laws of physics cannot be changed.

Holger Mueller

>> Correct.

>> And if data's expensive, you've got a problem.

Holger Mueller

>> Yes.

>> You're saying that means the edge device can't generate. So we're back down to, it reminds me of the tactical edge conversation. Highly availability or highly available data, there's nuance on data management. So I just don't-

Holger Mueller

>> But John, we have 5G now. It's all solved.

>> But this has to change. There has to be a changing of the model of the egress.

Holger Mueller

>> Right.

>> Your either egressing, and eliminating that completely, which it will never happen, unless you see some... Or an entrepreneur will come out and create new switches, or new systems that could handle it.

Holger Mueller

>> Yep. I think we saw some motion in the egress, largely thanks to Oracle, didn't charge much to it. There's been a reflection for Google, and AWS on the egress course. So this like, California hotel model, come in and you can never leave, by bringing your data but you can't take it out, I think that's weakening up to a certain point. And enterprise will make a big stink, if that's going to be the hindrance for them to create that one single place. Where, by the way, as we know in good IT, you should always have the backup place as well. Now to your edge problem, the bigger challenge I see as a software guy, on the edges, not necessarily the connectivity, it's that the edge has to do also something that is autonomous. Your thermal start in your home cannot make a decision, or to how it could because it lost connectivity, or it's still waiting for the data to download. So it's the autonomous part and then the augmented part, when there's connectivity, and how much data can you push to the edge to make the edge do intelligent decisions again, autonomously, because it might be disconnected again? That's the interesting question in my view.

>> That's a constraint that needs to be solved for.

Holger Mueller

>> Yes. Yes.

>> So that could either be solved by real data, or synthetic data, possibly.

Holger Mueller

>> That's .

>> Not only synthetic data is emerging. Maybe there's a copy of data. So we're starting to get into a whole... Again, this notion changes the game. So the question for you is, who wins and who loses? Because again, we're starting to see the line in the industry forming between the old way and new way.

Holger Mueller

>> Yes.

>> The fog is lifting. You start to see it. If you do things this way, you're going to be screwed. You better be on the right side of history. So we got to get on the right side of the street. So what does that look like to you? If you can say, can you start to see the formation of the line, between old way, new way? We're in a classic disruptive enablement market. It's an enablement but it's going to disrupt something. Something will change. It's either old and you're not new, and some old becomes newly transformed. What's your take on old and new? Give us your breakdown.

Holger Mueller

>> The old will be anybody who is stuck, and on premise is what I call finite computing architectures, where you would go, and go out and size your ERP system and say, "How much do we need for the three peaks of ERP?" Closing your books, running your manufacturing, and running payroll. God beware, and never run those together. The system will stop for everybody. You have to space them out. So anybody who thinks in finite computer architectures, will be in a big of problem. Anybody who also thinks small, and is trying to get wins an other pocket, as much as they might be. So I think in 2024, you will get lots of wins and enterprises, small models, right? If you're the CTO in charge, do some small models to keep everybody happy, but keep the prize in mind, which is larger is better. And anybody who's an infinite side, who's on the cloud side, who has elasticity, architecturally, commercially, who can run algorithms who are easy to use, or easy to use by business people, we will see the same democratization of AI that we saw for coding. So low code, no code of AI needs to happen this year, hopefully latest next year. Democratization of AI, those will be the winners.

>> So just to put it out there, small model and large models, they're not mutually exclusive in your mind. You can do both.

Holger Mueller

>> You should do both, but you need to keep in mind that the large model is ultimately going to win it. The sooner you are at the large model, it will make better decisions, better ultimate decisions for you.

>> It's like a brain, it's the bigger brain, but you have other synapses in there.

Holger Mueller

>> Correct.

>> I'll give you another debate here. So okay, I think they're not mutual. You can go all in on the company models, large. You have some central brain that you create, multimodal of course. And then you've got smaller models that will work almost like plugins. It's almost like connecting into the LLMs, working together, because there's small pockets of little proprietary things, but model integration will be a reality.

Holger Mueller

>> Model integration is going to be key, and that's where the incompatibilities might come out of these small models. You have the world saying it's hot in Boston now, and the small one saying, "In May, I don't put ice cream in my things." Who are we going to trust?

