Dell’s AI Journey: Implementing AI to Drive Meaningful ROI

Summary

Highlights

• John Rose, CTO at Dell, shares his unique journey in tech innovation and leadership 🌟.
• Dell takes a strategic three-market approach to AI: pre-generative, generative, and enterprise 👨‍💼.
• The role of data in AI: Clean and integrated data is imperative for successful AI application 📊.
• Dell evolves from building to buying AI solutions, focusing on reducing tech debt and speeding up AI adoption 🚀.
• AI factory concept at Dell aims to streamline and elevate AI infrastructure needs for clients 🏗️.
• John advises a top-down approach to AI, ensuring project priority and resource allocation are strategy-driven 🔝.
• Future gaze with John: The promising role of agent-based systems in digital enterprise transformation 👁️‍🗨️.

Key Takeaways

• John Rose shares his journey through various CTO roles leading to his position at Dell, highlighting personal and professional growth pathways 🚀.
• Dell's strategic use of AI focuses on maximizing business impact and scalability rather than just technological advancement ⚙️.
• Understanding the complexities of AI, John emphasizes the segmentation of AI markets, including pre-generative, generative, and enterprise AI sectors 📊.
• Data plays a critical role; identifying, managing, and preparing data is essential for successful AI deployment 📈.
• Dell's shift from building in-house AI solutions to buying ready-made solutions reflects their adaptive strategy to accelerate AI adoption ⏩.
• The introduction of AI factories aims to simplify AI infrastructure for Dell's clients, emphasizing speed and efficiency in AI integration 🏭.
• John emphasizes a top-down approach to manage AI resources effectively within organizations 🏢.
• John forecasts a future where agent-based AI systems significantly transform enterprise processes by digitally replicating skills and tasks 🔮.

Overview

Chapters

• 00:00 - 00:30: Introduction and Guest Introduction The chapter introduces the host, Dave Lyntham, who is an author and speaker with a focus on cloud computing. The podcast aims to discuss the realities of cloud computing and how enterprises can leverage it effectively. In this episode, Dave introduces the guest, John Rose, who is regarded as an innovator in the field. The podcast sets the stage for an insightful discussion with experts in the cloud computing space.
• 00:30 - 02:00: John Rose's Background This chapter provides an insight into John Rose's background, who is the chief technology officer and chief AI officer at Dell Technologies. With decades of experience in the tech industry, John spearheads Dell's innovation initiatives across various domains like artificial intelligence, cloud computing, edge computing, and other emerging technologies. The discussion begins with a friendly tone as the interviewer asks John to elaborate on his journey in the tech field, focusing on his origins and path in AI, fulfilling the curiosity of the audience about his professional growth.
• 03:00 - 04:00: AI Market Complexity The chapter delves into the complexity of the AI market by introducing a key figure in the industry—a serial CTO and current chief technology officer and chief AI officer of Dell. The individual narrates their extensive career journey, emphasizing their leadership experience at major tech companies like EMC, Broadcom, Terra's Cabeltron, and Nortell. The narrative highlights the diversity of professional journeys in the tech industry and suggests an inspiration for those aiming to replicate such success.
• 08:00 - 11:00: AI Implementation in Dell The chapter discusses the implementation of AI at Dell, highlighted by the experience and journey of a technologist who previously worked with Huawei through an entity called Futureway. The narrative emphasizes the diverse technological advancements and highlights the significant role of AI in today's technology cycle, underscoring Dell's position as a leading technology company during these transformative times.
• 15:00 - 31:00: Challenges and Learnings The chapter "Challenges and Learnings" begins with a discussion about swords visible behind the speaker. The speaker shares their recent travel from Austin, Texas, where the headquarters are located, to their home in the mountains of New Hampshire. The narrative touches on the impact of the COVID-19 pandemic, specifically mentioning how the speaker isolated in New Hampshire for 446 days during the pandemic.
• 41:00 - 47:00: Future of AI and Dell's Role The chapter discusses the unique office setup of an individual who practices martial arts, as they have relocated their office into a dojo. This setting reflects a blend of professional and personal interests. The individual humorously mentions the necessity of having a blurred background during virtual meetings to hide the martial arts paraphernalia that might be intimidating. The conversation seems to be setting the stage for a more in-depth discussion about Dell's role in the future of AI, although this part is not fully detailed in the provided transcript.
• 47:00 - 48:00: Conclusion and Closing Remarks The chapter discusses the complexity and importance of AI at Dell, highlighting that the AI market is composed of three concurrent segments. It emphasizes that traditional AI methods such as machine learning, computer vision, and reinforcement learning remain crucial, even with the advent of technologies like ChatGPT. These methods are described as the 'analog front end' of AI developments.

Dell’s AI Journey: Implementing AI to Drive Meaningful ROI Transcription

• 00:00 - 00:30 [Music] Hey guys, welcome back to the cloud computing insider where we talk about the realities of cloud computing and how to make it work for your enterprise. As always, I'm your host Dave Lyntham, author, speaker, beless geek and really am talking today about uh some great innovators in the space and really, you know, as we talk and invite guests on the show, you know, people are kind of moving and shaking in the space and today's no different. I got John Rose.
