AI Predictions: Reflecting on the Past, Shaping the Future - Dell’s AI & Us Series

How have last year’s AI predictions shaped the technology landscape, and what shifts can we expect in the coming year?

‍

Host David Nicholson is joined by Dell Technologies John Roese, Global CTO & Chief AI Officer, for a conversation on AI predictions: reflecting on the past, shaping the future. Together, they discuss the evolution of AI predictions, dissect lessons learned from previous forecasts, and explore new AI forecasts that defining the business, social, and ethical landscape.

‍

Key Takeaways Include:

‍

🔹Reflections on last year’s AI predictions: What hit, what missed, and why.

‍

🔹Emerging trends in AI: Agentic updates and the expanding influence of generative AI.

‍

🔹Anticipated challenges: Regulatory changes, talent shortages, ROI, societal pushback and more.

‍

🔹Predictions on Future Innovation: Guests discuss opportunities and wild-card predictions for AI and business transformation.

‍

Learn more at Dell Technologies.

‍

Watch the full video at sixfivemedia.com, and be sure to subscribe to our YouTube channel, so you never miss an episode.

‍

Or listen to the audio here:

‍

Disclaimer: Dell’s AI & Us Series is for information and entertainment purposes only. Over the course of this webcast, we may talk about companies that are publicly traded, and we may even reference that fact and their equity share price, but please do not take anything that we say as a recommendation about what you should do with your investment dollars. We are not investment advisors, and we ask that you do not treat us as such.

‍

David Nicholson:

Welcome to this installment of AI and Us. I'm Dave Nicholson with the Futurum Group, and we've got a special one on tap today. I'm joined with someone who audience members will be familiar with. John Roese, Global CTO, and I'm just going to say king of all things AI over at Dell Technologies. John, good to see you. Good to see you. And this is that time of the year where we get to tally up the scorecard from the prior year's predictions and then take a fresh sheet of paper and look forward. I've taken the liberty of going back and reviewing the video from last year, and let's just say you nailed every point. So instead of rehashing that, I would say to audience members, go back and watch the video from last year. It's actually pretty cool knowing what we know now. But John, anything that surprised you in 2025 in particular?

‍

John Roese:

Yeah, probably the biggest thing, you know, we predicted enterprise adoption would grow, but it was interesting how many. unplanned obstacles showed up in the enterprise and they weren't technical. They were mismarketing agentic, you know, organizational dynamics. And what we found was the journey of an enterprise on the AI path is a pretty complex one. And it's got a lot of kinds of, as you do it, you discover where there are friction points. And we found a lot more, whether it was government policy that was kind of oriented towards other stuff or organizational design issues, or or just navigating the marketing or the ecosystem. So I guess I predicted it would be a good year for enterprise, and it was. I didn't predict how many changes are going to have to happen in the enterprise journey to be successful with AI. And so we'll just work through those.

‍

David Nicholson:

Well, with that as the background, with friction points and areas that need work and change, let's get started. Calendar year 2026. We're going to stick to calendar years and not fiscal years. What are your thoughts?

‍

John Roese:

