SoftBank Group Corp. ($9984)

Thank you very much. We would like to start the Arm business briefing. I would like to introduce today's presenter. Vice President of Arm in charge of Investor Relations, Mr. Ian Thornton. This meeting is provided with a simultaneous interpretation service. On-site participants and those analysts are kind of requested and to select the channel. [Operator Instructions]. So a presentation by Ian Thornton. Over to you

Good morning. Thank you very much for coming to this presentation. It is wonderful to be back in Tokyo again. Although it's only been a year since I was last here, so much has happened. Arm has now reported 4 quarters with revenues over $1 billion. Demand for Arm technology has never been stronger, driven by excitement around AI. Arm has continued to introduce new products to meet this demand and has accelerated investment in R&D to speed up the development of the next generation of Arm technology. And as you may be aware, last week, we announced the first step in a major new strategy. For the first time, Arm is now selling complete chips. Firstly, I will talk about Arm business model, our products and our primary end markets. And then I'll talk about our new business strategy that we have just announced. And as part of that announcement, we also provided revenue and profit targets for the next 5 years. And then we will have time for your questions. I will start with Arms' original business model. Arm provides the design of the CPU, the brain of the chip. The purpose of the CPU is to run software. The software that runs in your smartphone, your PC, and your television, all run on a CPU. We license the design of the CPU to companies who develop chips, either chips to be sold on to OEM or to be used in their own products. The customer will pay on a license fee for access to the CPU design. Once our customer has the design, they can then focus on building their chip. These chips can be very complex, and it can take several years to complete the design of a chip. Once they have designed a chip, they can send it be manufactured at a foundry such as those managed by TSMC, Samsung or Intel and hopefully Rapidus next year. When the chips come back to the foundry, they can then be sold and Arm gets a royalty from every chip that contains our technology. Because each chip design can sell for many years, the royalty revenue becomes a very reliable future revenue stream. Arm is still receiving royalty revenue from technology that we developed more than 30 years ago. I mentioned earlier that the purpose of the CPU was to run software. Software is tied to the CPU that it is written for. And the success of the CPU is dependent upon the broad availability of software. Arm has by far the largest software ecosystem on the planet. We've been able to achieve this because over 50% of all chips with CPUs are Arm-based. We estimate that over 22 software developers are currently creating new software programs and apps for Arm-based devices. And that to date, they've invested over 1.5 billion hours in creating Arm software. I'll also invest in software. We invested around 10 million hours in software development for Armv8, and we're investing more than 3x that amount for Armv9. At today's salary rates 10 million hours is about $1 billion. And so for Armv9, we're investing about $3 billion in software development. Arms revenue will grow as the market grows and Arm gains share and as chips use more Arm technology. The demand for semiconductors continues apace. Analysts estimate that the CPU part of the semiconductor industry will grow at around a 10% CAGR between 2025 and 2030. Within the industry, Arm is gaining share in long-term growth markets such as automotive, cloud and IoT. As chips need to deliver more intelligence, they become more complex, which drives the need for more technology. For example, as more AI gets deployed in edge devices, this will drive the demand for more Armv9, which can accelerate these workloads. More complex chips also cost more to develop, which will help drive demand for Arm's compute subsystems, and this will result in higher royalty rates. Combined together more chips, more valuable chIps, higher royalty rates and share gains will help to grow arm's revenues for years and decades to come. Let's look at this in a little more detail. As you may know, we've doubled the royalty rates from Armv8 to Armv9, and we've doubled again for customers who use our compute subsystems. We can also charge a slightly higher royalty rate for each Armv9 generation and a higher royalty rate for each compute subsystem generation. On average, our royalty rates for Armv9 CPU and compute subsystems, increases around 20% generation to generation. And as we introduce a new generation every year, this becomes a 20% annual increase. The increasing complexity of chips of today and tomorrow is also contributing to our royalty revenue growth. In the data center, for example, including Agentic AI is driving our customers to design an increasing number of cores into their chips. Tracking both the increasing call count over recent years, and our customers' planned increase in future core counts, we can see that the number of cores per chip is increasing by about 20% per year. Typically, if the number of cores increases, we get a higher royalty per chip. And so our royalty revenue should again increase around 20% per year. The significant annual increase is in royalty revenue per core and in the number of cores per chip is a significant part of our confidence in continued robust growth in the data center royalty business. I will now turn to the end markets that we're driving Arm's royalty revenues. Starting first with Edge AI which includes smartphones, PCs and also other consumer electronic devices such as smart glasses, smart watches, televisions and game consoles. Al already has 100% market share of the main chip going into own smartphones and tablets, including all Apple iPhones and all the Android phones, also has a very high share in many other consumer electronic devices, and we're helping to make these more intelligent with our latest technologies. If you use AI on your smartphone today, maybe for improving picture quality, photograph or for live translation and transcription then you are already running an algorithm on the Arm CPU in the smartphone. More than 70% of the AI-enabled apps on the Android app store run entirely on the Arm CPU. They don't need any additional acceleration. Over the years, we have had many conversations with customers along the lines of our chip performance isn't where we want it to be. If you could help us achieve better performance, we could price our products higher, gain market share and ultimately perform better, and we'd be willing to share that upside with Arm. And this was really the genesis of our complete subsystem to build a world-class product, you need a full stack optimization from the system design to a CPU design to the physical implementation all the way down to the transistor level, and then all the way back up to the software the device drivers, the firmware, the operating system and the applications. And that's exactly what we've done with CSS. We work very closely with leading partners like Samsung and TSMC to optimize down to the transistor level, and we work closely with operating system vendors like Microsoft and Google and game engine companies like Unity and Unreal to optimize the software. Importantly, we go beyond providing individual IP blocks, the GPU or system IP. We integrate them into a cohesive system and provide our customers with a proven recipe. The newer customers, often OEMs entering silicon design for the first time, that's incredibly valuable. They can effectively take that recipe and implement it directly into their chips. More experienced customers, may