That was Sridhar Ramaswamy from Snowflake on the enterprise AI inflection point and why data is the moat. Now we are moving into AI infrastructure. Dell Technologies is making the case that agentic AI is not just an infrastructure story anymore, it is an operating layer story. What has to be true underneath for that to actually work in production is what this track is going to unpack. Matt, where does the infrastructure conversation actually need to shift for agentic AI to move from demo to production?
Yeah, it is a great question, Dave, and man, this is a really interesting topic, right? For the last few years we have been talking about training. The honest answer is that when you look at infrastructure for training and agentic in production, agentic is obviously an inference problem, and that does not look anything like training, right? Or it does not look like the inference, even we inference or we sized for even just a couple of years ago. There are a couple of things to think about, right? The first is memory becomes that binding resource in many ways. In training, it is about compute and interconnect. In agentic, inference it is about memory. An agent carries state across these long-running workflows that have many calls and call other agents, right? These live inside of KV cache.
Give it a large context window and 20 to 30 turns, and you run out of memory capacity pretty quickly, and memory bandwidth well before you run out of compute capabilities. That changes what you actually have to buy, and capacity per accelerator starts to matter a little bit more than peak throughput, right? You are not going for absolute speed. You are looking for more and more simultaneously. Memory is a real issue, and that ties back to something else. While we have been talking about GPUs for the last four years, and we all love GPUs, and they are near and dear to our hearts, CPU has come back into the loop, and it is a critical part of the equation. I do not think this fully gets priced in yet. The CPU, as I said, it becomes the orchestrator. It makes the calls.
It does all of the branching logic, calls the tools, as I mentioned, does retrieval, sandbox code execution, all of that thing, coordination between agents. It all happens, originates with the CPU. It is interesting, NVIDIA back at their GTC in Taiwan, talked about the impact of agentic on the CPU, and they talked about GPU to CPU ratios of somewhere, I think it was about eight to one or 16 -1, I cannot remember the exact number, in traditional AI. As you came into agentic as the workload of choice, that ratio went down to two to one, and then one to one in their modeling. CPU is a really big issue. The third, and I will make this quick because I know we want to get to the interesting stuff, but it is about latency, right?
When you're in a demo or you're doing a pilot, an agentic pilot, an agent makes a few calls, and it looks decent. You can see the utilization start to increase a little bit more and more. You drop that thing into production across hundreds of folks on a team, thousands of people in a company, tens of thousands, hundreds of thousands of agents working in coordination, latency becomes a real issue. It's not a capacity problem. It's a tail problem. You can't fix it just by adding nodes. I'm going to stop here because Alistair looks like he's chomping at the bit to weigh in.
You've hit a bunch of the important things as we're transitioning, that agentic and inference is very different from training, and one of the other ways that it's different is the types of data it wants and the velocity at which that data changes. With training, we work with a very big data set that is relatively static, but with inference, we want to be working with very current data and very specific data. So the ability to deliver very targeted bursts of data is really critical. The other thing is, as you are alluding to, that move from being able to demo one or two agents doing something useful to actually being able to run hundreds of agents.
That requires a platform approach to delivering agents rather than a science project approach, and that's a really significant shift that whilst you can build specific, highly tuned environments for training because it's a single workload that runs for a long period of time, when you're doing a platform approach to running agents, you need much more general-purpose, reusable, and reappliable technology that can go across all of these different agents that we might be using over time.
Yeah. Alistair, it's interesting you say that because we forget. To what you were saying, training is a sustained. You're constantly feeding data into that rack level or that farm of GPUs. When you get to agentic, it is sustained, but there's a lot of burstiness that goes on in there as well.
Yeah. Well, let's kick off the AI infrastructure track. I had a chance to sit down with John Roese, Global CTO and Chief AI Officer at Dell, on this question of what does it take to make agentic AI real when you're moving from AI assistance to a truly autonomous enterprise. Here's John . Hi everyone, and welcome to The Six Five Summit, AI Unleashed 2026. For this AI infrastructure track opener, we're exploring how enterprises move from AI experimentation to operational transformation. Joining me is John Roese, Global Chief Technology Officer and Chief AI Officer at Dell Technologies.
