AI is rapidly expanding its presence. The lines between mobile devices and robots are becoming more blurred. AI is gaining physical abilities. Morgan Stanley Research looks into how the intersection of AI and the physical economy is transforming industries and creating new markets. Watch this video to understand how embodied AI is rapidly advancing, from autonomous vehicles to humanoid robots. Learn more: Humanoids: A $5 Trillion Market May 14, 2025
In 1995, astronomer and planetary scientist Carl Sagan made a disturbing prophecy in his book The Demon Haunted World about the future of the United States when nearly all of its manufacturing industries had slipped away to other countries, leading to adverse geopolitical and societal outcomes. AI is all around us. AI listens to you. AI sees your face and body. AI knows where you are right now, training the latest chat bot. AI can read. AI can write. AI can talk. AI can make a picture of cats playing poker. But AI rarely ever actually moves. In nature, motility is an organism’s ability to move independently under its own power. According to fossil records, the earliest evidence of motility on Earth traces back to bacterial flagella in the Precambrian era. The lines between mobile device and robot are starting to blur. A new Cambrian explosion of organisms is dawning. AI is about to get physical. As software becomes agentic, robots become agents. Electric machines are the corporeal sockets for the AI brain. Any machine that can be automated will be automated, maybe even including you. If you solve for autonomy for cars, you solve autonomy for everything. A humanoid robot is just one of thousands of form factors of embodied AI, a broad definition. A broad definition: Any machine that collects photons, perceives the world around it, learns, navigates, or manipulates three dimensional space. Embodied AI and national security are inextricably linked through dual purpose. Embodied AI lends itself to natural monopolies and utility networks across decatrillion dollar TAMs, or total addressable markets. It’s December 31st, 1879, Thomas Edison makes the first public demonstration of his incandescent light bulb. The incredulous crowd laughs and asks, how could people buy light bulbs if they can’t afford electricity? Edison responds, “We will make electricity so cheap, that only the rich burn candles.” This is Fifth Avenue, New York City, Easter Sunday, 1900. Spot the car. Here. This is the same street in 1913. Spot the horse. Here. Between 1485 and 1490. An Italian polymath of the High Renaissance made some sketches in his Codex Atlanticus of some fantastical machines. Over 400 years later, in 1903, Two brothers from Dayton, Ohio, achieved first flight over the sand dunes of Kitty Hawk, North Carolina. By 1914, we have the first commercial airline flights between Saint Petersburg and Tampa. Fast forward to 1967, the Boeing 737 is introduced into service. Nearly 60 years later, we have a plane so similar in design they didn’t even bother changing the name from this to this is innovation. Da Vinci would be impressed. But from this to this, I think Leo and the Wright brothers would be super disappointed. Let’s draw a simple two axis chart together. In the y axis we have the knowledge economy. The economy of bits and bytes. In the x axis we have the physical economy, the economy of atoms and photons. AI is well on its way to consuming the knowledge economy. Moving rapidly up the y axis, disrupting occupations such as writing, accounting, tax, legal, CRM and equity research analysts. But what happens when all the digital data is captured and trained? When everyone has the same compute? How will the LMS differentiate themselves? We’ll need to move to the right into the physical economy. But even with unlimited quantum compute, you can’t train a vision language actuation model without vision data. The race for photons has begun. But first, let’s talk about fat tuna. Imagine you’re on a remote island looking out to sea. Three miles offshore is a plump 600 pound bluefin tuna hunting for squid. You have no boat and no fishing tackle. How much is that uncatchable tuna worth to you? Zero. Now imagine you have a boat, fishing tackle and the latest generation fish finder. What’s that tuna worth now? In 2019, a 612 pound bluefin tuna fetched $3.1 million in a Tokyo auction. Now, let’s turn to vision data. What’s the world’s visual data worth? If you have no way of collecting it, zero. Now imagine you have the ability to collect and process yacht TFLOPs and yacht TFLOPs of data. What’s that data worth now? More than zero. Biology is hyper efficient. The world is full of creatures doing the greatest amount of work, using the least amount of energy, subject to its environmental constraints. Take the example of the humble Drosophila. Just look at that, Putnam. Now these creatures are really, really good at navigating and orienting