believe that they can achieve similar results themselves, and that's fine. But then our solution serves as a benchmark when their engineers say 4 gigahertz, it isn't achievable, we can demonstrate that it is. And then that drives them to push further to get to 4.1, 4.2 gigahertz and beyond. This results in a virtuous cycle that helps our partners succeed in their markets, which in turn drive success are. So how do we grow royalty revenues from here? Armv9 is an important technology for edge devices as it provides the acceleration needed to run AI algorithms at the edge. Armv9 is also important for Arm as it has a much higher royalty rate than the previous generation of technology. Currently, Armv9 can be found in premium and high-end smartphones. But over time, we believe that almost all consumer electronic devices will want to run AI algorithms as they will need to move to Armv9. We are also providing our compute subsystems into this market. compute subsystem is particularly useful in chips, which can contain a lot of ARM technology such as in a smartphone or PC. again, this enables us to charge a higher royalty per chip. And finally, not all consumer electronics are currently are Arm based. We are still gaining share in the PC segment. Apple is 100% Arm based. Many Chromebooks are Arm based. Qualcomm and NVIDIA are now both providing chips into the PC space using ARM technology. Our second market is physical AI. Here, we are focusing on providing the intelligence to autonomous vehicles and robotics. If we are to share the street with autonomous vehicles and share our homes and workplaces with robots, we need to be certain that they will not accidentally hurt us. Therefore, we have worked with the industry to ensure that safety critical capability is built into all of our CPUs for this market. Nearly every company who is developing high-end chips for the automotive market is now using Arm CPUs in their chips. This includes semiconductor companies like NVIDIA, Qualcomm and Renesas and car OEMs like Tesla and Rivian. And it is the same technology used in cars to develop the first generation of humanoid robots too. As with Edge AI, we are also providing a complete compute subsystem that customers can take straight through to tape out and into production without modification. This delivers a significant economic value. Our CSS platforms for automotive are targeting autonomous functions such as driver assistance and also in-vehicle infotainment systems and can also be used for the main brain in a humanoid robot. Again, Armv9 and CSS are both being deployed across many chips that are going into autonomous vehicles. But here, the main growth driver is a significant increase in the amount of compute that is being deployed in each car. More advanced cars will need both more advanced Arm technology, but also more Arm-based chips. One car OEM has recently announced that they're deploying 433 Arm cores in their first generation of self-driving vehicle, and they expect even more in future generations. We calculate that this car will generate between $100 to $200 of royalty revenue. And we're still at the early days of robo taxes and various reports suggest that only 10,000 will be sold this year, but that could increase to more than 10 million by the end of the decade, providing a strong revenue driver for us. And finally, to robotics. This is still a relatively nascent market. We think that the compute content will be similar to the automotive market, but with important differences. They will both need a single independent brain to make decisions that will need to be very powerful. However, humanoid robots will need more sensors and actuators to move its limbs safely. But hopefully, they will not be traveling at 100 kilometers per hour. And finally to cloud AI. The biggest part of this market is the cloud data center, both in general purpose cloud compute and also in AI data centers. Enterprise data centers such as those being managed by telecoms companies is the next largest. And finally, we have wired and wireless networking equipment. Most of the new stories around AI have focused on large language models in the data center, such as open AIs, ChatGPT, running on NVIDIA's GPUs or Google Gemini running on their own custom ASICs. However, every GPU and every accelerator chip needs a CPU chip to run alongside it. The GPU is greater mathematics and doing the complex matrix multipliers needed for training, but a CPU is needed to control the GPU. The CPU runs the operating system and the applications of the GPU is accelerating. And increasingly, the CPU chip is based on our technology. Currently, most of the focus in the cloud has been on training new models. But increasingly, we are seeing inference taking over as enterprises and consumers start to use AI in their everyday work. Unlike training, inference requires more work from the CPU. Where was the GPU is great for mathematics, the CPU is greater decision-making, and this makes the CPU more important in inference applications. One of our goals has been to continue to lower the barrier to entry to make it easier and quicker for our customers to get their chips to market. Our first CSS customer told us that CSS gave them a better starting point for their chip design, and so save them about 89 years worth of engineering work. We had another customer from the time we handed them CSS to the time they had production silicon [indiscernible], including time in the fab was less than 18 months. That is the amount of time needed to build a simple microcontroller, not 128 core 20 gigahertz server chip. And in return for these great benefits, Arm is able to charge a higher royalty per chip. Armv9 is already widely deployed in data centers. However, increasing core count is helping to drive our royalty revenue. Most of the data center chips being shipped today have around 100 Arm cores, and we already have visibility into future data center chips with more than 200, 300 and even 500 Arm cores. And more cores means a higher royalty per chip. If you double the number of cores, you double the royalty per chip. And we're also seeing that royalty revenue per core is also doubling in future generations, giving per core price increases in both ArmV9 and CSS. The compounding growth of these 2 drivers plus market growth and share gains gives us a lot of confidence in data center royalty revenue growth. And we also have several customers for our compute subsystems. But the big revenue driver for us here is going to be our new chip strategy for data centers, which we announced early last week. And I'd like to tell you a little bit more about that now. First, let me give you some background. Let's start by looking at the role of the CPU in the cloud. Before the rise of AI, the cloud followed a relatively simple model. Demand for compute was already growing rapidly, driven by providers like AWS, Microsoft Azure and Google Cloud and most workloads were straightforward. You entered a query, the system processed that request efficiently and you received a response. In that world, the CPU did almost all the work. Throughout the growth of Software as a Service over the past 10 to 15 years, cloud infrastructure scaled on the back of CPU performance. Now with the introduction of AI, the model has fundamentally changed. All data centers are a mix of CPUs and accelerators. When training a new model, the accelerators do much of the work as they take in new data to adjust all the billions of parameters in the model. The CPU is just there to manage the process. However, as you move to inference workloads, the CPU does more of the work. A user end as a prompt, that request is sent to the cloud. And from there, CPUs orchestrate the process, managing workloads, coordinating data flows and returning the results. The accelerators generate tokens and each