Latest / Autonomy Insiders / $41 Billion in Physical AI: What Investors Need to Know About Robotaxis & Humanoid Robots

Transcript

$41 Billion in Physical AI: What Investors Need to Know About Robotaxis & Humanoid Robots · Autonomy Insiders

• Daniel Abreu Marques: Welcome to Autonomy Insiders. Today we are joined by Martin Briggs, Director at Bank of America Global Research, where he leads the investing team. Martin and his team just published a major primer on what they call physical AI. And the core thesis is, AI is leaving the chat and entering the real world, moving from screens into machines that perceive, reason and act. And the first large scale proving ground for this shift is autonomous driving. And that's what we hear talk about. Today with Martyn Briggs. Welcome to Autonomy Insiders, the show where global industry leaders in autonomous driving unpack their real-world insights. So Martin, welcome and so happy to have you on the show today. MB: Thank you, Daniel, great to be with you and I look forward to chatting. Daniel Abreu Marques: So your report is titled AI Left the Chat And for our audience, what is physical AI and why does your team frame autonomous driving as a first major proven crime? MB: Sure. Well, thanks, Daniel. think we went to CES early in the year in January, and as you probably noted, many of the discussions and the launches and the announcements that were coming out of that were around physical AI. It's something that we've written about in the past over the last year or so. to frame what we're talking about here, physical AI is the shift from AI that only generates digital outputs to AI that can perceive, reason, and act in the real world. ⁓ catchy a punchline, as you put it, is AI moving from screens to machines. What this enables are a combination of things, So autonomous vehicles we'll talk about extensively today, it also enables potentially things like humanoid robots, drones, and the notion of moving from pre-programmed automation to adaptable ⁓ intelligence. And I think that's the key important message here is think of it of giving AI a body, know, machines that don't just process information, but can safely interact with people, navigate complex environments, and deliver real-world productivity. So there's a number of different technologies that are enabling this that we went into in the report, notably the advancement of AI models, ⁓ language and world models in particular, ⁓ also with sensors, robotics and edge compute so that these machines can understand their environment and start to execute physical tasks autonomously. So this is all being accelerated now by a combination of the hardware getting cheaper, better simulation. and a massive wave of investment, course, which is why we wanted to look at it for our investor audience. saw about $41 billion that were invested last year in physical AI and robotics, and this trend is only set to continue, we think. We saw about $41 billion that were invested last year in physical AI and robotics, and this trend is only to continue, Daniel Abreu Marques: Yeah, and you quickly touched already on world models. ⁓ So this is kind of one key architectural shift beyond the LLMs. And can you explain to our audience what a world model does that a large language model cannot ⁓ why that is also essential for autonomous vehicles? MB: Absolutely. think, it feels like that. if we take a step back a bit, that we put wrote bit about the chat GPT moment from ability, right? It's the first point. And we'll talk about world models in that context, because it feels like the enabling forces are here are converging at once, right? You similar to what they did with generative AI. As I mentioned, got better AI models, more data compute, and that continues to bring more sophisticated AI applications. But two things have here. One is that the vehicles have now got some of the onboard sensors compute the software architectures needed to run some of these advanced AI models in real time. And that's only going to get better. And on the other, you've got huge volumes of driving data and simulation and synthetic data are coming through. So that needs to be brought together by something. And the difference with say, GenAI with LLMs versus world models and physical AI is that LLMs were using largely data and or text and data and, ⁓ you know, being able to update that in a physical AI world models are bringing that into the real world. So perceive reason and act. so a world model is fusing together different, ⁓ taking, guess, foundation models, VLAs and VLMs in particular. So vision language models and vision language action models to enable AI to be able to do that perception, and then operate in the real world. So that's the difference here. bringing together some of the capabilities and the tools needed to do that, which you didn't need in the generative AI ⁓ initially, which was just pointed towards digital applications. Daniel Abreu Marques: And who is leading in this world model race, so to speak? Is this big tech or are these also startups? MB: I think a bit of both. Yeah, we wrote a bit about it in the report is that you've got on the one hand, you've got more startups that are entering the world model market here. As I mentioned, you had about $41 billion invested in physical AI. But some of that, you know, a small proportion of that was was dedicated just towards world models. But the number of equity deals rose tenfold in the last five years. So you're seeing a number of different funding rounds ⁓ to world model developers now. To answer your question, though, it's a combination of dedicated startups, companies that are going after this ⁓ as a bespoke opportunity. But you've also got all of the big tech ⁓ that are developing world models as well, right? We saw several announcements of that at CES, in particular the likes of NVIDIA with Alpamayo. They've got, obviously that's for autonomous vehicles. They also have a number of different robot models. I say a lot of the hyperscalers see that as an to develop some of the tools that... that companies will need to develop physical AI on the one hand, and that can bring a lot more investment into the chips and the AI infrastructure that will only benefit the hyperscalers. at the moment, it's kind of both startups and smaller companies are developing these to try and become the first commercially viable, but all of the big tech companies are also developing those in parallel. And I think the final interesting part of that is that a lot of them are making world models available via open source channels, right? And that's speeding up the development, the flywheel of the models. And it also gives developers a number of different options of models they can trial, right? So it's no longer just closed applications for world to get access to these models. You've got a range of the Googles with DeepMind, Nvidia, DeepSeek in China, several of them that are pointing towards model applications. ⁓ all of the hyperscalers are looking at that as an opportunity as well, for sure. Daniel Abreu Marques: So the foundational technology stack is evolving fast. Let us maybe now look at where it is landing first, so the vehicle. report