and flight. Our fruit flies super intelligent. Well, the fruit flies brain is the size of a poppy seed. So compared to a sponge which has no brain. The fruit fly is highly intelligent. Now, I’m no insect ologist, but poppy seed brain beats no brain. Every time scientists think the secret to the Drosophila aeronautic abilities has something to do with these things. Two enormous compound eyes that are bigger than its entire head, each featuring around 400 hexagonal ocw.mit.edu. These insect lenses act as tiny computers. Light comes into the lens, which does a calculation, pre-processing the data before entering its brain. Again, folks, we’re talking poppy seed brain, not sesame seed. Lenses are amazing computers that don’t use any of the fly’s energy, and they don’t make mistakes. Hardware like this is the compounding product of Darwinian forces of biological survival. Mutation and procreation over hundreds of millions of years. But Tesla and Google don’t have hundreds of millions of years to solve autonomy. They need tools that can simulate billions, or septillion of years in just a few days, to get those Darwinian forces moving a little bit faster. An NBA point guard steps up to the free throw line for a foul shot. He takes three dribbles and closes his eyes, imagining the perfect path of the ball right into the hoop. Swish! That shot never happened. It’s all in his mind. When a U.S. open champion imagines a perfect serve didn’t happen, I. a Premier League star imagining a penalty kick. It didn’t physically happen. But when a robot dreams in simulation, it does so in a hyper realistic digital twin, complete with physics engines to replicate the laws of motion, thermodynamics, fluid dynamics, and the behavior of light as if it actually happened. As the robots collect more data, the sim to real gap continually narrows. When you’re driving a Tesla. You’re not just driving in physical space, you’re also playing a video game, feeding data into the simulated world to train Tesla’s latest FSD model. When you wear meta glasses, you are teaching the robot model how to play piano. Knit a sweater. Pour coffee or take out the garbage. Think of a timeline of the history of the modern internet from 1995 to 2025. Over this 30 year span. What’s the most valuable contiguous five minutes of data if you’re training a large language model? Well, the last five minutes of data only surpassed by the next five minutes of data. The power of recency is critical for the predictive capabilities of inference and reinforcement learning. Take this rather simple, if not esoteric, example from my office at Morgan Stanley’s headquarters in Midtown Manhattan. If I throw this pink highlighter in the direction of this Ferrari four, five, eight Italia on the coffee table and freeze time just after the highlighter leaves my hand, you will have a very good sense of the trajectory and speed of the highlighters, flight, and the sound it makes as it enters the car. You didn’t have to travel through time to know that your prediction benefits from your experience. You’ve seen it before and the context. You’re seeing it now. Historic data is important, but those who have the best allies, real time data have a major advantage. Industrial companies are rich and tech poor. Tech companies are tech rich and tamper. If you’re a $4 trillion market cap company, you’re not going to become a $10 trillion company by going after a $50 billion Tam. You got to target that $5 trillion Tam. That $50 trillion Tam. That’s not going to happen by getting people to spend one more hour of the day on a smart device. You need to go after the physical markets and grind out those atoms. If there was ever a Tam opportunity that could exceed the size of the global economy. Embodied AI is the one. Have any of you ridden in a Waymo there all over San Francisco, Los Angeles, Phoenix. They just launched in Austin and will soon be running around Atlanta, Miami and Tokyo. Our internet team forecasts Waymo’s fleet to grow from just over 1500 units today to 23,000 by 2030. Now, why would the world’s largest tech search engine company want to get into the autonomous car business? Because the tech is really good. It’s getting better all the time. And maybe because they want to spread their bets beyond the core search business, you may soon look back in astonishment that you ever got into a rideshare vehicle with a human driver. Meta is building some serious capability in AI Foundation models, simulation, and metaverse. But it’s Meta’s efforts in reality Labs that could really open up entirely new teams and transform the company. What if you had two ultra high definition cameras embedded in a device that you could wear on your face? Imagine those cameras capturing precious real world data of all the things you do with your hands. Now imagine 20 million of these things in operation within two years, nearly two times the number of