Unitree's Massive IPO: Following the Capital in Physical AI

Zach Martin

Picture the opening bell in Shanghai on August 19th. A humanoid robot maker from Hangzhou called Unitree Robotics is priced as a company worth roughly nine billion dollars. Then the market opens, and the number starts climbing. It climbs as much as 629 percent above the offer price during the session. By the closing bell, the shares sit 460 percent above where the offering priced, and a nine billion dollar company has become, on paper, a fifty billion dollar one. Retail demand for the offering had been oversubscribed by more than eight thousand times. Welcome to Robot Capital, from KMKT IR Hub Radio. I'm Zach Martin, and this week we're following the money in robotics, physical AI, and the machines that are starting to do real work.

Zach Martin

If you listened to the last episode, Physical AI's Mega Round Moment, you'll remember the thesis I ended on. The first half of 2026 was a funding boom. Crunchbase counted forty seven point four billion dollars of venture money into physical AI in those six months, which was nearly four times the second half of 2025. But when I looked at who had actually gotten out, who had turned private paper into public shares or cash, the answer was aerospace, defense, and drones. Robotics, the humanoid and warehouse kind, was waiting for its liquidity moment. Well, this week I owe you an update, because the fresh October data says that moment has shown up. It came through two doors. One is a public listing in Shanghai. The other is a set of large acquisitions by chip companies. We'll walk through both, and then we'll spend the rest of the hour on where the capital is actually going and whether the machines can earn it back.

Zach Martin

Let's start with the facts of the Unitree debut, and I want to keep fact and interpretation in separate buckets today. The facts, according to the Crunchbase Q3 report and InvestorPlace coverage: Unitree listed on Shanghai's STAR Market, raised about six point one billion yuan, which Crunchbase puts at nine hundred and five million dollars, and debuted at a valuation of about nine billion dollars. First day, it finished 460 percent above the IPO price after touching as much as 629 percent during the session. InvestorPlace described that as turning a nine billion dollar company into a fifty billion dollar one by the closing bell. Now the interpretation. The same InvestorPlace piece said, and I'm quoting the spirit of it, that this is probably too much, too fast, and that many humanoid robots still spend more time performing demonstrations than doing valuable work in homes and factories. I think that's the right instinct. A first day pop tells you how badly people wanted to own the story. It doesn't tell you what the story earns.

Zach Martin

Think about that retail oversubscription for a moment, more than eight thousand times. Here's an analogy. Imagine a concert with a thousand seats and eight million people trying to buy tickets. You learn something from that. You learn the artist is famous. You do not learn whether the band can play, whether the venue can be built on time, or whether the people who paid scalper prices for those seats are going to enjoy the show. Oversubscription is a measure of appetite, not of quality. And an appetite that large tends to push the first trade far above what a patient buyer would pay. That's not a prediction that it falls. It's a reminder that the price on day one and the value of the business are two different numbers, and your job as a disciplined investor, or as a business owner reading this market for signals, is to keep asking what the second number is.

Zach Martin

So let's go find the second number, because Unitree gives us more to work with than most robotics names do. According to the KraneShares guide to the listing, Unitree generated roughly 1.71 billion yuan in revenue in 2025, about 235 million dollars, and that represents 335 percent growth year over year. Humanoids became its biggest revenue driver, at 52 percent of sales, up from 30 percent in 2024, with quadrupeds, the robot dogs, making up the rest. Gross margins expanded to nearly 60 percent. Net profit margins ran in the mid teens, and the company has positive operating cash flow. Read that again, because it's unusual. This is a humanoid manufacturer that is actually profitable, and not just on a spreadsheet that assumes the future. Most Western humanoid companies I'll talk about later in this episode are valued on promise. Unitree is valued on promise too, but it brings real revenue and real margin to the table.

Zach Martin

Now let me pull out the four part underwriting rubric I introduced last time and run Unitree through it. The four questions are purchase orders, deployed unit economics, gross margins, and balance sheet runway. Gross margins, check. Nearly 60 percent is a number you can build a company on. Balance sheet runway, also a check, since the company just raised about nine hundred and five million dollars on top of positive operating cash flow, so this is not a business that has to fear next quarter. Deployed unit economics is a partial check. We know the revenue and the margin, which implies that the units sell above their cost, but the sources I have don't break out per robot pricing or what customers do with the machines once they own them. And purchase orders is the gap. I don't have a disclosed backlog or customer concentration in the material in front of me, so I won't invent one. If you're ever evaluating a company like this, that's the first thing you'd want to see in the prospectus, and I'd treat its absence from my notes as a to do item, not as a verdict.

