The $47 Billion Physical AI Wave and the Hardware Rebound

Zach Martin

Welcome to Robot Capital on KMKT IR Hub Radio. I'm Zach Martin, and it's the week of October 9, 2026. Let me open with a roster problem. Imagine a football team that spent three straight drafts taking nothing but quarterbacks. Great arms, great highlight reels, and nobody to block for them, nobody to catch the ball, nobody to snap it. For a stretch, that's a little bit how physical AI money behaved. The bipedal humanoid was the quarterback, and the demo video was the highlight reel.

Zach Martin

In Episode 7 we looked at the mega rounds in autonomous vehicles and defense. In Episode 8 we walked through the record $11 billion pouring into venture backed humanoid robotics, and we talked about how the middle of the market was getting hollowed out. A few enormous checks at the top, a lot of small ones at the bottom, and not much in between. This week's briefing picks up from there, because the money is doing something different now. It's moving down the stack, into the actuators, the dexterous hands, and the teleoperation data that make a humanoid possible. And it's moving out across the economy, into the power grid, into orbit, into chip fabrication plants, and into the software that lets a mixed fleet of tractors and haul trucks run themselves.

Zach Martin

Then, because this is a briefing and not a pep rally, we're going to hold all of that up against the people who actually have to buy and run the machines. Four hundred warehouse operators answered a survey, and their answers are a lot more sober than the funding headlines. We'll finish with the graveyard, the well funded hardware companies that raised hundreds of millions of dollars and still didn't make it. Capital is moving fast. The question for the hour is where it lands, and whether it can wait long enough for the payback to show up.

Zach Martin

Start with the scoreboard. Crunchbase counted global venture funding into physical AI at $47.4 billion across 521 deals in the first half of 2026. And a quick note on definitions, because it matters. In Crunchbase's criteria, physical AI includes robotics, autonomous vehicles, aerospace, drones, industrial automation and sensors. So this is a wide net. It is not a humanoid only number.

Zach Martin

Now the comparisons, and these are the ones that made me put the coffee down. That $47.4 billion is almost four times what physical AI startups raised in the second half of 2025, which was $12 billion across 470 deals. It's up nearly 80% from the first half of 2025, when the figure was $26.4 billion across 436 deals. And here's the one to hold onto. Across the three years from 2022 through 2024 combined, venture investors put $41.9 billion into physical AI companies. That's several billion dollars less than what went in during the first six months of this year alone. Three full seasons of spending fit inside one half, with room to spare.

Zach Martin

But be careful with a headline number like that, because a few giant rounds can carry it. One deal in particular accounted for nearly a third of all the venture dollars in that period: Waymo's $16 billion Series D in February, co led by Alphabet, Dragoneer Investment Group, DST Global and Sequoia Capital, at a valuation of $126 billion. Then you have Anduril, which raised another $5 billion in May at a $61 billion valuation. That's double the $30.5 billion valuation it got less than a year earlier. Shield AI landed a $2 billion Series G in March, co led by Advent International and JPMorgan Chase, at a $12.7 billion valuation. And Saronic, the autonomous sea vessel company, raised $1.75 billion in a Series D led by Kleiner Perkins at a $9.25 billion valuation, more than double its Series C level from 2025.

Zach Martin

So when you hear the big number, know what's inside it. Forbes contributor Sabbir Rangwala, writing on October 5, puts the first half at roughly $50 billion across roughly 500 deals, and he notes that his figure includes secondary investments in companies like Waymo and Anduril, not just fresh startup rounds. Two different counts, same direction. The point is not whether it's $47 billion or $50 billion. The point is that the year already looks like nothing the category has seen before.

Zach Martin

Zoom out to hardware generally and you get the Silicon Valley Bank view. Their 2026 report on physical AI and the future of robotics says total venture investment in hardware companies is on pace for $120 billion by year end, if the current pace holds, and that 2026 has already eclipsed the full year total for 2025. That's a projection, so treat it as a projection. But here's a number that's an observation, not a forecast. One in five venture investors now allocates at least 10% of their deals to US hardware. That share has more than doubled since 2022.

Zach Martin

Why would investors who built careers on software lean into motors and metal? The SVB report points to growing confidence that hardware's physical intellectual property is more resilient to AI disruption. Let me unpack that, and I'll flag that this is the report's reasoning, not mine. If an AI system can write a software product in an afternoon, then the software moat gets shallow. A machine that has to be built, certified, shipped, serviced and trusted in a real building is much harder to copy with a prompt. Hardware has been the unloved part of the venture world for a long time. This is a repricing of that view.

Zach Martin

The same report says that if you set aside the outlier investments in OpenAI and Anthropic, hardware companies account for nearly one third of all US venture capital invested this year. And within hardware, defense gets a premium. SVB found that defense oriented companies command a 31% valuation premium relative to the broader hardware space. Defense and aerospace are, in their words, areas seeing particular enthusiasm. Keep that 31% in your pocket, because we'll come back to it when we talk about whether these valuations are earned.