Tesla vehicles on the road. Every medic glasses user is training a humanoid avatar iterated in simulation across billions of scenarios in a digital omniverse. The glasses may be stylish, but your face your face is a battleground. It’s no secret Amazon is a major force in AI, but its vertically integrated physical infrastructure uniquely positions the company for pushing the boundaries of robotics. In 2017, Amazon had five human workers per robot in 2024. We estimate Amazon had around two human workers per robot. Amazon’s highly automated fulfillment center in Shreveport, Louisiana, is being transformed through robotics. Our internet team sees potential for $10 billion of annual savings from robotics and automation. At that scale, Amazon can turn robotics into its own business. AWS started out as an internal efficiency measure before it became nearly 60% of Amazon’s operating income. Could we see Amazon Bot Services in the next few years? Apple may have paused their autonomous car project, but once CarPlay gets access to the video data from inside and outside of the vehicle, then things could get very interesting in Cupertino. The skills transfer ability for Apple into embodied AI is pretty obvious, but we’ll list them anyway. Software. Hardware. Compute. Battery. Sensor. Infrastructure. Supply chain. The car of the future is essentially a giant iPhone wrapped around your body like an immersive Imax screen covered in carbon fiber reinforced plastic attached to an electronic skateboard driven by a supercomputer. Apple doesn’t want to make a car. They want to turn your car into a mobile Apple Store. Often asked, What is Tesla’s secret sauce? What’s their moat? Well, it’s not really one thing, but the combination of six attributes that set Tesla apart from its peers. Let’s look at each of them. Data. 7 million cars on the road today. Over 100 million by 2040. Robotics. In-house electric motors and actuators. Just the hands of the optimists have 22 degrees of freedom. Energy leading battery storage solutions at scale. AI, a world class AI team developing FSD dojo and custom silicon manufacturing. The most vertically integrated US local sourced auto company in the world. SpaceX redundant, resilient cyber secure comms space is the transport layer, the connective tissue of the AI ecosystem? So out of Tesla’s dreams, what does Elon Musk think is the single most critical component of the company’s moat? Without any doubt, manufacturing. You need to make the probes to collect the data, to improve the probes, to collect more data, to improve the probes. You get the idea here. Data defines the software. Software defines the hardware. Hardware defines the manufacturing. Elon Musk has used the car industry as a laboratory to develop competency in other areas. The car is to Tesla, but the book was to Amazon. There are 1.2 billion cars on Earth traveling around 12 trillion miles per year. That’s two light years annually, roughly the distance between the Earth and the sun. 130,000 times. With an average occupancy of 1.5 passengers per car and 25 miles an hour. That’s 720,000,000,000 hours of passenger time, or 82 million years. Humans spend 82 million years of time inside cars every year. 80 million years ago was the Late Cretaceous period. At the end of the Mesozoic era, the height of the dinosaurs. The Earth looked like this. Bees were pollinating the first flowering plants at the end of the Mesozoic era. Bees. Now, what’s the value of an hour of your time? Well, that depends who you ask. But 720,000,000,000 hours times anything is a very large number. Then there’s safety. With traffic fatalities still on the rise. It seems we’re getting dumber, faster than the cars are getting smarter. For those of you with a manual transmission steering wheel having whoopee in the garage, don’t sell it. These babies will be cherished by collectors as artisanal classics from an era when humans made the machines and operated the machines. And we still get asked why Ferrari trades at 50 times p. The next time you’re flying in an airplane on a clear day. Take a look down at the earth. Notice all the roads, the parking lots. See how much of our planet surface area is devoted to vehicle transport? If I can drive thousands of times safer than humans in two dimensions, in the pedestrian filled chaos of our cities, navigating in three dimensions is a walk in the park. Now consider what I can do for our antiquated air traffic control systems. We estimate the low altitude economy can eventually surpass the size of the global car market on our calculations. One EV stall can generate as much revenue as 10 to 15 Ubers advances and E motors, energy storage, material science, communications and compute will make flying cars ubiquitous and take some of the stress off the surface of our planet. No, not that AWS, this one. Elon Musk recently posted that China makes more drones in a day than the