John, welcome to Six Five.
Great to be here.
There's been no shortage of excitement around AI, and a lot of organizations are talking about what they're going to do, moving beyond pilots and proof of concepts. What convinces you that agentic AI is really ready for production?
Well, a big thing is that I'm actually doing it, so I feel comfortable that I actually have agents in production at the company. But the bigger thing is just we are seeing that pattern hold. We are seeing the technology mature. At the same time, just to be really clear, the amount of agent washing going on right now is staggering. If you ask a random person what an agent is, I don't know what they answer, but it's probably wrong. We are confusing chatbots and digital assistants and very traditional approaches to just unlocking data with autonomous agents. It is incredibly important that people realize that these are two different things. And the most important thing to recognize is we graduated from an era where almost all of the generative AI work in enterprises was about unlocking proprietary data with generative capabilities using chatbots.
That's great. You should do that. That's super important.
We've had a tremendous impact on Dell by just unlocking our proprietary data, completely decoupled revenue growth from cost structure, fantastic, worth doing. The second phase, where we're moving to agentic, is different. You're not just unlocking data. What you are doing is digitizing work. You are literally shifting work from a human being to a machine. That work may be very simple. It may be just an autonomous task, book my travel, summarize this thing, but it's done without any kind of human guidance or oversight, or it could be much more significant, clean up my CRM data, build this software for me. Those are very different worlds, and I believe that one of the challenges is people have not quite figured out that that breakpoint was pretty abrupt. It's different infrastructure. It's different technology stacks.
It's actually a different objective. One, unlock proprietary data to make humans more productive. The other, decouple human capacity from work capacity. That's what's going on in agentic. The reason I'm confident that exists is we built our first agents two years ago. We put them into production over the last year. We feel like there are enough examples of actually using these tools in very carefully targeted, specific ways, following a good governance process, that they actually have a material impact. I have agents cleaning up CRM data. I have agents writing code. I have agents doing all kinds of I have agents doing special pricing. I've found that if you find the right process and the right work, and you use the technology correctly, these do get into production. When they get into production, the biggest impact is in that first phase.
You could expect 20%, 30%, 40% productivity improvement around a task. In the second phase, when you start talking about agentic, it's orders of magnitude changes in how fast and how effectively you do the work, and that is the exciting part. We're very early, and we don't expect everybody in the world to be exactly at the same place, but it's a different era. There's demonstrable technology in place. If we sift through all the noise and are very precise about the definitions, it really does create an enormous opportunity for everybody.
Along with that opportunity is concerns when people read headlines, because often the terms task, job, work, employment, all sort of get conflated with one another.
Yeah.
When you talk about the things that agents can do that humans might have done in the past, I want to hear your perspective on it.
Yeah.
Are we talking about going in and doing the stereotypical decimation, one out of every 10 humans is no longer necessary?
Yeah. We didn't do ourselves a lot of favors as an industry with just kind of hyperbole, there's some reality underneath this. Let me try to cut through the noise. When you think about agents, they are a piece of technology that is capable of doing work, but I use that word very carefully. I didn't say a task, I didn't say a job, I said work. What I mean by that is agents are good at doing the effort to achieve an outcome in a certain category. They are designed to do work of a certain type. They might be very good at hygiene, low autonomy or high autonomy, low complexity, rinse and repeat work. They might be very good at coordination work. They're trained to do that. They might be good at productivity tasks.
The bottom line is they are a piece of technology that does work. By the way, there are other technologies that we've used before that do work for us, a word processor, any kind of automation tool, internal combustion engine. We've been on this journey of shifting work away from human beings and into the machine layer for a very long time, and AI agents are just another step in that. The reason that word is so important, though, is that we have to basically put it in the context of those other words you used. For instance, human beings do work, but they do work in the context of a job. A job is a container of work that the person does in their day. Inside of that container, there are multiple types of work.
In fact, we have done studies and realized there are five kinds of work that seem to happen in aggregate. There is productivity work, the act of doing a simple, repeatable task. Hygiene work, the act of doing a long-running below the noise floor task, clean something up, do this thing. Write code is a hygiene task. There is coordination work, make sure this process happens, drive this to completion.