Zach Martin

Here's another piece of the shift, and it's about who is writing the checks. In that Forbes piece, Rangwala says Wall Street is getting in on the action, and he names JPMorgan, SoftBank, Goldman Sachs, KKR and Blackstone. Remember that Shield AI round I mentioned, co led by JPMorgan Chase. Now think about what that implies, and this part is my interpretation. Traditional venture funds raise money on ten year cycles and expect a few big winners. Physical AI is capital hungry, it needs patience, and it needs big checks. Rangwala himself says physical AI is tough, that it takes patience, innovation, data and extensive testing before it can generate revenue, and that traditional venture capital stayed away from it. Large pools of institutional money are a natural fit for that profile. They can write the check that a seed fund can't.

Zach Martin

What do the investors themselves say they're buying? Crunchbase talked to two of them. Ryan Ziegler, a general partner at Edison Partners, argues the opportunity is much broader than robots and humanoids. In his view, physical AI is the convergence of software, hardware, sensors, and services across real world applications. What's changing is that AI can now process data from those systems at scale and speed, while the hardware gets cheaper. He made a point I liked. Even our mobile phones now have LIDAR scanners on them, he said, democratizing the ability to map objects and spaces. And he says the hardware becomes the distribution model for a data intelligence flywheel. In plain English, you sell or lease a machine, and the machine's real job is to collect the data that makes the software smarter, which makes the next machine more valuable.

Zach Martin

Joe Fath, a partner and head of growth at Eclipse Capital, offered a line I haven't stopped thinking about. Eclipse describes its focus as investing on the shoulders rather than the head. The head is the standalone large language model provider, which Eclipse generally avoids. The shoulders are the chips, compute, energy and data centers underneath, plus the companies using AI to build new businesses in the physical world. Fath says funding is shifting away from experimentation and toward companies that can hit production milestones, land customers and scale efficiently. And he expects the strongest moats to belong to companies that vertically integrate and own multiple layers. His summary of what customers want is worth quoting. They value operational efficiency, reliability, and revenue, not technical sophistication alone.

Zach Martin

Now let's talk about exits, because a boom isn't a boom until somebody gets paid. Crunchbase says the physical AI space has produced several notable exits in 2026, but activity has been concentrated in aerospace, defense and drones more than in robotics. SpaceX is the clear outlier. It raised $75 billion in its June IPO at a $1.77 trillion valuation. Space intelligence company HawkEye 360 went public and raised $416 million. Autonomous drone maker Aevex raised $320 million in its debut. On the acquisition side, Mobileye bought Tel Aviv humanoid startup Mentee Robotics for roughly $900 million, in a deal the company explicitly tied to its push into physical AI.

Zach Martin

One more data point on strategic buyers. PitchBook's Q4 2025 report, published in March, put 2025 robotics venture investment at roughly $27.6 billion across 1,009 deals, with Q4 alone at $5.8 billion across 249 deals. Defense and security robotics led all segments by capital invested, fueled in large part by uncrewed aerial systems, and PitchBook noted that robotics M&A rose sharply in 2025, suggesting strategic buyers are getting more active as startups move closer to scaled deployment. So the picture is this. The big exits so far are in the sectors where governments and giant corporations were already customers. Pure robotics is still waiting for its own version of that moment. Keep that gap in mind, because it's exactly the gap the rest of this hour is about.

Zach Martin

So that's the macro. Record dollars, a repricing of hardware, Wall Street at the table, and exits that so far favor defense and space. Now the question that matters for this week. When the money isn't going to a humanoid, where is it going? Let's go find out, and the first stop is the electric grid.

Zach Martin

Rangwala has a framework that helps here, and I want to give it to you because it organizes everything else in this chapter. He uses a term, Autonomy of Things, which he trademarks, so I'll credit him for it. The idea is that the internet of things connected static machines to computers about three decades ago, and autonomy connects both static and moving things, plus sensors and location data, to create autonomous movement. Then he sorts the world into three kinds of environment.

Zach Martin

The first is uncontrolled. Think autonomous cars in dense city streets. The second is semi controlled. Autonomous trucking on highways, the autonomy of blue collar vehicles in construction, agriculture and mining, drones in regulated airspace, robots and drones monitoring infrastructure. The third is controlled. Robots in industrial settings and campuses, rail and shipping yards, warehouse logistics. And here's the trend line for this week. The first wave of robotics business was concentrated in that third bucket, the controlled one. The capital is now flowing in volume into the middle bucket, the semi controlled world, where the weather, dust, vibration and shock are part of the job description.

Zach Martin

Why would anyone go to the harder environment? Rangwala lists the drivers as capital efficiency, productivity, quality of life, safety, and what he calls the acute shortage of skilled human labor in various physical industries. And he reminds us that this isn't new for the big incumbents. Caterpillar and John Deere have been investing in partial or full autonomy for at least three decades. The 2006 DARPA Grand Challenge spurred the driverless car movement, initially pioneered by Google, which became Waymo. What's new is that specialized venture funds are now backing the startups founded by alumni of those companies and universities, in driverless cars like Wayve, trucks like Aurora, Waabi and Gatik, drones, and infrastructure monitoring like BrightAI.