United States makes in a year, and that all future wars will be fought with drones. Let that sink in. With conventional technology, it takes five people to operate $130 million drone. With AI, one person can operate 100 drones. Redundant, resilient if necessary. A tribble working as a team in an autonomous swarm. Asymmetric capability that may call into question the very nature of defense budgets around the world. When Elijah Otis invented the safety elevator in Yonkers, New York, in 1852, the venture capitalists of his time may have struggled to imagine how elevators would change architecture forever. This is New York City before the elevator. This is New York City. After the elevator, we think of a reusable rocket as an elevator to space. SpaceX has reduced launch costs while increasing payload capacity while increasing satellite bandwidth 100 X. That’s a 10,000 x improvement in cost per gigabyte of salable capacity in orbit. SpaceX is creating an all new internet that provides downlink connectivity to everything, every car, every plane, every drone, every ship, every home, every business, every phone on our model. SpaceX grew revenue by 60% in 2024 to $14 billion, reaching $66 billion by 2030. On our estimates, the company’s latest tender offer valued SpaceX at approximately $350 billion, according to PitchBook, making it the most valuable private company in the world. On October 4th, 1957, shortly after 7:28 p.m. Eastern Time, the US Army Signal Corp in New Jersey detected an unusual beep beep signal at 20 and 40MHz. The Sputnik moment meant the Soviets beat the United States in putting a satellite into orbit, creating great anxiety in the Pentagon. In May of 1961, President John F Kennedy announced the goal of putting a man on the moon and bringing him safely back to Earth. By the end of the decade. And in that moment, countless eight year olds in the US wanted to become astronauts, and US math and science proficiency skyrocketed. National security is a powerful innovation catalyst. Competition with China is reawakening the Apollo spirit and catalyzing policy and public support for the next era of innovation in the fields of AI, cyber, space, robotics, and quantum. The United States didn’t put men on the moon just to collect rocks and do donuts in a lunar rover. They did it because if we didn’t do it, someone else would have. A tiger cub learns by watching its mother hunt. A human shaped robot learns by watching you drink coffee, flip a hamburger, hit a forehand, or miss a three foot putt. Why? Humanoids? Because the world was made for humans by humans, and there are over 8 billion of us to watch and imitate. First adoption. Starting now with the most boring, repetitive, dangerous tasks and environments where you can control for temperature, weather, light, humidity, and where humans are in predictable areas. A manufacturing line, fulfillment center, kitchen. Lithium mine, the Tam. There are nearly 4 billion people in the global workforce, with an average annual wage of $10,000. That’s a global labor market of $40 trillion. One humanoid, at least at $5 an hour, can replace two human workers making $25 an hour, supporting a net present value of approximately $200,000 per humanoid. The US labor market has 160 million people. Every 1% substitution by humanoid is worth over $300 billion, or around $100 per Tesla share. So how can our clients express a view on the embodied AI theme? Morgan Stanley presents the humanoid 100, a global mapping of equities across a range of sectors and regions that may have an important role in bringing robots from the lab to your living room. We divide the robot into the brain, body, and integrators. Sammy’s sensors. Batteries. Actuators. Encoders. Harmonic reducers. Screws. Bearings. Connectors. Magnets. Rare earths, and many more. This is manufacturing as a percentage of U.S. GDP since World War two. Just look at this chart. It’s not just the magnitude of the decline from nearly 30% to 10%, but the persistent, uninterrupted linearity of the decline. Decade after decade, for 80 years. How much further was this going to continue before we realize that it may have gone a little too far? Morgan Stanley is making a strategic commitment to telling the story of embodied AI, leveraging our platform and relationships to help our clients identify the next crop of multi-generational, compounders transforming industries and creating new markets we believe can exceed the size of today’s global GDP. The intersection of AI and the physical economy offers the chance to disprove Carl Sagan’s 1995 prophecy. History books will be written about this time and the next 5 or 10 years. The impacts across markets and geopolitics are likely to be profound and far reaching. The Morgan Stanley research team is here for you as we navigate these consequential times. We are grateful for your partnership and thankful for your business. Stay human.
