Latest / The AI Supercycle / Physical AI: The Race for Embodied Intelligence

Transcript

Physical AI: The Race for Embodied Intelligence · The AI Supercycle

• [0:10] Hello and welcome to the AI Supercycle Podcast, where we follow the capital flows, the physical constraints, and the commercial returns shaping the largest industrial build-out of our generation. From semiconductors, AI factories, and hyperscale data centers, to energy markets and power grids, critical materials, orbital compute, and embodied intelligence. We track where the capital is being deployed, where the binding constraints are emerging, and what it means for traders, investors, and the broader economy. [0:40] Today is Wednesday, the 12th of August, 2026. I'm Tim Hardwick, founder of Quantum Fields Market Intelligence, and your host for today's show. Before we get into today's episode, I just want to give you a brief update on the business developments and also the book. One of the main things I've been working on is Prism, the decision intelligence platform we're developing at Quantum Fields AI. Prism brings together macroeconomic conditions, company and industry fundamentals, capital formation and infrastructure development, as well as market positioning, sentiment and price behaviour. And what this does is to produce a more disciplined and probabilistic assessment of what is most likely to happen, what alternative scenarios could emerge, how confident we should be and what evidence would either strengthen or invalidate the underlying thesis. Now it does not replace human judgment as it is not designed for that purpose. It is designed to make that judgment more systematic, more transparent and responsive as the evidence changes. [1:40] Now, I've been testing this minimum viable product under live market conditions for about three months now, and the early results are very encouraging. So the next phase is to move from the analytical prototype towards a scalable, cloud-based platform. We are also exploring how the same approach can be applied to the space sector and orbital infrastructure. Now, this is a natural extension of the wider work we are doing across the AI industrial economy. And of course, the space sector is a key part of the AI continuum, continuum as data centers are set to extend into orbit in the future. And we've talked about this on a previous episode on this podcast. So let's talk a little bit about the book as well, because there's been good progress with this, Trading the AI Supercycle, which is now in the editing phase, and this should be completed during August. So the publication schedule has slipped a little bit, as these projects often do, and I'm now expecting the book to be released in September. So here is what the book's about. Trading the AI supercycle gives you a cross-asset framework for capital intelligence across the entire AI industrial economy. [2:46] The book follows the money through the whole chain, hyperscaler capex, semiconductor supply chains, data center build-out, power generation and transmission, critical minerals, logistics, physical AI, and the emerging orbital layer. But it also examines the demand side. There's a dedicated chapter on enterprise ROI and token budgeting, the first demand-side constraint in the AI supercycle, and why enterprise monetization may matter just as much as physical infrastructure in determining where the cycle goes next. It draws on historical precedents, like the railroads and electrification, to show where the AI supercycle is likely to rhyme with the past and where it won't. [3:29] So that is the central thesis behind this podcast, right? AI does not begin and end with models and software. It connects digital intelligence to semiconductors, power, materials, machines, and increasingly, the physical world. And that brings us directly to today's subject, physical AI, robotics, and the race for embodied intelligence. In the previous episode, I examined the enterprise side of the AI equation, whether companies can convert rising expenditure on models, compute, and tokens into measurable business value. This episode moves that discussion from the digital world into the physical one. [4:10] What happens when an AI system no longer simply generates an answer, creates an image, or analyzes a document? What happens when it can see the world around it, reason about what it sees, and take physical action? And what happens when that intelligence is placed inside a machine that can walk, lift, and manipulate objects. A machine that can navigate a warehouse, work on a production line, or eventually help build infrastructure in orbit or on the moon. This is the emerging world of physical AI. It includes the glamorous, attention-grabbing humanoid robots that increasingly resemble something from the film iRobot. It also includes the more unsettling associations created by the Terminator and Skynet, but it encompasses is something much larger. It includes industrial robots and autonomous mobile systems, as well as drones, self-driving vehicles, and warehouse automation. The central question is therefore not whether robots are going to arrive sometime in the future. Millions of robots are already with us. The more important question is whether AI is about to transform what those machines can do, how widely can they be deployed, and where the economic value will ultimately land. [5:25] And hopefully you'll stay with me to the end because this story doesn't stay on Earth. Physical AI has a role to play in orbit and eventually on the moon, so we'll get to this a little bit later on. So let's start with what actually makes this different from the AI we've been talking about in the series so far. The current AI cycle has largely been defined by generative AI. A user provides a prompt, a model interprets it and generates an output. That output might be text, software code, an image, music, or a recommendation. Physical AI introduces another dimension. The system must