Arm CEO Rene Haas 谈 AI：Nvidia 的启示、Intel 的衰落与美中芯片战

There's a company nearly every chipmaker relies on that doesn't actually make anything tangible. Yet its blockbuster IPO in September valued it above $54 billion. It's the largest public offering in over 2 years, and the valuation of the company has tripled.

If you have a smartphone in your pocket or in front of you, you have an Arm CPU somewhere inside of it. We are the CPU, the heart of everything. They're the winner on the CPU side. The foundation models and the software are moving far faster than the hardware, so what we're seeing is people investing faster and faster in new hardware, which ends up being a good thing for us.

Thank you so much.

How are you? Welcome. Welcome, David. Hey, good to see you. Hi. Hello, Rene. What are you using these days—3-milligram Zyn pouches, or are you up to 9?

I know you're competing with Nvidia, so you probably want to go with the 9, right?

I will go with the 9 with Jensen. You have to go big.

You have to go big with Jensen. What's that like, competing against Nvidia?

I will say Nvidia is a customer of ours, so I'm not going to say Jensen is my competitor today. But I worked for Nvidia for many, many years, as you know, and I learned so much working there, working for him, and working with him. Then Nvidia almost acquired Arm in 2020, so I almost had a chance to work with him again.

What did you learn from Jensen?

One of the things about Jensen that is amazing—and I think it's also true for people like Michael Dell and Masayoshi Son—is that you have these entrepreneurs who started their companies 30 or 40 years ago, and they're still running them. You have this amazing set of characteristics: vision, speed, fearlessness, taking risks, and an ability to pivot very, very fast.

I saw that a lot at Nvidia. When I was there, we were only about $4 billion in sales. At that time, we were looking at lots of different ways to grow—business models and such. I remember one story: We were at a strategic off-site that was supposed to be a review of road maps, where each of the general managers would go through what they projected in their business.

What was intended to be a road-map review turned into, “We're changing the strategy. We're abolishing this product line. We're going to move 2,000 engineers off of Project X onto Project Y.” By the way, we were only about 6,000 people at the time.

What was Project X? What was Project Y?

We were involved at that time in trying to develop mobile chipsets connecting to an Intel processor, right? Back in the day, for those who remember PC architecture, developing these chipsets and competing with Intel was really difficult. Intel was making it very, very hard to compete, relative to the integration that they had done.

In fact, that was the genesis of starting to pivot to Arm in a very big way inside Nvidia. At that time, Jensen looked at what was going on with SoCs and Arm-based architecture and moved everybody onto the program.

Let's take a step back and level-set for the audience. Just to give some background, Masayoshi Son and SoftBank took Arm private—

Took it private, yeah, for $32 billion.

$32 billion, and then tried to sell it famously.

Yes.

Couldn't find a bidder.

Could not find a bidder.

They hung on to it and took it public. It's now a $150 billion market-cap company.

That's right.

You were telling us backstage that he famously refuses to sell a share. It's been a slow process of building the shareholder base, but you've done phenomenally well as a business.

Just set the landscape for people who want to understand Nvidia, the most valuable company in the world, but also a window into understanding AI. What do they make that's so powerful? Why aren't there other competitive solutions at that level of scale yet? How do you think that changes over the next 5 to 10 years?

There's a lot there to describe. The way to think about Nvidia—and, to some extent, even though I'm the CEO of Arm, I don't want to tie it necessarily back to Arm—is that in our world, what really drives demand is compute workloads. At the end of the day, it's compute workloads. When a new workload is either identified or invented, it comes down to what the best processor architecture is for addressing that workload.

Let's look at AI. The lightning-bolt moment of AlexNet, and the work that the DeepMind team was doing, showed that AI—particularly training—is a very, very complex parallel problem that is well suited for a GPU. In fact, the very first work done by the engineers on AlexNet was not with Blackwell. It wasn't with an AI processor; it was with a gaming GPU, a gaming card.

Nvidia was in a very good place to seize that moment relative to the DeepMind moment, AlexNet, and transformer training. Fast-forward to today: Training these complex AI models, as Demis was just talking about, is a huge amount of work.

Every one of these workloads requires a CPU not only to run the computer but also to help the accelerator run. That's where Nvidia is a customer today. Its most advanced chip, called Grace Blackwell, has 72 Arm CPUs alongside the Blackwell architecture, and that's where Nvidia plays today.

There is competition. Demis talked earlier about Google, which makes its own chip called the TPU. Nvidia is obviously the leader in general-purpose computing, but right now we're in this interesting world where people are asking whether they should use a general-purpose chip, a custom chip, and so on. It's a fascinating time to be in this industry.