highlights also a shift from software-defined vehicles to AI-defined vehicles. What the ⁓ practical for an OEM ⁓ making decisions today? MB: There's a few things that are happening here. I the notion is that ⁓ of the key breakthroughs we're seeing happen at the moment is, you know, talk about AI defined vehicles, but I think in general, a lot of the companies are progressing towards this end to end autonomy, right? We spoke a bit about this in the report, how this infrastructure layer could be shared across vehicles and a number of other form factors. But guess physical AI requires redesigning the vehicle more as a robotic platform, not as a mechanical one. you know, just that it's going to be dependent on electronics. So OEMs are doing a number of related things to this. ⁓ guess the first is a fundamental rearchitecture of the rearchitecting the vehicle. seen this from moving from hundreds of distributed ECUs to the more centralized compute and zonal electronic architectures or E-architectures, ⁓ mirroring what Tesla pioneered a few years ago. ⁓ And then, of course, this shift isn't just because of autonomy. It's for a combination of things, right? To run the performance AI models, fuse the required sensor data that's going to be created and enable better features that are upgradable over time. And of course, autonomy is one of those. So several car companies are partnering with chip makers to build some of the AI ready platforms and software stacks in-house. And we saw some of those at CES, for example, Mercedes with NVIDIA and Geely as an example. guess the big question is going to be the pace of that, right, which we can talk a bit about. But many of the OEMs have been constrained by legacy software stacks and the validation processes. And guess the players that are treating this as ⁓ initially platforms are moving towards these AI-native devices, are going to be better positioned for physical AI. Daniel Abreu Marques: Yeah, you mentioned this shift distributed ECUs to centralized compute is kind of at the heart of this. How does this reshape supplier landscape? MB: Yeah, there's a few things, right? you think initially a lot of the traditional tier one suppliers are already starting to take note of that. But it also gives rise to some of the newer tech focused, specifically AI SoC companies, right? So all of the companies such as Qualcomm, Nvidia, Ambarella, many others are trying to get into that market to firstly have the better AI native chips that can service this shift that's going on towards AI native vehicles. But equally, a lot of the tier ones are still, whether it's the Contis or Amovios of the world, and several others, are also ⁓ utilizing that as an opportunity to ⁓ higher content per vehicle. many that that could totally ⁓ disrupt supply chain and it is leading to some new tech-focused players that are entering the market. But ultimately, the OEMs are still kind of in control of what products and services they want to deploy in these vehicles. ⁓ but not many, but some are trying to vertically integrate some of that. Most will still depend on a range of suppliers to deliver this shift towards more centralized compute. I guess long story short, it brings more AI-focused players into the equation on the tech and the SOC front, but the traditional tier one, tier two suppliers are still very relevant in bringing all of that together. But instead of 150 distributed ECUs, will have, you know, maybe it's one, two, four different... a zonal high-power compute and high-performance computers that can bring some of that together and enable the different functionality and feature upgrades over time. So the other thing that this is important for, of course, is that can start to upgrade a vehicle's capability once it leaves the dealership, not once it's just been manufactured. And that's, course, the key unlock potentially towards higher levels of autonomy over time, but even just towards better AI features, better infotainment. more related features. So the supply chain could be impacted by that for sure, but I don't think it's a complete disruption of the whole supply chain. It just means more tech focused players that are entering the market here. Daniel Abreu Marques: And at tech focused we also see more and more OEM and AV company partnerships. For example, Wayve with Nissan or also Waymo had announced something with Toyota. It was more like a letter of so to speak. So not a real deal. But goal is always to integrate the self-driving system into MB: Mm. Daniel Abreu Marques: production cars. is in your opinion the strategic rationale behind that, especially also from the OEM perspective? MB: Sure, I think the strategic logic is simple here, Because both sides need each other to commercialize autonomy at scale. on the OEM side, you've got access to self-driving systems without spending billions of dollars on an AI stack from scratch, they can integrate autonomy directly into mass production platforms. The AV companies can gain what they can't build alone, right? So manufacturing scalability, safety-certified vehicle platforms, global distribution. ⁓ and the ability to test and deploy across multiple models. So both sides here can benefit. OEMs can future-proof their vehicles for higher levels of autonomy and stay competitive in this software-defined era that we just discussed. But the AV companies can gain that distribution, cost reduction and the regulatory credibility. So together they can move far faster towards this end-to-end autonomy stacks than either could do independently. So that's the simple logic. I think what's less clear to me at the moment is if each one or each OEM has one AV partner or multiple, right? So you mentioned Wayve there, we've got Nissan last week, investment as well from Mercedes and Stellantis. You've got several OEMs that are partnering with, well, some are partnering with one AV company, some are partnering with multiple now. So I think it's less clear to see whether it's, you know, each OEM needs one AV partner or whether we'll see multiple. And of course what happens across regions, right? You might end up with one AV partner for the US, one for Europe, one for China or one for Asia, right? So I think there's a combination of things at play here, whether OEMs going to go after one specific AV partner. And if so, if that precludes them from working with others, either because of capacity ⁓ or geopolitical reasons. Daniel Abreu Marques: The report ⁓ also argues that the physical AI technology that is built for the passenger vehicles can also then transfer directly to trucks, mining, construction and also defense. How real is that transfer today and ⁓ also what the limits? MB: Yeah, sure. Well, we wrote about this in this report and also one previously we wrote last year on autonomous vehicles is that, you know, most of the headlines in the recent times have been around autonomy for passenger cars and robo taxis. And, know, we'll talk about that a lot today. And, you know, probably be the largest town overall. But I think there's also this this notion of a shared infrastructure layer where you can utilize physical AI, the same systems being deployed in automotive for