• 00:30 - 01:00 He's the chief technology officer and chief AI officer at Dell Technologies with decades of experience in the in the tech industry. John leads Dell's innovation efforts across artificial intelligence, cloud computing, edge computing, and emerging technologies. Hey, John, how are you? I'm doing well. How about yourself? So, fill in the blanks. In other words, uh, you know, part of this is about your journey, you know, and how you got to where you are, you know, what you did, how you started with AI. Um, my listeners always ask me to, you know, dig deeper into where
• 01:00 - 01:30 these folks came from because they all want they want to recon, they want to, uh, replicate your success. Okay. Well, uh, good luck with that. I Everybody's got their own unique journey. I I am a um serial CTO. I'm I currently I'm the chief technology officer and also the chief AI officer of Dell. Been at Dell Technologies since we created the company when we combined with EMC about eight years ago. Prior to that, I was the CTO of EMC. I was the CTO of Broadcom and Terra's Cabeltron, Nortell
• 01:30 - 02:00 and ran an entity called Futureway, which is did a lot of advanced development for a little Chinese company called Huawei about a 15 years ago. Uh you know, punch line is look a multi-disiplinary technologist passing around a bunch of different industries and and here we are today at uh you know, one of the biggest large broadest technology companies in the world at one of the most interesting times in the technology industry which is this AI cycle. So, you know, that's that's the journey you have to go on if you want to do this. Yeah. And it's you just don't hear enough about AI these days. So, I'm
• 02:00 - 02:30 glad we're bringing it back. You know, it's just one of those things. So, I'm going to go to the next stage and I'm sure my listeners are thinking about this right now. What's going on with those swords behind you? I actually just flew back from Austin. I was down I our headquarters down Round Rock, Texas. I was down there for most of the week. Uh uh landed last night. Now I'm up in my my place in in the mountains of New Hampshire. But uh I this is a byproduct of co when co happened I kind of parked myself here for 446 days and I didn't actually have
• 02:30 - 03:00 a proper office but what I did have because I've done martial arts my whole life is a proper dojo and so I just relocated my office into my dojo which normally when the screens are blurred nobody notices but in this case we have them unblurred and so you can see a little bit of paraphernalia behind me. So uh that that's that's the context here. I I would never do a Zoom call without without blurred. So I want them to know they're there. So yeah, some intimidation things like that. So what is what is like the Dell before we get
• 03:00 - 03:30 into the into the journey? Um you know what is the definition of AI at Dell and how you know what's the elevator pitch in terms of what you guys are doing at Dell? Yeah, I'll give you two elevator pitches that are important. First is look the AI market is incredibly complex but it's really three separate markets that are happening at the same time. There is the world of pre-generative AI reinforcement learning machine learning computer vision did not go away when chat GPT happened. In fact, it's become more important because it's it's almost the analog front end. It's the place
• 03:30 - 04:00 where data gets pre-processed before it ever gets to a generative system. That's a big market for us. We've been playing in that space for a very long time and it's actually becoming more and more interesting. The second market which is the one that's probably most visible is we have emerged as the provider of infrastructure for the people building out these gigantic training clusters. Colossus at XAI and Cororey and other folks and that's turned into a you know 10 plus billion dollar business from nothing a couple of years ago. Uh it's a
• 04:00 - 04:30 great it's a great industry but it it's a collection of a very small number of companies on the search for AGI and ASI and who's going to win search and social network and advertising. It's not the enterprise. It's it does create models that are used by the enterprise and it is a great place for us to learn what technology will be for the enterprise. In fact, the products we're building to so solve computing problems at X are exactly the products that we're shipping to customers in the enterprise. Which brings us to the third market which is just emerging even though it's probably
• 04:30 - 05:00 the biggest which is the enterprise. It's the place where you actually take these technologies and apply them to do something interesting in an enterprise. And so that's the first part of the elevator pitch. The second is well okay what is that enterprise market? How do you define that? Is it just random use of AI or is it something more specific? And so I created, you know, this kind of very uh I'll call it accurate but boring definition, but I think it's worth having. And that is that the enterprise AI market is the application of artificial intelligence against your most impactful processes in the most
• 05:00 - 05:30 important parts of your business to improve the product productivity of your organization. So I mean that's very it's kind of a long definition, but what it says is this is not about random acts of AI. This is about knowing what will change the trajectory of your business from a process perspective and then applying AI to that to make you more profitable, more scalable, more efficient, more competitive and and that's a very different discipline than that second market which is we're going to go build AGI and ASI and so you know
• 05:30 - 06:00 it's a little less sexy but that's where the money is. And and interestingly enough we're only really just starting to see that market ramp up now as because candidly the first year we didn't have any tools. The second year it was a do-it-yourself project and only this year do we have enough standardization off-the-shelf technology archetypes to follow. But honestly, that's the thing that's going to transform the industrial base of the world. And you know, that's what we're entirely focused on at Dell. You know, I love that explanation because I, you know, I teach a generative AI architecture course and one of the
• 06:00 - 06:30 things I tell my students, the reason we're doing this is we're creating a strategic innovative differentiator for the business. That's what we're doing. We're trying to make back, we're trying to make value come back from the business. If you're trying to, you know, create uh, you know, create the next, uh, you know, autonomous, you know, autonomous engine that's going to do everything, you're in the wrong place. We have to find these use cases for it. And typically, those are going to be tactical use cases. It's going to be very small successes and how you build things up. And that's why I love I love you guys have a very pragmatic uh,
• 06:30 - 07:00 explanation for it at Dell. I love that. So, let's talk about build versus buy versus compose approach for enterprise AI. and you know ultimately the best path for business and you know how you're teaching the businesses and teaching the organization you know conceptually how to use this stuff to get to a successful instate. Yeah, there's kind of two parts of that story. The the build versus buy versus compose paradigm is something that's been around for a long time. you know, you have a choice about how you implement technology. And and funny enough, in the Gen AI space for enterprise, you
• 07:00 - 07:30 remember we're in year three. You know, we had year one, which was chat GPT popped out and everybody had this cool tool that completely forced us to rethink everything. Of course, we couldn't do anything because you couldn't use that tool to go transform an enterprise. And so, it was the year of ideiation. In fact, at Dell, we had 800 projects get proposed around, you know, what could we do with AI? And there's no way we could do these. And they were fantastical projects, but they but but that was very normal because year one you had this new technology that people could actually understand to some degree and the ideation went off