Yeah, I mean, you know, first of all, the disclaimer, these are some predictions more than what I say will happen. But these are ones that we think are worth paying attention to. There is a theme behind them all just for you know, and it's really that this is becoming real. This is not you know, some magic thing that you've never experienced will happen next year. You're already on this journey. AI is real. And it's really about the maturation or early maturation of the AI market as a complete ecosystem. So I walked through a handful of I think things that are going to happen next year or that need to happen. Top on the list is governance. We have not established strong governance frameworks even inside of enterprises. There are very few tools that people are still doing from the bottom up grassroots. Let's do a thousand projects and hope for the best. And I do feel like governance in general is going to be a big deal. in 2026, and it'll have two dimensions. One is inside of the enterprise, investment in having a structured approach to AI will become a requirement. There will be no more of this. We're just going to give everybody a tool and hope for the best. That is not going to hunt by the end of next year. The second is governance outside of the enterprise is going to align to the enterprise. This year, governance was all about trying to govern AGI in the consumer market. It wasn't about enterprise. Enterprise was kind of an afterthought, or maybe even the regulators didn't understand it. We will see a precipitous shift towards basically the external governance, realizing proper governance at a national level is an enabler for your industrial base to win in AI, and bad governance is a disenabler. And we've already seen that subtle shift in Europe and Asia with their new policies. It'll hit home next year. We're going to have a wrapper of governance around AI, hopefully the right kinds of governance that will help enterprises navigate this forward in a safe and secure but effective and fast moving way. The second is around data. We've been obsessed as an industry around the compute part of AI. You can see how many GPUs are shipping, but what we're finding is as AIs go into production, in enterprises specifically, they don't just consume data, they produce it, and they create entirely different layers. So for instance, when you move to agentic, what we've learned is you're no longer just taking your primary data and just kind of turning it into math and using it as a transaction. You are creating a knowledge layer and that knowledge layer includes things like knowledge graphs and graph databases and vector databases and other ways to express data for the AI system. You're also changing the APIs into your systems of records so they can be more friendly. and used by an agent, for instance, with an MCP interface. And so we're going to see a necessary investment in not just the compute side, but the data ecosystem that surrounds it. And without that, it will be very difficult to do enterprise AI or AI in general. And so we do think data management is going to become kind of really the backbone of definitely agentic AI and most AI innovation going forward. Compute will be somewhat taken for granted, still critically important, but more well understood. The third is around agentic in general. Last year, I think I got it right. I said, agentic would be the word of the year in 2025. I said that in November and December of 2024, no one knew what I was talking about. I think we got that one. But now we're going into another level that we've gone through this era of agentic becoming a term. As I mentioned, it's wildly misused, it's confusing, but when you settle it down into an autonomous agent built as a software system with tool use and perception and the ability to interwork with each other and the ability to have persistent memory and long-term memory and access to LLMs, when you understand what it really is, it becomes very clear this is not the AI that we started with. This is not the same as a chatbot. And so what we expect to happen is not that that will occur. We already know that will occur. Most people are getting to the point that they will understand what an agent is versus a chatbot. What will happen is we're starting to think about new ways to use it. And one of the big aha moments we had at the end of this year, which we think will bleed into next year significantly, is that agents don't play just a role as a tool. And they don't just play a role as a user in isolation. They have a very powerful capability because of the fact that they run continuously and they can keep state and they can interact with the ecosystem around them to become what we call a continuity manager. The idea that an agent may in fact allow us to do really complex tasks with AIs, not just delegating it to it, but having humans and AIs work together in a coordinated way. And it turns out the agent is the ideal orchestrator for that. That can revolutionize how our factories work, how we do long-term processes, how customer care works, and it's a very powerful tool. And sometimes I say it, one of our big problems in long-lasting processes is human chaos. We have shift changes, we have people going on vacation. We have people translating information incorrectly, and just adding tools to that doesn't take that away. But using agents as a way to kind of keep everything moving in an orderly way can actually mitigate a lot of that human chaos, and that's a significant source of inefficiency. So we think there's going to be some very big use cases where agents play a more prominent role. Action is kind of a set of scaffolding around us, an orchestration. You know, all the other stuff's going to happen too. We're going to use them as tools. We're going to use them as users. We're going to use them collectively. But we also think they'll play a role in coordinating and helping us work better with AIs than we ever have. The fourth is... Before I go to the fourth.

‍

David Nicholson:

Quick comment on agentic. I will tell you that I would say 10 to 20% of folks still, when they hear agent, they say, oh, I don't want to run any agents on my servers. And it's like, no, no, no, no, no, different conversation. But back to number two, we used to talk about this in terms of data gravity. Do you see the trajectory of movement toward or away from cloud being changed by AI? This is sort of a little detour into the infrastructure side of things before you hit it.