passenger cars or even robo taxis. across the commercial vehicle market. And equally, some of that might move faster in terms of commercialization, either because there's more acute labor shortages like you've got in trucks, for example, or because autonomy could be easier to execute in some of those markets, right? In trucks, for example, you know, you've got mostly highway operations. There's some off-highway applications could be brought in on public land. If it's an agricultural application, you've got higher margins for error there in a field with no people around versus an urban area next to a school, for example. So I think in realistic is it to deploy across those different form factors? If we'd have had this conversation maybe five years ago, I'd say it's unlikely one will move to the other. But more recently, every ⁓ developer that you speak to are talking more in the terms of they see themselves as a robotics lab. And therefore, this technology could be ⁓ you know, shifted across form factors with a combination of different sensors or slightly different hardware to sense what's around. ultimately being able to perceive act and apply that across a truck or a car or a tractor or, ⁓ know, anything ⁓ ⁓ to be ⁓ terms that these companies are talking about. So the important part of that from our perspective, of course, is you're dramatically increasing the addressable market. We talked about a a potential doubling of the total addressable market in dollar terms that could bring. And as I say, some of those other areas might move commercially faster, even if they're in smaller total markets in dollar terms. Daniel Abreu Marques: So now maybe let's get ⁓ to the reality today. So your report counts 158, think, active robotaxi deployments globally. How do you the current pace of the industry and where's the real inflection point happen? MB: I think that was some stats that we took from a Bloomberg report on the Robo taxi front. But if you include the commercial vehicle applications as well, there was over 200 operational pilots now for both the Robo taxi and vehicle applications, but also truck delivery vehicles, et cetera. So I think the big for us is that ⁓ this now, I guess, moving from a moonshot to commercial applications, especially within the Robo taxis, but even in trucks and delivery vehicles as well. And I'd say that the two, know, comparing commercial vehicles for robo taxi applications, they're not necessarily competing, if you like, in terms of one moving faster than another, but maybe just maturing across different paths. And I think on the robo taxi front, you know, we've already seen the fully commercial operations in nine or 10 cities now. I think that's likely to significantly grow this year. The fully commercial include, you know, running in all weathers, all ⁓ no safety driver or driver out operations on and obviously operations that charging customers fares. whilst we've seen 150 or so pilots on public roads today, only a small proportion of those, as I say, a class is fully commercial, but that will change this year. And you see in the announcements being made almost by the day now from Waymo and Baidu and Pony and several others that the number of cities this year will be significantly higher than 10 or so. So the commercial inflection point is starting to happen on the robo taxi front. The trucks and delivery side is maybe taking a little bit longer, I think already we have a few driver out operations this year. Most of them are talking about 2027 for more commercial scale of scaled operations. And at that point, I think we'll start to have more driver out operations and well, much more trucking applications coming from 2027 onwards. I think The reason why maybe we've talked about the two comparatively is that on the trucking side, you already have some of the operations that are live now, some hybrid operating networks, right? So you've got hub and spoke networks that could be partially human in the loop before they go fully autonomous. Of course, that doesn't transfer as well to robo taxis, whereas kind of all or nothing in terms of autonomy to get the cost savings. So big developments in both commercial vehicles and robo taxi markets. This year we'll see far more scalability on the, or more cities coming to market on robo-taxis. And I think next year we'll see more of the trucking applications getting started. Daniel Abreu Marques: Yes, and this momentum is also of reflected in the valuations levels. ⁓ example, Waymo's latest fundraise values it 126 billion. ⁓ And an investor perspective, ⁓ how we think about valuing AV companies today? So what metrics or milestones should underlie those valuations? MB: Yeah, I think for valuing AV companies today, requires, thinking very differently from a traditional auto or even tech multiples. And I think there's a few things here. The companies are building infrastructure. We talked a bit about world models earlier on, simulation engines, data, and fully autonomous platforms. so on one hand, this can scale across form factors, as we've discussed, right? Cars, trucks, robotics, drones, et cetera. I think for ⁓ AV companies today, requires, well, very differently from a traditional auto or even tech multiples. So the valuations need to reflect the optionality of powering entire physical AI ecosystems across these form factors, not maybe just in one specific product. so investors are starting to evaluate this to track the maturity of the autonomy of the stack that these companies are building, some of the evidence of the operational scale, the unit economic improvement that this can bring. valuations need to reflect the optionality of powering entire physical AI ecosystems across these form factors, not maybe just in one specific product. finally, the integration of autonomy into OEM production platforms to be able to progress towards these driver out commercial operations. I think ultimately milestones like reducing the hardware BOM costs towards automotive thresholds, that's one thing that's becoming important, especially with the autonomous car ⁓ investments. Expanding approvals as well is still relevant here. That's something out of the control of some of the companies, but regulations still a key part of the as to who's got the rights to operate in some of these areas and starting to demonstrate positive fleet level economics rather than just the revenue that can be gained these platforms in the short term. I guess investors are starting to price both the platform optionality balanced with execution risk as to how and where this is going to work and what is likelihood that it won't. We think that autonomy is probably like a winner takes most technology. It's not going to be a monopolistic one, don't think you're going to, given the high barriers to entry that I've already seen some consolidation in the market. I don't think this is a situation where you'll see as many AV developers as there are OEMs in the world today, for example. And a combination of the commercial deployment milestones, the fleet scale and utilization are going to become more important. the cost curves and the maturity of going down those cost curves, which we've seen that