• 07:30 - 08:00 the charts. Year two was the first year we could actually do anything with it. But unfortunately during year two it was a do-it-yourself project. The only people that could really put this stuff into production were people like us that had immense technical resources and expertise because you didn't have anything to buy. You had to literally go build all this. you to build your own rag system, build your own load balancing architecture, build your own model farm. And while that actually produced a lot of value for us, it also creates tremendous technical debt. And it is just simply not possible to do it that way if you're a normal enterprise
• 08:00 - 08:30 that isn't a tech company. We're now in year three, which ultimately is the first time we've seen the ISV ecosystem, the standards materialize. And today, honestly, we have a thing called an AI factory, which is this approach to how to build the infrastructure. You don't have to think about how you're going to build your infrastructure. We've already figured that out for you. It is you can buy one of those. Uh even above it, what you run on it, you're be you're able to buy. There are people who deliver off-the-shelf ragbased chat bots that are enterprisegrade. Uh there are people
• 08:30 - 09:00 that will deliver you content engines, coding assistants. No one would build their own coding assistant. There's 10 different good choices about how you could choose and implement a coding assistant. And all that creates less friction for the enterprise to move forward. So back to that build by compose. We were in build and we were there last year and it was a very slow and painful process for anybody who did it but you could get there. We are now squarely in the buyin composed world where generally in the case of a coding assistant you're just buying it. In the case of a content engine it's a little
• 09:00 - 09:30 more finesse. You buy the tools. You still have to do a little bit of composition of kind of put the things together to be unique for you. But the one thing that we've seen is the commodity line just continues to go up. more and more of this stuff is now repeatably consumable by the industry as opposed to you having to do it yourself. And you know, two things I' I've been an enterprise CIO before and I I run it and I look the reality is there are two things that have to be true for an enterprise to really adopt something. The first is you have to be able to adopt it without accumulating tremendous technical debt which means you should
• 09:30 - 10:00 you really want to buy it. You want somebody else to have that technical debt. That's why industry exists. The second is you don't want to go first. No one wants to go first. We as Dell have the privilege of being able to go first. In fact, we have a program called Customer Zero because it's our job to go first before our customers figure it out so that they don't have to go first. And we're now in a year where a bunch of us, not just Dell, but a bunch of my peers and big companies have declared victory. We've gotten over the finish line. We've had ROI. We've made these things work and we're now sharing it. And at the
• 10:00 - 10:30 same time, the industry is making this more and more consumable. So that's the paradigm shift that people are going through and it it's what kind of excites me about this year because I can see in year three this is a year where actually a whole bunch of enterprises are going to get out of what we call PC prison and into production and so that's very encouraging because the effect of that back to you know what you said about your course this is all about solving a business problem if you only theoretically solve the business problem that's not very interesting but once you
• 10:30 - 11:00 actually solve the business problem something happens happens to your business and it's probably a good thing and it probably makes you more competitive and it makes the world better and generates more revenue or more impact and so getting things tipped into production and out of the PC prison is really what is likely going to start happening at scale this year mostly because the industry is just kind of matured and that's actually a good thing you know I love about your explanation is you talk about the journey in terms of how you're thinking change at Dell so you know in other words you're evolving with the customers evolving with the technology and kind of feeling their
• 11:00 - 11:30 pain and then figuring out of their p best path via your technology, you know, versus declaring success and and then and then monkeying around with trying to argue what success is. You basically evolve around the needs of the customer, which I think is something that uh a lot of uh a lot of others can can uh can can learn from. So, one of the things I do tell my class is uh AI is data data data. In other words, you have to understand where the foundation of the data is, how you're going to extract the data. You have to have your own training
• 11:30 - 12:00 data. It's not about you just leveraging LLM, you know, capabilities. It's about leveraging small language model, agentic based systems that are based on your information. If your information is not up to par, your AI engine is not going to be up to par. In fact, it has to be fixed before you implement it. So, you guys focus on data discovery preparation, you know, ultimately the implementation strategies as using data as a foundation for AI. Tell me about that. Yeah, I mean well you know fact you cannot do an AI project without
• 12:00 - 12:30 data. There's no AI without data. Full stop. And you know if you've heard that before but let me you know give us some examples. Um you know in our use cases inside of Dell we picked four big areas that are now fully in production around sales services uh supply chain and engineering. The path and pace in which those four emerged directly correlated to the quality of data that we had. Um the places where we had very high quality data that was very much in a modern state, we moved fast. The places
• 12:30 - 13:00 that we did not, we had to fix the data before we could move fast. You know, in places like our coding assistant deployment, uh our source code repositories are quite good. Our our IDs are quite good. There wasn't a data problem. So, we could literally take an off-the-shelf tool and implement it into our environment and and fundamentally be up and running and have productive engineers almost immediately. In the case of things like our content, not to pick on ourselves, but yeah, we we learned a lot of things. We realized that content, you know, there's a lot of
• 13:00 - 13:30 content out there and it's in all kinds of different formats and different repositories. and the idea of being able to take uh and use AI to make it easy for our sellers to access and use that content to prepare for sales calls to be in front of customers which is a enormous opportunity like 40% of their time was that in order to do that if we didn't have our content in order if there was like bad content outofdate content and we wrapped an AI system around it that would not end well and so it required us to actually do a whole bunch of content preparation work before
• 13:30 - 14:00 we actually went into the cycle of deploying AI But the other thing that happened which is very interesting I I I'm I'm both the CTO and the chief AI officer in the CTO role about three or four year and actually maybe four or five years ago um we started look we've always been on a data improvement journey but the bottom line we realized that we had certain areas of data that would be really important to us in the future that weren't in a good place and one of the ones we identified was telemetry. It was you know we build products those products emit telemetry we use that
• 14:00 - 14:30 telemetry to service our customers to improve the products. It's a very powerful tool. It's very valuable data. It's very big data. There's a lot of it. Uh and what we realized it was quite fragmented. It was not necessarily in a consistent state. So we began a journey way before Genai to improve telemetry data to basically get it into a common format. We we adopted a data mesh architecture. We decided everything would be a data product. We started to consolidate this stuff. We actually took a lot of cost out by not doing it in 12 different ways. And ultimately the benefit of that is that because we had