token, which is essentially a word or unit of output is part of a much more complex system response. So while accelerators remain important, CPUs are central, both in traditional cloud infrastructure and now within AI data centers. Today, a typical data center may contain on the order of 30 million CPU cores per gigawatts of capacity. These CPUs sit across the system, supporting AI clusters, managing head nodes and coordinating activity across racks. At least that was the state of play until very recently. What has changed is the emergence of AI agents. Agents are autonomous programs that can undertake multiple tasks utilizing AI. Agents are not simply answering questions. They are executing workflows on your behalf. They can run payroll processes, scheduled tasks, perform analysis and return fully developed outputs. Importantly, they operate asynchronously and continuously. This shift has significant implications. First, the number of tokens generated per user increases dramatically, potentially by 15x or even more. And that's because agents generate requests far faster than humans and they operate around the clock. As a result, demand on the data center increases sharply. Accelerators continue to generate tokens, but these tokens must be managed, scheduled and routed. That work falls to the CPU. And as agent-driven workloads grow, CPUs become an increasingly critical bottleneck. A useful way to think of this is that the accelerators produce the output, but the CPUs manage the system. They handle the coordination, the scheduling, the execution of complex workflows and the agentic AI significantly increases that burden. The consequence is clear. Data centers now require far more CPU capacity. By our estimates, CPU demand could increase by roughly 4x, rising from around 30 million cores per gigawatt to approximately 120 cores per gigawatt. That is 4x the number of cores, and that creates a new challenge. We're now trying to fit 4x as many CPU cores into the same power envelope. And that's a problem. Happily, we have a solution. I'm holding up here. This is Arm's first commercial chip. It is called the Arm AGI CPU, and it has been designed to be very flexible and can be used for general purpose cloud workloads for agentic AI workloads on its own or in support of an accelerator. We designed the chip in close collaboration with Meta, and they have big plans to widely deploy the Arm AGI CPU in their data centers. We also have visibility of orders from many other companies, and I'll introduce you to some of them in a moment. The Arm AGI CPU has been designed around 3 core principles. These are fundamental to how we think, how we design and how we execute. First, performance. In an environment defined by massive parallelism, millions of threads, constant orchestration, continuous workflows, there is no margin for latency. These systems operate 24 hours a day. If performance falls short, the entire infrastructure slows down. So performance is not optional, it is foundational. Second, scale. The magnitude of investment in AI data centers is unprecedented, potentially trillions of dollars building out tens, if not hundreds of gigawatts of power. That scale extends down to the data center, to the racks, to the boards, to the chips and to the individual CPU cores. Every layer must scale seamlessly to support the system as a whole. And third, efficiency. Ultimately, none of this is viable without efficiency. Power is constrained, capital is constrained. To deliver the required performance at the required scale, it must be done within a highly efficient envelope. Efficiency is what makes the system sustainable. These 3 principles: performance, scale and efficiency define our approach. And critically, we do not trade one-off against another. The design challenge was to deliver all 3 simultaneously. The Arm AGI CPU is based on our Neoverse V3 compute subsystem platform. This is the same CPU that is used in Amazon's Graviton 5 and Microsoft Cobalt 200 chips. Our V3 has been designed to not only be efficient, but also the best-in-class performance. We have added to it a dedicated 2-megabyte Level 2 cache for every core and support for up to 3.7 gigahertz in performance. Okay. We have matched the 136 cores to 96 lanes of PCI Express, which means that the ARM AGI CPU can easily connect to any accelerator, device or expansion memory port. On the memory side, we are leveraging DDR5 for maximum capacity. And to minimize latency, we went with a multi-chiplet design where I/O and memory controllers are on the same die as the compute. Finally, we packaged it all up into a power sipping 300-watt power envelope, tapping into that legendary Arm efficiency. This is a standard 36-kilowatt rack. You will find these standard racks in data centers all over the world. These are the building blocks that companies like Amazon, Google and Meta use to build out their data centers. This one is a bit special. First, it doesn't draw 36 kilowatts. Even filled with 30 trays of Arm server chips, which is 60 chips or 8,160 cores, it still uses much less than 36 kilowatts. Secondly, it delivers 2x the performance when compared to the same rack filled with chips based on x86. This means that for the same CapEx and the same power budget, you double the amount of compute output. Although we only announced the first generation of AGI CPUs last week, we are already completing the second generation, which should come to market in just over a year's time. And we started working on the next generation after that. With a new product established, let's look at what this is going to do for Arm's revenue and profits over the next 5 years. Note that we're calling this the first phase of our market expansion because there is more to come. Let me start with the framework. Customer demand for Arm created chips and the size of the opportunity has led us to ramp investment in R&D over the past 2 to 3 years. After years of exploring the market and developing our own internal capabilities, we're now introducing the first family -- the first in a family of chips for the data center. Furthermore, Arm's IP business is going from strength to strength. Royalty revenue growth and license revenue growth both have multiple, long-term structural growth drivers, which we expect to continue for years to come. Finally, the financial consequence. By FY '31, the combined model is materially accretive to revenue, gross profit dollars, operating profit dollars and EPS. Importantly, much of the R&D investment is already in the business. So additional chip gross profit dollars has meaningful drop-through to earnings. Arm's opportunity is huge. We are growing into the largest market in history, one that is just getting bigger and bigger. It is over $500 billion today, and we think this grows to more than $1.5 trillion in FY '31. And just to be clear, this is just semiconductor logic, CPUs and XPUs, no memory, no optical, just chips where you might Arm technology either today or in the future. Breaking that down a little more, cloud AI is a $330 billion market today, growing at around 35% a year to more than $1 trillion in FY '31. Edge AI is a $180 billion market, which we expect to grow at around a 7% CAGR over the next 5 years to $250 billion. And physical AI, which includes automotive applications and robotics, that is a $25 billion market today, doubling in size over the same period. And this then adds up to more than $1.5 trillion. Double-clicking on the cloud AI market, we are forecasting more than $100 billion of cloud AI and enterprise data center CPUs and $55 billion for wired and wireless networking. The remaining $1 trillion includes data center accelerator chips, and we will come back to that opportunity another day. Today, we are just focused on the CPUs, and there are around $500 billion of CPUs that are in Arm's