coming out in the BOM cost of AVs and LIDAR and sensors in particular. finally, the business model clarity, like what are the companies trying to solve for? Is it a vertically integrated autonomy stack or something that it's going to be a licensing play that companies can hook into? And of course, that could be far more profitable. Daniel Abreu Marques: Yeah, you mentioned already the cost reductions of the sensors and so on in your report. I think especially LiDAR got like 90 % cost reduction or something like that. And ⁓ what also can see is that the Chinese platforms, for example, the RT6 from Apollo Go, ⁓ they significant ⁓ lower costs than probably the US competitors. ⁓ So are MB: Mmm. Daniel Abreu Marques: cost reductions primarily then also seen in different regions or is this more like a global thing? MB: think so far we've seen maybe more disclosures from some of the Chinese platforms in terms of those cost reductions. think overall it will be a global phenomenon, It's ⁓ ultimately a lot of the cost reductions are coming from scale and the Chinese partners that ⁓ guess the big shift and we wrote a bit about this in the report is the shift from EVs to AVs in terms of the strategic priority from a government level but also at an OEM level and the AV partners that you know, starting to overlap the EV and AV supply chains more closely, sensors being the obvious one. if you, you know, whilst none of this is all confirmed, it's all kind of from media reports and what some of the companies have said. I think Bloomberg had about an additional 30 to $40,000 sensor cost requirement for a level four robo taxi in a Waymo or a Western AV for about a year or so ago. Some of the Chinese companies, particularly Baidu and Pony have stated BOM cost of the entire vehicle of less than that, right? Of $28,000 $40,000 in total, not the incremental sensor cost, but the total vehicle platform cost. So that is obviously a, you know, could potentially be a game changer here. And, you know, I guess once you start to get the, well, firstly, the automotive grade equipment and therefore the scale and bringing that cost reduction, that can get over a lot of the incremental center costs. But secondly, moving from kind of retrofitting AV technologies into vehicle platforms and bringing that into the factory is going to be the key unlock. So I think for now, it looks like a lot of the developed BOM will be lower based on what's been initially disclosed. But equally, you know, certainly the Western and the US ⁓ AV ⁓ are more partnerships. You know, most recently Waymo with Hyundai and Ojai or the Zeekr platform. I think you'll start to see several more of those in the future. And even if the vehicle cost is slightly more expensive for a European or a Western, a US developed vehicle, given the cost reductions you're going to see from autonomy and taking the driver out and the various operational savings, it's not to say that it's going to be just the lowest cost wins necessarily in the short term. Daniel Abreu Marques: Yeah. And maybe staying in China, you also ⁓ reported that penetration rate of the automated driving is to reach 93 % by 2030 in China. So do you this is also then partly due to the low cost that all these vehicles have? MB: I think there's a couple of things in play here, right? So China and the US are scaling autonomy in very different ways because of partially because of that cost, but also the operating models, the regulatory structures and the market dynamics being very different, right? So China's scaling AV is like almost like a national industrial strategy, fast, vertically integrated, hardware efficient, low BOM costs as we just discussed. But the US and some of the European developers are scaling more like safety critical technology sectors, right? cautious, but also regulated city by city, which is one of the key differences. So that sense, ⁓ the model might sound like it's slower, safety case driven and software centric, China's model being a bit more infrastructure back, hardware efficient and deployment first. ⁓ Both pros and cons and it looks like China's might be quicker, given the of development that we've seen so far. ⁓ But whether scales internationally is maybe the key question here, right? ⁓ I guess the interesting part is going to be happens in these international expansion of AV companies that starting to play out because companies need to both prioritize the safety validation, the geographies that are going to be possible in the short term, as well as having to have high fidelity operations and uptime before scaling. ⁓ But final point on this, looking at the numbers that you referenced, ⁓ our industrial's team believe that 85 % of passenger cars are going to have a level two plus or level three capabilities by 2030 of trying the passenger car market. But then a further 8 % of level four capabilities, right? Whether that's in robo taxis that are fleet owned or privately owned. So 93%, as you say, autonomy in just four years versus 30 % or so in 2025. So it's huge shift in the autonomous capabilities of the the passenger car market in China. Obviously, the rest of the world probably be moving slower than that at an overall level. But I the key thing to watch will be that level four robotaxi number and how quickly that can move, which will be dependent bit on both the cost and regulatory appetite. Daniel Abreu Marques: And we also see a lot of these Chinese players entering Europe more and more. How ⁓ do view the geopolitical dimension of this? Could AV Tech also become a trade or national security issue? MB: I think it already has been, right? We wrote a bit about this in the report that, you you've got, rising trade and tech tensions between the US and China might preclude China's AV companies operating in the US and vice versa, or at least in the near term, you've got this race underway to build the required partnerships, tech and approvals to launch the AV services internationally. And you said, right, starting here in Europe, but also the Middle East as well in particular, you've got I guess, both these areas that... On the one hand, you need to be sensitive around what partners they choose, whether there's national security concerns, and if you're choosing one partner, does that upset another? But equally, going to depend a little bit on markets that are going to value the advanced AI capabilities ⁓ CAPEX, right? If it does move to a lowest cost opportunity in technology, obviously, we're still some way from having all these things validated, but that's a different equation. to what we're seeing in the short term, where you could argue that the US AI capabilities are still a few months ahead, and certainly of the Chinese for the Gen AI capabilities. And I think that probably extends into some of the AV market as well, given the levels of investment and the chip requirements and all the partnerships on the chip side that are underpinning this. But in the future where that becomes less of gap potentially, then you'll start to see maybe looking at the scalability of deployment based on cost. And I think the Middle East is going to be a pivotal neutral test bed for this given they're already starting to