• 14:30 - 15:00 already done that when we started looking at our services interaction of how we would serve our customers better what we realized is that many of the AI projects in an enterprise cannot be done with just generative AI. Even if I because normally the way people approach services is they say I'm going to take a uh all of my service information and throw it in a vector database and then I'm going to put a chatbot on top of it and that's going to become the new service interface. That's great. You should do that. It could be even better if in addition to all of the static
• 15:00 - 15:30 content that you have in your vector database, the actual real-time telemetry of the product the customer is using is part of that experience. But if you haven't fixed that telemetry data and gotten into into a state that you can apply machine learning to to extract the information from it, you can't do that. Well, we did that and the result was a dramatic improvement in the efficiency and effectiveness of our services engagement across our entire portfolio. And so, you know, the the moral of the story is look, the you're if you start an AI journey thinking it's all about
• 15:30 - 16:00 the compute and the AI side, and you don't really take seriously the data side, you're probably going to run into unintended consequences and failures. The bonus prize though, which is very interesting because when we started the data modernization journey, we looked at all data in the company and said it all has to be modernized. So if I tell you, okay, I would like you to move that mountain from there to there with a spoon. That is a incredibly daunting proposition. You can do it. Give it enough time. But where do you start? And
• 16:00 - 16:30 that's kind of where we were at. Telemetry was an obvious one, but there were thousands of different areas that we could modernize our data. The one bonus prize of AI is it has given me a prioritization vehicle. So that I haven't modernized all my data, but today my priority is any data that is a dependency for a critical AI project is at the top of the queue. I know how to do it. I have a data mesh. I have a data product architecture. I know how to modernize it. But instead of me sifting through my hundreds and hundreds of pabytes of data and just randomly trying
• 16:30 - 17:00 to modernize it ad hoc, I have this very clear view of the data that's most important to my AI AI outcomes has now become the priority data to modernize. And that's allowed me a sense of kind of relief that I don't have to do it all. I have to do the stuff that matters to AI first and the other stuff I can leave alone for a while and ultimately that allows me to create better value. So everything I just said justifies what you were saying. The the intersection of data and AI. These are completely inseparable, complete dependency and any
• 17:00 - 17:30 AI strategy that doesn't fully contemplate the data dependencies and data strategy is incomplete. It's that simple. Yeah. Yeah. Yeah. And it's the primary cause for AI project failure that I'm seeing out there is it's the data data data. And it's something I think people are reluctant to fix because nor sometimes the data is something they don't want to uncover and actually have to, you know, get into restructuring things and cleaning the data up and dealing with data hygiene, but it's absolutely necessity before you're successful with AI. So, I love I love the way you said it. So, you guys
• 17:30 - 18:00 use patterns across use cases to identify some of the critical AI capabilities that you're looking for. Um, I love that. Tell me about that. I'm gonna learn something. Yeah, that was fun. Um, so I think every company I've talked to has the same scenario we had when we started this journey and that was, you know, Gen AI happens and there's hundreds of ideas and those ideas are small ideas to I want to build the holiday. You know, it's just it's crazy stuff. But it's you have this enormous set of ideas that people think
• 18:00 - 18:30 you could use this for and many of them are perfectly valid. Now, I already talked about you have to prioritize them and you have to pick the right battles. But what was interesting is even though we had 800 theoretical use cases and we didn't choose to implement them. We basically flipped it around and chose what we'd implement top down in these four big areas. We still kept all that data. We're very interested in what ideas people have. But one of the things that we did is we let the data help us inform our infrastructure choices because our goal was not to have 800 AI projects with 800 different tools. That
• 18:30 - 19:00 would be a disaster. And so we actually went through those 800 sometimes we we say it's 360 because we could kind of distill them down but hundreds of projects and we asked what we used a technical team to ask one simple question. What fundamental AI technology do I need to have in place to do what this use case needs to make that happen even if I'm not going to do it. It was interesting to see what would be the answer. And so as you go through, you know, you start finding there's patterns. And it turns out, you know, after doing 360 of these things or 800
• 19:00 - 19:30 of them, a pattern emerged that there were like five capabilities that you needed. And, you know, you still have to implement them and figure out exactly what you're going to do, but you don't have 360 ways to do stuff. And those five in no particular order are pretty straightforward. The Swiss Army knife tool that you must have, if you're an enterprise today and you don't have one of these in place, you're doing it wrong. And that is a ragbased chatbot. It is a a a generative system that can actually access retrieval augmented generated data which is basically data that has been vectorized that is
• 19:30 - 20:00 accessible to a large language model to ultimately combine with ad hoc in real time to interact with users and really it's just an engine that converts proprietary data into generative experiences chatbots other services that became I would say 50% of the use cases that's the only tool you needed to do it's like the building block and you know if I told people 101 AI I from a technical level learned that get good at that. You can solve a lot of problems with just that one tool. But there were more. The second one is we realized that
• 20:00 - 20:30 coding assistants even though they look like chatbots are different. And they're different because they interact with IDEs and they interact with repositories. And that was a if you have an environment that does software development, you need a foundational coding assistant in place. If you don't, you're doing it wrong. And so we have tens of thousands of engineers now using coding assistants. Fantastic impact they've had and only getting better. So second foundational tool then you get to the third one uh content you need content engine or a content engines and
• 20:30 - 21:00 the problem with it is is that there isn't like one universal content engine today like coding assistance you can buy one and use it for everything. Content engines are a little different because there's kind of two halves of that story. At the bottom, there are some really uh good companies that focus on content curation and content management. And many of them are adding significant amounts of AI to those tools to really help you get your content under control and kind of act as a future repository and a platform. But what they don't typically do is emit the content in ways you can use it. And so above it there's
• 21:00 - 21:30 all these kind of point products that are really good at, you know, taking your content and merging it into a marketing workflow or a sales workflow or something else workflow. And some of those tools take content and synthesize it into video and and so it's still a little bit of a wild west, but what we did say is we need a strategy to kind of limit that to have architecture for content for AI as one of the foundational tools. Then we got to the fourth and fifth, which were the ones in front of us, but now they're very real. The fourth was we now saw that agentic
• 21:30 - 22:00 or the use of agents which is a very different approach to AI than a traditional firstgen reactive AI system like a chatbot uh were emerging quickly and they were likely to be a very complimentary and powerful tool. We can have a long discussion about it at Gentic. We probably will in this conversation. Uh but in order to do that you got to have a platform. Agents are not built out of random stuff. They run on a platform just like a ragbased chatbot is a platform and you're going to need something to do that. And then the fifth one which is very interesting