There is expert work, do something that requires a specialized skill to achieve an outcome. Then there is something called human element work, which is all the stuff people do, interact with this person, influence this person, sell to this person. You take those five categories of work, independent of agents, and look at any job in the world and take it apart and ask how much of that job are those five categories, you will find some very interesting things.
You will find that every job in the world has probably at least three different types of work inside of it. There is no job that is just productivity or just hygiene because you are talking to people and you are doing other things. There are many jobs that do not require any expert level skills. That is not required. But there are many jobs that require deep human element work to be successful. The reality is, when we think about what agents are going to do to the workforce, they do not take your job, they change your job, and they do that because now they can do some of the work that is within the job as a machine. So they have removed that work from your container called a job, and they have put it below the noise floor, and they have done it for you.
The effect that has on your job, if suddenly your job included all these different kinds of work. Afterwards, it only included this much work because everything else was being done for you. What happens to the job? It does not go away. It evolves. It changes. You concentrate on the expert and human element work, and you stop doing productivity, hygiene, and coordination work is a very good example of that. That shift is incredibly important to understand because if you keep it at the job level and think agents are going to take your job, first, you are factually incorrect. There are no agents in the world that can do all of the work in a job simultaneously. They just do not do that.
There are many agents that can do pieces of the job, meaning certain types of work, and when they do that work drops below the noise floor, and the jobs that remain, the work that still is within the job, becomes your focus. Let me give you a real tangible example of how this has already happened. If you look at software development before agentic and after agentic, before agentic, we were using coding assistants. They were not agents, and what we did when we applied an early coding assistants to our engineering organization, if you think about those five types of work, turns out the only thing they really affected was the productivity piece. They took a bunch of kind of productivity pieces of being an engineer, code annotation, comments, and they made those go away. It was worth doing.
It took a lot of time away from the engineers wasting it on this boring stuff so they could focus on coding. That's what we said. They were going to focus on coding and not do the other stuff. They weren't doing productivity work, but they were absolutely doing hygiene work, which is coding. They were doing coordination work. CI/CD was still a manual process. They were doing expert work, developing architecture. They weren't doing a lot of human element work because they're kind of introverted people. Fast-forward into today where we have spec-driven development agentic coding, what has changed? Well, the productivity work's still gone. It's actually more gone. The hygiene work is gone. Hygiene is the act of coding. The reality is the agents are doing that for you.
Sure.
The coordination work is gone. The CI/CD pipeline is run by an agentic workflow, not you anymore. The only thing left for you is you are now the spec-driven development architect, and your job is now to have expertise to build the spec. But the funny thing that happened is you actually have to pick up human element skills because it turns out to write a spec, it's not just the technical parts that are in the spec, it's product requirements, it's market requirements, it's interaction with customers. So the very definition of an engineering architecture before an architect before and after was in the old world, they did kind of all of this stuff partially, and they were mostly a technical expert.
In the new world, they did none of this stuff, and they were mostly an expert that had good human interaction skills and could interpret requirements and turn them into a spec. Are those the same job? They are not.
Right.
They are the same person that has evolved, and they evolved because agents had materialized and removed part of the work from the job and forced the job to evolve. Last thing I'll leave you with, though, is if you believe what I just said, which is exactly what we're doing at Dell Technologies, exactly what seems to be playing out, one thing it really does is debunk this idea of anthropomorphizing agents. This idea that agents are part of your workforce, that agents are digital humans. They are not. In fact, it makes about as much sense to put an agent as I described it in your org chart as it would have to put a word processor in your org chart in the past. That is a very different way of thinking about it, but we think that is actually what's happening.
Because we are getting to the point that these are becoming real and we are mature about how we think about them, I, for the first time in three years, actually have a reasonable view of what the workforce at Dell Technologies looks like, what the jobs of the future are. What I will tell you is every job is going to change because all jobs have work that agents can extract and do, but the jobs that remain, and there will be lots of them, are an evolution of the jobs that exist today, accommodating for the fact that a whole bunch of the kind of boring stuff and the things that are not necessary to be done anymore disappear below the noise floor and humans shift towards the high-value work.