>> Yeah, yeah.

Holger Mueller

>> What will we going to do? So I only see small long-term, and I'm talking two, three years exceeding, when you're on the satellite on the way to Mars, because nothing is going to happen around you anymore. If it happens on this very planet, so many things happen. The Iranian President has a helicopter accident, what does it mean for my investments, for the oil price, right? How does the oil price effect my filling up? Take an Uber or do I take public transport to come here to speak on theCUBE? But it's all connected, right? It's the famous bike in China falls over and the world changes.

>> Well, first of all, a helicopter crashed. It was reported as an accident by state television from Iran. Of course, ironic, right? Probably not.

Holger Mueller

>> But the thing is-

>> ....

Holger Mueller

>> how does this influence us? Is this not important? Is it non-event? Is my large model going to run wild? Or does my large model need the data if I'm in Iran? Because there will be probably a holiday. Don't wishing anybody to crash in a helicopter.

>> Yeah.

Holger Mueller

>> But again, I'm in retail in Iran, I'm a business in Iran. There will be 10 days of morning or something like this. I'm going to be shut down. What am I going to give my cash? The cash model, the cash prediction model, CFOs like to have them under their desk as a very small model.

>> Well a lot of people want to get a win in the enterprise. So small right now is to get confidence. So if I'm the big boss, I want to show that I have the AI moves.

Holger Mueller

>> Right, yeah.

>> So I'll throw a little rag in there, I'll do a little modeling. But, ultimately, it's a talent game as well. You need the talent, you need the skills.

Holger Mueller

>> Correct.

>> . I see little wins to give confidence and budget, by the way, frankly.

Holger Mueller

>> Yes. Yes, of course.

>> So budgets are now being allocated for AI, and so you got to get some reps.

Holger Mueller

>> There's no doubt, John, little wins are key. Like successes on the way is important for any CXO adopting any new technology. That is not changing for AI. It's more important so than all the other technologies that you have the endgame in mind, which is, in my view, and I can be wrong, talk to me in a few years, or a few months. The large side is winning. Everybody who is saying do small stuff, all the people saying, "Oh, we're doing AI since 20 years." Well kind of like machine learning or predictive analytics, nobody new generative AI two years ago.

>> Well, I've been saying, "Do the small stuff." So we're debating that now.

Holger Mueller

>> Perfect, let's debate it.

>> Okay, let's go. All right, so we're going to put that-

Holger Mueller

>> Another time, another time.

>> We're going to put a pin in that. But I do want to ask you two more questions. IBM's prospects, but first before that, we compare the internet and web to this AI movement, and a much different compressed scale, obviously. But transformative, all aspects of it, the web is kind of like AI now, everything changes. Question for you, and like the web, seeing consumer lead, consumer models are out there, OpenAI. The consumer providers, like Amazon and Meta, they're buying all the chips. I thought the broad common Nvidia guys are saying, "The consumer AI is here first. Enterprise is lagging." Dave calls it chat box 2.0, not a compelling enterprise AI direction. Will the AI gap between consumer and enterprise, which has traditionally been years in adoption matching, compress, shrink? Or is it shrinking and is it closer? Because there's data involved now, and there's actually connective tissue between consumerization of IT, which we've been talking about for over a decade, is actually happened. So if consumerization of IT has happened, you'd think that the lag between consumer AI and enterprise AI would be months, not years. What's your take on this?

Holger Mueller

>> Well definitely, lag is getting smaller, because everything's moving faster. My research enterprise acceleration, you have to move faster, you have to be agile. The interesting thing so is, that there's something missing in the current generative AI, to make it really work in the enterprise. Here's the example. How much of the consumer of my relevant data is in documents? Pretty much all of it. Pictures, videos, really documents, which generative AI, the transform algorithm, can make sense of. Come to the enterprise. How much of the relevant data of an enterprise and documents versus transaction databases? It's all in transactional database. If you think about as a consumer, well, I have some transaction database information too. It's always when you interact with an enterprise. It's your paycheck, your employment date, right? It's for government, right? Birthdays, tax things, and so on. You interact with an enterprise, then as a consumer you have data, structured data. Structured data needs to be understood by transformer 2.0, or it's going to be called whatever. Then we'll see all hell break loose, because right now you cannot make sense of what is happening in the enterprise for real. Once that happens, we'll see the next part, and then we're not going to talk, "Oh, it's not relevant for the consumer." For the first time, the enterprise leads here, because technology can do what the enterprise needs to do. So I think that is the interesting thing to look, the people who are researching algorithms and research developing universities at the start-ups, that's the key price to look.