sweet spot. And I want to direct particular attention to the $100 billion of data center CPUs as that is the market being addressed today by Arm AGI CPU. These numbers may seem high. However, we do have visibility into our hyperscalers road map. And with the rise of agentic AI across all data centers, we are confident in this estimate. Okay. So how are we landing the new chip business? This business is entirely based on customer demand from multiple companies across hyperscalers and large enterprises. These are companies who prefer to buy a chip from us over building one from our IP. We believe that Arm is uniquely positioned to build a CPU for the data center. If you look at all of the Arm-based chip CPUs from Amazon, from Microsoft and from Google, most of the technology in their chip already comes from Arm. Because we have such a strong initial demand, we've been able to quickly turn customer interest into actual business. And you'll see on the next slide, we already have multiple customers lined up, and we have line of sight to over $1 billion in chip revenue over the next 2 years. We expect material revenue from the Arm AGI CPUs starting in FY '28 with an exponential ramp to around $15 billion in FY '31. The first driver is naturally increasing chip volumes. We also expect rising chip complexity and core count to lift ASP significantly by FY '31. On this slide, we have just some of the customers that are placing orders for our new chip. Meta, I already mentioned. We also have OpenAI, SAP, Cloudflare, SK Telecom and F5. There are a similar number of companies that did not want their names to be public, but who are also interested. The chip business is additive to our existing license and royalty business, which remain very robust. Arm's royalty revenue has multiple secular growth drivers. The end markets into which our technology is being deployed are growing. We are gaining share as our customers deploy more Arm-based chips. Increasing complexity is driving up core count, especially in the data center and in high-end automotive chips, which leads to a higher royalty rate. And our most advanced technology commands a higher royalty rate. Over the past 5 years, royalty revenue has grown at about a 14% CAGR, and this has accelerated to over 20% in the past 2 years as Armv9 and CSS have started to ramp. Looking forward, we expect that royalty revenue will remain 20% for the next -- sorry, royalty revenue growth will remain at 20% for the next 5 years. Maybe not every year as there will still be the occasional market downtick or inventory correction. But looking through those, we think the long-term average for royalty revenue will be 20%. One of the questions we often get is around our visibility into future revenue trajectory. I think this helps to tell the story. Many of the contracts that underpin our royalty revenue forecasts are already signed and the royalty rates are already agreed in the contract. We've also delivered the technology. Our customers are building the chips. And in many cases, they're already shipping the chip in high volume. Looking out over the years, '27 to '31, over 70% of the revenues we are forecasting to collect are already covered with royalty rates set in the contract. Even by fiscal ' 31, the contracted base is still around 60%. And the remaining 40% is almost all with existing customers who we are confident will want access to the next generation of Arm technology and typically at higher royalty rates, higher core counts and at higher volumes than they do today. Over the next 5 years, we expect that cloud AI will be our fastest-growing revenue driver even without the contribution from AGI CPU chips. And when you add in chip revenue, it will surpass Edge AI in FY '30 and become by far the majority of our revenue in 5 years' time. And finally, to licensing. This has been growing well ahead of our expectations. At the IPO, we said that it would grow low single digits, which we quickly uplifted to mid-single digits, and it's then been growing over 20% per year. And this is being driven by a combination of the AI cycle, more customers getting access to Arm technology through subscription licenses and compute subsystem agreements and also the expansion of our license and design service agreements with SoftBank. We think all these drivers will continue. We expect the AI cycle will continue to provide the majority of the growth through more demand for next-generation CPU IP and compute subsystems at higher royalty rates. And we expect that the SoftBank license and design services revenue will grow at high single digits. And of course, the strong license revenue growth should lead to higher royalty revenue growth in the years and decades to come. As I mentioned right at the start, the Arm AGI CP business, the royalty streams and the license revenues all compound one on top of the other. The chip business is targeting customers who either don't have the internal resources nor have the desire to develop their own chips. So these are all new customers to Arm. We do not expect the chip business to cannibalize the IP business, although if some companies do ultimately choose to switch, that is accretive to earnings as we previously discussed. With around $15 billion of revenue from the chip business expecting in FY '31 and another $10 billion of IP revenue, we are forecasting very strong revenue growth from the combined business over the next few years. And the good news is that we've already done a significant part of the heavy lifting when it comes to hiring the engineers needing to hit our plan. If you've been following our financials for the past few years, you will know that we've already ramped R&D to support our product road maps. Increasing R&D, combined with good execution creates a virtuous cycle of new products driving revenue growth. From here, we are forecasting mid-teens OpEx growth through to FY '31. Most of the incremental spending is on R&D investment going into new technologies. We expect our revenue by FY '31 to have grown by more than 2.5x faster than our non-GAAP total costs. So as revenue and gross profit scale, particularly in chips, much of that incremental gross profit can drop through. That is the operating leverage in our model. For [ FY '31 ], we then see 2 meaningful profit engines. First, we expect the IP business to reach about $10 billion of revenue, reach its 99% gross profit margin and deliver over 65% non-GAAP operating margin. Second, we expect the Arm AGI CPU business to reach about $15 billion of revenue with a gross margin of at least 50% and a non-GAAP operating margin of over 30%. Put those together and the consolidated business has $25 billion of revenue, industry-leading blended gross profit and operating margins and yielding more than $9 of non-GAAP EPS power in FYE '31. This is not a story of choosing between IP and chips. It is a story of combining a very high-margin IP model with a large, fast-growing and accretive chip business. Let me close on the 3 points I started this section with. First, customer demand is allowing us to materially expand our opportunity through selling chips. We already have line of sight to more than $1 billion of revenues from some of the companies I mentioned earlier, and we are forecasting $15 billion of incremental revenues in FYE '31. Second, our existing IP business continues to have strong underlying growth drivers with the chip business compounding and not cannibalizing the IP business. We expect this to deliver around $10 billion of revenue in FY '21. And third, the combined model is significantly accretive to revenue, gross profit dollars and operating profit dollars with more than $9 of EPS power. And because much of the R&D investment is already in the base, the incremental economics are very attractive. And with that, I'd like to turn to questions