Zach Martin

Now the grid, which is the most concrete example in the research this week. Two European venture firms, SET Ventures in Amsterdam and Future Energy Ventures, called FEV, in Berlin, published a white paper on how robotics and physical AI can support the energy grid. They researched and mapped 600 companies. A little background on the firms, because it tells you who's reading this map. SET was founded in 2007. Its Fund III, at 100 million euros, and Fund IV, at 200 million euros, are both active. It invests early. FEV was founded in 2016 as a spin off of Innogy Ventures, which was backed by E.ON, the major European utility, and it has $235 million in funding across the energy space. FEV tends to focus on later stage deals. So this is an early stage investor and a later stage investor putting their names on one map.

Zach Martin

The map ranks four priorities, and the ranking is the story. Number one is operations and orchestration. As robots deploy into the field, the layer that coordinates them rises in value, and the stickiness comes from logic embedded in each company's specific workflows. Number two is physical work. Increasingly customers want robots to actually do something, like repair a piece of equipment. The firms say companies with proprietary technology, recurring robots as a service models, and clear operational advantages are an investment priority. Number three is inspection and monitoring, where drones and robots have progressed dramatically, which means there are fewer opportunities to find high quality investments. It's an area of selective focus. And number four is asset intelligence software, which is not an area of focus at all.

Zach Martin

Let me translate that last one, because it's the sharpest edge in the paper. Asset intelligence software is the layer that takes data from sensors on a power line or a turbine and tells you what's wrong. For years that looked like a great business. The white paper says advancements in AI have eliminated barriers to entry and increased commoditization for that base layer. If everyone can build the dashboard, nobody can charge much for the dashboard. And inspection by drone, which was the hot category not long ago, is now crowded enough that the investors are picky. The money wants to move one step further, from seeing the problem to fixing it.

Zach Martin

The two investors put that in their own words. Moritz Jungmann, general partner at FEV, said physical AI in energy is about machines that act, not just observe. Drones have gotten very good at flying around and capturing data, he said, but the next generation will interact with what they find, so robots might actually fix something, not just flag repairs. And Austin Wood at SET said the next wave of physical AI in energy will be won on integration. The winners, he said, will take proven advances in autonomy, batteries, and actuators, and engineer fit to task systems that are reliable in the field and profitable from day one. He added that with hardware and compute costs at historic lows, the real challenge is finding the environments where a purpose built system can beat human labor on cost or productivity, and then optimizing performance relentlessly.

Zach Martin

Notice the discipline in that. Profitable from day one. Beat human labor on cost or productivity. That's not a science fair pitch. And the examples in the paper are very physical. One image shows a COSMIC robot doing real work in solar panel assembly, installing the panels. The idea going forward is that robots will repair infrastructure, like repairing a valve on an oil pipeline, or repairing and reattaching electric power cables. The paper points out that this kind of physical work is already happening in agriculture, construction and mining, and that as robots, sensors, compute and AI mature and get cheaper, it becomes practical elsewhere.

Zach Martin

The portfolios give you a feel for the scale of the companies involved. SET has two current investments in the AI bucket. Spoor, from Norway, uses computer vision to detect birds near wind turbines. Flexidao, from Spain, aggregates global energy and emissions data. FEV backs Prisma Photonics, which delivers monitoring using lasers and fiber optic sensors. These are specific, narrow, almost unglamorous businesses. That's the pattern. In this part of the market, boring and specific tends to beat grand and general.

Zach Martin

Here's a second place the money is moving that I find fascinating, because it solves a problem anyone who's ever managed a crew will recognize. Rangwala highlights a key emerging trend in deep tech venture investing, companies like ASI and AIM that build OEM agnostic physical AI stacks. OEM means original equipment manufacturer, the company that built the machine. An OEM agnostic stack delivers autonomy and fleet management to mixed fleet operations, in trucking, construction, agriculture, mining and logistics.

Zach Martin

Why does that matter? Picture a contractor or a mine operator. Their yard isn't full of one brand. It's a few machines from one manufacturer, a few from another, a couple of older units that are still earning their keep. If each manufacturer sells you its own autonomy, you get an autonomous island here and a human driven island there, and the whole operation can't be coordinated. Rangwala says OEM agnostic autonomy lets the operator run the whole fleet autonomously rather than piece meal, enabling a coherent, productive and efficient process. This is my reading, but it's the same logic as why a team wants a coordinator who can run any formation with any players, not a system that only works with one specific roster.

Zach Martin

And look at how that echoes on the humanoid side. In the humanoid funding tracker I'll talk about later, a European startup called Motion raised a $2 million seed round in August, led by Extantia Capital. Its pitch is OEM agnostic humanoid robots delivered as a managed monthly service for factories, warehouses and logistics operations. Small check, but same idea. The customer doesn't want to bet on a single hardware brand. The customer wants the work done, and the layer that orchestrates the machines is where the value sits.