I actually feel like I can see through the fog for the first time because we have done this work. By the way, we did this by analyzing 6,800 jobs, by basically using agents to analyze this data, by getting a lot of empirical data, and then by going and doing it. This is not a theory. This is actually what we are doing, and it tells me that the nuance around agentic, specifically vis-à-vis things like jobs and work, is incredibly important.
Yeah. I have to say that is the best explanation of from an individual perspective, the direction that AI, specifically agents, is actually going. Beyond that, you sort of touched on this a little bit, Dell Technologies as Customer Zero.
Yeah.
Dell Technologies figuring out for itself how to deploy these technologies effectively so that you can then work with clients to help them develop. What about this idea of organizational change? Because
Yeah
Great, you've just laid out an amazing explanation for how agentic workflows will change our individual relationships with work. What about the organizational change that needs to take place?
Yeah.
What is Dell seeing, and how is Dell helping customers on that journey?
The entire AI journey for the last, let's say, two and a half years, has required significant organizational and cultural change in every company, and we are no different. It started with the first thing we learned a couple of years ago. You've got to be able to prioritize. You've got to have governance. You've got to do this top-down. You cannot do this as a suggestion box. You cannot have random chaos and hope you get to an AI outcome. It's just not possible. You have to pick your battles. You have to target them. Because of that, you have to understand your business process. You have to understand where your value is. You have to be very precise. Then you have to realize that you're not going to be able to serve everybody. I had 900 projects when I started.
We canceled them all and did about 13. Those 13 turned into decoupling revenue growth from cost structure and a significant impact to the company. That was a cultural and organizational change. Top-down culture and organization versus bottom-up culture and organization. Now we shift into the agentic era, and now we're proliferating this technology directly into jobs. All those things were kind of tools. Now we're changing where work is done, and that will require significant change. The biggest changes that are going to happen there are the ability for people to understand that the effect of these autonomous systems is not incrementalization. It is rethinking of the entire organization and the work. That's why, by the way, you used the word task. I don't use that word to describe agents.
If you narrow the agent to do a task as you define it today, you're underestimating what it can do. You're limiting it in a way that is actually counterproductive. If you want to do RPA, go find a task and apply a script to it. But if you want to do agents, you go find work. You find higher level abstractions that are outcome driven, and you apply agents to do that work. Think about what culturally and organizationally that requires. It requires you to be extremely open-minded about the fact that every job is going to change, how you do work is going to change, who does that work is going to change. That is a significant effort to go through that exercise. It is not easy to do. That's one of the reasons why we did not go and randomly throw agents all over the company.
We've been very precise about targeting them in places where we can work through these changes with the organizations. We're now at the point that we are scaling that, but we are doing it based on the description I just gave you with a repeatable framework. We have clarity about what this means. It's understandable because you're talking about fundamentally changing the structures of how a company is organized and what people do within that company. Don't take that lightly, but realize that it is going to change. If you proactively change it, what ends up at the other end is a highly productive company with people doing work that exists on top of an agentic and AI foundation that is much more efficient and powerful than you've ever had, and usually an environment that is much more fulfilling, much more productive, and much more effective and successful.
That's a great goal to go after. You don't get there by incrementalizing on the past. You have to be willing to rethink every process, be open-minded to changing every job, and changing the organizations around it. That's a big deal. We've never had a technology that did this to us, maybe since the Industrial Revolution. We have to be in a mindset that we're willing to do that. But again, good governance, targeting it, working on trying to figure it out before you go broad, but don't take two years figuring it out. All of these things are learnings that we've had, and the result is when you do them, again, you decouple human capacity from work capacity in your company, and that causes just explosive growth to occur. Orders of magnitude improvement in productivity in the dimensions where you apply it.
Let's talk about the investment in all of this moving forward. There's a saying in financial circles that if you borrow $10,000 from a bank, you're a customer.
Yeah.
If you borrow $10 billion from a bank, you're a partner. Increasingly, just the raw number of agents that are going to exist in an organization make this more of a partnership
Yeah
than a customer-vendor relationship.
Yeah.
When people are considering ROI, I is the investment.
Yeah.
How do you quantify that investment? We start talking about things like tokens.