Now until 11:30 Japan time, we are going to open the floor for questions. Questions can be accepted either in English or Japanese. We would like to entertain as many questions as possible. [Operator Instructions] First, we will take questions from the sell-side analyst on site and then we are going to entertain questions from online participants. [Operator Instructions] Tokunaga of Daiwa Securities.

Two questions. First question is the semiconductor business of SoftBank Group, Ampere, Graphcore ], there are different companies that belong to the same group and that you are going to advance into the chip business from the viewpoint of the semiconductor business of the SoftBank Group, is it be instrumental in bringing synergetic effects? Or are they supposed to be separated and [ segmented ] in terms of the business strategy?

Yes. within, I guess, SoftBank's AI compute group, you have Ampere, Graphcore and now Arm. However, the chip that we've just announced is independent of anything being worked on by SoftBank's semiconductor companies, Graphcore and Ampere. This is entirely an Arm initiative, and there is no link through to what SoftBank is doing. Obviously, there is a lot of work being done. I think the Graphcore has had anyway about 500 engineers. I believe they're expanding their capacity. Ampere had about 1,000 engineers, and they may also be expanding their capacity as well. But those engineers are working on something completely different. As to when that will be announced, what they're working on, that's probably a question for my colleague over here. But for now, the only thing we can talk about is the CPU chip that Arm has just announced.

There is a second question on Page 37, CPU business. And in FY '29 and '30, there's going to be a huge jump in terms of the sales. Is that because of the inference demand is going to rise or capturing customer demand or mass production system will be kicked in. What is the reason behind this huge jump you are expecting in those years?

Combination of both of those. By the time we get to FY '31, we expect that our third chip will be in ramping into high volume. So the first sort of bumps are the first chip and then it starts to ramp, the second chip starts to ramp. And then by the time you get to the final bar in FY '31, you now have 3 chips from Arm all ramping. That is also combined then with, we believe, the sort of expansion of inference within the large data centers and with particularly the opportunity from Agentic AI. So it's a combination of both multiple chips from Arm and also an expansion of the opportunity with more agentic AI.

I need to let you know before. But this is Arm business briefing. So we appreciate it if you could just ask questions regards the ARM business. So the second person in the second row.

Thank you very much for your question. That is very comprehensive and really great opportunity for us to understand what A basic going forward. And my question is that setting your own ARM AGI CPU mean you are now competing with your big customers like AWS, Graviton or Google, Axion. In my understanding that beauty of the Arm business model is neutrality. So I just a little bit worried about that concern. So I'd just like to hear your thoughts on this.

Yes. Thank you. So AWS is building chips to be deployed only in AWS' data center. Microsoft is building chips to go into Azure and Google is building chips going into GCP. None of those companies are selling their chips to anybody else. It is only for their own internal use. maybe NVIDIA is selling chips in this area. But even NVIDIA seems to be wanting to sell systems, not individual chips. So if you're a Meta, if you're Cloudflare, if you're SAP, you have no choice. There is nobody to sell you a chip. And so we've kind of found a market that is very underserved. And we think one of the reasons why it is so underserved is that these chips contain largely Arm IP. So if you are another -- if you're a chip developer, there's very little opportunity for you to add value or to differentiate. So it's not been an attractive market for someone to go after. But for Arm, because already it's all based on Arm technology, it's a very easy and attractive market for us to address.

And my second question is margin, which is a bit early to think about, but generally speaking, saying hardware usually has lower profit margin than licensing IP. So how will this change your long-term profit target or the mix of profit margins going forward?