Zach Martin

Now to the part of the Forbes piece that, to me, best illustrates the broadening. Linse Capital. It has backed deep tech and physical AI companies for two decades, with early bets on Skydio, Waabi and Wayve. Historically, Rangwala writes, its actual funding was limited to late rounds where it took controlling positions and wrote big checks, roughly $100 million a deal and about $1.7 billion in total. But Linse recently announced an early stage fund called Linse Ignition, which has raised and deployed $75 million since its launch in 2025. There are also co investment vehicles of $75 million to double down on those investments. All told, the fund spans 14 startups, picked from an initial funnel of 400 opportunities.

Zach Martin

Bastiaan Janmaat, a general partner at Linse, put the logic plainly. While they historically focused on leading growth rounds, they always met founders early on to build relationships. Now, he said, they can put their money where their mouth is and deploy early checks, working with the next generation of deep tech companies early and putting real heft behind the ones they believe will become industry leaders. Think about what that says about market structure. A firm that used to wait for the proof is now paying to get in before it. That's what a competitive market for deals looks like.

Zach Martin

So what's in the Ignition portfolio? Five examples, and each one is a piece of the physical AI supply chain rather than a robot. First, Amca, which designs and manufactures critical components for planes, military vehicles, and core infrastructure, using physics and geometry based systems that combine design, prototyping, testing and production at once. Rangwala frames that as a supply chain resiliency play for physical AI. Second, Varda, which lets customers process materials in orbit and return them to Earth. It designs and builds in orbit production equipment and reentry capsules for low Earth orbit, and the source describes it as a life sciences company.

Zach Martin

Third, semiconductors. Fab2 builds compact, modular chip fabrication facilities and the machinery inside them. Lace Lithography is developing what it calls beyond EUV, or BEUV, atomic scale lithography systems meant to enable quantum technologies and physical AI. Fourth, Mythic, which is developing an analog compute in memory architecture that it says can deliver up to 100 times greater efficiency for edge computing and AI. Edge computing just means the thinking happens on the machine itself, not in a distant data center. And fifth, Thea Energy, which is pursuing what it calls a faster, simpler approach to commercializing fusion power, which Rangwala notes is critical for powering the energy hungry data centers that support AI.

Zach Martin

Step back and look at that list. Orbit, chips, fusion, aerospace parts. Not one of those is a humanoid. But every one is a bottleneck that a world full of working robots would hit. Robots need chips, chips need fabs, fabs need advanced lithography, edge machines need efficient compute, all of it needs power, and a national industrial base needs parts it can trust. Rangwala names supply chain resiliency and sovereignty as another driver of these investments, creating opportunities for startups to compete in physical AI industries. That's the same shoulders idea from Fath. When the head of the body is crowded, the money goes to the shoulders.

Zach Martin

I want to be fair about what we do and don't know here. These are early stage bets, and the Forbes piece is describing a thesis and a portfolio, not reporting profits or customer contracts. Nothing in the source tells us how any of these companies are performing commercially. So treat this chapter as a map of where sophisticated investors are pointing their capital, not as proof it'll pay off. The proof is exactly what the next chapter is about, because the people who have to buy the machines have opinions.

Zach Martin

Every football season has a moment when the preseason hype meets the actual schedule. Everyone's undefeated in August. Then the games start. Warehouse robotics is in that moment, and the best evidence comes from a survey that Silicon Valley Bank ran together with Prologis, the global logistics real estate company. They benchmarked automation adoption across 400 of Prologis's customers. These are people who run real buildings full of real inventory.

Zach Martin

Here are the headline findings. One in three large companies have automated fewer than 25% of their warehouses. Let that sit. After all the hype about autonomous mobile robots, picking arms and robotic everything, a third of the big operators have automated less than a quarter of their buildings. SVB describes this as significant room for growth, which is the polite way of saying the market is early. And the stated reason is not that the technology doesn't exist. The report says high upfront costs and long payback periods remain the biggest barriers to broader adoption.

Zach Martin

Let me explain why payback is such a big deal, and I'm going to explain the logic rather than put numbers on it, because the survey doesn't supply a payback figure. When you buy automation, you spend the money now. The savings arrive slowly, month by month, as you need fewer hours of labor or move more volume with the same crew. If it takes a long time to earn that money back, you're exposed to everything that can change in the meantime. Your customer volume could shift. Your lease could end. The technology could improve and make your purchase look dated. A chief financial officer looks at a long payback and sees risk, not a robot. That's why the robots as a service and lease models matter, because they turn a big upfront check into an operating expense.

Zach Martin

The survey has another finding that tells you how operators actually behave. Fewer than half of warehouses larger than 250,000 square feet use AI in day to day operations. AI adoption goes up with facility size, but even among the very big buildings it's under half. And what do operators do first? The report says most are starting with low friction use cases, such as reporting, labor planning and quality monitoring, before expanding into more complex physical AI deployments. So the sequence is software first, then the robots. They want the dashboard and the schedule before they want the machine that lifts the box.

Zach Martin

I think that's a healthy sign, honestly, even though it cools the headlines. It's like a team that installs the simple plays before the trick plays. Operators who learn their own data first will be much better buyers of physical systems later. But it also means that startups selling complex robots into warehouses are selling to customers who haven't finished the first steps.