have fast approvals in the UAE and Saudi in particular, more supportive regulation for autonomy and maybe less of the US China tech barriers. So it's one of the few regions where you'll see that both ecosystems can potentially prove how large their scalability or their operations will be. Here in Europe, think already seen maybe more of the Well, emergence of both the Chinese players in the short term, but now knocking on the door of several of the US partners as well, not just Waymo, know, Wayve here in the UK, you've got May Mobility I think, that have announced a few partnerships, Lyft with FreeNow and some of the partnerships they're developing as well. So becoming a bit more of a, you know, much more diversity of the AV players that are intending to launch across Europe. ⁓ long story short. I think you will have some national security concerns and question marks, but ultimately we're still at that pilot stage of where and how autonomy should be fitting into the landscape. And it's likely not just in passenger cars, of course, but in more shuttles, public transit and more of the shared mobility in Europe than you'll see maybe in the US. Daniel Abreu Marques: maybe let's also talk about another player in the AV industry, ⁓ Uber. And Uber's new CFO outlined that ⁓ Uber will pursue off-take agreements and also invest in AV infrastructure and take equity stakes in AV makers. ⁓ So basically its free cash flow to secure future vehicle supply. How important are off-take deals and long-term purchase commitments in this industry in your opinion? MB: Yeah, I think offtake agreements are going to become more important in this industry, right? Because they risk or de-risk some of the supply chain or the vehicle supply in particular. You've got lower cost potentially, and it gives platforms priority access to capacity and what is currently a capacity constrained industry. So, you know, on the one hand, we've got AV fleets are still supply constrained and building a level four vehicle, despite all the cost reductions we just discussed, building a level four vehicle is still expensive, capacity is limited. hardware costs are only going to drop when you can commit to some of this manufacturing at scale. But at the same time, demand is clearly going to rise. And you've seen that evident from the AV commercial operations that are there now. And if we're going to plan to grow from a few thousand vehicles today towards hundreds of thousands of level four capable vehicles in the next five to 10 years or so, platforms like Uber and others are going to want to secure their spot in the queue in the supply chain, right? And the supplier vehicles in particular. So off take deals could give AV makers the volume they need to reduce costs, but they also give ride-hours platforms guaranteed access to vehicles in a market where supply is likely to be tight for a while. So it's the same, you could argue a similar playbook to what we saw in EV batteries, right, a few years ago, locking in some of that future supply has been drastically, you know, really strategically important and necessary. It's not just a luxury item. And mentioned infrastructure as well, which I think is going to be important here. We've seen a few deals in the last few weeks pointing towards that around charging, parking, maintenance, et cetera. ⁓ to answer your question, think access to AVs and the supply of vehicles has been one of the bottlenecks in the short term. So you will see, I think, more of these announcements to secure some of the vehicle supply that these platforms are going to need in the short term, at least. Daniel Abreu Marques: Uber also said in an earnings call that ⁓ AV fleets might eventually be then financed like REITs, so fleets of vehicles owned by financial entities. then the executive basically said outside capital will come in once revenue are proven. Do you think we will see such autonomous vehicle REITs MB: Mm. I think we will. It's harder to predict who's going to be or who's going to play that position at the moment. But if you think of the financing structures for EV fleets, think we are likely to see some of that, whether it's REIT style or infrastructure style financing structures, because ⁓ economics, the capital intensity and the operating model all kind of push the industry towards third party ownership in the long term, I think. as we just discussed, the vehicles are still expensive to build. The hardware alone can can run into tens of thousands of dollars per car or whatever. So scaling robo-taxis to hundreds of thousands of units is going to require a lot of capital. I mean, even if it was just a normal car, irrespective of having AV capabilities, that's going to be expensive, right? But at the same time, the revenue model is likely to be fairly predictable you earn a per, especially those that have like a per mile service fee attached. So becomes a bit more just like other infrastructure assets, but it only works once AV show stable per mile economics. high levels of utilization and uptime and that regulatory certainty. So clearly we don't have all of those things today, but let's assume that those things start to come once we've got the utilization and safety performance proven in more cities, outside capital is going to want to come in. So I'd expect that, you know, maybe some of that, whether it's AV leasing funds or autonomy REITs, whatever it's going to be called, maybe towards the back end of this decade, as you see larger fleets want to deploy across multiple cities. That's going to need some of the offtake agreements that you mentioned. yeah, they might look more like infrastructure funds or aircraft and rail vehicle leasing platforms than traditional REITs, ⁓ assume. But underlying principle is the same, right? Third party capital that's going to ⁓ own generating AV fleets. Just as I said, it's hard to predict exactly who that will be that steps into that yet. We can only speculate. Daniel Abreu Marques: And now Uber is normally an asset light company and also ⁓ wanted stay asset light long term and short term is willing to buy assets. And Lyft ⁓ has said it buy thousands of AVs ⁓ to bootstrap the more or less. So fundamentally different strategies. ⁓ How do you evaluate asset light versus asset heavy approaches for these AV platforms? Maybe what are the pros and cons? MB: Yeah, I mean, the asset light, as we've seen with ride hail and asset light would always scale faster and cheaper. But equally, asset heavy gives you a control and utilization, right? So ⁓ right model and the pros and cons, I guess it depends how quickly the AV supply becomes abundant. And, you know, these are ⁓ two different philosophies and asset light models, you know, it's almost to stay flexible and avoid that huge capex that we just discussed and ⁓ buying thousands AVs. So therefore potentially lower risk and works well in markets where the AV supply is limited and regulations fragmented. an asset heavy approach is going to give you advantages ⁓ controlling the fleet and giving you maybe a bigger share of the long term margin opportunity here, right? And owning the vehicles ensures that you're not going to maybe be supply constrained when autonomy