‘Like Disneyland when it just opened’: Arm CEO Rene Haas on the AI revolution

RANA EL KALIOUBY: Among the many hats I wear – as an investor, fund manager, mom, podcaster – I also co-chair the Fortune Brainstorm AI conference in San Francisco. The annual two-day event brings together luminaries from across the industry to dive into the exact topics we cover on this show. I’m just back from the event and still buzzing from it. I learned so much about how leading companies are leveraging AI and where money is flowing next.

I got to host several conversations on stage, and we’re bringing those to Pioneers of AI.

This week, my stage session with ARM CEO Rene Haas. Many of you have probably heard of ARM, but even if you haven’t, you’ve no doubt used devices powered by their chips. ARM-designed CPUs are in basically every smartphone and a ton of other devices. I realize ARM is likely in my laptop … my car … even my refrigerator. And ARM’s chips are helping power the revolution in AI by providing CPUs to data centers.

For his role as ARM CEO, Haas was included in Time Magazine’s list of the 100 most influential people in AI. An electrical engineer by training, he also worked at Nvidia and Mythic before joining ARM. Our conversation touches on the role CPUs are playing in both AI training AND inference, the fragilities in chip supply chains, and – yet again – whether we are in an AI bubble.

EL KALIOUBY: Rene, welcome to the Fortune stage. Thank you. So excited to have you. So before we dive in, I wanna give the room a sense of scale because I don’t think everyone here is familiar with how pervasive ARM is. You’re in our phones, laptops, fridges, stoves, you name it. So give us a sense of how pervasive ARM is.

RENE: Yeah. So what we do is a product called a CPU, which is the brain of every modern electronic device. So I would say that each person probably uses 50 to a hundred ARM CPUs on their person or in their home. I can assure you everyone has one today because we are in literally every single mobile phone built.

We’re in all the iPhones, we’re in all the Android phones, but we’re in Ford F-150s. We’re in Teslas, we’re in data centers, we’re in washing machines. The brain that powers everything. So when you turn on your smart TV and all those apps come up, you’re basically trying to pick whether it’s YouTube or Amazon Prime, that whole operating system that’s running on the TV, that’s all running on an ARM CPU. The company’s been around 35 plus years. We’re just talking.

EL KALIOUBY: Yeah. It’s Cambridge University.

RENE: Started in a barn in Cambridge with ambition to be the global standard for CPUs and that mission, they accomplished it.

Copy LinkHow Arm expanded from smartphones to the cloud

EL KALIOUBY: Yeah. I also wanna double click on the fact that it’s not just devices. You talked about data centers, and you recently announced that in 2025, 50% of all compute shipped to top hyperscalers will be ARM-based. Tell us more about that.

RENE: Yeah, because that’s kind of counterintuitive. And also back to the history of ARM, we started in a barn in Cambridge. It was a joint venture between a chip company called VLSI Technology and Apple.

And they were looking for the first chip for a first PDA, known as the Newton. If people remember that device, which was a PDA way ahead of its time relative to running off batteries, had a display. There were a million things wrong with it, but what it needed was a low power CPU, something that could run off a battery, and that’s how ARM was conceived, was a chip to run off batteries, which.

Fast forwards serve us incredibly well because back to the mobile phones, because the battery power is so key there and you’re running off batteries, energy consumption is key. Power efficiency. So that is really the application space where we’re quite strong. Fast forward to these modern data centers that are running a hundred megawatts, 500 megawatts, gigawatts.

Energy efficiency is everything. So ARM had gotten into the cloud data centers a number of years ago. General Purpose Cloud with Amazon and Microsoft and Google. But when Nvidia made the choice with their Grace Hopper Chip and then Grace Blackwell, they needed a CPU to pair up with the accelerator and they chose ARM.

And because Grace Blackwell is literally every single AI data center compute being done, we’re there too. So hence our big number, 50%.

Copy LinkWhy power efficiency gives Arm an edge in AI

EL KALIOUBY: That’s incredible. You’ve made a bold statement that ARM is the only compute platform delivering AI everywhere from milliwatts at the edge to gigawatts in the world’s largest data center. What makes ARM uniquely positioned to play in this space?