Where do you think companies like Tesla fit? Tesla recently taped out AI4, and now they're working on AI5 and AI6. Then there are emerging companies like Cerebras, Groq, and a whole slew of others that have raised enormous amounts of money.

Do you believe the role of Arm should be, for lack of a better phrase, the arms dealer to all of those folks who need that capability? Or, at some point, do you see enough of it where you think, “I could just do this better”?

Maybe a little bit of both. Today, the role we play is that we're increasingly the microprocessor that connects to these accelerators, whether it's something done by Cerebras, Nvidia, or Google. They're connected.

Could we do something ourselves, custom? It's possible. Could we also supply the intellectual property to somebody building a custom chip? We're doing that today.

To some extent, we're in a very unique position. We can provide the solution, whether it's standard or custom. But as AI moves from gigawatt data centers to running in headsets, wearables, or something that needs to be energy-efficient, you still need to run the compute workload. Now you need to run the AI workload as well, and that's a place where I think only Arm is uniquely positioned to address.

So you're going to make chips and compete with Nvidia.

I'm not going to say that today, but could we do that? I hinted in the last conference call that we're looking at going a little bit further than we do today.

Could we see, in the next few years, a divergence in the market between training and inference? What I've noticed is that xAI and OpenAI, and Google already with its TPUs, are building their own chips for inference, which might be—I don't know—99% of the workloads.

They seem to acknowledge that Nvidia is the best at training, and they haven't, at least publicly, announced an effort to challenge Nvidia for training. Is there a possibility that the market could bifurcate into training chips and inference chips, with inference becoming much more competitive?

Yes. I also think you have a third bucket, where training distills down to simpler training chips that don't need to run a trillion-parameter model. You could have a giant model that trains and teaches smaller models—mixture-of-experts models with 20 billion parameters—that can be a mix of inference and training, doing reinforcement learning, where the chip is helping with learning and training.

It's almost like the professor teaching a student who can also be a student-teacher, right? It can do a little bit of both.

Then there's inference, which over time will be very dedicated, particularly as you get to endpoints where you can't have a GPU running at 1 kilowatt of power. It's impossible.

Right. So if you have robots in the field—we have 500 million robots—what is the chip market going to look like for robotics? How is it different from what we have today on the embedded side versus the data-center side for AI?

Physical AI is going to be a gigantic market. Today, quite candidly, I think it's bigger than data centers.

Yeah, I think so.

I think today they largely use repurposed automotive chips—things that have functional-safety compliance around ADAS—but they're not specific to actuators or to smaller parts of the joint. Physical AI, particularly AI that can learn, is going to be a giant market because the robots themselves will have tens of chips, hundreds of chips.

So, from a unit standpoint, it could be huge. The numbers are going to be well beyond what we see today.

You started the business—or ARM started—really making reference designs and then working with partners. Does that give you a different perspective on things like export controls and export restrictions, and the role that China plays in this ecosystem, than, say, a different kind of vendor who would actually be originating and trying to tape out themselves?

To some extent. Although we don't build anything, our business model is that we do the design and someone else has the chip built—mostly at TSMC, some at Samsung, even Intel. But because we are early in the value chain relative to the software ecosystem—in other words, we probably see what people are doing earlier than anybody else, because ultimately we're the link between the hardware and the software—on export control, yes, to some extent, we have a very big lens into it.

Today, the China ecosystem actually follows the global ecosystem, which is good from the standpoint that every mobile phone in China—it doesn't run Google Android, but it runs a version of Android, and it leverages the app ecosystem that comes off of Android. The same thing is true with autonomous vehicles. They leverage the ADAS stack that was created by Arm, and then Qualcomm and NVIDIA.

So right now, the China ecosystem on software looks a lot like the West, which for us is obviously great. We have a very clear opinion in terms of where we want things to go: it's great if the global ecosystem remains open.

What's your take on President Trump taking a 9% or 10% stake in Intel, and how did that company miss this entire revolution so badly?

Semiconductors are what I've spent my entire career in. I started at TI in 1984, and I've just been in semiconductors my whole career. There are long product cycles. It takes a long time to develop chips, invest in fabs, and define architectures and ecosystems. If you miss a few, you will be punished for that.

I think Intel has unfortunately been punished in a few areas. They were punished in mobile; obviously, they missed that completely. They were also punished in terms of manufacturing, specifically for not going to EUV. EUV is an advanced methodology for building the smallest chips on the planet. They decided not to invest in that probably a decade ago at the rate that TSMC did, and they fell behind.

Once you fall behind in chips, it's very, very difficult to catch up because the cycle gets on top of you. TSMC now has the best fabs in the world. The leading-edge companies—Apple, NVIDIA, and AMD—they all build at TSMC. TSMC gets better at what they're building.

