00:00ARM coming out with its very first AI CPU, ARM deciding to become a chip maker itself,
00:06having been a provider of intellectual property and blueprints for chip makers,
00:09it is now becoming a chip maker. How significant is this for ARM?
00:13How significant is this for the semiconductor industry?
00:16Morning Tom, it's great to be here.
00:18And we think it's very significant because ARM is now a full stack provider of the CPU.
00:25And what we see time and time again with technology platform transitions is that the lion's share of economics
00:31goes to a handful of players and these are full stack players.
00:35Think of Microsoft in the PC era, Apple for the mobile, AWS for the cloud.
00:41And ARM has got the software ecosystem, 22 million developers, that's over 80% of the global total.
00:49It's got the IP and now they've got the hardware, they've got the CPU.
00:54Now the key question is what's more important, designing the CPU first and establishing leadership,
01:01think of Intel, AMD, or building out the ecosystem first and then designing the chip?
01:07And we actually think it's the latter because designing the chip, okay, it's hard,
01:12but it's not the hardest part. Building an ecosystem takes years, it actually takes decades.
01:19If you think about CUDA, NVIDIA's software ecosystem, that took two decades to build out.
01:26It's got over 5 million developers working within CUDA today.
01:31ARM has done the same over 35 years and that ecosystem is actually four times larger.
01:35So we think this is very, very important for the ecosystem.
01:39It places ARM in a pole position, in our view, for agential AI and AI at the edge.
01:46So you don't think, because the criticism from some might be, while the hyperscalers are already doing custom silicon,
01:52you've got AMD, you've got Intel, you've got NVIDIA, of course NVIDIA, Jensen Huang announcing their own CPU.
01:58The criticism could be that ARM is too late, that they're late to the game on this.
02:04So we think that NVIDIA announcing their CPU signals that this is a large and growing addressable market.
02:12Because what we've seen so far over the past two years is that the focus has been on the GPU.
02:17And that is largely because the first three scaling laws that's been driving compute demand are focused on the GPU.
02:26That's pre-training, that's inference time reasoning and reinforcement learning.
02:30We're now entering the fourth scaling law, agential AI.
02:34And this is where the CPU actually has a pivotal role.
02:38Now, the hyperscalers, yes, they're designing their own custom silicon.
02:42But when it comes to the CPU, that is almost wholly on ARM-based architectures.
02:49Think of AWS's Graviton.
02:52ARM is delivered for them 40 to 60 percent better performance per watt.
02:56The same goes for Meta, who was obviously the launch customer for ARM's AGI CPU recently, and Microsoft as well.
03:07They're all building on ARM.
03:09And ARM's CEO, René Haas, has spoken to us, of course, and he said that Argentic, to your point, is
03:15a key driver, a key catalyst.
03:17And he's going to be a key driver for that market share expansion.
03:22Where are we in this journey when it comes to Argentic AI?
03:25It's an excellent question, Tom.
03:27We are scratching the surface.
03:30And so the past two and a half years have all been about those first three scaling laws.
03:35Argentic AI has really only started to take off over the past three months.
03:40Claude Cowork and OpenClaw have really served as the catalyst for Argentic AI.
03:46And so we're in the very, very early innings.
03:49We do not think the market is prepared for the run-on inference compute that we're about to see.
03:54Because Argentic workloads, they're not just giving you answers.
03:57They're going away and automating tasks end-to-end.
04:01They are calling tools, calling APIs, and operating the task for you.
04:08Now, all of that work, that orchestration work, is CPU work.
04:13And so the GPU remains the workhorse, but the CPU attach rate actually increases.
04:21And this is also because the GPU performance has increased by such an order of magnitude over the past three
04:29years.
04:30And the CPU hasn't caught up.
04:32You need more CPUs per GPU in order to keep the data fed and orchestrated.
04:39What are you modelling then in terms of potential additional revenue streams for ARM as a result of this new
04:46chip?
04:46The most important statistic that Rene announced at ARM Everywhere was that Argentic workloads require four times more custom CPU
05:00cores per gigawatt in the AI data centre.
05:03That is an order of magnitude higher and speaks to the pivotal role that the CPU plays within Argentic AI.
05:15The other point which is worth mentioning here and not baked into numbers at all is physical AI.
05:25So AI at the edge.
05:27If you think about why ARM is so prevalent today in smartphones, underpins 99% of smartphone architectures, there's no
05:35comparable monopoly position in technology today, not even Google search.
05:39This really comes down to energy efficiency.
05:42And the same low power architectures are required in every robot, in every autonomous vehicle.
05:49So for every autonomous vehicle, we estimate there are probably around 100 ARM chips required.
05:56And even Tesla, the ultimate vertically integrated company, designing its own silicon for its optimist humanoid robot, that's being built
06:07on a custom system-on-chip ARM architecture.