Zach Martin

Now let's look at what the deployed numbers say, because this is where I have to wave a flag. RCSV Research, which publishes a State of Robotics report, says logistics remains the largest single robotics vertical, with over 41,000 units deployed globally, up 28% year over year. And in the United States alone, it reports logistics robot deployment exceeding 780,000 units, up 40%, heavily anchored by Amazon's operational fleet of more than 750,000 units.

Zach Martin

Did you catch the problem? The US number is bigger than the global number. Forty one thousand worldwide, and 780,000 in one country. Those figures can't both describe the same thing. My guess, and it's only a guess, is that they count different categories, maybe the global figure counts a narrower class of robot and the US figure includes Amazon's huge fleet of simpler mobile units. But I don't have that from the source, so I won't pretend to reconcile it. What I can say with confidence is the direction. Logistics is the biggest vertical, it's growing quickly, and a single company, Amazon, accounts for the overwhelming share of the American count. That concentration is itself a warning. If one buyer is most of the market, then the market for everyone else is much smaller than the headline.

Zach Martin

Also, a word on the source, since we're leaning on it. RCSV Research publishes the State of Robotics report as part of the Robotics Center, which also runs a store selling robots, arms and datasets. That doesn't make the numbers wrong. But it's worth knowing that the publisher has a commercial interest in a vibrant market, and some figures in these reports are labeled as estimates. I'll attribute them by name each time and we'll keep a careful ear on what's a measurement and what's a projection.

Zach Martin

With that said, the RCSV report gives us the best picture I've seen of where the machines actually are. Its headline number is a global robotics market of $38 billion in 2026, up 34% from the year before, which it calls the fastest growth rate the sector has seen in a decade. The US market, it says, hit $11.4 billion, up 29%. And it describes the year as an industry moving from hardware to infrastructure, where what matters is whether an operator can collect data, train a policy, and redeploy it repeatedly, economically, at scale.

Zach Martin

The more interesting part is the long tail of verticals that don't make the funding headlines. Food service robots, 8,200 units deployed globally, up 61%, operating in over 340 quick service restaurant locations across the United States, Japan and South Korea. The report says those units amortize over three to four years at labor rates above $18 an hour. That's the payback logic made visible. A restaurant can look at an hourly wage, look at a machine, and do the arithmetic.

Zach Martin

Healthcare and lab automation, 1,200 units deployed, up 94%, and projected to reach 3,500 units by the end of 2026. The report attributes that to updated FDA and EU medical device regulation guidelines for systems that don't make contact with patients. Agriculture and harvesting, 3,400 units, up 47%. And semiconductor manufacturing, 22,500 units, up 18%, driven by CHIPS Act fabs that need flexible, reprogrammable arms. Notice that semiconductor line, because it ties directly to what Linse is funding with Fab2 and Lace. The chip factories are both a customer for robots and a product of the robotics economy.

Zach Martin

Here's what I take from the long tail. The verticals growing fastest by percentage are the ones where someone has a clear answer to the question, why would I buy this. Restaurants have a wage. Labs have a regulatory path that just got easier. Fabs have a subsidy driven building boom. Where the answer is fuzzy, adoption drags. That's the same message the Prologis survey sends from the other direction.

Zach Martin

One more contrast from the RCSV form factor table, and it's my favorite number in the whole file. In 2025, an estimated 18,400 six degree of freedom robotic arms priced under $10,000 shipped, and about 3,100 bimanual arm systems, and 2,200 mobile manipulators. And full humanoids? An estimated 410. Four hundred and ten units, at prices from $85,000 to $245,000, with the primary use cases listed as factory pilots and media or demo. The thing that dominates the conversation is, by unit count, a rounding error next to the arms. That doesn't mean it won't grow. It means the capital story and the deployment story are running at different speeds, and the gap between them is the risk.

Zach Martin

So when you hold the survey against the funding, you get a more honest picture. Investors are, as Fath said, moving toward companies that hit production milestones and land customers. Operators are, as the survey shows, moving carefully, software first, wary of payback. Those two can fit together. But only if the companies that raise the big rounds can show a customer a payback that fits inside a budget cycle. That's what the next chapter is about, because the cost of the thing that teaches robots to work has just collapsed, and that changes the math.

Zach Martin

Let's talk about something that sounds dull and might be the most important number in this briefing. The cost of teaching a robot what to do. In robotics, you can build a beautiful arm, but it's useless until it knows how to pick up a thousand different objects, in different light, with different grips. A common way to teach it is teleoperation. A human operator controls the robot, performs the task, and the system records what happened. Those recordings are called demonstrations, and they become training data.

Zach Martin

RCSV Research reports that the fully loaded cost of collecting teleoperation data fell about 60%, from roughly $340 per hour in early 2024 to $118 per hour in March 2026. It also says a standard manipulation task requires somewhere between 300 and 1,200 demonstrations to reach 80% task generalization, which means the robot handles the task correctly in new situations about four times out of five. Put those together and the report concludes that an enterprise pilot data budget now lands in the range of $50,000 to $150,000, within reach for most enterprises. Think about the difference between a $5 million science project and a $100,000 pilot. A company can get approval for the second one without a board meeting.