‍

John Roese:

No, no, I think it's a great question. I think, well, clearly it moves away from high degrees of centralization. The knowledge layer is the real-time data access of the AIs, and the AIs more and more are living closer to where the data and users are, which is out at the edge, is on the device, is in the real world. So I don't see a lot of value in having your knowledge layer centralized on the other side of the planet, because it's the transactional layer. In fact, it's interesting. All of our transactional systems today just become data sources to the knowledge layer. And in fact, you know, when you think about the resiliency of an AI system, if you took a just a rag based chat bot, and the rag was fed by your CRM and ERP systems and other databases, and then the RAG created a vector database and the vector database was used by a large language model and ultimately created the chatbot, you could actually lose all of the primary data. It could just shut down and the AI would keep running. Because once you've converted it to math, in many cases, it's not dependent on these non-real-time systems of record. So we do actually see that the knowledge layer is more tightly coupled to where the compute lives and the compute is definitely being distributed. And more importantly, it's a real-time environment. There is no concept of cold data in the knowledge layer. It doesn't make sense. If it's cold, it shouldn't be there. It's not necessary. And so I do think this will, you know, I would call it the backbone of AI innovation is going to be the AI data layer, the knowledge layer, and that belongs where the action is. Sure, there'll be some of that in clouds, and there'll be some of it in data centers, and some of it at edges, and some of it on devices, and some of it out in the real world. That is absolutely true. But the idea that it could all be centralized or that it would track with where your systems of record, your ERP and CRM systems are is nuts. That they're just data sources for the knowledge layer and the knowledge layer needs to be close to where the action is. So yeah, I think that's a guaranteed trend. Otherwise you'll end up with a very inefficient and difficult system to operate.

‍

David Nicholson:

Yeah, yeah. Makes sense. So let's hit point four then.

‍

John Roese:

Okay. Point four is actually an interesting continuity of that conversation. We are building AI factories all over the world today. We have about 3,000 customers that are on that journey with Dell and more every day. There are different stages of innovation. Some of them, like us, have fairly big AI factories running our infrastructure, and some of them are just starting. That's great. But once you build something, you got to make sure it works and it works and survives. And so I think in 2026, we're going to see a significant investment and focus on disaster recovery and resiliency for AI factories. How do you back up an AI factory? What's the best way to do that? And there's concepts out there where you don't do it the same way. If you have an AI factory with a thousand GPUs, do you need to have a thousand GPUs as a hot standby to have an AI factory? You really don't. You just need the data and the tools in the right place and you need access to potentially 1,000 GPUs when you need them.

‍

David Nicholson:

I'm sorry, John. I just got a text from Jensen Wong and he said, yes, you do. You do need access to many GPUs. But please continue.

‍

John Roese:

Well, you could. It would be very expensive. But it turns out we're actually seeing that there's going to be novel approaches to resiliency. And the biggest issue when you try to create a hot standby or a backup or a resiliency environment for an AI factory is GPUs. Can you afford an extra thousand just sitting there idly, or do you go hot, hot, or do you do something else? And something else is fascinating because there are large CSPs forming, there are sovereign infrastructures forming, there are places where GPUs are aggregating. that if you put the necessary, funny enough, data management layer there and replicated it, so your knowledge layer was there, and you put the right kind of access controls and security there, if you had a failure, there are places where in a matter of minutes, thousands of GPUs could be at your disposal. They aren't necessarily public clouds. In many cases in the AI cycle, they're very different alternatives like sovereign infrastructure and CSPs. And so we don't know exactly how it plays out, but we do know that once you build an AI factor, you become dependent on it. And backing it up and making it resilient is a non-trivial exercise. But unlike the past, where we kind of had one way of doing it, just take the whole thing and create a second version of it, you have new tools about where those GPUs will be and how to access them. So we think it's going to be a very interesting innovation cycle that will impact not just enterprise, but the CSP ecosystem, the sovereign infrastructures, the world, and lots of other parties. But the goal will be, I don't want my AI factory to go down once I have it in place because it runs my business if you do it right. It is the system of record of the future.