perceive a real environment and understand what's happening, as well as decide what to do, and then convert that decision into physical movement. [6:12] That creates a continuous loop. Perception, reasoning, action, feedback, learning. A physical AI system might use cameras, LIDAR, radar and touch sensors to perceive its surroundings. It then has to build some picture of that environment. It has to identify objects, people, obstacles, distances and movement. It needs to decide what action to take, plan that action and execute it through motors, joints and grippers. Then it has to interpret the consequences. Did the robot pick up the correct object? Did it apply too much force? Did the object move unexpectedly? Has a person entered its path? Does this task need to be replanned? [6:58] That continuous loop between intelligence, the body, and the environment is the essence of embodied intelligence. The body is not merely packaging for the AI model. Its physical characteristics shape what the system can perceive, learn, and do. A robot with wheels experiences the world differently from one with legs. A robot with a simple two-finger gripper has different capabilities from one with a dexterous, five-fingered hand. Where its cameras sit, how sensitive its touch sensors are, how far its joints can move, all of that shapes the intelligence the machine can actually express. This is where the difference between traditional automation and physical AI becomes important. Traditional industrial automation is highly deterministic. A robot performs a predefined movement in a structured environment. The production environment is designed around the robot, and the robot is programmed around one particular task. That model works extremely well when the task is repetitive and the environment stays stable. It doesn't work as well when objects vary, when people move through the workspace or the robot meets something it hasn't encountered before. Physical AI is meant to make machines more adaptive. The aspiration isn't to program every individual movement, is to give the machine enough understanding of the world to complete a task even as conditions change. [8:23] Now that's a profound shift. It moves robotics from programmed motion towards machine intelligence operating in the physical world. [8:32] That raises an obvious question. Why humanoid, specifically? So let's take a closer look at that. Humanoid robotics is the most ambitious expression of embodied intelligence. There's a practical argument for the humanoid form that goes beyond appearance. Most of the human world has been designed around the human body. Factories contain workstations built for human arms and hands. Warehouses contain shelves built for human reach. Buildings contain stairs, corridors and doors. Tools, control panels and vehicles have all been designed for human use. A sufficiently capable humanoid robot could, in theory, enter these environments without any of them needing to be redesigned. It could use existing tools, navigate existing buildings and move between tasks that currently require several different kinds of specialised automation. That's the general purpose proposition. But there's also a very compelling counter-argument. If a fixed robotic arm can do a particular task faster, more safely, and more cheaply, there may be no reason to give that machine legs at all. If an autonomous mobile robot can move goods through a warehouse on wheels, forcing it to walk may just add cost and reduce reliability. Humanoids maximise potential versatility. Specialised robots frequently maximise near-term productivity. Where the commercial development of humanoid robots ends up will depend on where the balance falls. [10:02] For this episode, let's concentrate on five benchmark platforms. Tesla Optimus, Boston Dynamics Electric Athlas, Figure 03, Agility Robotics Digit, and Agibot A2. This isn't the only credible selection. Companies including Unitree, Aptronic, and OneX are also developing important platforms. But these five give us a useful benchmark across the different business models, the different technologies, and different stages of commercial maturity. Let's start with a platform most people already know, Tesla Optimus. Tesla describes Optimus as a general-purpose, autonomous, bipedal humanoid intended to perform work that's unsafe, repetitive, or boring. Its strategic importance comes from the assets Tesla can potentially bring together. Tesla brings deep experience in computer vision and autonomous navigation, – in batteries and power electronics, and in high-volume manufacturing. It also owns the factories in which robots can be trained, tested, and eventually deployed. That combination creates the possibility of a vertically integrated development loop. Tesla can build the robot and develop its AI in-house, it can test that robot inside its own production system, and if the technology becomes reliable enough, it can manufacture it at scale most robotic startups would struggle to match. [11:26] Optimus also sits within Tesla's broader autonomy thesis. The company is trying to build machines that perceive and navigate the physical world, whether that machine is a car or a humanoid robot. [11:37] The investment case is significant then. If Tesla can transfer parts of its autonomy, compute and manufacturing capability into a general-purpose robot, Optimus could become far more than a factory automation product. But ambition and deployment evidence need to be kept separate. Tesla has talked about very large production ambitions. What investors actually need to examine is different. What the robots can do autonomously, how long they can operate, how often people have to step in, and whether the whole system delivers productive work at an acceptable cost. Now here's a number to keep in your mind until later. Each Optimus unit needs something like 3.5kg of high-performance magnets spread across more than 40 servo motors in its arms, hands and