• 22:00 - 22:30 is is you know you mentioned that we build our own models. We actually don't we've never built a foundational model at Dell but what we do do is we synthesize models by using things like retrieval augmented generation and a little bit of fine-tuning to kind of bring some of our sensitivity into open models like what we do with people like mistrol and and meta and others. Uh and but what we started to see happen as we looked at these 800 use cases is some of them actually started to convince us maybe we do have to have more of our own DNA in these models. And so the the way
• 22:30 - 23:00 to do that is something called fine-tuning which is where you take an open model and you distill into it a set of your proprietary data and now your model has become that open model has become your model. You're probably familiar with the phrase cattle and pets. Uh c you know cattle is you have cows and you can replace any cow. You don't really care about the cow. It's just a cow. Uh and then as soon as you give the cow a name, it's your pet which means you own it for the rest of its life and you have to care for it. And so that's what happens. You know, if you use an open model and you don't change
• 23:00 - 23:30 it with something like rag, you can change from one model to another and it's not a big deal. The minute you fine-tune it, hey, you got to refine tune the next model to keep that thing continuous. And so now it sounds like a lot of work, but when you start talking about small language models about Agentic, the idea of having a self-contained model with additional proprietary information in it is very powerful. And so we're actively we just rolled out our first internal platform for Agentic. We are evaluating platforms for for fine-tuning. And but but the
• 23:30 - 24:00 moral of the story, and this is something I tell a lot of people to do, is if you have a big long list of theoretical workloads and ideas, don't throw it away. It's very valuable because it gives you a hint of the art of the possible and then the technique we used was okay let's just go answer the question what do I need to do each of these each of these things that people want to do and does it require a onetoone bespoke infrastructure or is there a pattern and with us the pattern told us hey you know what I have to build and put in place in my infrastructure to do everything these five things if I have those in place
• 24:00 - 24:30 that's the foundation I still have to do stuff around them but I don't have you know 500 different ways of doing this I don't have 500 tools which would be completely untenable and so so that that was the approach by the way you know back to learning we're pretty self-deprecating we did not know that when we started this journey we just kind of discovered it as we went through it and we had this open question of what do we have to do to build out our infrastructure to be future proof and it turns out that we had a data set that allowed us to figure that out by looking at that data set and ask asking a pretty
• 24:30 - 25:00 simple question but asking it 800 times and seeing what it turned out so it was a you know that by the you're in the future maybe I just ask an AI system to do that for me but we did it manually that first time around. Yeah, I love the fact you guys are using uh patterns and then finding common reuse reusable patterns across these different architectures. You know, one of the things that drives me nuts uh when I was working you know with large enterprises on their AI AI frameworks is that many of them were reinventing the wheel. They weren't sharing common
• 25:00 - 25:30 patterns. they were basically solving the same problems and that was a waste of resources and a waste of time and the ability to kind of combine them and have a common pattern repository that people could share was absolutely a gamecher for them. So what about the uh the uh you know looking at some of the projects out there industry challenges and adopting AI you said you were self-deprecating you you guys are telling us about your journey and uh you know where you you know where you
• 25:30 - 26:00 changed uh changed direction and pivoted things like that I love that so what were some of the lessons learned from your journey including the risk of overcustomization uh in and uh fleeting platform life cycles love to hear about that yeah we I mean we've just learned so much uh by being customer zero by being willing to go first uh you know I wouldn't recommend it to everybody but it's it's a fun adventure if you can fail forward faster you know that's kind of an important thing um the the several key
• 26:00 - 26:30 learnings I think are informative the first was it is very hard to have a coordinated effective AI strategy in a modern enterprise bottom up it's almost impossible And the reason for it is these projects require significant resources, people, dollars, and they they and and if they do not produce a return on that investment, they drain your resources, they slow your company down, and they actually stall your ability to do real things. And so, so we made a decision early on to do this top
• 26:30 - 27:00 down. The reason I exist in my role, I'm a president of the company. I'm the CTO of the company. I I I I am in a position where I I can make a decision. I can try to help make sure that decision actually happens. Uh it was because we knew that even though there were abundance of good ideas, that wasn't what we were trying to do. It wasn't a volume thing. This was quality, not quantity that we were going after. And in order to do that, you know, doing that bottom up, everybody thinks their project is special. And if they have the right to implement their project because they see
• 27:00 - 27:30 it's special, you will run out of GPUs very quickly and have no idea whether or not you'll be successful with the right projects. In fact, I had this moment early on in the process where I kind of had this, you know, this bit of terror. I'm like, you know, if we don't change the way we operate, the way our IT structure was set up is it's a service center like most IT groups are. A business unit comes to them with a project and dollars and they just do it. And we had a finite number of GPUs. Remember, there was a GPU shortage. It was we'd rather sell our GPUs to our
• 27:30 - 28:00 customers than use them oursel. And so these were not populated resources. And I had this fear that like the first business units, maybe the first three would show up with projects that nobody actually really wanted to do, but they had a budget and the resources to do it and the IT group would just put them into production and the fourth project that we didn't have any capacity for now would actually be the one that transformed the company and it never happened and we basically stopped. We basically said we h we can't do this. We have to do this top down and so the first learning was well maybe somebody
• 28:00 - 28:30 can accomplish this bottom up in a very small company or it doesn't work. It's going to be very slow and you have to be willing to take a top down pos position which means you're not going to make everybody happy. You're going to have to prioritize winners and people are going to have to wait and ultimately that that decision I think has has paid off for us. But it was a very hard decision and it's been a very difficult thing to implement in terms of just the culture of companies. People are held accountable. You you're supposed to be autonomous. That doesn't work in AI. You have to have a prioritization vehicle and you can't do that without a top-
• 28:30 - 29:00 down structure. That would be number one. Um the second is really that you have to be willing to change you know very quickly your your core principles to some degree. So think about this you know in year two of the AI cycle we built everything ourselves and they were decent implementations but we also knew that we could not maintain an advantage building this oursel competing with the market that is out there that does this for a living. We are not in the business of building ragbased chat bots as the software. we'll build the infrastructure