Zach Martin

Let me do some rough arithmetic, and I want to be clear this is my math and not a reported figure. At the nine billion dollar IPO valuation, against 235 million dollars of 2025 revenue, you're paying somewhere around 38 times revenue. At the fifty billion dollar close on day one, you're paying more than 200 times revenue. Both of those are rich. But notice the difference in kind. When we get to the American robot brain companies in a few minutes, you'll see multiples north of 400 times revenue. Those multiples sit on top of revenue that is measured in the tens of millions, with little margin history to speak of. Unitree's multiple sits on top of a business that already has hundreds of millions in sales and a profit line. A high multiple on a real business and a high multiple on a hope are not the same animal, even when the headline numbers look similar.

Zach Martin

There's one more piece of Unitree's profile that matters, and it's the supply chain. According to the KraneShares guide, more than 90 percent of Unitree's components, the motors, the reducers, the controllers, the sensors, are sourced domestically in China. The argument, and this is the analysis in the research I'm reading, not an audited fact, is that this gives Unitree a cost and lead time advantage over Western integrators who are assembling robots from parts scattered across several countries. Think about a kitchen where every ingredient comes from the farm next door. You can change the recipe in the morning and have the new ingredients by lunch. A kitchen that imports half its pantry has to plan weeks ahead. In a hardware business, the speed at which you can iterate on a design is a form of margin, because every month saved is a month of revenue earlier and a month of learning sooner. That's the real asset the public market was bidding on, and it's hard to copy with money alone.

Zach Martin

Zoom out to the other exits in Q3, because Unitree didn't list in a vacuum. The Crunchbase report says China dominated the IPO side of the quarter, with semiconductor and robotics listings surging on their first day of trading. The largest was ChangXin Memory Technologies, a DRAM chip developer, raising 8.6 billion dollars at an 85.5 billion dollar value, and its stock jumped more than 500 percent on debut. Shanghai chipmaker Enflame, a competitor to Nvidia, raised 912 million dollars at a 25.5 billion dollar valuation. Singapore based Shein listed at 26.3 billion, well below its 2022 peak of 100 billion. Milan based Bending Spoons raised 1.6 billion at 18.4 billion. And Unitree, 905 million at 9 billion. Put those side by side and you see the pattern: the public market in China has an enormous appetite for the hardware layer of the AI buildout, chips and robots together. That's the first door to liquidity.

Zach Martin

The second door is M&A, and here the buyers were American and the sellers were the brains. Crunchbase reports that semiconductor companies were the largest acquirers in Q3. AMD announced its intent to acquire World Labs, Fei Fei Li's San Francisco company that builds models for physical intelligence, for 8.2 billion dollars. Note the wording there, announced intent, so treat it as a planned deal rather than a closed one. Nvidia made a bid to acquire the open source AI platform Hugging Face for 12.9 billion dollars, and Hugging Face is relevant to robotics because it partners on robotics toolkits like LeRobot. And earlier in the year, Mobileye agreed to buy the Tel Aviv humanoid startup Mentee Robotics for roughly 900 million dollars, a deal the company tied explicitly to its push into physical AI, while it outlines plans for a robotaxi rollout of 17,000 vehicles by 2027.

Zach Martin

Here's my interpretation of that pattern, and flag it as interpretation. The chip companies are buying the software that tells their silicon what to do in the physical world. A processor with no model that understands space, motion, and objects is an engine with no driver. By buying spatial and physical intelligence teams, AMD and Mobileye are trying to own more of the stack above the chip, so the chip is harder to swap out. If that's right, it tells venture investors something important: the exit for a robot brain company may not be an IPO at all. It may be a check from a semiconductor company that needs a story for its next generation of hardware. That's a very different underwriting question, and we'll come back to it when we talk about Skild and its peers. For now, hold on to the big picture. Last episode, the problem was no exits. This episode, there are two proven exit doors in the same quarter, and every private robotics valuation on earth just got a new reference point.