‍

David Nicholson:

So I have to say that in the not too distant future, what you just said is going to seem so completely obvious. But right now, it isn't because people are looking at, OK, what are we going to do with this new thing? We haven't even imagined all of those things. But it's easy for us, you and I, to think back to a time when that first time the email system went down and the CEO's email was affected, all of a sudden, it became mission critical. And yes, imagining the mission criticality of AI moving forward, it's like, well, yeah, I guess that's solid. We don't know how to do it yet.

‍

John Roese:

Yeah, well, sadly, you and I are old enough and have been in this industry long enough to know that this repeats itself. Every new technology, there's excitement in the early days, rapid adoption, broad adoption. And then, oh boy, we got a problem. We didn't think about what happens when something bad occurs and this isn't available anymore. So once you have that dependency, you got to have the resiliency. We haven't gotten to that point to be perfectly honest on AI factories on a global basis. But like I said, we have some fantastic tools at our disposal. which kind of gets me to my last of the five, which is sovereign AI. You know, this wasn't even a term a year ago, and then all of a sudden it's a term in an entire industry. And I've spent a lot of time talking to governments around the world. And we have a pretty straightforward, you know, definition that there's three different ways to do it. There's government for government, where it's the infrastructure that powers your government services. There's a government for industry, which is, you know, you're basically allowing your industry to use your infrastructure, building data centers in your jurisdiction so that the industry can move faster. government with industry, where it's more the government and industry are on the journey together to make sure their industrial base and society benefit from AI. That's all obvious. We've talked about it a lot. But as we go into this year, into 2026, I think the big impact will be we have all that stuff. These infrastructures are being built now. And based on the previous conversation we just had, there are all kinds of enterprise gaps that can be filled by this. And if a government is smart and they realize, boy, I have, I just laid out a whole bunch of GPUs and have all this capability and all this networking and infrastructure. And I could use it just to, I don't know, do my, you know, my basic government services or I could get deep with my enterprise customers, I could help them move faster. And I just see a few countries around the world already doing that. And it just is one of these things, if you got an asset, and you don't know what to do with it yourself, but you have friends and dependencies, in this case, your industrial base that actually want to move faster. But one of the problems is they don't have access to infrastructure that can be private and focused on their actual specific needs in a confidential way. That is a vacuum that gets filled. And so, we do see this intersection between sovereign AI and the enterprise acceleration being something very real, just because one side has demand, the other side has supply, they're in the same country, they have the same ultimate goal, they're going to figure out ways to work together. And whether it's to create disaster recovery, to create fine tuning infrastructure, robots, you know, world brains, agent farms, there's all kinds of things that could happen, and they only happen when you connect the dots, and the dots are already in place. By the time we exit this year, there'll be a whole bunch of enterprises doing AI, and a whole bunch of sovereignties building infrastructure, and they haven't quite got together yet, but it seems like that's inevitable.

‍

David Nicholson:

Okay, John, a lot of people have said that the race for dominance that's underway, when we characterize it as a race toward AI, in fact, the underlying level that needs to be addressed most significantly is power, electricity. What are your thoughts about that? Some have said that as of this moment, we don't really have line of sight to sources that can power this AI revolution. What have you seen in your travels around the globe?