06:10So we're modelling revenues in excess of what's being baked into the price right now.
06:16Are the markets, are investors, underestimating ARM still?
06:22Yes, we believe so.
06:24And the reasons are threefold.
06:27First is the full stack ecosystem advantage.
06:32And this, by the end of this decade, we think will propel significant network effects.
06:37AMD and Intel, we think, possibly might lose market share.
06:42But then again, Rene's total addressable market was about double what Lisa's put out there for the CPU and the
06:49data center.
06:50So it all will benefit to an extent from the takeoff of agential AI.
06:54The second thing that's being underestimated is how large agential AI will become.
07:00And this will benefit ARM, we believe, to a greater extent than peers because of the energy efficiency.
07:10As long as power remains the binding constraint on the AI day to send a build out, which it is,
07:18there are a number, but it's an important one, the best technology wins.
07:22So that's NVIDIA for GPUs, that's ARM for CPUs, because this has a direct impact on my total cost of
07:30ownership as a hyperscaler, but building my data center.
07:35And agential AI is only scratching the surface, particularly in Europe, but even in the US, we're in the very,
07:43very early innings here.
07:44And the final point being underestimated is AI at the edge, physical AI.
07:49Is the cost to this for ARM that it can no longer claim to be the Switzerland of the semiconductor
07:56industry, that it can no longer claim to be a neutral party that can supply all chip makers, as it
08:03has done up until this point?
08:05Is that a risk for ARM?
08:06This was the key risk, and we think that ARM made the right call because they were essentially trapped.
08:17ARM's key customers were downgrading their relationship with ARM in order to pay them less.
08:26So moving from pre-designed ARM CPU cores to designing their own CPU cores and just paying ARM the bare
08:36architecture license fee.
08:38Qualcomm was the canary in the coal mine there with their acquisition of NVIDIA.
08:42So if they weren't to make this move and launch the AGI CPU, we think they could have seen just
08:51a gentle erosion in royalty rates.
08:53Compete subsystems, which have been gaining traction, where ARM sell the entire compute engine, not just the IP for one
09:03custom core.
09:04It's all sort of pre-tested, and this speeds up the go-to-market for customers by about a year.
09:09That was one move in the right direction in order to monetize their IP better.
09:16But the ultimate step is producing their own AGI chip.
09:21And crucially, this is for companies that don't have their own silicon design teams.
09:27You'll note that the launch customers at ARM Everywhere with the AGI CPU, these are customers that are not designing
09:39their own CPU in-house.
09:41Meta is the perfect example, but notably absent from the list, the likes of Amazon, the likes of Microsoft.
09:48So we don't think this is cannibalization, we think this is a net new revenue stream.
09:53Just stepping back, is it still fair to think of ARM as a UK company?
09:59Headquartered in Cambridge, 90% owned by SoftBank, a Japanese company of course, publicly listed in the US.
10:08Is this still a consequential UK company?
10:12How significant is ARM to the UK and the European semiconductor ecosystem?
10:18We do think that ARM has a pivotal role in the UK ecosystem, but part of the company's success has
10:26been migrating onto the global stage.
10:28And in order to become the default full stack provider of the CPU, which we think is the path the
10:36company is taking, you've got to be global.
10:41Almost all companies, particularly in the semiconductor industry, the most global supply chain in the world,
10:50their customer base, their supply chain doesn't really reflect where they're located.
10:55Are the ingredients there in the UK, in Europe, to germinate, to build another company, another tech company, another company
11:05within the chip ecosystem,
11:06with the kind of size and scale that ARM currently has?
11:10Are there lessons to be learned?
11:12Definitely lessons to be learned.
11:14And the way we sort of view Europe when it comes to AIs is there are three layers.
11:20And you've got the supply chain, which is world-leading, ASML, Bezzy, Aikstron.
11:26And that's the most important contribution that Europe really provides for AI on a global scale.
11:34But then you've got the model and the startup ecosystem, which is anchored by Mistral.
11:41And that's looking pretty promising.
11:44And then finally, you've got a very unpenetrated adoption ecosystem.
11:51And this is probably the cause for concern.
11:54Only about 3% of European companies, according to AWS at the end of December 2025,
12:01who has a very good look on who's using AI because they're the largest cloud,
12:07that only 3% are using agential AI.
12:10So we can close that gap for sure.
12:13But the model and startup layer is looking very encouraging.
12:16Here in the UK, we've got WAVE, which is end-to-end autonomous driving.
12:21You've got Black Forest Labs in Germany.
12:24There are a number of companies coming to market over the next decade,
12:28we think, in Europe that are going to add a lot of vibrancy to Europe's AI ecosystem.
12:33testing.
Comments