RENE: So let’s think about the data center application we just talked about, right? These giant AI data centers that are using GPU acceleration, they all need a CPU as well because the CPU does all of the management of the data, pairs the data that goes in the accelerator.

There’s just a huge codependency there. You can’t have a GPU without the CPU. That’s fine when you’ve got 50 megawatts to deal with, a hundred megawatts. But when you go to a very small application like wearable glasses where you have maybe two watts of power, you can’t put a GPU there. Simply put, you have to find a way to run that AI accelerator on the CPU platform.

So that’s where we play. So we not only have a solution up in the cloud, but we have a solution in the smallest devices.

Copy LinkWhy the AI opportunity is bigger than the bubble debate

EL KALIOUBY: Yeah. Are we in an AI bubble?

RENE: I think it depends on how you define bubble. If the definition of AI bubble is that the stocks are overvalued and the PEs are running way ahead of themselves historically.

Maybe. That’s not something I frankly spend a lot of attention on. If you ask the question of are we in an AI bubble relative to the technology being relevant and powerful, absolutely not. 1000%. I feel that in my bones, and you might ask, well, why are you so convicted about that?

You simply just have to look at the areas that AI tends to help us with already. Well, imagine a world where you go to AI and say, write me a business plan.

Or write me a way that I could come up with a novel concept to build a company that could solve this level of problem. Take life sciences. I’ve always thought that the killer app for AI is around health. If you look at one of the bottlenecks in drug discovery today, it’s actually around trials, human trials.

If you could replace all of that with AI and have AI instantly tell you from the time you’re starting R and D that you know exactly what the trials are gonna look like, and you’ve accelerated drug discovery by 10, 15 years, that by the way, all of what I described is going to happen. None of that is an if statement. It’s a when statement. So when I think about AI and the changes it can make for humanity and the planet and where we are relative to its capability, we’re so early and the benefits will be so profound. So I absolutely do not believe we’re in any kind of bubble.

EL KALIOUBY: We’ll be right back with more from ARM CEO Rene Haas after a short break.

[AD BREAK]

Copy LinkHow inference could become the biggest driver of AI compute

EL KALIOUBY: A lot of attention goes into the training of these AI models, but as demand for AI increases, the prediction is we’re gonna need more and more AI inference. And this is the shift that’s gonna benefit ARM. So tell us more about how you think about the training-inference dynamic.

RENE: Yeah. And again, another example of why I don’t think we’re in an AI bubble. So just backing up for a moment, what is training? What is inference? Training is teaching a model to do something, whether that is teaching it to recognize images, answer questions about history, solving an enterprise problem, et cetera, et cetera. Months and months and months are spent training a model. Training ChatGPT at OpenAI takes three months, six months, eight months to train this giant model.

That’s training. Once the model is done, it’s released to the world and on our phones we can ask a question and the question comes back with an answer. Sometimes it thinks, it takes the amount of time to give you the answer. That’s inference. Okay. So the giant use case, as you think about it, is of course inference.

Right? The more sophisticated the training model is, the more inference we’re going to use. First, things today are still researching. Companies spend more time training than they do on inference. So all this compute, these hundred megawatt data centers, they actually spend more time training than doing inference.

And you say, well, why is that? Because there’s just so much still that the models can get smart at. Inference is the use case that happens everywhere. It happens in the cloud, it happens in your car, it happens in your earbuds. And again, back to what is good for ARM to your question, those will all run in areas that generally will have an ARM processor.

So we think as inference explodes, and it will, as more and more of these models get sophisticated, you’re gonna run these AI workloads everywhere. It will be ubiquitous. We will not think about AI running in a cloud or AI running in an edge device. It’s just software — that’s how it works.

Copy LinkWhat fragile chip supply chains mean for the industry

EL KALIOUBY: Yeah. Do we have any questions from the audience? Yes. There’s a question at the back. Please share your name and organization.