Yes. Thank you. So you're absolutely correct. And as we indicated on one of the earlier slides, I think is it 2, no? Here we go. Yes. So the IP business in this time frame will probably have about a 99% gross profit margin. and a greater than 65% operating margin. And the chip business, we think will have in this time frame a better than 50% gross profit margin and a non-GAAP operating margin of 30%. So you're absolutely correct. The gross profit margin is lower. But what we are expecting is that we will provide information so you can do a sum of the parts calculation on valuing the IP business as you do today and also then the new chip business. The other thing I'd point out is that today, we have an IP revenue stream. We have the cost of an IP business, and we've got the cost of a chip business already in the company today, and we're still delivering around a 40% operating margin. So any additional revenues from the chip business, even though there will be a lower gross profit impact, but those gross profit dollars, most of those will drop down to operating profit. And so to a certain extent, we have the benefit of already having all the costs associated with the chip business in the business today. Going forward, when I talked about the sort of mid-teens growth in costs, other than inflation and a bit of an extra investment that we're going to be doing into the chip business, most of the remaining costs are actually into new technology areas that we're not talking about today. And so if those are successful, that will hopefully result in more revenue. And if they're unsuccessful, then the margins could be even greater.

So in the second row, the third from the left.

Masuno of Nomura Securities. First of all, CPU, a basic question. Other than Arm architecture, Graviton [indiscernible] compared to those competitors CPUs, AGI CPUs performance, how much is it? Twice as good as by X86. But compared to the other players, what is the performance level? The AGI chip is going to be applied to GPU of Google and Amazon and [indiscernible] Open AI accelerator Cbus accelerator. Can it be applied to all the accelerators? Would you please share with us some technical information on this aspect?

Yes. Thank you. What we have -- looking across the different chips that are available and how they fit with the accelerators from the cloud companies, most of them have been highly optimized to work in their systems. And so for example, Google's Axion chip has been targeted to work with very, very well with the TPU, NVIDIA's Vera chip is being optimized to run with Rubin, their GPU. And because they have been -- the 2 have been built to work very well together, although you could put an Arm AGI CPU in there instead of an AXION, it may not perform as well because the Axion has been optimized to work with the TPU. But if you were to take an Axion and have it working maybe stand-alone, then we think we'd be very competitive. And so if you are Meta, you will have a competitive product to Axion or to Graviton or to Cobalt. So we think it's very competitive. We're not being paired with their own proprietary accelerator. But stand-alone or with somebody else's accelerator, we think it would be very, very similar in terms of the capability. But the key thing is that for most companies, they don't have access to these technologies. So just being able to provide them with the equivalent is very powerful for them.

So if that is the case without own CPUs and AI accelerators alone, so Meta and OpenAI and [indiscernible], these are those players. So this product is going to best fit those players. Is my understanding correct?

Yes, we work very closely with Meta on the design of the AGI CPU. They work as a lead partner, helping us to fine-tune the design to make sure that it fits very well for what they needed. But we didn't want to make something that was only applicable to one company. We wanted something that we could also sell to everybody. So we were trying to do 2 things at once, and we think we've done a pretty good job of producing something that will work well for them as a companion to their accelerator, but also to be broadly used across general purpose cloud workloads, AI -- running AI workloads on itself in the cloud as well as being a companion to an accelerator chip.

The second question is that going forward, license revenue -- guidance on the license revenue, you are talking about 3 elements. The factors for the drive AI cycle for one thing and next generation and then the -- and then the third one is licensing to SoftBank. And I think you were talking about all these things. They are going to continue into the future. Is my understanding correct?

Yes, your understanding is the same as mine. I mean we have not yet seen any slowdown in the AI technology cycle. So we are seeing many companies wanting access to the next generation of Arm technology because they need to run more advanced workloads in their next generation of chips. And we're seeing that in the data center. We're seeing that in people building chips for cars and for robotics, for smartphones, pretty much every end market you can think of companies are building more advanced capability into their next chip. And some of that is going to be driven by Agentic AI. We saw how OpenClaw, which is a local agent that runs in a MacBook Mini or in a Raspberry Pi can be run locally and can effectively work as like a digital assistant, but that's not run in the cloud. That's run on a box next to you. And so we see that going into many, many different types of end product, consumer electronics product. And so that will continue to drive licensing. Similarly, we have only signed up about 30 of our top 50 customers for our subscription licenses. We think that more of them will want to access. And so over time, that will increase the license revenues. CSS has only so far been licensed to a small number of our customers. There are many more we think will be wanting to take it. And so that will help drive licensing. And then the third point is we have visibility into SoftBank's road map and their plans. And so we expect that the licensing and design services that we're doing with them will not only continue, but will continue to increase over the next few years as well. So those 3 drivers of license revenue, we expect to continue for multiple years. If you look on the chart, I can't quite see, we are more confident in the first few years worth of growth. The outer years are harder to see, but that is normal. Whenever we are forecasting license revenue growth, we can see the next 2 years, but then it starts to get harder to see. But by the time we get there, we'll probably have more visibility of more growth.

SoftBank Group after the project is complete. It doesn't grow after the project is complete. But do you still expect the royalty to continue?

Well, most semiconductor companies in general, don't build one chip. They build a series of chips. They have a product portfolio that will continue for many years. They plan for success. And so there will usually be a version like we had earlier, we showed AGI 1 and 2 and 3. So assuming if SoftBank Semiconductor Group has a road map, then they will need more technologies to come in the future. So we're anticipating that there will be multiple projects, not just one.

Then the second person from the first row.

My name is Yasui from UBS Securities. I have 2 questions. First question about AI architecture, about computation algorithm. I think there was a diagram on that. And you mentioned about important software. So for smartphone, iOS, Android, I think there are basically 2 softwares, which you basically monopolize. But what about AI software, the core is could be CUDA or PyTorch or [ Meta ] or the third party, another third pillar. So from your perspective, from Arm's perspective, who will be your software ecosystem partner? I would be curious to know that.