• 29:00 - 29:30 underneath it better than anybody else. But that package runs on our stuff. It's not our stuff. And so we we but but we had to pivot because honestly you had a lot of people that were very proud of their work and and they did good work. But then the world changed and you moved from the build cycle to the buy cycle and you know our CIO and I got together and we said look it's now time. We just have to declare we are no longer in the build phase. We cannot do that. our bias will be to buy off the shelf and implement that'll allow us to move faster. But that was an example of and
• 29:30 - 30:00 and and we will probably have to do this several times because fund and actually we have I won't share all the details. Several times we've had to rethink our core principles. We had a we had a strong opinion and then the conditions change. Now we have a principle at Dell that is very good. We are data driven. We we use the data to inform our decisions and if the data changes we will change our decision. We are not stubborn in that regard. And so this was an example where we really had to walk the walk on that that it was it was unacceptable to be stubborn because even
• 30:00 - 30:30 if you thought you were right six months ago the conditions changed so fast in AI that you have to be flexible keeping that northstar you're trying to make the company successful you're trying to be efficient. Um that was number number two on the list. The third obviously was uh that you know you have to have a very clear picture of why you are doing this. And what I mean by that is you're not in the and we talked about a few times already, you're not doing this because it's interesting technology. You're not even doing it to learn. You're probably not
• 30:30 - 31:00 even doing it for happiness and goodwill. You are doing it because you want something to happen to your company that actually means something. If you're us, it's increased profitability, increased revenue, decreased cost, lower reduction of risk or lower risk in our regulatory space. If you're a university, it's better students, it's better faculty, it's better educational outcomes. But if you don't have clarity about the thing that you're actually trying to do without that northstar, every AI project is kind of unobjective
• 31:00 - 31:30 in in its ability to kind of assess value. And so, you know, but once you got there, you got there. The the the fourth learning, and there's a million of them, I'll stop at four. These are highlevel ones. The fourth learning was this this thing that if you're going to do a top- down approach and you're going to strategically implement AI, you have to realize that while there will be people who get to do AI and people that don't get to do AI, you have to communicate that to the entire company. Even the people that you're telling them
• 31:30 - 32:00 no, they need to understand why. And so, as a good example, we did a I do a quarterly review broadcast. After every quarterly earnings call, we do a update to the company. And I typically do like a two or three minute what's going on with AI at the company. And the first couple of times we did it, it was more like status updates. Here's what's going on. These are the projects. You know, the last one I did, it was after we had completed our roll out of Dell sales chat, which was the fourth of four. We had picked these four big areas. We had identified very clear projects and we had implemented them and they were
• 32:00 - 32:30 rolled out that literally a vast majority but not everybody at Dell had transformed the way they worked and had a material impact because of AI. So success we got the fourth one out. And so my message to in that quarterly review broadcast is I I thanked a bunch of people. I thanked the people who built the products and did the technology. I thanked the users of these projects because they're now putting them into production. And then I thanked everybody else because if you weren't one of those first two groups, you still played a role. You let us do this by
• 32:30 - 33:00 allowing us to focus. Even if your job was to not do something, you were still part of the strategy, which is we have to prioritize. And if I do 800 projects, I would have been an inch deep and a mile wide and accomplished nothing. But by picking the ones that mattered and getting the whole company to accept that we're going to start with these, that will be the win. And we'll get to you later. We'll get to you when it makes sense. But you're still part of this process is critical. So it's, you know, it's a it's a this is a transformational technology in industry and you got to
• 33:00 - 33:30 think of it that way. And a lot of our learnings weren't purely around the technology. We learned a lot there, but it was more around this this psychology and the culture and how to get people working together with one goal. You there any AI project that doesn't get into production uh that's meaningful is is a travesty. And so, you know, and we use this phrase, you know, most of the enterprises in the world are in PC prison. They are still kicking the tires. They haven't tipped over into production because they can't get over that line because there's too many
• 33:30 - 34:00 things going on. It's too complex. They don't have prioritization. And our guidance and what we ended up going through is the exercise to get out of PC prison and get into production. I will tell you once you're on the other side of that, life is way easier. You know, it's really hard to go back to the boss and say, "I need more money." And they say, "What have you done?" Oh, nothing. Just a bunch of projects. versus, hey, we need to make an investment. And by the way, I've already put gigantic impact on the P&L of the company by getting a bunch of things into production. Which dialogue do you want to have? And it all boils down to getting something into production that that matters. If you do that, things get
• 34:00 - 34:30 easier. Yeah. I always I used to tell uh, you know, tell people the the job of a CTO is to learn how to say no. Yeah. And you know, I had half a dozen CTO roles, very similar background to you. Um, and it's uh it became kind of a uh an executive guide to success and and unfortunately you have to be the designated buzzkill sometimes but it's very important that you set expectations and put people on the right path. At the end of the day you're trying to configure the resources you have to
• 34:30 - 35:00 bring the best benefit to the business in this case Dell. And so it sounds like you guys have figured out how to do it. So tell me about uh Dell AI factory framework and how you guys are using this to uh to pump out some good good systems. Yeah. Yeah. It goes back to that you know as we enter year three of you know AI generative AI you know one of the levers that's improving is that customers don't have to figure all this out themselves. The AI factory is is roughly a year old. We we came up with
• 35:00 - 35:30 that term and launched it about a year ago with Nvidia and and it's now kind of propagated. Everybody's building an AI factory which is kind of gratifying and interesting but um anyway punch line is uh what what it was was it's not a product per se in the sense that it's a specific product even though they are offers you can consume these but really it is this approach to if I could basically define the infrastructure necessary to achieve an AI outcome to a point where you could consume it as if it is a product and even though you know
• 35:30 - 36:00 there are multiple ways to do this we have a we have several different AI factory templates some with that are open, some that are all Nvidia. Uh, and and obviously there's choices in the world. We're big believers in open and choice. But the idea is imagine if that line, you know, think about this is the surface area of doing AI and it's a big journey. Lots of things you have to do. We talked about data, get your data stuff in order, and then there's this build an infrastructure. That's a lot of work. Storage, compute, networking, configuration, model farms. And then above that, you have to figure out what
• 36:00 - 36:30 AI platforms. Then on top of that you have to figure out basically the actual workload and on top of that all the people and politics layer you just got a lot of work to do. And so our vision was what if we move the bar up where everything below the line we just kind of take care of for you. You you still buy it and you can choose to run it onrem or collocation or wherever you want but all that hard work of putting the system together to get to a point where you have a stable resilient platform that's comprehensive is just not something you have to think about anymore. It's something we take care of.