AUDIENCE: Sure. Julie Michelle Morris. I’m always here to ask about cybersecurity and the security of this. The single source, the supply chain of course makes me a bit nervous. Does it keep you up at night to know that the weight of the world is on your shoulders?

RENE: We have a very unique situation in the semi industry and where chips were not so interesting for a number of years. And then they got really interesting and one of the events that made it so interesting was COVID.

And when you lost your key fob, it took 52 weeks to get a new one. It was a bit of what is going on here. Well, that was just a function of the fact that the semiconductor supply chain has many single points of failure. Not just ARM, there’s TSMC, which is in a very obviously interesting part of the world geopolitically, where it sits. TSMC is by far and away the provider of all these AI compute chips.

Everything Nvidia builds is from there. And AMD, Broadcom — that’s a single point of failure. There is a very sophisticated device that has to go into these fabs that comes from one company on the planet for EUV. It’s a company called ASML, a Dutch company. They’re the only people in the world who build EUV machines.

There are a set of mirrors that go into that EUV machine made by a company in Germany called Zeiss. They’re the only company in the world that makes those mirrors. So I dunno if it keeps me up at night, but it is something the entire semiconductor supply chain has been learning to live with.

But I will say, and you’ve seen this partially with the Trump administration, this acceleration to move manufacturing back to the United States, I think is a step in the right direction. There’s a lot of criticism in terms of how long it will take. And there’s no question that it will take decades to get to where it was because it took decades for it to kind of get to the point that it is. But we have to start, and I’m glad to see that it’s starting because it’s necessary.

Copy LinkWhere competition in AI chips is coming from next

EL KALIOUBY: Well, there’s a question here. When we were prepping for our call, I mentioned that I’m an investor in an early stage chip company that’s very focused on AI inference, so perhaps you can give us a sense of what the competitive landscape looks like?

RENE: Well, right now the leader is Nvidia, by far and away. And they do almost everything relative to training and inference. Now when you have such a large competitor in such a large market, that of course begets lots of competition. And there will be lots of competition from a combination of sources.

You’ll have hyperscalers doing their own chips, whether that’s Google TPUs and that’s going on for a number of generations. Microsoft, Meta, AWS, and then you’ll have specialized companies trying to do chips around inference. Not only for the cloud, but also for these smaller edge devices. So I would never try to second guess or underestimate Jensen and Nvidia. I worked there for many years. He’s an amazing leader and they’re an amazing company, but the opportunity is so large and the scale is so broad. You will see other players. No question about it.

AUDIENCE: Animanin Kumar. So I do agree on the Nvidia side. Building off of that, how do you see the hardware landscape evolve?

RENE: Right. So Grace Hoppers for instance, the bandwidth between CPU and GPU is increasing. So as these models get bigger, agents, long context, it’s all memory and bandwidth centric. So how do you see the ecosystem evolve, and what would the CPO of the future look like? Yeah, it’s a wonderful question.

Because ultimately when you’re handling these large AI models, and if you enter ChatGPT or enter a question into grok and it says thinking, and you wonder, well, why is it thinking? It’s not that the computer is thinking, it’s that the computer is waiting for data to come into it to figure out how to make the answer.

That’s memory bandwidth. So I think what’s going to happen over time is you’re gonna see some innovations in memory technology. You’re gonna see innovations in packaging. You’re gonna see innovations in interconnect. In other words, what’s the material used to connect between the devices? Is it light, is it different kind of photonics?

Is it a different area in terms of other interfaces? So I think back to your question on innovation, if you’re investing in this space, it’s a bit of the wild west, because people are looking for all kinds of different levels of innovation there, because it all has to get solved.

Selfishly I think ARM’s gonna be in a wonderful place there because most of that traffic needs to run through or around our devices, and we’re very involved in a number of activities in that space.

Copy LinkHow physical AI could reshape cars factories and wearables

EL KALIOUBY: Can we talk about physical AI? That could be robotics, autonomous vehicles, drones, but also these small embedded AI-first human-machine interfaces or wearables. And again, what does that mean for ARM?

RENE: Well, again, because just about everything that you described is gonna require something from a CPU standpoint to manage the activity, we will be involved. So then you get into physical AI. What’s taking place there? And you see this a little bit with autonomous vehicles.

