00:00As you outlined for us, a major shift for this company then.
00:03René Haas, thank you very much indeed for your time.
00:05You've gone from a company selling IP, selling the blueprints to your customers,
00:09to a company now that for the first time in its history will be designing and making chips.
00:15It's a significant moment. How significant is it and why move in this direction?
00:19Thank you, Tom. It is a very big day for Arm. There's no doubt about that.
00:23We've been an IP provider since we started.
00:26We started selling compute subsystems a few years ago, which was sort of the next step,
00:30but it's not a physical thing.
00:32But as we grew into the CSS business, we had customers that kept asking for more and more and more.
00:38And here we are today with Arm's first chip product, the Arm AGI CPU,
00:44which Meta is our lead partner for.
00:46And Meta asked us to do it with them. So we did.
00:49Talk to me about Meta and the partnership then.
00:52One of your customers, OpenAI, also a customer.
00:55What are the commitments from Meta, from OpenAI to take this chip on board?
00:59When do you start shipping?
01:00So as Mohamed said, it's available now and we'll start shipping at the end of this year.
01:05We're not saying anything publicly about the volumes associated with it,
01:09but they are material enough that we have to start talking about the product
01:14because it's going to start showing up on our revenue this year.
01:17The revenue starts showing up this year.
01:19And in your presentation earlier here in San Francisco,
01:22you talked about potentially just this one revenue stream,
01:26just this one part of the business, potentially $100 billion by 2030.
01:30Talk us through how you get to that $100 billion.
01:32So we're looking here then, Tom, at the general purpose CPU, Tam,
01:37which today in 2026 is about $60 billion, $70 billion, depending on your maths.
01:45I talked about Agentec AI driving a 4X CPU workload even today.
01:50So $60 billion going to $100 billion by the end of the decade
01:54in terms of market opportunity is not really that much of a stretch.
01:57How much of this is about that huge demand we're seeing,
02:00particularly here in Silicon Valley, for Agentec AI?
02:03How much of this is aligned to that shift?
02:05I think you have something key that's underpinning everything,
02:07and that's the demand for compute, period.
02:10And compute has been perceived in AI about just generating tokens.
02:14But those tokens need to be distributed, managed, orchestrated, scheduled.
02:19That's all work that only CPUs can do.
02:21So the more and more compute that drives Agentec AI drives huge, huge CPU demand,
02:27which is great for the product we just announced today.
02:29The other number that you announced was $1 trillion,
02:31and that's the entire business by 2030.
02:33You were a $4 billion revenue company in 2025,
02:38and you're looking at $1 trillion potentially by 2030.
02:40Just unpack that for us.
02:42How do you get to $1 trillion potentially by 2030?
02:44Yeah, so what I showed there was a TAM,
02:47total available market of $1 trillion by the end of the decade.
02:49And I showed a picture that had some other dye photographs of things that we could do
02:55with the tagline, we're not stopping here.
02:58So more to come, but today is all about the ARM AGI CPU.
03:02What to come?
03:02So you have the pipeline of CPUs.
03:04Do you move into GPUs?
03:05What else is involved in more to come?
03:07The compute demand for AI is off the charts,
03:12and ARM is at the heart of literally every AI compute workload.
03:16So we have a lot of different ways to go address that.
03:19Can you give me some detail, Rene,
03:21in terms of revenues in the next 12 to 24 months coming from this product?
03:25So we're going to be talking about it a little bit later,
03:27but the market potential for this product is game-changing for the company.
03:31It puts us in a whole different zip code relative to the size of market we can address
03:36and the kind of revenues we can drive from it.
03:38Does it come at the cost of margins, the spend on R&D?
03:42Is that the pain point potentially?
03:43So one of the things to think about when you get into the chip business is the gross margin percentage
03:50may be less,
03:51but the gross margin dollars may be much, much larger.
03:53So if you think about, let's use the TAM example.
03:57$3 billion TAM, 95% gross margin, you'd call that $2.8 billion.
04:03Take $100 billion TAM and apply a 50% gross margin, you're at $50 billion.
04:09So $50 billion of gross margin versus slightly under $3.
04:13So percentages less, dollars much, much, much more.
04:18The critics might say the hyperscalers are doing custom silicon.
04:23Jensen Huang talked about CPUs, their own CPUs, that maybe ARM is late to the game.
04:29Are you late on this?
04:31No, I don't think we're late at all.
04:33And I think if you just look at what we've seen with Agentec AI and the fact that CPUs are
04:38being seen again as first-class customers,
04:41it's a proof point that there is going to be huge demand for this product.
04:44And by the way, Tom, it's not just in the data center.
04:47As I talked about, ARM is the heart of physical AI.
04:50We're the heart of autonomous vehicles.
04:52We're at the heart of mobile phones, edge devices that fit inside your ear, wearables.
04:57That all needs compute in a very big way with AI.
05:00This is bad news for Intel, presumably.
05:04I don't know if it's bad news for Intel.
05:06I think it's good news for the market that really needs to have extra performance at the same power.
05:11So I think for customers who need that extra performance but don't want to spend more in power, it's really
05:16good for them.