Zach Martin

The method of teaching has shifted too. For the first time, the report says, 61% of developers cite imitation learning as their primary manipulation training method, compared with 31% using reinforcement learning. Let me define those. Imitation learning is learning by watching and copying demonstrations, like a rookie studying film of a veteran. Reinforcement learning is learning by trial and error against a reward, which can take enormous numbers of attempts. If imitation learning is winning, then demonstrations, meaning human operators and data collection, are the fuel. And the cost of that fuel just dropped by more than half.

Zach Martin

Then there's the model layer. The report says adoption of vision language action models, called VLAs, tripled over the past year and now backs 40% of all new commercial robot deployments. A VLA is a model that takes in what a camera sees plus an instruction in plain language and outputs the robot's movements. The report names OpenVLA, Pi0, and RDT 1B as leading models, and says they run quantized at 10 to 25 hertz on consumer grade GPUs. Quantized means compressed so it needs less computing power. And here's the surprising claim. RCSV says fine tuning a 7 billion parameter VLA on 200 to 500 curated demonstrations consistently outperforms 70 billion parameter models. Its view is that efficiency trumps size. I'd call that a claim from one report, and I'd want to see it replicated, but it points to the same place. Good data beats a giant model.

Zach Martin

Now connect that to where venture money went this year. The humanoid funding tracker from Newmarketpitch lists a handful of deals that are really about this data layer. Avatar Robotics, a North American startup, raised a $6.5 million seed round in August led by AlleyCorp. Its approach combines two handed robots, remote human operators, and AI autonomy for warehouse and industrial tasks. In other words, a human is in the loop at first, doing the task, while the system learns to do it alone. Proception raised an $11 million seed in June, led by First Round Capital, to build human like dexterous robotic hands and sensor gloves for humanoid manipulation and training data collection. And Noitom Robotics, in a September 2025 deal, provides humanoid teleoperation, multimodal data acquisition and embodied AI data infrastructure.

Zach Martin

The RCSV report says the quiet part directly. The defensibility layer has moved up the stack. Companies that will look back at 2026 as pivotal, it says, are those that built repeatable data collection workflows, rigorous policy evaluation systems, and genuine vertical depth, not the ones that chased the latest hardware launch. And the F-Prime Capital report we'll get to shortly backs that up with pricing. Hold that thought.

Zach Martin

Now the humanoid money itself, using the Newmarketpitch tracker, last updated September 22. I'll caution that this is a single analyst tracker, it counts reported capital, and it says its figures are at least numbers because many rounds don't disclose amounts. With that said, here's what it found. In 2026 so far, 59 humanoid robotics financings versus 54 over the same period last year. Deal count barely moved. But reported capital jumped from $2.39 billion to at least $5.32 billion, more than double. The median round tripled, from $14 million to $43 million.

Zach Martin

Now look at the structure, because this is where I connect back to the middle market theme from Episode 8. The tracker says the boom is much less dependent on one superstar round than it used to be. The largest 2026 financing represents under 10% of measured capital, versus about 42% last year and about 65% in the comparable 2024 period. There are 24 rounds above $50 million and 21 at or above $100 million. Your creator's brief notes that disclosed rounds over $100 million rose from four in 2025 to 21 this year. And yet, below the headline rounds, growth is more ordinary. Capital excluding financings above $100 million rose only about 9%, and funding under the $50 million cutoff increased roughly 23%. So the top got deeper, and the small end grew modestly. The squeeze in the middle that we talked about is still visible. Don't expect this tracker's totals to line up with the $11 billion from Episode 8. Different trackers define scope and timeframes differently, and I'm not going to force them together.

Zach Martin

The tracker also says new startups can still get funded, but the really large checks are going to companies that have already raised before. Follow on financings account for about $5.12 billion of current capital, and their median size has risen to $64 million. That's a repeat player's market. And geography matters enormously. China is the center of gravity in this dataset, with 52 of the 59 financings and roughly $4.5 billion of measured capital. The tracker cites manufacturing depth, a large startup pipeline and explicit deployment programs. It also warns that many Chinese round labels don't map cleanly onto a standard seed to Series D framework, so the stage data is messy.

Zach Martin

Now, the part that matches this chapter's title. The tracker says the money is spreading across the stack. General purpose humanoid makers still attract the most capital, but AI platforms, actuation, manipulation, industrial robots and consumer systems are taking a much larger share than last year. Let me give you the names. On the headline side, Apptronik's $520 million round in February, tagged Series A in the tracker, with B Capital, Google and Mercedes-Benz involved. Sunday Robotics raised a $165 million Series B in March led by Coatue, for a household robot called Memo. TARS Robotics raised $455 million in April.

Zach Martin

And down the stack, here's where it gets interesting. In dexterous hands and tactile sensing, AGILINK raised roughly $139 million in June, after a Series A of more than $29 million in May. Linkerbot raised roughly $215 million in a Series B in February. Daimeng Robotics raised roughly $14 million for tactile sensors and dexterous hands. In joints and actuators, ENCOS, which makes integrated joint actuator modules, raised more than $44.7 million in a Series B in September. Xynova raised more than $29.6 million in a Series A in May, with Li Auto's strategic investment arm among the backers. And Hengxuan Robot Joint Technology in Shenzhen, which makes harmonic reducers and integrated rotary joint modules, raised an undisclosed amount in August. On the AI platform side, Spirit AI raised roughly $209 million in a Series A in June, and X Square Robot raised roughly $290 million in a Series B in April.

