Dell Technologies Inc. The Six Five Summit: AI Unleashed 2026
Review the key takeaways and the transcript of this earnings call.
Transcript
Preview the first fifteen paragraphs, organized by speaker.
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 8 to 1 or 16 to 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 2 to 1, and then 1 to 1 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 2 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.
FULL TRANSCRIPT
Continue the full translated transcript in StockNow.
Access every statement, the English original, and speaker-by-speaker history with StockNow Pro.
View the full transcript with ProCall participants
3 people spoke on this call — only 2 are shown here.
PARTICIPANT LIST
View participant details in StockNow.
Log in to see executives and analysts, their roles, and complete speaking history.
Log in to view all participantsKeep exploring