Well, I think you mentioned a couple there. So obviously, NVIDIA's CUDA runs -- the CUDA runs on the part on the Arm CPU and part on the GPU, Meta with PyTorch and ExecuTorch obviously, is an important partner. But also with Google just announced as part of their Axion 2 launch that they just ported 10,000 workloads -- so 30,000 workloads over to run on Arm. Microsoft have announced how they're porting Teams across to run on Arm as well. And all of the software that needs to run in the cloud has also now runs on Arm. So I think Amazon recently announced of their top 1,000 customers, over 99% of them now run some of their software on Arm. So pretty much all the software that you can think of for the cloud now runs on Arm. There's very little that is now commercially available on our base chips.

I guess the image I had was you have this big NVIDIA ecosystem and then you have the cloud service ecosystem, maybe like orchestrators like or SS, maybe it's container-based or test based, it seems like it's running there. So from your perspective, who is the important software ecosystem partner from your perspective? So for example, you have NVIDIA CUDA, right? That's one. This is a cloud service provider OSS? Or do you think there are multiple, many more OSS kernel base? Would you be able to elaborate on that as well?

I mean I think all of it really. I mean all of the software that runs in the cloud is now already running on Arm. So I think there is nothing that's missing. So you mentioned containers. So yes, you can run Docker or Kubernetes on Arm. Python and Perl interpreters are optimized and running on Arm. So it's all of it.

And then my second question is about 1 gigawatt basis. So what's the assumption? And in 2030, it's $15 billion, that's your assumption. So would you be able to give us a rough picture until 2030. So if it's 1 gigawatt, I think it's 131 per chip is what, maybe 200 or 300 core, right? And if you actually translate that into CPU counts, it's about 200 core, it's about [indiscernible] it's about 0.4 million CPU. Maybe that would be necessary for 1 gigawatt data centers. So I want to know the assumption between 1 gigawatt and 2030, 15 billion because I guess the intention is that what's the CPU shipment when you say 2030, 15 billion.

Yes. I wondered if you were going -- if your question was leading to how many chips? We're not disclosing the number of chips, mainly because obviously, pricing -- per chip pricing is very much related to volume. And we see multiple ways of getting to 15 billion. It could be just through a small number of large customers or it could be through many smaller customers. So we've decided not to disclose chip pricing and chip volumes in the same way that most semiconductor companies also don't disclose chip volume and chip pricing. We'll keep that secret.

Then in 2030, the number of chiplets is 2, right? So could it be 4? So per CPU, how much performance improvement do you expect? Could you give us a ballpark figure on this?

So just taking a step back for a moment. So indeed, that is a -- our first chip is a 2-chiplet solution. One of the things that we are trying to create at the moment with other companies within the semiconductor industry is to create a chiplet marketplace. So at the moment, anybody building a chip based on chiplets, so including AMD and Intel and NVIDIA are doing so using proprietary interfaces that they have had to develop themselves. What we're trying to create right now is in the same way that our 30 years ago, we created IP standards so that multiple companies could integrate IP from -- IP components from multiple companies into a single SoC. But with standard interfaces, standard busing, standard protocols inside the chip, we're now trying to help the industry to coalesce around standard interfaces between chiplets. And so you could take a CPU chiplet from Arm, an accelerator chiplet from an NVIDIA or Rebellions or whoever, maybe an I/O chiplet from somebody else and then integrate those all into a single system-on-chip design or a single package, sorry, I should say. And that then could mean that you could then choose as to whether you have 2 CPU chiplets as we've done with our first chip or 4 or 6 or 8. As I mentioned earlier, we are already seeing some companies who are planning 500 core chips, which would be multiple chiplets. And we've seen PowerPoint presentations, I should say, and plans, where the top right-hand chip is over 700 cores. And so significant expansion in the cores. As I mentioned earlier, our first chip has 136 cores, but we are expecting that over time, the core count in the products that we also produce will increase as the sort of the need for more CPU performance [indiscernible] inference data center increases. I can't tell you what that chip count will be in FY '31 or core count in FY '31. But certainly, we're expecting it to be higher.

We would like to take questions from the online participants, if there are any. [Operator Instructions] SMBC Nikko Securities, Harad-san. [Operator Instructions] Well, if it's difficult, in the interest of time, we would like to move on. Any other questions online? Then, on-site, we would like to come back to the on-site questions. CLSA.

Congratulations. So just to clarify, your EPS this year is tracking at less than $2, and you're saying it's going to be $9. So it's going up by about 5x. And that does not include anything from potential accelerator market expansion. And you mentioned $1 trillion. Could you just say what you thought about the TAM of that accelerated market? I think you mentioned $1 trillion, but you kind of skipped over it. So my question is, just to clarify, that's the magnitude of profit increase? And then could you tell us a bit more about this potential accelerator market? And any comments on what you would do with all that cash?