‍

John Roese:

Yeah, I think there's a lot of angst about it. I don't know if I'm doom and gloom about it. I think if you look at AI as a linear path that we're going to do it in one way and in one topology with one set of technology, yeah, you run out of energy because you can do that math. At least you run out of it until you build more. So two things to take away from that. The first is in that linear path, one of the advantages of what's going on is we've elevated AI to be so important and so strategic on a global level that it's actually catalyzing a build out and a modernization of the energy infrastructure. We are seeing the acceleration of small modular reactors. Even fusion is getting investments that it wouldn't have gotten otherwise. Renewables are still heavily invested in, especially around the world. If you get out, you know, in certain countries, there's gigantic build outs for this purpose. And so I think we'll have a bumpy couple of years. I think it's going to be, you know, struggling to find enough energy, but in parallel to that, massive investment in building out not just new energy generation sources, but fixing the distribution problem, building out the grid, making sure it's more resilient. And if you take a long view, yeah, a couple of years of us thrashing around trying to find the necessary energy to build out the data centers, we'll probably figure it out. But at the same time, accelerated investment goes out about three years into the future, and you end up with probably a pretty good power grid and actually a pretty stable and significant diverse set of energy sources that look more sustainable long term. That's part A. Part B is not a linear path. Not everything looks like a giant data center in the AI world. These are distributed systems. You can do all kinds of things to change the energy consumption of an AI system. You can move it all the way out to an AI PC and run it locally. That's within the current energy footprint. You can be smart about how you build agents and you can build an agent that is just there to coordinate and do the logic, but uses existing tools to get most of the work done. That's incredibly energy efficient and much smaller and can be distributed. I think what we'll find is even if we do hit an energy wall in aggregate, we have a lot of levers about using small language models, distributed architectures, intelligent agents, using tools instead of creating AIs to recreate the tools. There's a lot of things that we can do. And our experience, at least on the enterprise side, has been able to work within our energy budget. But it hasn't hit a wall because we're smart about doing the right thing in the right way with AI to make sure that we sit within the constraints that we have. But long term, I'm actually pretty optimistic that the build out which scares everybody is also the catalyst that fixes the long term problem, which is our transmission grid and our energy production environments were not sufficient for whatever was going to happen. And in a couple of years, we'll probably be in a much better place. So for a few years, we'll have to be smart. And I think we can handle it. And then things will get better. And I think we'll create a long term path that's much more aligned to a sustainable, efficient and predictable energy ecosystem to power the world.

‍

David Nicholson:

Well, you've actually eased a bit of my power panic, John. That's a great It's a very it's a it's a positive view, which is likely the way it's going to turn out. You've talked about public private partnerships in the past and you've emphasized the importance of the private sector leading. But what about funding? Is this something that warrants a massive nation state investment? Of course, it depends, you know, various nation states around the world. But is this something that is enough of a common good to warrant serious government investment in dollars?

‍

John Roese:

It's almost as if the investment's already there. The sovereign infrastructure build-outs are happening very rapidly now. We have infrastructure available. It's being built by smart countries that realize that they're going to need this. And the enterprises are discovering access to at-scale infrastructure or private infrastructure or infrastructure that's highly secure. and in their country is very valuable. And so it's a weird phenomenon. Normally it's like the demand exists and then the supply reacts. This is like the supply is already moving and the demand is starting to form. But for those of us who are in the industry, we can see the two intersecting. I mean, if you think about where to spend your government money, spending it that gives you a gigantic return on investment of catalyzing your industrial base and growing your economy seems like a good thing. And I think as more enterprise use cases benefit from sovereign AI, it will create more incentive to do more of that. That may, in fact, be one of the catalysts that actually continues to drive sovereign AI forward in a way much bigger than just government services. But it is interesting. There's a lot of sovereign infrastructure out there. It's a question of what we do with it? How do we connect it to enterprise use cases? And that's kind of the inverse problem that we've normally had.

‍

David Nicholson:

Okay, well, unless you wanted to put yourself on the hook for a sixth point, Are you good with five? Because I'm going to offer you a wild card, crazy, you're not going to be held to this science fiction John opportunity in a moment, but go ahead.