legs. Now that sounds a bit like a footnote about materials science, but it isn't, and we'll come back to exactly why this is the case later in the episode, because it's one of the more overlooked constraints on the entire sector. The indicators that matter aren't the fluidity of a demo or the size of a production target, their reliability, utilization, and cost per productive hour, and that standard needs to apply to every humanoid company here, not just Tesla. Next up is a very different starting point, Boston Dynamics. Boston Dynamics approaches the problem from a very different starting point. It has spent decades building some of the most mechanically capable mobile robots in the world. [13:06] Atlas became famous for running, jumping and performing complex dynamic movements. Those demonstrations established Boston Dynamics as a clear leader in mobility and whole body control. There's a really good story in how this new version arrived. Back in April 2024, Boston Dynamics published a farewell video for the old hydraulic Atlas, a highlight reel of a decade of running, falling over and getting back up again. It looked a bit like a retirement notice. It was actually a magic trick. That very next day, the company unveiled the new electric atlas, standing up off the floor in a way no human spine would ever allow. The hydraulic robot wasn't being shelved, it was being replaced, on a one-day turnaround designed purely for maximum theatre. The new electric atlas is meant to turn that research heritage into an industrial product. [14:00] Boston Dynamics unveiled this production version in January 2026 and began manufacturing immediately. Initial deployments were scheduled with Hyundai and Google DeepMind with automotive material handling as the first industrial focus. Atlas is designed to operate in environments built for people, but Boston Dynamics hasn't limited the robot to the human range of motion. Its joints can move in ways that would be impossible or just uncomfortable for a person And that would matter commercially There's no economic reason for a humanoid robot to inherit every limitation of human anatomy, The platform also builds in features needed for sustained enterprise use Boston Dynamics says Atlas can find its own way to a charging station and swap its own battery That means it can get back to work without waiting for a person to recharge it manually you. [14:51] The company is working with Google DeepMind on an AI foundations model. Hyundai, meanwhile, provides manufacturing capacity, component expertise, and industrial deployment environments. That's a powerful combination. Boston Dynamics' mechanical capability with DeepMind's AI research and Hyundai's manufacturing ecosystem. The real question is whether Atlas can turn extraordinary mobility into repeatable, economically useful work. Enterprise customers don't need a robot that produces the most impressive video. They need one that can complete the same task thousands of times, safely and predictably. [15:30] Then there's Figure, which takes yet another approach again. Figure represents the AI-first startup model. Its objective is to build a general-purpose humanoid where the intelligence and the body are developed in a single system. Figure 03 was introduced in October 2025. The company redesigned the robot around its Helix Vision Language Action architecture, with the longer-term goal of operating in both industrial and domestic environments. The Vision Language Action model tries to connect what the robot sees and what it's told in plain language directly to physical control. Let's take a simple instruction, like place an object in a drawer. The robot has to interpret that instruction and identify the right object. It has to work out where the drawer is, and it has to plan and control the movement itself, adjusting as it goes based on physical feedback. Now that's considerably more complex than generating a written answer. [16:30] One demonstration is worth mentioning here, because it points at where this is actually heading. Figure showed two of its robots running the same shared vision language action model, working together to tidy a bedroom, making the bed, hanging clothes, moving around each other with no central controller telling either one what to do next. Neither robot was told to avoid the other. They inferred it, the same way two people quietly negotiate space in a small kitchen. FIGURE has also moved well beyond the lab. Its FIGURE-02 predecessor operated at BMW's Spartanburg plant for more than 1,250 hours, contributing to the production of more than 30,000 vehicles. FIGURE reported in late 2026 that it had delivered more than 350 FIGURE-03 robots, and that it had increased output at its BotQ facility for one robot a day to one robot an hour, a 24-fold jump in under four months. By early May 2026, that momentum had turned into something more concrete. BMW signed on for an initial fleet of 40-figure 03 robots at Spartanburg, which figure describes as the first paid, commercial-scale humanoid contract anywhere in the industry. Not a trial, not a pilot, a purchase order. [17:50] By the summer, Figur had crossed 1,000 robots built at Bot Q, marking the occasion with a special gold-finished unit that Brett Adcock only half-jokingly compared to C-3PO. And for those that don't know who or what C-3PO is, it is a robot that is in the Star Wars films. Banaty aside, the numbers matter because it is the clearest evidence yet that Figur has turned manufacturing itself, not just the AI, into a repeatable process. These are company-reported figures, but they matter because manufacturing a robot is in itself one of the sector's main challenges. A prototype can be carefully hand-assembled by a team of engineers. A commercial product has to be manufactured repeatedly, serviced, updated, and supported, at a cost that actually makes sense. Figures' challenge is therefore twofold. It has to prove general embodied intelligence, and it has