05:17There was a time when ARM was characterized as the Switzerland of the semiconductor industry.
05:23Neutrality.
05:24Can you really claim to be neutral now that you have your own fully designed manufacture chip?
05:29I think so.
05:30The market and world is complicated relative to customers who compete with each other.
05:35You've got Google building phones, yet Samsung is a large Android customer.
05:39You've got Microsoft building laptops for things like Surface, yet you've got companies like HP building the same.
05:46The thing to keep in mind is why ARM building chips is beneficial is that it lifts all boats for
05:52the ecosystem.
05:53The better the software is on ARM, anyone building an ARM-based chip can benefit from that opportunity.
05:59At GTC, Jensen Huang, NVIDIA, at their annual conference, outlines the picture for inference, and they've obviously bought the IP
06:07and the licensing around GROC, which plays a big role in inference.
06:10The way we interact with these models, the questions and the answers.
06:13Do we move now from a place where this is less about training and more about inference, and what does
06:19that mean?
06:19Does that mean more demand for your products or less demand as we start to shift from training to inference?
06:24Well, I think the shift to inference from training is an intellectually obvious one, right?
06:30The more models that are trained, the more sophisticated they are, the greater the demand to actually use them.
06:36So, as workloads shift more and more to needing inference, I think, again, all boats go up.
06:42I think we're not done yet with training by any stretch of the imagination, but with inference, you're really servicing
06:47the models that have been created.
06:50So, I think both demands go up, which, again, is good for ARM.
06:53I want to zoom out a little bit to what's happening geopolitically, the Middle East, the war in Iran.
06:57What is that supply chain impact having on your customers in terms of their ability to operate in this environment?
07:05I've not heard anything yet.
07:07I've read about things in the news about that space, but when I think about customers coming to us or
07:12looking at our financials and saying,
07:14oh, my gosh, we're going to see some impact due to what's going on in the Middle East, we're not
07:18seeing it.
07:19When it comes to memory, high bandwidth memory, NAND and other parts of the memory space, massive supply shortages,
07:26do you see a contraction of the smartphone market this year?
07:30Memory is going to be tough to get, period.
07:33I think it's going to be a driver of growth for the industry, but to the point we were talking
07:38about,
07:38more and more of these inference models, more and more compute, means demand for a lot of memory.
07:42So, do I think smartphones are going to go down?
07:45I don't have a good projection on what that looks like, but I do think that the memory supply cycle
07:49is going to be pretty tight.
07:50It remains pretty tight until when do you see that tightness in particularly HBM memory, high bandwidth memory, tightness?
07:57When do you see that being resolved?
07:58How many more years of this?
07:59I think the memory market is going to be a tough market if you listen to SK Hynix or Samsung
08:06or Micron the next year to 18 months.
08:09I think it's going to be quite supply-constrained.
08:11Okay.
08:12Is it fair to conclude that we're going to see smartphone prices generally go up and fewer products as a
08:18result of this, at least in the short to medium term?
08:20Yeah, I've not seen it yet, Tom.
08:22I think for what I've seen for ARM is we've had a lot of questions about it.
08:25How is our business going to be impacted by the memory issue with smartphones?
08:29And because we've gone to CSS and much higher royalty rates, even a reduction of 15% to 20%
08:35in smartphone volume has probably a 1% to 2% impact on our royalties.
08:39And that's just because of the mix of where we play.
08:41So, right now, as ARM's business is going, we're not seeing it, and we've modeled in some bearish figures.
08:47You, in your role, you have to be looking to the future, and I know you do this, Rene, years
08:52out.
08:53Jensen Huang has talked about this.
08:55Elon Musk has talked about this.
08:56Jeff Bezos has talked about this.
08:57Data centers in space.
08:59Do you take that seriously, Rene?
09:01And are you and the team looking maybe at servicing at some point with your CPUs, with your architecture, data
09:07centers in space?
09:08I think it's quite real.
09:09There's some huge benefits you get from it.
09:11One of the obvious ones just being power.
09:13The benefits you get from power are quite profound.
09:15Now, when you think about putting things in space, swapping out boards that have gone bad, a system that needs
09:23to be made redundant because something's not working, obviously all those issues come into play.
09:27And then there's obviously issues around latency.
09:29Can you get the performance you need?
09:30But do I think we'll see data centers in space someday?
09:33Absolutely.
09:35And I want to loop it back.
09:36We can play there.
09:37You can play that.
09:37You can play in data centers in space.
09:39There we go.
09:40The ambition remains.
09:43Another question on data centers, given what we're seeing in the Middle East.
09:45We've seen data centers run and owned by the likes of Amazon in the Middle East being targeted, being impacted.
09:50Do we need to rethink security around data centers?
09:52Is that a topic that's being discussed in the industry?
09:54I think it is.
09:55And, yes, for sure.
09:56When you think about the scale of these data centers and the value of the intellectual property inside, the amount
10:02of compute that it services, absolutely.
10:04If you go to data centers in North America, for example, they are very heavily security.
10:08There are security devices and personnel around all these data centers.
10:14So certainly it's something that has to be done.
Comments