Zach Martin

Why would investors fund a hand maker? Here's my interpretation. If humanoids become a real product, the scarce things are the parts that are hard to make well. A hand with many joints, touch sensing and durability is a hard thing to manufacture. A joint that survives millions of cycles is a hard thing to manufacture. These suppliers can sell to many robot builders, so they don't need to win the humanoid race. They need the race to exist. That's the same shoulders logic again, and it's the oldest trick in the business. Sell to everybody on the field.

Zach Martin

Now the second reality check in this chapter, and it's about who can actually build the parts. RCSV Research has a United States market report that describes what it calls the American paradox. American companies lead in software, AI research and private valuations, and capture 52% of global robotics venture capital. Figure AI is valued at $39 billion following its September 2025 round, Apptronik at roughly $5 billion, and Skild AI at roughly $4 billion, per the brief we're working from. But in physical manufacturing volume, China dominates. In 2025, Unitree and AgiBot shipped 5,500 and 5,168 humanoid units respectively, compared with roughly 150 each for Figure, Agility and Tesla.

Zach Martin

Read that again. About 5,500 and 5,168 on one side, about 150 on the other. And the supply chain behind it leans the same way. Citing Goldman Sachs data, the report says China controls roughly 26% of the global actuator market versus about 5% for the United States, around 70% of harmonic drives, and about 60% of rare earth production. An actuator is the motor and gearing combination that makes a joint move. Harmonic drives are the precision gears inside many robot joints. Rare earths go into the magnets in the motors. So an American humanoid builder, however well funded, may depend on Chinese components at several levels.

Zach Martin

Put that next to what we heard in the grid chapter. Rangwala names supply chain resiliency and sovereignty as a driver of venture investment. Linse backs Amca for exactly that reason. And SET's Austin Wood talks about taking proven advances in autonomy, batteries and actuators and integrating them. That's the American paradox seen from the investor side. The valuation leadership is in software and design. The scarce thing, and increasingly the strategic thing, is the ability to build the physical parts at scale and at a cost that works. I'd say that's an interpretation, but it's one the numbers lean toward.

Zach Martin

Some other data on the hardware itself. The RCSV report says at least 14 manufacturers across five countries now sell six or seven degree of freedom robotic arms under $10,000, and eight of those 14 are Chinese. Lead times from Chinese manufacturers have compressed from 14 weeks to as few as three weeks for standard configurations. The open source OpenArm platform shipped more than 2,400 units in 2025. US suppliers, the report says, compete on support density, software integration and certification rather than component cost. Twelve commercial humanoid platforms are now available for purchase or structured lease, up from three in 2024. Four are bipedal full humanoids, three are upper body only torsos, and five are mobile bases with two or more dexterous arms. Average prices run from about $28,000 for the lightest torso only systems to $245,000 for full bipeds with onboard compute, and several manufacturers offer lease first programs at $3,500 to $8,000 a month.

Zach Martin

Look at what that does to the business model. The arm is becoming a commodity. When hardware commoditizes, the margin moves to whoever owns the training pipeline, the policy library and the support contract. The RCSV report calls that out directly. Companies that built advantage on hardware exclusivity are repositioning toward training pipelines, policy libraries and support contracts. So at the very moment investors are pouring money into humanoid makers, the underlying hardware is getting cheaper and more available. That's great for adoption. It's a puzzle for anyone whose valuation depends on hardware being scarce.

Zach Martin

So now the question that every episode like this has to face. Do the prices make sense? For that I want to use the F-Prime Capital State of Robotics 2026 report, which analyzes public market multiples, venture deal sizes and some high profile failures. Start with the public side, because public markets give you a benchmark for what profitable robotics actually earns.

Zach Martin

F-Prime shows public robotics companies trading at high revenue multiples. Intuitive Surgical, the surgical robotics giant, has a market capitalization of $201 billion, trades at 20.3 times enterprise value to revenue, and has a 35% EBITDA margin. Symbotic, which automates warehouses, has a $33 billion market cap and trades at 14.0 times enterprise value to revenue. AeroVironment, the drone and defense company, has a $12 billion market cap and trades at 13.8 times. For listeners who don't live in spreadsheets, enterprise value to revenue is how many dollars investors will pay for each dollar of annual sales. Twenty times is a premium price. Intuitive earns it with a 35% margin, which means real profit on those sales.

Zach Martin

Look at who's on that list. A surgical system, a warehouse automation company, and a defense drone maker. Those are exactly the three zones where we've seen capital concentrate in this episode: healthcare, logistics, and defense. And it matches the SVB finding that defense oriented companies command a 31% valuation premium over the broader hardware space. The market is pricing the robotics businesses that have real customers, real revenue and repeat purchases at a premium. That's a verdict, and it's not a verdict on bipedal demos.