to build an industrial company capable of producing and supporting a large robotic fleet. Now let's take a look at a platform that's taken a much more functional approach. Agility Robotics Digit. Digit is a humanoid in that it has legs, arms, and a body built to work in human environments, but agility hasn't tried to make it look like a person. Its legs have a distinctive backward-jointed look, which is more ostrich than human, and its hands are designed around industrial handling rather than human imitation. [19:19] Now that's a useful reminder. Humanoid robotics isn't necessarily about recreating the human body exactly. It's about creating a form that can work within human infrastructure. [19:31] Digit also gives us some of the strongest evidence of sustained humanoid deployment anywhere in the sector. Agility reported that Digit moved more than 100,000 totes during a commercial deployment at GXO's logistics facility in Georgia. So you might be asking, what's a tote? And they are basically standardized plastic bins or crates used to store or organize things, or maybe transport things around. The robot has also been part of programs involving Toyota, Schaeffler, and Amazon. Moving containers from one place to another might seem unexciting next to a robot dancing or doing household chores. Economically, though, that's exactly the point. The early humanoid market is likely to be built through workflows that are bounded, repetitive, and easy to measure. The task needs to be worth automating. It needs to be predictable enough to control, and it needs to be unpleasant, unsafe, or difficult enough that the organisation has a real reason to invest. Digit's important isn't that it appears to have general intelligence, it's that it's starting to build real evidence for the commercial use of a humanoid form in live operating environments. [20:43] And the fifth platform takes us somewhere completely different geographically. China's AjiBot A2. AjiBot, also known as Juen Robotics, and I hope I pronounced that right, gives us an important window into China's fast-moving embodied intelligence ecosystem. The company has built a broad portfolio, full-size humanoids, smaller interactive robots, wheeled industrial systems, and quadrupeds. [21:09] The A2 series has focused on multimodal interaction and autonomous navigation, mostly for reception, presentation and service work AGBOT's separate G-series, by contrast, focuses on manipulation and factory deployment, In December 2025, AGBOT announced that its 5,000th mass-produced humanoid across its whole portfolio had left the production facility, And they broke that number down for us 1,742 of the full-sized A-series 1,846 of the smaller, more agile X-series and 1,412 of the industrial G-series built for manipulation work, That's not one company betting on one robot It's three separate product lines running in production simultaneously, Again, company-reported production shouldn't be equated with productive commercial deployment A robot built for research or demonstration isn't the same as one completing thousands of paid working hours in a factory. [22:13] But the production figure still shows a strategic difference. China isn't approaching humanoid robotics as a collection of individual startups. It's building an entire ecosystem. Robot companies, component manufacturers, research bodies, and government support. All tied together through highly integrated industrial supply chains. China accounted for 54% of the global industrial robot installations in 2024. And Chinese manufacturers took a majority share of their own domestic industrial robot market for the first time. That gives the sector a huge base of manufacturing knowledge, components, customers and operating data to build from. Agibot has since introduced newer A3 variants, but the A2 remains a useful benchmark for its shift from interactive humanoids toward a wider embodied AI platform. The broader investment implication is this. The humanoid race isn't simply Tesla against a handful of US startups, it's becoming a competition between entire industrial ecosystems. Now let's step back from those five platforms for a moment, because humanoids are not the whole story here. Although humanoids dominate the current discussion, the physical AI economy is already much bigger than the humanoid market. [23:32] The International Federation of Robotics reported that 542,000 industrial robots were installed during 2024. Now that took the global operational stock to roughly 4.6 million robots. Annual installations have stayed above half a million for four years running. The professional service robot market added almost 200,000 units in 2024, more than half of them in transportation and logistics. The existing installed base really does matter because physical AI doesn't have to replace conventional robotics to create economic value. It can simply make existing robot categories more flexible and easier to deploy. [24:13] A machine vision system might recognize a wider variety of products. A collaborative robot might learn a new handling task by demonstration rather than lengthy programming. An autonomous mobile robot might respond more intelligently to a changing warehouse floor. This may be where some of the earliest returns from physical AI actually show up. The industry doesn't need to wait for a fully general-purpose humanoid before AI starts improving the productivity of physical machines already at work. The market is likely to develop along both paths at once. Specialized systems will get smarter, while humanoid systems gradually try to take on a wider range of tasks. So let's follow the money here. One figure sometimes attached to robotics is more than $300 billion of venture funding. That doesn't look like a credible estimate for robotics venture capital alone. It probably comes from a much wider measure of AI investment, or overall technology funding. The real robotics estimates are much smaller. Crunchbase