‍

John Roese:

Well, I would throw a bonus one out and I feel like it's like a PSA. I have to inform you that this is coming every year and it's quantum. We are seeing significant advancements in the quantum ecosystem. And it's not just around QPUs and qubit density. Those are definitely improving dramatically. We are seeing software innovation. I saw a company that was doing a high level software abstraction and compilation that made it possible to run very complex algorithms on significantly less qubits. And so just as we predicted, this would be full stack innovation, and we're seeing it move very fast. I'm not ready to declare quantum supremacy yet, but it's coming. And when it comes, mark my words, the disruption could be bigger than when chat GPT popped up, because what it will do is give us the ability to do math at a level that we've never been able to do, orders of magnitude more efficient. And one of the primary areas that will be affected by it is the AI ecosystem. So I've always said, when quantum becomes real, whatever the state of the art of AI is that day suddenly becomes orders of magnitudes more powerful, more efficient, and better. So keep an eye on it. I'm not sure 2026 is the year of quantum breakthroughs, but there will be some. But it will definitely be acceleration, and we'll start seeing quantum utility examples. We'll definitely see algorithmic improvement. And so one day we'll have this, and we'll be talking about that quantum thing that happened, and it disrupted everything. I'm predicting that's still true. We still don't have the exact date. but it's coming sooner than you think.

‍

David Nicholson:

All right. Well, I, for one, I'm going to hope for no quantum, no quantum effective movement in 2026, because I think that I'd like to think of 2026 as the sort of we're going to stretch and warm up with AI. I think if quantum hit us right now, we'd all pull muscles and it wouldn't be good. So but what about what about kind of I say crazy, but that's not what I mean. I don't mean crazy. I mean, what if sort of predictions, what might happen? What unusual thing?

‍

John Roese:

Yeah, I think probably the biggest thing that could happen that people may think has already happened, but I don't see a path to get there is breakthroughs in the robotics ecosystem. We've got the mechanics of robots relatively worked out, and we've got some of the software worked out, but what people don't realize is when we actually get robotics right, and we use AI, for instance, to program robots, because that's the biggest expense. Programming a robot is incredibly difficult to do for humans. When I can actually take the software effort to make a robot a robot and make it autonomous, most of them are not autonomous today. If we can take what we learned with agents and apply them to robots in real time speed and do it in an automated fashion and continue to drive the mechanical costs down, the impact on our world gets really interesting because it's no, I've always said, you know, AI was about shifting the cognitive work of the world below the machine line. Physical AI brings us right back to moving the mechanical work again below the machine line. And I don't think we're going to get there in 2026 fully, but I'm optimistic that if something's going to surprise me next year, it's probably in the robotics space that people get really good at driving the cost and complexity to build robots down. And those robots become more pervasive than we thought. And they actually start to impact not just our cognitive work, but they start to really transform the physical world again in a way. Imagine if you have the AI evolution happening and the industrial revolution happening at the same time. That could be very interesting. Hopefully we don't throw the quantum revolution on top of that, but that's one that I think is more likely than not to see some surprising disruptions occur in that space. And if it does, it cascades all over the economy and all over our industries and ecosystems in a way that we're probably not fully prepared for.

‍

David Nicholson:

Okay, so no one can say they were not warned. John Rose said a robot may be walking up behind you and tapping you on your shoulder in 2026 as a possibility. Might be, might be. I don't know if we told you at the beginning of this, but we're recording this actually. Okay. So do you want to go back and change any of the predictions that you made? Not at all.

‍

John Roese:

I think we're good. Last year went well. I think I'm confident most of these are going to happen in some way and I'm pretty sure that it's going to be a fantastic year for AI moving forward and becoming much more real again, which is a great thing.

‍

David Nicholson:

All right, well, here's to you and yours as we get towards the end of 2025 and we look into 2026. John Rose, it's always a pleasure. Thanks for getting together with me and sharing your insights. I'm Dave Nicholson with the Futurum Group. This has been an ongoing series that's been incredibly interesting, informative, and I think important. Because when we talk about AI, it's not just about the bits and the bytes and the bobs. It really is about AI and us. Stay tuned. Look for installments in the future. And have a great and happy new year.

‍