ramps up. in reality, for now, I think we're still in something of a ⁓ will remain bit of a hybrid model. platforms start asset light, but when cities are more mature and predictable, we could eventually use more structured financing or external fleet operators, owners to control some of those, more of the supply stack. But I think for now, you know, you're seeing almost parties starting to decide which will work. And even those that claim to be asset light, as you say with Uber and a few others, having to invest in parts of the supply where they see a gap. charging and real estate in a way you park and maintain the vehicles is an obvious one, at least in terms of the announcements we've seen in the last few weeks. I think it's maybe not as easy to see is are these models going to be entirely asset light versus asset heavy anymore, because that would have just meant before, do you own the vehicles now because of all these different fleet management investments that are needed to be made, you might have something where somebody else owns and finances the car. or the vehicle, whatever the form factor is going to be, but you need to invest in a depot to charge, run, clean and maintain them. So there is a bit of a contradiction that's starting to play out. I think the notion of just a clear separation between asset light and asset heavy business models is going to be a bit trickier to understand now, given all of the different dynamics that are starting to play out here. Daniel Abreu Marques: And there's also another trend or another that ⁓ lot of people talk about and this is ⁓ privately owned AVs. ⁓ And do you see privately owned AVs and share robotaxis converging or do they remain separate markets? MB: Mm. I think they could become, the lines here could get a bit more blurred. And what we mean by that, we wrote in the report is that, you we've got a lot of the AVs now could increasingly move between personal use and commercial service, depending on the demand time of day in the economics. Clearly we're not going to be there for a while. And this is going to be dependent on regulation and the viability of the technology and the ODDs in which they can operate. But as autonomy continues to mature, a privately owned AV is unlikely to get either because of the cost involved or the tech capability is unlikely to sit there idle for the famous 95 % of the time that today's cars do. Right. So instead it could either automatically join a shared robot taxi network when you're not using it as the owner generate income, but then return to you as the private when you need it. And at the same time, shared fleets could look more like personal vehicles in terms of the personalization. in-cabin AI agents, all the features, user profiles that the AV companies are not only talking about, but starting to integrate now, certainly in the likes of Waymo that you've seen. So, you we end up in a world where cars are not strictly private or shared anymore, as you can clearly define them today. It's a bit more flexible and dynamically switch roles between whether it's your control and your vehicle or into one of these shared mobility models. And that's what blurs the line. So ownership becomes I'd argue less important than access and your car becomes more of a digital identity that moves with you rather than a physical object that sits on your driveway. Now that's more of the, I guess, the utopia position, the what if in the mid to long term. In the short term, you've got a few different business models of is this autonomous vehicle that is fleet operated and fully autonomous the whole time or something that is a bit more. gradually improving ADAS like a Tesla FSD that could maybe one day be in a shared platform. And of course, companies like Tensor that launched the vehicle at CES a few weeks ago, they obviously want to be the world's first privately owned robo car. So the lines are starting to blur between all of these models. But ultimately in a world where autonomy works, at least in urban and even suburban and rural scale, you've got many different business models that could emerge here. And I think ultimately that reduces the total number of private cars on the road. It could reduce the private car ownership, but maybe increase the kind of total miles driven, even in a world where you reduce the total number of cars on the road, because the utilization could improve somewhat. Daniel Abreu Marques: And now there's also another dimension or several other dimension to physical AI that you ⁓ briefly on before. But one is directly connects also to the automotive industry, ⁓ the robots. And they're entering the factory floor. So your report shows that MB: Mm. Daniel Abreu Marques: Auto and also industrial companies are entering this humanoid race and Tesla is also reallocating car production space for Optimus. You have Hyundai, have BMW and others humanoids on assembly lines. What driving this conversion of automotive and robotics? MB: Yeah, it's a really interesting one. It goes back to a little bit what we discussing earlier on is that the worlds are starting to collide a bit between the technologies needed for deploying autonomous vehicles and humanoids. So humanoids could become more of an extension of the broader autonomy landscape than what we would have previously thought of a few years ago, mainly because it uses the same physical AI stack that you have in AVs, And higher degrees of freedom in a robot as we'll come to. But in terms of the world models, the perception, humanoids could become more of an extension of the broader autonomy landscape than what we would have previously thought of a few years ago, mainly because it uses the same physical AI stack that you have in AVs, simulation, edge compute, some of the underlying hardware here is being applied to manipulation and obviously humanoid robots in human designed environments. So I guess the overlap is becoming more about the hardware and the supply chain between automotive and humanoids. You've seen several of the tier one suppliers that see that as a huge opportunity to retool. Tesla, as you just said, have announced almost retooling. car production to humanoids, but also you've got more vertical integration, the likes of Hyundai we saw at CES, obviously owned Boston Dynamics. I thought their announcement of having humanoids on production lines of cars was very interesting and several other OEMs have done that, even if they don't own the robot that's on the supply chain. think, or on the assembly line rather, Figure AI, that I think BMW claims, Figure robots have helped to produce 30,000 BMWs last year already. So you do see some of the overlap as to where humanoids could be involved to help on assembly tasks. But to answer your question at the supply chain level, know, the underlying, there's a lot more equipment there that would overlap than you'd expect, right? Things like the batteries and the energy and the thermal management systems, the actuators, there's, think Scheffler told us about 28 more times more content in terms of actuators you would have in a humanoid robot than you'd have in an electric vehicle. So there's a huge, ⁓ ⁓ content opportunity for these suppliers