puts global robotics fundraising at approximately $15 billion in 2025. F-prime, using a slightly different definition covering the US, Europe and Israel, calculated around $16.3 billion. [25:30] By June 2026, Crunchbase estimated that global robotics funding had already reached $18.8 billion for the year. Now the totals vary because the category itself is hard to define. Should autonomous vehicles count? What about drones, defence systems or foundation models built to control many kinds of machine? The exact number isn't really the point. What matters is that robotics investment is measured in the tens of billions, not the hundreds of billions. And capital is increasingly concentrating around companies trying to build general-purpose robot intelligence, humanoid systems, and the infrastructure needed to train them. The other figure that gets repeated a lot comes from NVIDIA. Jensen Huang has described physical AI as an opportunity tied to the $50 trillion manufacturing and logistics economy. Now that should not be read as $50 trillion of robotics revenue. It describes the scale of economic activity physical AI could potentially touch. The actual revenue available to robot makers, component suppliers, and software platforms is a fraction of that. Now this distinction matters because the investment narratives often confuse the size of the industry being transformed with the revenue actually available to the companies selling that technology. The opportunity is still potentially enormous, but it has to be measured through deployment, productivity, and value capture, not just by attaching a technology to a very large GDP number. [27:00] Okay, so let's move on to talk about the Physical AI stack, because the capital isn't flowing only to robot manufacturers. Physical AI needs an entire technology and industrial stack behind it. At the bottom sits the materials and components, such as motors, magnets and actuators, batteries and power electronics, and increasingly sophisticated hands that can actually grip and manipulate things. Above that sit perception systems, like cameras, LiDAR, radar and touch sensors. The robot then needs onboard compute fast enough to control movement in real-time, along with connectivity, fleet management and cyber security. Above the machine itself sits the intelligence layer, the computer vision, world models, motion planning and orchestration software. Much of the training happens in simulation. It's slow and expensive to collect every failure and edge case using physical robots. Simulation lets companies build synthetic environments and test robot behavior without damaging equipment or putting anyone at risk. NVIDIA is positioning itself across much of this stack. Its role runs from the compute used to train models, through to its Omniverse simulation platform and ISAC robotics software, all of the way to its Jetson edge computing hardware. [28:16] The company is working with the industrial robot manufacturers, automation companies, and embodied AI developers. In 2026, NVIDIA highlighted a partnership in which Skilled AI worked with ABB Robotics and Universal Robots. It also pointed to work with Foxconn on high-precision assembly for NVIDIA's own production lines. This illustrates the ecosystem model at work. Few companies control everything needed to build, train, manufacture, deploy and maintain physical AI. Partnerships connect model developers, robot manufacturers, component suppliers and industrial customers. The value chain may eventually prove more investable than the headline robot brands themselves. All of this technology has to answer to a much less glamorous set of questions, so let's talk about the CFO and the COO test. [29:10] Enterprise adoption will ultimately come down to that. The chief financial officer wants to know the total cost, the payback period, and the financial risk. The chief operating officer wants to know whether the robot actually works, reliably, inside the real operating process. How many hours can it work? How often does it fail? How many tasks does it complete successfully? How often must a person step in? Can it hold the required production rate? How long does deployment take? What has to change in the factory or warehouse to accommodate it? And who fixes it when it breaks? Here's why that intervention question matters more than it sounds like it should. Cisco has done the maths on this. A 95% success rate sounds excellent, right up until you're running a 100,000 cycles a day, at which point it quietly leaves you with 5,000 failures that still need a human. At industrial scale, the last few percentage points of reliability are worth more than almost anything else in the business case. The basic return equation includes more than just labour substitution. The potential benefits are real. Labour savings, extra throughput, better quality, fewer injuries and the ability to work in places or at times that people can't. Against that sits the costs, the robot itself, integration, training, energy, maintenance and downtime. [30:40] The most revealing metric may be cost per productive hour. A robot can look relatively cheap once its purchase price is spread across several years, but if it only operates for a limited number of hours or needs frequent intervention, its true cost per productive hour can end up far higher than it first appears. This is why robotics as a service could become important. The International Federation of Robotics reported strong growth in subscription and rental models during 2024. A service model lowers the upfront capital requirement and shifts some of the technology and maintenance risk onto the provider. It turns a large capital decision into an operating expense tied more closely to actual use. But the provider then carries more of the financing risk. A robotics company may need to fund the manufacture of a large fleet well before the matching subscription revenue arrives. That means the eventual capital stack could