Yeah.
The generation and consumption of tokens. How should we think about that?
Yeah
About the investment in AI?
Yeah. Let's be really clear. Nothing is free in the world. AI is no exception to that. AI is powered by compute. It is powered by data. It is powered by processing, and that processing has to exist somewhere. There is a cost to it. Hopefully everybody gets that. But when you think about applying AI now to the broad and diverse work that makes a company run, you have to have a diverse approach to this. You have to have choices. You cannot do that with a monoculture. Because think about it. If I describe, I do not know, five different kinds of work that happen in a company, and then I ascribe economic value to them, okay? Let's compare two.
If I build an agent and that agent lets the CEO of the company make better decisions in real time to guide the company into the future, how valuable is that agent? Pretty high. If it costs me $100,000 a month to run that agent, it is probably worth doing, right? On the other hand, if I build an agent and its job is to clean up CRM records, and each CRM record it cleans up has an economic value of $0.50, then while there are millions of them, I better have an environment in which it is cost effective to do that, and it does not cost $12 in tokens to do a $0.50 task. That is the spectrum that we are dealing with here.
What we've realized is the first principle of token economics in the AI agentic era is you have got to have a diverse source of intelligence. You need to have sources that range from different compliance levels, different capabilities, but also different economic strata. At Dell Technologies, I have four moving to five different fundamental sources of tokens. They have very different economics. I run open models on-prem in my data centers. I run frontier models in my data centers. I run frontier models in VPCs I control. I use APIs. I now run models on devices using things like personal agent frameworks. Each of those are not arbitrary. When you look at them from either their economics, their performance, their regulatory and compliance risk, and their functionality, they are different. Now you would say, "Oh, this is really complex." It's not. It's complexity by design.
Because when I look at a piece of work comparing that CRM agent to the agent that's powering the CEO, I now have a choice. I don't have just one answer. Imagine if you were the customer that wanted to do those two things, but you had adopted a single provider with a single set of models over a single economic model. You fundamentally wouldn't be able to do the CRM agent because it's just not affordable in that model. We've come full circle back to the hybrid architectures of the world, which basically say hybrid is not arbitrary. Hybrid is choice. Hybrid is diversity. When you're applying AI technology to jobs and work, they're affecting the jobs and they're changing the work. Is your work and your job structure homogenous? Is every job equal? Is all work the same? Of course it isn't.
Is it done in the same place? So having an infrastructure and an AI capability that gives you the ability to map to the right economics, the right control, the right framework, the right compliance, these are all things that are absolutely essential. For us I sound like a broken record. You've known me a very long time. I can show you presentations from almost two decades ago where we said hybrid is the only answer. It makes no sense to think of whether it was the cloud era, one cloud to rule them all, or a single IT infrastructure, and AI has just amplified that. I cannot imagine trying to execute AI in a monoculture, because what you're executing against is a highly diverse set of activities that make a company a company.
Yeah. No, it makes perfect sense. As I sit here as sort of a proxy for the CIOs and CTOs that I work with, you've taken me through this journey where, okay, we've taken care of this idea that no, we're not firing all the humans.
No.
We've addressed organizational concerns. Now, cost concerns, the economic model, the kind of hybridization, the idea of there isn't one fit for all functions. The thing that keeps me up at night as the proxy CIO
Yeah
is this idea that, I need to deploy infrastructure today that will support all of these things.
Yeah.
But I'm being told that we are on the cusp of the post-quantum era from a security perspective.
Right.
How do I make sure that I'm living up to my fiduciary responsibility from a security perspective today?
Yeah
but also setting myself up for success in the future?
Yeah, there's two pieces of that. Let me take the post-quantum one. That's the easy one. Look, we have done a good job as an industry developing post-quantum algorithms. To be very precise, quantum risk from a cyber perspective is very specific. It is the ability to factor prime numbers in a quantum system with the right algorithms and right scale becomes very easy. The foundation of asymmetric key management protocols in cryptography are based on that. If that math becomes easy, those things break, RSA breaks, a bunch of other stuff breaks. Four years ago, we started building the algorithms. The algorithms are available. They are starting to roll out. We still have a reasonable amount of time. There are some issues around capture now, harvest later, so it's a very real thing, but you shouldn't panic over that.