to address. And you've got more of the underlying AI and SOC trends that we discussed earlier on, right? Not just the world models and simulation, but more the actual hardware and AI chips that could be deployed across an autonomous vehicle or across a humanoid robot. So the hardware is becoming a bit more overlapping. The technologies to bring them to market in terms of world models and simulation and data are also quite similar. But actually the difficulty of commercializing a humanoid robot is probably or will be higher than an autonomous vehicle, given the kind of controlled nature of where you'll operate a car relative to what you'd need a humanoid robot for. So in the short term, you're seeing a bit more separation of where you deploy humanoids versus with humans, being able to cordon off certain areas for where they can be tested and piloted. But yeah, overlaps are becoming more significant. Daniel Abreu Marques: So basically, could see automotive OEMs becoming not buyers of humanoid robots, but also maybe becoming major producers of humanoid robots. So basically, could see automotive OEMs becoming not buyers of humanoid robots, but also maybe becoming major producers of humanoid robots. MB: Exactly. Exactly. Yeah, that's the where some of the market seems to be heading. think initially it's more testing as to useful the robots are going to be. Right. And you've seen a bit of a almost a two stage curve emerging here. Right. Humanoids can start with automotive companies as incrementally or incremental efficiency boosters on the production side. ⁓ over time, as you say, it could be more fundamental driver of retooling ⁓ manufacturing to improve the competitiveness at a much larger scale. So I think in the near term, you've seen a lot of the car companies are looking to manufacture them initially for their own benefit, but could move towards more of wanting to manufacture them to sell them to third parties. In terms of the reasons and the rationale behind that, think in the short term, it's being pointed a bit more towards that labor bottlenecks and also to cover some of the repetitive difficult tasks and using humanoids to just lift tote boxes and move them around an assembly line. But as the dexterity and the ⁓ and certainly the cost of these humanoids reduces, then it could ⁓ higher uptime, more flexible production lines and almost a shift towards near continuous operations, which is expensive for automotive OEMs when they move to three shift ⁓ for ⁓ cars, right? If you're deploying more humanoids, you could logically reduce ⁓ cost of doing that. So at that point, the change has almost reshaped some of the manufacturing capabilities here. course, countries that have got more deep robotic supply chains can see some scale advantages there. So incremental for now, but I think in the next few years, you'll see ⁓ evidence as to what kind of cost reductions and productivity breakthroughs humanoids can enable in automotive. Daniel Abreu Marques: And can you also give us a short rundown on the major players in the Humanity Robot game? MB: Yeah, so you've got a few things emerging, right? We've kind of separated it into body and soul of the humanoid as to what that means in the supply chain. ⁓ And in terms of the ⁓ side of it, that's more of the AI developers, the chips, the ⁓ the AI chain that's going to be as relevant in an autonomous car as it might be in a humanoid robot. But in terms of bringing together the integrators of humanoids, That's where you see a bit of a divergence between vertically integrated players that want to produce the whole humanoid themselves, know, the likes and on the OEM side, that's mainly Tesla and XPeng. You've got a few of the other Chinese companies that want to do something similar. I GAC and Cherry have got in-house humanoids as well as Hyundai, as we mentioned. So you've got a number of automotive players that are involved that want to produce their own humanoids for internal purposes or external. But you also have, guess, Pure Play humanoid developers, there's a couple of them that are listed equities, most of them are still private, likes of agility, robotics, figure AI, etc. There's about 50 humanoid developers now that's really ⁓ significantly in the last few years. But category of who's behind some of these initiatives you've got some Pure Plays that want to develop humanoids for a number of different use cases. You've also got some of the more automotive OEM focused investments as we've alluded to because either it improves their own or lowers their production cost ultimately they could scale externally. And also it's important to note, you know, going back to the geopolitics we discussed in AVs, you've seen about 20,000 humanoids that were produced last year, 80 % of which were manufactured in China. And there's a huge supply chain emerging around the Chinese humanoid industry. equally, you've got a load of suppliers that could scale in Europe the US as well around both automotive retooling for humanoids. But it seem like the next kind of technological race kind of linked autonomous vehicles, but also a separate product across US China tech competition, a range of not automotive, but also more industrial robotics, you know, manufacturers that are behind these companies now. Daniel Abreu Marques: Yeah, super interesting topic. MB: I think, sorry, finally, Daniel, the scale is one thing to watch here. It's not just whether this is going to be relevant in an automotive context or not. you think of some of the numbers that are being talked about now, there's still a lot of hurdles to overcome. But as I just said, there's 20,000 humanoid that were produced last year, which is peanuts in relative terms. But with our China industrial team think that by 2030, you could have just over a million humanoid that are manufactured and by 2035, it could jump to 10. And you could even be in a situation where 2050 or 60, kind of bull case scenarios that are being talked about are billions of robots that are being produced. you know, that number, you we can decide maybe in a few more years, but the short term that growth to one or 10 million human was being produced in the next, you know, within 10 years. I think that's going to be, you know, the important thing to watch here, because if the cost of a humanoid today, most claim is about 90 to $100,000 are made in China humanoid robot. It's already sub $40,000. Daniel Abreu Marques: Okay. you MB: I think Tesla are talking 50, 60,000 at the moment for Optimus that could reduce even further. So once you get to those levels of finished robot costs that are only going to reduce with scale, you could be talking sub $20,000 per robot. So I think the industrial applications are clearer to visualize as to where that ROI can be, whether that's on a $20,000 robot or $50,000. But equally, the cost is reducing. It's still going to take a lot longer to move or progress those capabilities into the commercial and domestic applications that some of these companies talk about. But even in the industrial applications, there's enough of a, you know, a TAM and an ROI to justify those type of in the next five to 10 years. Daniel Abreu Marques: If we would buy a