include venture equity, equipment leasing, and recurring software revenue, all layered on top of each other. The development of physical AI is therefore not simply a technology story, it's also an industrial finance story. [31:52] So what could actually slow all of this down? There's a real argument happening in the industry right now about exactly this question. Brian Moore, CEO of the physical AI data company Voxel51, said this July that the sector is closing in on its own chat GPT moment. [32:11] Kumar Sokka, who previously led the AI and autonomy strategy at Rockwell Automation, pushed back on that framing directly around the same time. His point was blunt. A model that hits 95% success in the lab often drops to something like 60% once it meets the real world. And the test that actually matters isn't a polished demo. It's whether the system can run 10,000 cycles in a row without somebody stepping in. Now that gap between what works in a demo and what works 10,000 times in a row is basically what this whole section is about. Let's keep one simple example running through it. A warehouse robot that has to pick a single egg out of a carton without cracking it, then a moment later heave a loaded tote onto a shelf above its head. Start with the intelligence itself. Robots need data about physical interactions, not just text and images pulled from the internet. And real-world robotic data is really expensive to collect. And interestingly, nobody has a data set of a million ways an egg can crack. And even where there is data that exists, it's often locked away rather than missing. Cisco has pointed to one aerospace supplier running 20,000 CNC machines, or computer numerical control, of which only around 5% are actually connected and generating usable data. [33:36] Increasingly, the problem isn't that the data doesn't exist, it's that none of it is accessible at scale. Actions unfold over time. Failures can damage equipment. Demonstrations often need a human operator standing right there. Simulation and synthetic data can fill that gap, but that creates its own problem. The simulation to reality gap. A robotic policy that handles the egg perfectly inside a digital twin can still crush it once it meets a real carton with slightly different friction or slightly different lighting. Then there's dexterity, and this is exactly where our egg comes back into the story. Walking gets the attention, but reliable manipulation may matter more commercially. Human hands combine strength, sensitivity and precision all at once. We grip a heavy tool, then pick up something fragile a moment later, Adjusting force constantly, without ever consciously calculating it, We truly are remarkable, really, when we think about it But back to the robot, because that's the egg and the tote in the same hand, seconds apart Reproducing that mechanically, at an economic cost, is extremely hard. [34:48] Power and heat come next. A useful mobile robot needs enough operating time. But more batteries mean more weight. More weight means bigger actuators and stronger structures just to move the thing around. On top of that, compute, motors and power electronics all generate heat. So thermal management becomes its own engineering problem. Then there's industrialisation. Moving from a prototype to repeatable manufacturing is really hard. Components have to meet demanding requirements for torque, weight and durability. Supply chains have to deliver consistent parts by the thousand, and eventually the million. Maintenance networks have to grow right alongside the fleet. Integration is another important point. A robot doesn't create value simply by arriving at the factory. It has to connect into warehouse systems, into safety controls and into existing workflows. The people working alongside the robot also need to understand how as well. Safety, liability and cybersecurity sit in a different category of risk entirely. A generative AI error might produce a wrong document or it might hallucinate something. A physical AI error can damage equipment or hurt somebody. A robot working around people has to fail safely and predictably, and it has to be protected from unauthorized control, not just mechanical failure. [36:14] And finally, materials. Remember that number I asked you to hold in your mind earlier? Three and a half kilograms of magnets in every single Optimus unit. This is why. High-performance actuators mostly depend on neodymium magnets, which is what gives a compact joint enough torque to actually be useful. The IEA expects robotics to become a much larger source of rare-earth magnet demand beyond 2030. And this isn't just theoretical. Tesla said in 2025 that Chinese export restrictions on rare-earth magnets had already affected Optimus production, the same magnets in the same 40-plus motors we talked about earlier. Rare Earths won't be the binding constraint for every robotics program this year. In the near term, intelligence, reliability and integration probably matter more. But if humanoid production ever moves from thousands of units to millions, magnets and specialist electronics could become genuine bottlenecks. Once again, the AI story comes back down to the physical economy. [37:17] So let's zoom out even further now, because physical AI isn't only a terrestrial story. Space is a natural environment for robotics, because human labour there is extraordinarily expensive, limited and vulnerable. Astronauts need life support, radiation protection, food, water and a safe way home. A robotic system can be designed around the task instead, without any of that. In orbit, physical AI could support spacecraft inspection, maintenance and refuelling. Robotic systems could help assemble structures that are too large to launch, fully built. Things like antennas, telescopes and solar arrays. On the Moon, robots could survey terrain, transport cargo and prepare landing areas before a crew even arrives. They could inspect equipment and carry out