Zach Martin

On the private side, F-Prime says Series A private valuations reached a median pre money of $42 million in 2025. And it breaks out what earns a premium. Proprietary data collection commands a 1.4 to 1.8 times premium. Commercial traction, meaning more than 10 paying customers, adds a 1.3 times premium. Connect that to chapter four. The data layer that we just discussed isn't only a technical advantage. According to F-Prime's analysis it shows up as a price premium. And customers, paying customers, are worth a premium on top. Ten paying customers. Not ten pilot agreements, not ten letters of intent. Paying customers.

Zach Martin

Now the counterweight, and it's an important one. F-Prime also tracks the failures. Monarch Tractor, which raised more than $250 million. Attabotics, which raised more than $170 million. FarmWise, more than $80 million. Naïo Technologies, more than $60 million. Rapid Robotics. And a public company restructuring, Luminar. Your creator's brief frames these as well funded hardware companies that shut down or filed for bankruptcy due to unit economic failure, and that framing is the one I'll lean on. I won't invent the details of each company's collapse, because I don't have them from the source. What the list establishes is simple. Money alone doesn't make a machine business work.

Zach Martin

Think about what unit economics means in this context, and I'll keep it to the general logic. Every robot you sell or lease has a cost to build, deliver, install, support and repair. It earns revenue or saves money for the customer. If the gap between those two isn't comfortably positive, then selling more robots makes the problem bigger, not smaller. A software company can sometimes ignore that for years, because adding a customer costs almost nothing. A robot company can't. Each new customer needs a physical thing, and sometimes a person on site. That's why the Prologis survey's talk of high upfront costs and long payback isn't a footnote. It's the whole game.

Zach Martin

Notice, too, where several of those failed names sit. Agriculture and warehousing. FarmWise, Naïo and Monarch are farming businesses. Attabotics is warehouse automation. Those are the same verticals that Rangwala describes as semi controlled or controlled environments, and the same ones where the Prologis survey shows operators moving cautiously. I'm not saying that tells us the sectors are doomed. RCSV counts 3,400 agricultural robots deployed, up 47%. I'm saying that being early in a hard market is costly, and some of the companies that were early did not survive to be right.

Zach Martin

So how do you hold both pictures in your head at once? The record capital and the graveyard. Here's my attempt, labeled clearly as interpretation. First, the capital is real, and it's being deployed with more discipline than the headlines suggest. Fath says funding is shifting toward companies that hit production milestones and land customers. F-Prime shows the market paying premiums for data and customers, not for demos. The SVB survey shows operators starting with low friction software. That's a market feeling its way toward businesses that work.

Zach Martin

Second, the structure of the stack is changing the odds. When arms drop under $10,000, when teleoperation data falls from $340 an hour to $118, when a small team can fine tune a model on a few hundred demonstrations, the cost to try something falls. Ziegler says the costs to build these companies have come down, comparing the shift to what cloud infrastructure did for software. Lower cost to build means more attempts, more survivors, and also more crowded competition. Cheap hardware helps buyers and squeezes sellers.

Zach Martin

Third, there's a time mismatch. Venture capital is arriving at a rate that assumes the commercial payoff is near. The operators say payback is long. Exits so far are weighted toward space, defense and drones, not toward robotics. The patient capital that Rangwala points to, from the big Wall Street names, can bridge that gap. But the bridge only holds if customers eventually pay. A valuation that needs the best case to be true is a valuation that can break.

Zach Martin

So here's what I'd watch in the coming weeks and months, and this is not investment advice, just a list of things to track. One, whether the median humanoid round of $43 million keeps climbing, and whether the middle of the market fills in or keeps thinning. Two, whether more of the large rounds are paired with named, paying customers, because F-Prime says traction earns a premium. Three, the component makers, the hands, the joints, the actuators, and whether non Chinese supply chains start to build real capacity, given that 26% to 5% actuator gap. Four, the grid and energy robots from SET and FEV's map, and whether they move from inspection to repair in actual field deployments. Five, the operators. If that Prologis survey looks different next year, with fewer warehouses under 25% automation, then the capital has found its customers.

Zach Martin

Let me bring it back to the roster. The old picture was a team with all quarterbacks. The new picture looks more like a full depth chart. Linemen in the form of chips, power and parts. Receivers in the form of hands and sensors. Coaches in the form of orchestration software that runs a mixed fleet. And, yes, a few quarterbacks still getting the headline rounds. A full roster is better than a lopsided one. But a roster doesn't win anything. Games do. And in robotics the games are played in customer buildings, over multiple years, on payback math.

Zach Martin

That's Robot Capital for the week of October 9, 2026. The money is moving to the grid, to orbit, to the chip fab, to the fleet software, to the hand and the joint, and to the data that teaches a machine to work. The reality check is just as real. A third of big operators have automated under a quarter of their warehouses, and the graveyard has some very expensive headstones. Both are true at once, and the job of anyone watching this market is to keep both in view. I'm Zach Martin. Thanks for spending the hour with me on KMKT IR Hub Radio, and we'll see where the capital lands next time.