humanoid robot for our factory line, then we have the upfront cost of, I don't know, 40, 50, 60,000 euros or dollars. But do they then also charge per operating hour, or is it a flat fee, or is it no additional fee? Do you have some insights into them? MB: A few, yes, so it depends. All of these options are still depending on what the customer wants. Some of the robot companies we met at CES and over the last couple of years said that whilst you would think most of the companies would move, would opt for as a service model, pay by the hour for the robot, for the labor, that is possible. But ⁓ least what we've been told, most of the in the short term were buying the robots outright to almost... test a number of different pilots as to how they can be used and then justify the ROI based on the either the relative labor cost savings if you're removing human ⁓ input from that specific task or if you're making it quicker and speeding up the or whatever it is you're pointing it towards. So it feels like ⁓ at least what we've been told a lot of the companies deploying them so far have just bought the robots outright to test and pilot where they can be operated and how. But equally this as a service robot as a service model, as it's called, is something that could become more relevant as the scale grows, where they're almost leased by over a number of years or by the hour and charged accordingly, and maybe even based on tasks in the future. So for now, it seems more like acquiring robots out, buying them outright and justifying the ROI depending on the productivity gains in the mid to long term. You could see more robots as a service operating depending on the paper. our paper task, et cetera, those type of models that could start to emerge. Daniel Abreu Marques: And scaling physical AI faces also real bottlenecks like energy, compute ceilings at the edge, chip politics. Which constraints worries you the most for autonomous driving specifically? MB: Yeah, I think there's a few, right? think in terms of the constraints for AVs, you've got a number of, I'd say the thing that we discussed earlier on was probably the shortage of vehicles in the short term. A few that started to emerge is either the pace of regulation, you know, and is that going to slow things down? Do we have the available vehicles and the cost of them? you've got a few others like the public acceptance that isn't talked about enough, you know, it still divides opinion as to will people be comfortable driving or riding in autonomous vehicles? I think that's maybe less of a worry in the mid to long term, but something that needs to be looked at in the short term. So I think for now it's probably more the bottlenecks around either is regulation going to move fast enough and will the cost of autonomy reduce fast enough relative to what we've seen today? Obviously we've seen the sensor costs reduce a lot and the BOM costs like we discussed earlier on that seems to be the biggest enabler, but equally Regulation is still fragmented, utilization could be quite tricky and access to vehicles is not necessarily a given. also the, I'd say maybe the final one, which is still to be confirmed is the willingness of OEMs to integrate that technology into factory production lines, right? A lot of it, as we said earlier, is retrofitted into companies like Waymo or whoever that would then take ⁓ vehicle ⁓ add autonomy stacks to them retrospectively, which is obviously adding a lot of cost. You do see some evidence of OEMs that are moving that into production lines. Most recently, the likes of the deals with Uber, Lucid and Nuro Hyundai deal with Waymo, a of the Chinese OEMs, cetera, working with Pony AI and Baidu, et cetera. But you don't maybe have as much of a proof point yet as to what that will lead to in terms of overall units in the next few years. We were kind of speculating and making assumptions. So think they're some of the things that could maybe be the remaining to scaling and adoption. yeah, it feels we are quite quickly towards the commercial trajectories here, but they're probably some of the hurdles that are likely to preclude some of that or be the biggest risks, I'd say. one other thing is that how quickly that could be adopted across or moved across geographies, right? know, initially you saw a of the robo taxis in particular that were being city by city deployed. that we've seen end to end models and more generalization of the technology, I think that will now move much faster into new cities and new geographies quicker. But that, you know, that's also still that needs to be proven out. How adaptable is autonomous technology from London to Paris or New York to Doha or whatever, right? So adaptability internationally is, I think will go much quicker now with the advent of new and more scalable, generalizable AI models. But yeah, that's still something that the difficulty of scaling across geographies needs to be looked at. Daniel Abreu Marques: Maybe let us close with ⁓ a looking ⁓ view. you had to name ⁓ the most important development in physical AI or autonomous driving to watch over the next 12 months, what would that be? MB: Yeah, it's a good question. think that now, the things that have excited us the most and made us write the primer now have been around the technology developments, right? So, know, specifically around the emergence of simulation and synthetic data. We spoke a bit about world models earlier on, but just on synthetic data and simulation, that is allowing companies to now test and validate these systems at a much faster pace, right? So Nvidia said that just one second of using simulation for humanoid robots or autonomous vehicles would have taken 27 minutes of testing in the real world without that, right, without simulations. That is what would have been much slower, much more expensive. Now you can accelerate the development of physical AI, whether it's in robots, AVs, drones, all these different applications. So I think the key unlock in the short term is going to be the of more high fidelity simulation and synthetic data, meaning that you can test and validate these products. before they come into the real world and cause a potential safety risks and of course the cost of bringing autonomy to market. I think that's going to be the big thing to watch in the next few years is just how much could ⁓ accelerate commercial development and who ⁓ bring these to market faster. Daniel Abreu Marques: Martin, this has been a super deep and insightful conversation with you. So thank you ⁓ much for walking through your team's work. And your team publishes the next update, whether on physical AI, autonomous vehicles, or the broader thematic investing landscape, you have a standing invitation to come back on ⁓ Autonomy So yeah, thanks again for joining us today. MB: Yeah, thank you, Daniel. And thank you to you and the team as well for all of the great newsletters and content that you guys put out. It certainly helped us not in just this report, but several others over the years. yeah, thanks for having us on and look forward to seeing you very soon. And yeah, we'll keep in touch, of course. Daniel Abreu Marques: Yeah, see you soon, bye. MB: Bye bye.