• 36:30 - 37:00 And by the way, there's some pretty complex things in there like how do you incorporate direct liquid cooling? How do you deal with, you know, uh, new network topologies? I mean, these are very resource intensive systems. How do you make them efficient? How do you make them fit into your environmental footprint? How do you deal with your data center strategy? All of those things have gotten absorbed into this concept of an AI factory. But the way the customer should see it is it's a way to consume or choose how to build out your infrastructure as if you're just buying a product. Even though it isn't just a product, it's got lots of pieces
• 37:00 - 37:30 in it. And I I feel like that's the roles and responsibilities of companies. We are an infrastructure provider. We provide infrastructure technology. We could provide it at the atomic level or we could apply it at a much higher level. We think applying it at a much higher level is an accelerant for most customers that they now have an ability to shift their work upstream. maybe figuring out what application they're going to run or what business problem they're going to solve and not spending all their time trying to figure out how to connect to a a switch to a storage array to a computer. I mean, that's the
• 37:30 - 38:00 AI factory story. And, you know, it's surprising how it's resonated because it's really allowed customers to stop thinking about doing things that aren't really high value. They have a prerequisite to have advanced, very advanced, very sophisticated infrastructure that is different. AI infrastructure is not the same infrastructure you have. It's almost a green field environment, but it's so complex that if you tried to do it yourself, you would spend a tremendous amount of time and energy on things that aren't necessarily value added. The minute the factory story came out, people stopped thinking about that. They
• 38:00 - 38:30 just said, "Okay, I can take for granted that once I figure out my workload, I will be able to consume and implement an infrastructure that, you know, hey, those folks at Dell have figured out for me that's pretty comprehensive and kind of covers all the bases and that allows me to move faster." Remember the the the one of the principles in AI is value creation. The other is speed. Even if you have the right value proposition, but it takes you five years to get there, it's too late. And so the speed angle the AI factory really does address because it basically takes a whole bunch of the work necessary to build the stack
• 38:30 - 39:00 and shifts that to us. In fact, early on in the dialogue, I said our purpose in this is we are shifting the intellectual burden to us and away from our customers on this AI journey. the more we do that and the things that don't differentiate them that allow them to just get started the faster they will be able to move and that is our responsibility in our job and a year into it it's been pretty successful I think we our last earnings call we have 2200 enterprises that are adopting these technologies and clearly we have a lot of the big uh you know giant uh farms that are building
• 39:00 - 39:30 foundational models using our technology and I think we did 10 billion plus in incremental AI revenue last year and announced that we're probably going to do 15 billion this year. So it seems to be working but the real model is look AI runs on infrastructure. Infrastructure could be this ad hoc collection of randomness or it could be curated and if it's curated let's turn it into AI factories and let's make that real. The other half of the story is though that you know I like the phrase that like Jensen and others will use is you're doing this not because you're just
• 39:30 - 40:00 building random infrastructure. You're doing this because the future of the enterprise actually requires an infrastructure that powers AI and that means that you are building a factory that really prod is able to deliver tokens which are the unit of currency of compute in this case and act as a platform to enable this new class of applications which are really the transformational part of your enterprise. You're building a factory under your enterprise to power the AI that is going to make your enterprise different. And so there's all kinds of, you know, terms you can use to describe
• 40:00 - 40:30 it, but at its essence, it's simply saying, we're going to just absorb the complexity of infrastructure building into Dell and our ecosystem and just allow you to move faster. And that so far has worked pretty well. Yeah, I love that. The ability to kind of blow by the complexity and uh, you know, get directly to the solution. I think, you know, I'm working with a lot of enterprises are walking around in circles right now trying to figure that out. And I think it's not necessarily an unsolvable problem, but they're kind of making it unsolvable because they're trying to, you know, go through it in whatever incremental ways and they're
• 40:30 - 41:00 trying to, you know, get it wrong before they get it right. And I don't think you necessarily have to do that. All right. Put on your future hat uh there. So, let's talk about the future of AI and Dell's role. So what's your p perspective on AI the evolving grow of enterprise transformation and what advice would you give to enterprises today that they should be focused on? Uh no pressure but okay thousands of people are listening to this. Well let me let me um let me give you a disclaimer there. Uh I will not predict the future in AI
• 41:00 - 41:30 more than two years out. Um and I won't do that. And by the way it's it's kind of a IQ test. If you find anybody who wants to predict five years out in the AI world, I don't you probably shouldn't listen to them. It's it's not possible unless it's a very generalized statement. Um this technology changes quite fast. Most of us don't believe we can see more than two years out in the future. And and that that that's kind of humbling because we we're used to being able to see a bit bit longer. Um that being said, let's look at that two-year cycle. So and so everything I'm going to say is kind of relevant to the near the
• 41:30 - 42:00 what we would normally call the near term, but the future of AI is literally next year or maybe the year after. that's the distant future. Um, a couple of principles. The first is look, we most enterprises in the world are not in production. And so, but we're seeing very good signs that say within a year the vast majority of enterprises will probably be in production with real AI workloads that produce real ROI. That is a profound thing. If you are not one of them or don't have a plan to get there, you probably are going to find yourself at a pretty strategic strategic disadvantage uh within the next year.