Intel and Samsung don't get the opportunities. It just compounds. And that flywheel, once it compounds and compounds and compounds, makes it very hard to catch up.

So, if you think about Intel having lost its footing, you did mention EUV and the leaders there, like ASML, and then, even one step back, companies like Carl Zeiss that make these lenses. Those are critical pieces of infrastructure that the West needs. Is there a role for the government to be spending more capital to incubate those kinds of things so that we have a little bit more diversity in the supply chain?

If you contrast and compare, there's the Intel investment, but then there are these other things that maybe we should also be doing.

Oh, 100%. If you look at one of the most critical components in building chips, it's these rare-earth compounds. There's a belief that China has cornered the market because it has all the access to these rare-earth minerals. Access to the minerals is global. There's no issue in getting access to materials.

Yeah.

The issue is in the refinement and actually building the factories that can refine the materials. Again, that's a decades-level investment.

I'll tell you one thing: when I lived in China for a number of years, one of the things that I was very impressed with when I lived there, and still am, is the industrial policy that sits inside the central government, which will last irrespective of an election cycle. It essentially requires a lot of the folks who are in the Ministry of Industry and Information Technology to be engineers, thinking about a policy on building.

So, to your question, should the U.S. do it? Absolutely.

Okay. So, Rene, let me put you on the spot. Between the Korea trade deal, the Japanese trade deal, and the European trade deal, we now have close to $2 trillion of investment capital that these countries will put into the United States. How do we go about creating an ASML-type company or capability, or these lenses? How do we do that? What universities do we go to, or what labs do we go to? What do we do?

I think there probably needs to be more of the U.S. companies working together. I'll say this because Arm is not a U.S. company, but we would do the same if we were working together—pooling capital for some of these initiatives to essentially get some type of grounding.

You need universities, but you need corporations to get behind this as well, as well as financing—private equity, all kinds of different capital—because this is a huge capital investment that also requires investment from companies and private equity, but at the same time needs to last for years.

Just talking about the fabs, TSMC has built this facility in Arizona. There were reports about the inability to get labor, train labor, and get a workforce that—I don't know what the right term to use is—would operate culturally the same way as they do back in Taiwan. They were really challenged, and they had to bring folks over to Arizona to work at the facility. These were news reports, so we don't know this firsthand.

Do you think we have the capacity to do fabs in the United States, onshore here? What's it going to take? If you were in the administration—let's say you were the AI czar, for example—what would you advise the president to do to ensure that happens successfully?

I don't want to take anything away from David. He's doing an amazing job as the AI czar. You've hit a very key tenet, though, relative to world-class manufacturing inside the United States and what is required to make that happen.

We had it decades ago, believe it or not. There was a time when the leading contract manufacturers in the world were U.S.-based companies, and we knew how to do that. If you go back 30 years ago, when Apple and Compaq used to build their own PCs and had their own factories, believe it or not, then all of that went to companies like Flextronics and SCI, et cetera, et cetera. So, we had that. Ultimately, for cost reasons, that began to move all the way to the Far East, to Foxconn in China, et cetera, et cetera.

There's a great book, “Apple in China,” that documents a lot of this. To your point, in terms of whether we could get that back in some ways, there's no reason why we couldn't. But it is a mindset. TSMC is a 24/7 operation where, if a line goes down or a customer has a problem, not only do the technicians need to be ready to go, the engineers need to be ready to go. And that is something that I think we've lost—the muscle memory inside the United States, quite frankly, on how to do that.

We may have had it a generation or so ago. I don't know that we have it now. And we certainly haven't trained a generation of folks to look at manufacturing jobs as being something that is as lucrative and prestigious. They're sort of thinking, “Oh, it's a blue-collar job. I don't want to go into that.” It's not viewed that way in Taiwan, right? In Taiwan, if you say you're working for TSMC or studying to go off and do that, it's a highly prestigious kind of thing.

So, it's not just the AI czar's problem. I think it's deeper than that in terms of us getting—

So, you've diagnosed the problem. Do you have a solution or recommendation? Is there a short form that you could highlight?

I think we've seen a huge amount of work already done by universities. I was at Carnegie Mellon a couple weeks ago. They now have microelectronics classes for chip design. That was gone a number of years ago. There weren't even people designing chips.

So, I think getting manufacturing operations excellence into the universities, making that a field of discipline that the universities get behind to build up that capacity in the U.S.—I think that's required.

Let me go back to export controls, which Chamath mentioned. I'm not sure people here know exactly how these things work, but basically, if a product like an advanced semiconductor is put on the export control list, it means that the company that's selling it, or the buyer, has to apply for a license from the Commerce Department to get their purchase order fulfilled.