In fact, I was reading an article in the New York Times this Sunday. It was rather amazing. It was looking at Waymo and the number of accidents that Waymo had, conventional accidents compared to human drivers. And it was about one tenth to one twentieth. Just looking at different kinds of accidents, right?

In terms of lane incursion or intersections. So there you have a data point that says, okay, autonomous physical AI is starting to see benefits in the cars. And Waymo, wonderful company. They’re all over San Francisco — that’s kind of autonomous 1.0. It’s got all kinds of equipment on there. It’s got lidar, it’s got radar.

It’s looking all over the place relative to how to drive autonomously. The next generation using LLMs needs less cameras, needs less devices. And why is that? Because they’re actually using AI as opposed to looking at all the physical inputs to make decisions on how to drive and how to move forward.

What does that mean? It’s going to really accelerate robotics, humanoids, things that can do factories. I think in the next five to 10 years, and again, ARM will be involved, you’re gonna see large sections of factory work replaced with robots. And part of the reason for that is these physical AI robots can be reprogrammed to do different tasks. One of the issues you’ve had with factory robots in the past is if it was a pick and place machine for a factory.

EL KALIOUBY: They’re just optimized for one task.

RENE: For one task, right. The software was for one task. The hardware is for one task. Now, if you can design a general purpose humanoid where the software is all AI and it learns by doing, it’s going to completely replace a large set of factory workers.

So I think in five to 10 years, and back to the geopolitical comment, I think it will level the playing field for countries because you will start to see physical AI be a great enabler for other economies.

AUDIENCE: Quick question — ARM is dominant. x86 is struggling. RISC-V, we’ll see. You talked about so many customers, right? Why then is ARM moving up the stack, going into building chips? Aren’t you competing with customers? What’s the calculation there?

RENE: Yeah, I don’t think I said anything about that.

AUDIENCE: It’s a Silicon Valley rumor.

RENE: Yeah. One of the things that we see is huge demand for our technology, and one of the other things we see is that the speed required to develop products is just getting faster and faster and faster. Part of that, AI exacerbates it in the sense that the models are moving far faster than the hardware can keep up.

So when you sit where we do in the platform, in other words, the compute sort of starts with ARM. It’s not unnatural for customers to come to us and say, if the compute and models are all running on your platform anyway, instead of waiting for your IP, why don’t you just build something for us? It could be quicker.

That is something we do listen to. So as I’ve said in our earnings calls, very carefully, that is something we continue to explore. Largely, the biggest customers are pushing us for it, and why are they pushing us for it? They just wanna get their end products out faster.

Copy LinkWhy Arm may move closer to building complete chip solutions

EL KALIOUBY: Rene, last question.

What are you most excited about for 2026?

RENE: I think that this continuance of AI and the acceleration of what we’re doing — to me, it’s like being in Disneyland. It just opened. We have hours and hours and hours to spend. I was at a panel where someone said, give me a bold prediction for this time next year.

And I said, no one will be talking about an AI bubble a year from now. Yeah. And then someone asked, is it because it popped? I said, well, I’m not gonna say big into that, but I just don’t think we will. I think people have embraced a year from now that we’re well on our way.

EL KALIOUBY: I love that — Disneyland just opened. Thank you for the conversation, Rene.

RENE: Thank you. Thank you.

I really enjoyed my conversation with Rene. The chip industry didn’t used to be sexy. But now chips are central to the conversation, as they become even more integral to AI.

I have two main takeaways.

One: as demand for AI increases, inference — not training — will dominate compute demand. And much of it will happen on CPUs at the edge as AI moves into more devices, vehicles, wearables, and robots.

Two: In hyperscale data centers, more GPUs doesn’t reduce the need for CPUs — it increases it. Because, while the GPUs do all the heavy processing, the CPUs handle everything else like storage, networking, data, and security.

But beyond the technical takeaways, my conversation with Rene reminded me that we are in the early days of AI, full of possibility… or like Rene said, Disneyland when it just opened.