maintenance between human missions. NASA's Lunar Technology programs explicitly include autonomous robotics, excavation, surface power and thermal management. [38:18] This doesn't mean a conventional humanoid gets simply replaced on the moon. The optimal space robot might have wheels, tracks or a body designed specifically for low gravity. But some of the capabilities being developed for humanoids are directly relevant anyway. Dexterous manipulation, autonomous navigation and the ability to operate within infrastructure built for people. Physical AI may also enable a different way of working. Instead of controlling every robotic movement from Earth, People could simply specify the objective and let the robot handle local navigation and execution itself. That becomes especially important as communication delays grow and real-time remote control stops being practical. The long-term lunar economy will therefore depend not only on launch vehicles and landers, but on machines able to do useful physical work once they arrive. So let's bring this back down to earth for a moment and talk about what investors should actually be watching. I'd divide the physical AI investment opportunity into three horizons. The first is the existing automation economy, industrial robots, machine vision, warehouse automation, and surgical systems. These are established markets with real revenue and real operating histories. [39:36] The second is the physical AI enablement layer, simulation, edge compute, sensors, foundation models, and systems integration. The opportunity here comes from making existing machines more adaptable and enabling entirely new robot categories. [39:54] The third is general-purpose embodied intelligence, and this is an area that I'm really interested in. This includes humanoid platforms and machines that can learn or perform multiple tasks. This horizon potentially offers the greatest optionality, but it also carries the greatest technical, commercial, and valuation risk. Now, across all three, there's a handful of things worth watching closely. Start with conversion. Are demonstrations actually turning into paid deployments? Are small pilots growing into larger fleets or staying stuck in pilot scale? Cisco's own July 2026 survey of industrial leaders found this playing out at the plant level already. 61% of industrial organizations were actively deploying AI in some form, but only around 20% had reached what it called mature, scaled adoption. That's the conversion question happening across the entire industry, not just in robotics specifically. [40:55] Also, watch the operating numbers too, runtime, uptime and how often a human has to step in. Watch the cost per productive hour against human labour or existing automation. That last one is probably the single most useful number in the whole sector. Watch the supply side as well, production yield and manufacturing cost at the robot makers, order growth further down the chain at the component suppliers. And watch the things that could constrain the whole sector, not just one company, such as rare earth availability and export controls, the development of safety standards and insurance, and repeat orders and recurring service revenue, since that tells you whether customers are coming back rather than just trying the technology once. The most interesting signal, though, might be whether one robot's experience can actually improve the performance of an entire fleet. If every robot has to be trained individually, the economics stay limited. If learning from one machine can improve thousands of deployed machines, robotics could start to pick up some of the scaling characteristics we associate with software. [42:05] There's an argument doing the rounds in the industry that pushes this idea further. An executive at Grey Orange, which sells fleet orchestration software, argued this year that the model and the hardware will both eventually commoditise, and that the durable value actually sits one layer up, in the software that runs a fleet day-to-day and compounds proprietary data from every failure it's ever recovered from. Data a new competitor simply can't buy or simulate its way into. And I think it's worth pointing out here that this is a self-interested argument coming from a company that sells exactly that layer, so take it with an appropriate pinch of salt. But the underlying logic holds up on its own merits. If fleet learning effects are real, whoever owns the operational data from thousands of live recoveries may end up in a more defensible position than whoever owns the underlying model, or even the robot itself. But those machines will still need to be manufactured, powered, maintained and integrated into the physical economy. Physical AI may therefore end up combining software-style learning effects with industrial-style capital requirements. Now that combination is exactly what makes the sector so compelling and so difficult to analyse. [43:20] So let's pull all of this together into a base case. And my base case for physical AI is not that humanoid robots are about to flood into factories and warehouses at scale, it's a slower, more uneven build than the demonstration videos suggest, running alongside a deepening of the existing robotics economy. [43:39] Over the next 12 to 18 months, I expect the loudest headlines to keep coming from humanoid demonstrations and production announcements. But I expect the more important progress to happen somewhere less visible. Machine vision getting better at recognising unfamiliar objects. Collaborative robots learning new tasks faster. Existing robot fleets becoming easier to redeploy. That's where I expect the earliest measurable enterprise returns to show up, not in general purpose humanoids. [44:09] The critical variable, the one that I want graded against the data as this plays out, is the gap between company-reported production numbers and independently verifiable fleet utilization. I call this the delivery