The Commerce Department will then process that license request, and it goes through some interagency committee. Five different departments will basically have to sign off on it. Best-case scenario, it takes months, but there are license applications that literally have been in the hopper for 2 years, by which time the chip is obsolete.

And believe it or not, there are a lot of people in Washington right now who are calling for literally every sale of an advanced semiconductor worldwide to be a licensed sale, because they think that GPUs are like plutonium or something and they're inherently scary. I mean, this is seriously the discourse that's going on right now.

In fact, there was a major rule that was put forward called the AI Diffusion Rule in the last 5 days of the Biden administration that basically did require every sale of a GPU worldwide to be licensed, subject to some carve-outs.

We rescinded that, but there is a never-ending clamor and pressure in Washington to bring back these sorts of rules. The irony is that the people who are advocating for these things call themselves China hawks.

But it seems to me that the whole basis of the semiconductor industry—the reason why it has moved so fast and why you get new chips every year—is that it has really been left alone by the government for the most part. It hasn't been a highly regulated industry. I'm curious: What do you think will happen to the industry and the pace of innovation if the government now makes it heavily regulated in the way that I'm describing?

You brought up a great point, and I think we may even have a couple of those in the queue that haven't been approved for a couple of years. You're right. Semiconductors have not traditionally been regulated. Because of that, if you look at the real heart of what drives semiconductor growth—compute—whether it's Intel, whether it's Arm, whether it's NVIDIA, that's the West.

And why is that the West? Because it requires both innovation at the chip level and a global software ecosystem. The world works really well when it's flat and there aren't constraints relative to whom you sell to or how ecosystems get built. If you shut off the supply of a computing architecture to other parts of the world, what will happen?

Certain parts of the world that have the capabilities, either in terms of people, technology, or innovation, will find a way around the problem. Once that happens, you've now created 2 parallel universes. The U.S. and the West would then be at risk of that other ecosystem becoming an ecosystem of choice.

So, if you can navigate those licenses being expedited, the world works really well in semiconductors when it's flat and a global ecosystem. May the best company win.

Rene, the company started in Cambridge, and originally all the employees were there. But now, I think 50% of the employees are in the UK. Tell us about building a company there, multiculturally, and where you're going based on where technology is going.

The company was started in the UK, in Cambridge, in a barn, as part of a joint venture for the Apple Newton—building a processor as part of a joint venture between Apple and VLSI Technology. They needed a low-cost chip that could run off a battery. They contracted a company to build the chip. The chip wasn't so good, but a bunch of guys said, “You know what? The design's pretty good. Why don't we try to build a business from it?” And that's how Arm was born.

I'm the fourth CEO, and I'm the first one who is not from the UK. What I've been trying to do in the 3 and a half years since I took over is keep the great scientists and technology innovation that we have in Cambridge, but inject a bit of Silicon Valley aggressiveness and a twist toward moving faster and going quicker.

As you said, half the employees are in the UK, but we've got folks globally: 2,000 people in Bangalore, probably over 1,000 in the United States, and others in different parts of Europe. So it's a highly global company. We go where the talent is, and we look for great engineers.

Are you able to find great STEM talent still here, or do you now need more investment in core EE and chip design?

We need far more investment. Our business is not yet one where I can say I'm hiring fewer people because of AI. I'm certainly hiring fewer finance people and legal people. Sorry, Jason and Spencer, if you're in the audience.

But for engineers, AI for development, AI for creation, and AI for science—that's still a hard problem to solve. That's why we need more engineers to develop chips, which is great. I think back to whether there's more demand for compute and whether this AI wave that we're seeing is going to continue in the world of generative AI for science and creation. I think there's a ways to go.

Leveling up for a second and looking at our relationship with China—to get a little geopolitical here—how do you view China versus America? Is this going to be winner-take-all with AI, or can these 2 powers get along? Are we competitors? Are we collaborators? Are we destined to fight and go to war in Taiwan, like we talked about last year on this stage? What's your take on it? Is there a path to us having a great collaboration with China?

I'm going to be an optimist here, Jason, and say I think yes. I think China views some of the things around AI—whether these are things like guardrails or policies—as ways to keep things in such a way that we've got the right level of safety checks. I think their minds are in the right space, and I say this based on conversations I've had with folks over there.

I wouldn't necessarily compare it to the nuclear arms race, but in some ways it's not dissimilar, in the sense that you need the countries that have the capabilities to be willing to sit at the table and have the conversations. China, in my experience, has shown that so far.

Ladies and gentlemen, Rene Haas. Thank you.

Thanks, Rene.