There is a fairly good top-down industry-wide effort to bring improved key management protocols into the places that matter. If your data's living in your data center and doesn't ever egress, you don't have a problem right now. If your data's flowing into a public cloud across a public interface and you're using very weak encryption protocols and key management, you probably ought to fix that, because that's a target. The reality of it is that one's easy. The cyber discussion in general, when you move into the AI era, is a bigger problem. Because just like agents are a different technology that does work in a different way and exist in your environment in a different outcome, guess what? They have a different security model. They are not exactly the same.
In fact, what we've learned, and this is one of our Customer Zero experiences, early on, beginning of last year, we started bringing the industry together to say, "We don't have any idea how to make agents talk to each other, how to make them secure. Our assumption that they're magically secure is a poor one." We spent the last year and a half working with our security partners and with the industry to try to figure out ways to standardize protocols. We now have protocols like A2A and MCP. Those are, in general, questionable in terms of how secure they are, but they're at least standard, so we can now secure them.
The biggest learning for us, though, is we realized that if agents do work, and that work is done on behalf of people or organizations, then it is essential that we be able to control them. One of the decisions Dell Technologies made last October, and we now enforce as part of our agentic guidelines, is that all agents that are running autonomously, touching our data, whether they are internal or external, carry a Dell-issued digital identity. We give them the identity. You have permission to touch my data, to interact with me based on that identity I grant you. The advantage of doing that is that if you go berserk and I revoke your identity, because tied to that identity is fine-grained access control authorization, I can make you disappear, even though you are running on a third-party platform or you are outside of my environment.
We made that decision a long time ago. That decision is starting to become fairly common because it gives you the kill switch. In fact, in the EU AI Act, there was a requirement for kill switches. If you ever wondered how to do it, I will give you the easy answer. Have a universal control over agentic identity that is consistent with your overall identity management framework, and if you control the identity of agents anywhere that they work on your behalf, you have the ability to introduce a kill switch. In addition to that, we realized there was a lot of fragmentation around telemetry. Even the difference between two types of agents. There are agents that are inherited. An agent that works for Dave is one kind of agent. From a security perspective, it uses your credentials, your authorization. It works on your behalf.
That is fairly easy to understand. You are accountable for it. We can monitor it as if it is Dave. That is an easy one. But then there are these things called headless agents, which do not work for Dave.
Right.
Dave might make them come into existence, but then they just go clean up CRM data, or they run a process, or they do something. Those are the ones that are far more dangerous, that probably represent 70% of the agents that will ultimately be deployed. Those headless agents actually were quite hard to incorporate into the security frameworks, because our security frameworks work on a human hierarchy. Remember my comment about the word processor? Agents are not necessarily going to be in your org chart, but if the way you ascribe identity and access control is to the organization of the company and the agent is not in there, you have to come up with a different way. We actually worked with partners like Okta and Palo Alto Networks and others and came up with ways to say how do you deal with headless agents.
All of this is now starting to become available and real, but the principle here is, just like as you move into the post-quantum world, something changed that you had to adapt to, post-quantum cryptography. As you move into the agentic world, something changed, agents that now do work independent of humans, and you have to adapt your security architecture for it. The good news, all of this is doable. I think we're early. I think we're ahead of it to some degree, but the one answer that will absolutely fail is to just do nothing, to just assume that these new technologies show up and your existing approach to however you've run your IT organization forever is sufficient. It is not. It requires a movement to a different kind of hybrid architecture, an evolution of your security architecture, a rethinking of your organization.
Sure
A rethinking of fundamentals of work. Nothing going on here. Just everything's changing. But if you change it programmatically, if you work through it and you use people like us and others in the industry that have done some of this stuff, it is navigatable, and the result is you disconnect this relationship between human capacity and actual work and the work capacity of the world, and the effect are things like this happened to Dell Technologies, where suddenly your revenue is doing this, and your cost structure is doing that at the same time, which has never happened in the history of business. Yet now it starts to become something that's accessible to a lot of the industrial world.
Fantastic. Thanks so much, John, for joining me.