>> That's a tipping point as well. That when that happens, when that transformation of database gets converted to AI, then a whole new level kicks in.

Holger Mueller

>> Correct. Because we say it's all about the data. All my data is unstructured data. Right? And how can I get structured data into this? We can see the opposite right now, right? Every database under the sun is getting vector support. That is to make the unstructured data query-able in the structured world. Let's bring it the other way around. We have to get the transaction data out, the transactional way to get it out is to put it in the lake house. There's no question about it. If you're seeing it's all listening, you have no lake house strategy, you're fumbling on your storage.

>> You're fumbling.

Holger Mueller

>> a small models-

>> You're going to sink in the mud....

Holger Mueller

>> medium models, or large, or super large models.

>> That was an incredible amount of data you just transferred here on theCUBE, putting it out there out in the open. Final question for you as we wrap up here. By the way, this is the pop-up CUBE. We will do whatever it takes to get the content. We're in Boston for IBM Think. IBM, they're back. IBM, the storied franchise, everyone knows it. They got a good comeback, then they kind of got lost a little bit. Now they're back. Watson 1.0 is not a winner but WatsonX is a winner. It's looking good. They got Red Hat, now they got HashiCorp.

Holger Mueller

>> Yep.

>> They got a gen AI gift dropped in their lap.

Holger Mueller

>> Yes.

>> Gen AI is a gift to IBM,, in my opinion, because they have an install base. They have large-scale, they've got customers that have a lot of data, and probably aren't that mature on the "lake house front". Your take on IBM and what they need to do as a company? If you had to look at IBM, their partners, their platforms, their opportunity? Arvind was on stage, the CESN.

Holger Mueller

>> Saying?

>> "The pie's big for everyone." Never heard him say that before. Can they pull a partner ecosystem together that's not a blue washing, or just IBM Blue? What's your take on IBM, their prospects? And what do they need to do to get back into that, into the winner's circle?

Holger Mueller

>> Right. So first of all, congratulations for IBM to be back, despite all the challenges. IBM is the only large old guard, as the friends at Amazon call them, IT vendor who had a million plus revenue in the '90s and is still relevant today. They're the only one who also failed at building cloud. Oracle's the other one. They have not failed at building cloud. But their main butter and water is services. And funny, you know Tony Bear who was on the call?

>> Yeah.

Holger Mueller

>> Tony Bear, that before it happened, and me, we did this on YouTube, a call like what should IBM do? They should get rid of the technical services, because they're not helping them on the modern side. They're going to get commoditized, all success code relations to Kindle right now. So I think that's the next step. And Arvind tends to decide, is IBM a software company? Or is IBM a service company with software, which they've been for a long time? And that has frankly hurt their products. Because if the product needed a little more services, that was an okay outcome for an IBM customer and for IBM. So the question is really, how much software revenue can the new IBM produce? And then they have some interesting bets on the hardware side. Not to talk about the hardware refer cycle, but this is the big year. This should, last year would've been the year of quantum, if there wouldn't have been this gene AI thing. We saw the first time quantum to scale, and the model which IBM is doing, with the eagle process of commoditization, this putting things in the rag, something well-understood, something which can have lots of problems. But we could see the quantum breakthrough coming from this very company. I'm not going to talk about this here. But that's something to keep tabs on, because that was another key disruptive thing. We talk about gen AI, we don't talk about quantum much.

>> Yeah, quantum's right around the corner, for sure.

Holger Mueller

>> Yeah.

>> So you're saying that they should be a services company first and product second? Or product first?

Holger Mueller

>> No. I think, Arvind, I'm a product guy, I love combination product companies.

>> Yeah, I do too.