Okay. All right. So 3 questions there, I think. So yes, we are anticipating, forecasting more than $9 worth of EPS -- in non-GAAP EPS in FY '31, so in 5 years' time. About $6 of those comes from our existing IP business and 3 of those come from the new chip business. The TAM for the accelerated market is indeed, we anticipate growing to more than $1 trillion from about $250 billion this year, so quadrupling in size. We do think that's quite an attractive number. And last week, when we announced our CPU chip, we did sort of have and watch this space for future, but nothing to announce just yet. And in terms of what should we do with the cash from that, we have nothing to update you on right now. So we generate at the moment about $1 billion a year in free cash flow. We aren't doing a buyback at the moment. Most of the requests from investors are for more shares, not fewer. And so doing a big buyback doesn't seem to be a good idea. We remain open to considering a dividend, but that is something that has not -- no decision taken to do one just yet, but we remain open to investor requests there. And we do, do the occasional acquisition. Most of our acquisitions, however, are really accelerating hiring. We've been hiring between 1,000 and 1,500 engineers per year. And as you hire them in sort of 1s and 2s, it takes 6 to 12 months to build an engineering team that can build new technologies for you. If you acquire a company as a whole, you end up with an actual team coming in that's already set up and running. It may be that you don't want them to develop their existing product, but they can -- you can give them a project to go and do and you've got the senior engineers, you've got the junior engineers, project managers, they often come with a building and with facilities managers. So the whole thing comes in a nice package that you can then go and get ramping up very quickly. So that's something that we have done from time to time and we'll probably continue to do. But those typically be a few hundred million dollars. They're not going to make a huge dent in the billion.

Any other questions? Anyone? Third person from the corridor on the front row..

My name is Matsu. So on Page 37, you have a revenue projection. So the big 2 customers, OpenAI and Meta, how much would they comprise this revenue? And what is your visibility? How good of a visibility do you have, like, for example, firm commitment?

I'm not going to give specific customer information. Contracts with them obviously are confidential. But as I mentioned, we have visibility to more than $1 billion worth of revenues. And if you look on Slide 37, there's a little bit in FY '27 and a bigger amount in FY '28. The sum of those 2 parts is more than $1 billion. And we have very strong interest to deliver on that. So we are as confident as you can be in revenues that far out.

The second question. So when you're able to actually book a revenue, what would the CPU's P&L trend? So for example, secular, you're expecting $3 of EPS. But of course, you have upfront investment. So is there anything that's accretive that could actually push down the EPS?

Well, I guess there's 2 things that could push down EPS. Firstly, if the revenue doesn't come through in the way that we expected or if we increase investments. So staring into our crystal ball, I think that we have -- if Rene was here today, he would be telling you that he is very confident in the $15 billion and can see multiple ways that it could be -- we can achieve that with -- just with the customer set that we have today. And in reality, we'll probably be building a bigger group of customers. And so therefore, it could be even greater. But it's 5 years out. So there is obviously lots of uncertainties, and so it could be lower. If we end up having another war, which seem to be brewing all the time or another credit crisis, could be -- lots of, obviously, uncertainties, but that's based on the customers and the customer demand we can see today.

Maybe it was lost in translation. My question was -- that's, of course, helpful. But my question was for the next couple of years, what should we expect from the chip CPU business? Is it going to be disruptive to EPS or accretive to EPS while you're running business at the early stage?

On the basis that almost all of the costs associated with that business are already in the business today, any gross profit is going to be -- is going to help grow EPS. So it is going to be accretive in the near term, assuming we are executing on that plan.

We are getting closer to the ending in time. We only take 2 more questions. If you still have a question, please raise your hand. In the third row from the front, the second from the corridor side.

My name is [indiscernible] of Goldman Sachs. One question for me. In terms of CPU, the power envelope, it has to be lower going forward in terms of the demand. And what is often talked about is GPUs, the consumption of electricity is relatively high. CPU consumption rate of electricity, what is the rate for CPUs? I think CPUs, the power consumption needs will be lowered in addition to GPU. Do you have any ideas about how to lower the power consumption of CPUs?

Yes. Thank you. On a chip-to-chip basis, a CPU will be lower than a GPU. As it is showed earlier, the maximum power used by one of our CPUs is 300 watts. That is a fraction of a GPU. However, as we also saw earlier in agentic data center, you may have many more CPUs than GPUs, whereas in a training data center, you have more GPUs and CPUs. So I think as we move towards more agentic workloads as we all start using AI more in a day-to-day life, then the data centers are going to have many, many more CPUs deployed with them. And that's going to mean that the energy efficiency of the CPU is going to become increasingly important. So we'll be having to deliver more and more compute in as minimum power budget as possible. And I think as you saw earlier, we have the 36-kilowatt rack that's the sort of the maximum energy draw that that can draw. We don't even draw all 36 kilowatts. And so we are delivering much greater power efficiency than anything that is currently being installed in the data centers, which is helpful. But clearly, this is going to become a multiyear focus, trying to get more and more compute out of the same amount of energy. And that happily has always been Arm's focus. Ever since we were started, we always were trying to create more compute out of fewer electrons, and that is what we'll continue to do going forward.

Do we have any other questions? The person in the front row. So can you just limit yourself to one question because of the time constraint.

My name is Satoru Kikuchi from SMBC Nikko Securities. I only have one time to ask one question. Related somewhat close to the question earlier, but so AGI CPU target of $15 billion, the breakdown of that? What is the specific contribution of OpenAI [ and Meta ]? I guess you are not able to give the specifics, but can you give me in terms of the amount -- like the breakdown of the hyperscalers? And on Page 40, you said that it's deployable to the new customer segment, which you were not able to capture before. So what is the specific customer segment that you're targeting? And is that included in this $15 billion in target?

So the $15 billion assumes that our existing customers in this area, so Amazon, Microsoft and Google do not buy chips from us. We assume that they continue to develop their own chips. And so the $15 billion is only made up of new customers, some of the logos that we showed before. And for the $15 billion, we see multiple ways of getting there. It does not necessarily need to be just Meta and OpenAI, it can be the other companies as well. We see opportunities much greater than 15. We're just setting $15 billion as a target as going from nothing to something is always hard to forecast, but we think it could be even higher than $15 billion

Then with this, I think we would like to end today's Arm's business briefing. And this will be streamed on-demand basis on our IR website. So thank you very much for your participation.