gap. Tesla, Figure, and Agibot have all published robot production or delivery figures over the last year. None of that on its own tells us how many of those units are doing sustained, paid, productive work. It doesn't tell us the intervention rate or the cost per productive hour. Figur's paid BMW contract is a genuine first step towards closing that gap, since a purchase order is a different category of evidence to a production headline. But it's the one contract for 40 robots at one plant. It's a data point and not a trend yet. The delivery gap is the thing to watch across the whole sector, not just at Figur. If it closes more broadly, meaning verified utilization data starts catching up with the production headlines, the humanoid investment case strengthens materially. If it stays wide or it widens further, I'd treat the production numbers as manufacturing progress rather than commercial proof. [45:19] The bull case is that embodied intelligence starts to show real learning transfer effects, where lessons from one deployed robot really improve the performance of the wider fleet. Now if that happens, physical AI starts to pick up some of the scaling economics normally associated with software, layered on top of an industrial capital base. That would be a powerful combination. [45:41] Now the bear case is different. Dexterity, reliability and integration prove harder to solve than the demonstrations suggest. Humanoid deployment stays confined to a smaller number of narrow, bounded logistics tasks, and the bulk of near-term physical AI value ends up going to the existing robotics and automation economy, rather than to the humanoid platforms currently getting most of the attention and the capital. The QFMI base case sits closer to that second outcome, at least for now. The existing robotics economy, things like industrial robots, machine vision and warehouse automation, already has real value, already has real revenue and real operating histories behind it. General purpose embodied intelligence remains the highest optionality and the highest risk of the three horizons. Until the delivery gap actually starts closing, I'd treat any near-term valuation built mainly on humanoid potential rather than on today's deployed and revenue-generating robotics, with real caution. So let's pull all of this together and wrap it up. Physical AI is sometimes presented as the next stage of artificial intelligence. I think it's more accurate to describe it as the point where the digital AI economy begins to converge with the physical industrial economy. [46:59] The humanoid race is real. Tesla Optimus, Electric Atlus, Figure 03, Digit and Agibot A2 all represent different approaches to building embodied intelligence. Some start with AI, some start with mechanical engineering, some start with a defined enterprise workflow, others start with manufacturing scale and an industrial ecosystem. But humanoids are only one part of the opportunity. Physical AI also includes the millions of industrial and service robots already working in factories, warehouses, hospitals and farms. [47:34] The near-term value may come from making those existing machines smarter and from automating bounded workflows where the return can actually be demonstrated. The longer-term prize is a general-purpose machine that can learn, adapt and perform a wide range of useful physical tasks. Between the two sits an extensive value chain, compute and simulation, sensors and actuators, magnets and batteries, software, integration and robotic services. Now the winners here won't be decided by who produces the most impressive demonstration. They'll be decided by who can convert intelligence into reliable physical work, completed safely, repeatedly and at a cost that actually creates a return for the customer. And as the sector scales, the constraints will reach far beyond the AI models themselves. They'll include dexterity and power, rare earth magnets and precision components, manufacturing capacity, safety and enterprise integration. That's why physical AI belongs within the AI supercycle. It connects models and compute to machines, materials, factories and supply chains, and eventually to orbital and lunar infrastructure. [48:44] The robot may be the visible product, but the investment opportunity lies across the entire system. Okay, so that's where I'll leave the physical AI thesis, because we've covered a lot of ground today. This also brings us to the end of this run of solo episodes. Over the last nine episodes, we've worked through the physical constraints behind the AI supercycle, including the memory constraint, the power constraint, the critical materials constraint, the orbital compute thesis, also photonics as an emerging constraint, capital formation, and enterprise ROI with token budgeting. Putting all of that together, it's essentially the full chain of the AI supercycle thesis, the capital-intensive industrial transformation spanning the entire AI continuum, from mines and megawatts to models and orbit. [49:36] And also it's that time of year again where we need to take some time out and recharge the batteries in the last few weeks of summer. So I'll be taking a short break from the podcast and when the show comes back in early September we'll have guests on the show to talk about certain aspects of the AI super cycle. If you found this episode useful please follow or subscribe on your preferred podcast platform and share with anyone interested in AI, robotics and the industrial economy developing around them. If you want to know more you can find all of our work at qfmi.substack.com, the market pulse and in the spotlight articles are free and always will be the weekly outlook the weekend debrief and the monthly strategic research report sit behind a paid subscription subscribe and you get the full picture, thank you for listening to the ai super cycle podcast enjoy the rest of your summer and i'll see you when we're back until next time.