• 42:00 - 42:30 So, that's there's a sense of urgency there. You know, you're not alone. there's a whole bunch of people tipping into production, figuring this out right now, but it's going to happen pretty quickly. The second is from a journey perspective, um there are two things that are occurring that I think are positive from an enterprise perspective. The first is that the amount of capabilities that enterprises are now able to deliver as private, meaning that they have control of the stack there. There's two ways to consume AI today. One is through an API to some blackbox
• 42:30 - 43:00 of magic, and the other is that you can actually define the tech stack. You can choose the model. You can you have control of it. By the way, both are useful in the enterprise. There are plenty of things where I can just use an API and that's fine. But the really important stuff, the things that differentiate you have really confidential information and there is an ability to create sustainable differentiation if you can control which model you're using, if you can fine-tune, if you can control the data. Uh two years ago, there weren't a lot of tools available. Now there's more and more. And so for instance, clearly we've
• 43:00 - 43:30 seen a lot of off-the-shelf tools like rag-based chat bots and coding assistants. But even yesterday at Google's big event, they announced that, you know, Gemini coming on prem and private instances of it. That's that was not something people were expecting. Uh Sam Alman talked about, you know, open AI. Yep. They're going to probably have some open models coming up. They wouldn't say that six months ago. And so all those are good indicators that people are realizing, you know, if you want to play in the enterprise. Enterprises are not willing to give up their data. They're not willing to use your infrastructure without
• 43:30 - 44:00 understanding it. And so if you want to play, you got to play by enterprise rules and that means you got to be transparent. You got to be controllable by the enterprise. You got to meet the enterprise criteria. That's a big shift. We also saw that in the startup ecosystem. A year ago, all the startups are consumer startups. Today, when I go to Silicon Valley, almost all the startups I talked about are enterprise focused because they're realizing that's where the money is. That's where the real problems are. There's a lot of value there. And so so fast forward again out over the next year more and more of the AI ecosystem is pivoting not just to talk about enterprise but to do
• 44:00 - 44:30 it in a way that we actually could consume which is a really good thing. The third thing which is probably the biggest and profound prediction that's happening is around something called agentic. And I I won't I won't really go deep on agents other than to tell you they're very different than a chatbot. And we'll talk a little bit about that in a second. But the big thing to see is on the end enterprise journey and what we're actually trying to do, the real goal is to apply AI to your enterprise to create differentiation. Turns out there are two sources of differentiation
• 44:30 - 45:00 in an enterprise that matter over the long term. The first is your proprietary data and first generation AI systems did a great job. A chatbot is a thing that literally extracts proprietary data and makes it generative. It unlocks your proprietary data in ways you could never do before. Same thing with a coding assistant. And we we've got that one covered. But the second dimension that makes an enterprise differentiable is the unique skills in the enterprise. The things your people know how to do that other people don't know how to do. So your proprietary data and the things you
• 45:00 - 45:30 do better than other people make the enterprise unique. Well, we got the first one covered with chat bots. The second one is not a chatbot. It's an agent. Because agents are this digitization of a full skill. They're not just a tool that a human uses. They're literally moving that skill into a computing environment to run autonomously. The skill can be very narrow or the skill can be quite broad if you it can be really broad if you believe in AGI and ASI. But in the near term, the fact that we are just beginning this shift to actually start
• 45:30 - 46:00 to think about not just unlocking our data, but actually taking the unique skills that our people and our company has and making them a function of the compute environment means that we can no longer just unlock the data and get value. We have an ability to scale our skills at a digital speed. You know, imagine being able to spin up a skill that used to require hiring a person and training a person and to be able to do that instantaneously. That's what agents do. And we could spend hours talking about agents, but the bottom line is predicting the future. That is clearly
• 46:00 - 46:30 coming. We are doing it already. Over the next year, we need to start thinking about the fact that the evolution of AI in the enterprise isn't just about unlocking data and creating a new set of tools. It's also about a fundamental shifting of work of where will the work happen and the work can happen to the point that it is in fact an autonomous entity doing that work. The beauty of it is is that just like everything in in in the enterprise world, there's gray areas. We're not going to put all the work into agents. We couldn't do that if we tried. But we're going to have
• 46:30 - 47:00 collections and fast forward to your workforce a year from now. It will be hybrid. You will have agents working with people in ensembles collectively making your business better. And that is incredibly powerful. By the way, those agents might actually be using the same AI tools that your people use today. It's different, but it's comp complimentary and important. So those are kind of three big things to think about long term. We could talk about quantum computing and a bunch of other stuff, but but the bottom line is, you know, the future of AI, the distant future is a year from now. And so it
• 47:00 - 47:30 requires us to be pretty aware of what's coming at us. Uh the only bonus prize is look, that might sound terrifying, but hey, you're in a big club. Everybody is trying to keep up with this. No one has won yet, and it's going to be a journey that we all go on together. Well, John, we just got about four hours worth of content in 45 minutes. So, it's a it's awesome. I I love the I love the conversation, the fact you're willing to share so much in terms of how your journey occurred at Dell and how people can learn from that and you don't see a
• 47:30 - 48:00 lot of that now. People are kind of holding things a little close to the vest and I don't think for any good reason in doing that. We need to share our experiences, share our knowledge and get better at it as a group of people. So, um don't don't forget to follow John's AI insights YouTube series. uh John Rose's AI in insights. It's going to be up on the screen now as well as you're going to see it in the description. Also, check out AI services and accelerator workshop which is also up in the screen now. Links to that and don't forget to follow this YouTube channel and this podcast. Don't forget
• 48:00 - 48:30 to check out my other stuff, my 150 LinkedIn learning courses, check out my info world blog, and uh come back here and look at the videos and also check out the other content we have with some of the other Dell executives. It's awesome stuff and I think I've gotten great great compliments on just the quality of u the quality of guests on those things. So until next time, you guys stay very safe. Cheers.