Holger Mueller

>> This is why attracting the partners, or the partners plug-ins, or the partners who are making 2, 3, 4 times the revenue because they provide the services, and I'd be honest is providing the software license? That's going to be the key question for me. The HashiCorp acquisition points and the otherwise side, I'm not a big fan of the HashiCorp acquisition in general, because we see that generative AI is doing phenomenally on the IT ops, DevOps parts. Those scripts are 90, 95, 100% there, because they're all the public domain. I don't know, five minutes, Arvind said, "What were you thinking? How can you get the multiple?" It's like, HashiCorp's a great company, no question.

>> Yeah.

Holger Mueller

>> But it's a great asset, but isn't an asset for someone who's trying to become a software company? I'm not sure. Who's someone who's trying to stabilize service revenue for another decade? I think IT ops, DevOps scripts-

>> I think it's a service play....

Holger Mueller

>> are the first to be commoditized.

>> Yeah. It's definitely a service player. It's a Red Hat. HashiCorp Red is probably the closest thing to the original Red Hat. You give it away for free and you service the enterprise.

Holger Mueller

>> Yeah, but Red Hat had so much, we see that today. All the code of Red Hat, forget the Linux, , but think of open shift. All that is not going to be touched by generative AI. In contrary, it's going to be a platform for general AI across multi-cloud. So it's a genius acquisition from that perspective. The HashiCorp code, which allows me to division things in the cloud and moving things, in my view, that was a problem of HashiCorp, and all the open source companies. They started for free and they can never find a way to charge appropriately for that. So I don't see them for software. I see them for service, and I don't know how much service pipeline HashiCorp will have.

>> I think that brings up a good point. We'll have to put another pin in that. The idea of agents coming will solve a lot of those services inside the platforms.

Holger Mueller

>> They're already there.

>> Whether its... They're already there. You're starting to see it.

Holger Mueller

>> If I write an IT Ops DevOps script from scratch, to move my data from worked into Sales Force on Amazon, I'm stupid.

>> Yeah.

Holger Mueller

>> I'm stupid. I shouldn't be there. I'm wasting my employer's money.

>> You bullish on agents being relevant part of the future.

Holger Mueller

>> A hundred percent. A hundred percent. We see this on the tiny things of many Copilots which have created, the assistance that Google Cloud showed us a few weeks ago. There's no question, we need more help. We are running out of people. Right? Finally, the Wall Street Journal talked about how the lack of more birds in the first world is going to change dramatically our hand to machine ratio, where we need any innovation that we need. Remember the discussion around, "Oh, we're going to get all the truck drivers obsolete?"

>> Yeah.

Holger Mueller

>> And then we had the backlog in Long Beach, close to where I live, of enough containers to be stuck from Long Beach to New York because there's not enough truck drivers. Please bring the self-driving truck to get my delivery of my container.

>> We need to re haul the highway system. Holger, great to have you on theCUBE. Again, great conversation, because what's happening is, ultimately, we're living in a whole new world now, where you're starting to see aging technology. The big get bigger, and the rich get richer, and what's the opportunity for entrepreneurs?

Holger Mueller

>> I don't buy into this whole, the big get bigger, the rich get richer. If we on this podcast figure out a better algorithm for search than Google has, Microsoft, Amazon would love to host us. Google, the founders, Larry Page and , had a great idea, but it took them years to even build the infrastructure to index the internet. Today, we can index the internet with a new algorithm any given time. So I think the barriers of entry are lower. We see that with the new AI start-ups. You have a European one, with Alex Alpha. You have a French one. French one. I don't know how long they're going to be around, but the barriers of entry, everybody likes to have loads, and we have this infinite compute system which enables startups with a great and bright idea to do something.

>> Well, I think we're both in agreement, but also might debate some of the key points around big and small. But the one thing that is certain, is that there's a platform shift happening. A combination of models are going to be interacting and intersecting, sharing data through APIs, sharing data through common platforms. Maybe they look like label up blocks, but ultimately, big systems will work on big processing things. Data will be the key. The CUBE is bringing the data. Holger, thanks for coming back on theCUBE.

Holger Mueller

>> Thanks for having me.

>> We are at IMB Think. I'm John Furrier, your host of theCUBE. Dave Velante's at Dell Tech World, we have everyone all over the world, this week and next week is conference season, bringing all the action. We're on theCUBE. We'll be right back. Stay with us for more coverage from Boston.