Zach Martin

Let's widen the lens. Before we go deep on robots, I want to give you the macro tape, because physical AI doesn't raise money in a vacuum. It raises money inside a venture market, and that market just printed some of the most extreme numbers in its history. Crunchbase released its Q3 2026 report in early October, with data as of October 2nd. Global venture funding totaled 159 billion dollars in the third quarter across close to six thousand startups. That was the lowest quarter of the year so far, down 25 percent from the 212 billion in Q2. But it was up 53 percent from the 104 billion in Q3 of last year, and it still topped every quarter since Q2 of 2022 other than the first two quarters of this year.

Zach Martin

Add the quarters up and you get 679 billion dollars of venture funding globally from January through September 2026. According to Crunchbase, that's the highest amount in the first three quarters of any year on record. I like to translate numbers like that into something you can feel. If venture funding were a marathon, the 2026 runners have covered more distance in nine months than any field has covered in a full year, and there's still a quarter of the race left. Whatever your view on valuations, this is a market where capital is not scarce. The scarce thing is the ability to turn that capital into products customers actually keep paying for, and I'll keep pushing that theme through the whole show.

Zach Martin

Now, the concentration. This is the part I want you to really sit with. Crunchbase counted an all-time record of 27 companies raising billion dollar plus rounds in Q3. The previous record was 16 in Q2, and Q1 had 14. Those 27 companies captured around a third of all venture capital raised in the quarter. Eight of them raised three billion dollars or more, and the largest deals went to Databricks and Safe Superintelligence at five billion each, with Crusoe, Moonshot AI, Mistral AI, Nscale, The Boring Company, and Kling AI each raising more than three billion. Frontier labs and data centers led the big raises by count. So when you hear that venture funding is booming, the accurate sentence is that a few dozen companies are absorbing a third of the money, and a very long tail of startups is fighting over the rest.

Zach Martin

The AlleyWatch numbers for U.S. startups in July make that tail visible. U.S. startups raised 19.44 billion dollars across 492 companies that month. The 65 deals of fifty million dollars or more took 15.87 billion of it, which is 81.6 percent of all capital. Meanwhile the median deal was 6 million dollars. So the average deal size was 39.5 million, and the typical deal was 6 million. When the average is more than six times the median, you're looking at a market with two very different economies sitting inside it. One economy writes checks in the billions to companies with sovereign scale ambitions. The other is the traditional startup world, where six million dollars is a real round and a lot of founders are still counting runway in months.

Zach Martin

Where did the money go sector by sector? Crunchbase says the physical side of AI has attracted significant capital in 2026, and that continued in Q3, with aerospace, robotics, data centers, semiconductors, and energy each raising ten billion dollars or more in the quarter. Pause on the breadth of that list. It's not a robotics wave. It's a build out of the physical infrastructure around AI, the machines, the chips, the buildings that hold the chips, and the power plants that feed the buildings. Crunchbase's own conclusion is that venture funding has expanded beyond software into physical infrastructure and robotics, which it describes as reindustrializing economies. I'd add the caution that Crunchbase groups these sectors broadly, and that a robotics company's slice of that ten billion is not the same as the headline physical AI number in the first half report, which defined the category to include autonomous vehicles, drones, industrial automation, and sensors as well. Different definitions, different totals. Keep your definitions straight when you compare them.

Zach Martin

And here's a stat I find telling about who's winning. Half of all startup capital invested in Q3 went to companies founded since 2022. Four years old or younger. In a sector like robotics, where hardware cycles are slow and factories take time to build, that's a bold statement about how much investors are willing to underwrite from a standing start. U.S. companies took 91 billion dollars, about 57 percent of the global total, and the San Francisco Bay Area alone accounted for 24 percent of global venture investment in the quarter. AI driven startups across the stack raised 102 billion dollars, or 64 percent of global venture, a lower share than the two previous quarters but still 14 percentage points above Q3 of last year. So AI is not a slice of venture anymore. It's the majority of it.

Zach Martin

Let me give you the first half context again, because it helps you see how fast the ground moved. In the first six months of 2026, according to Crunchbase, physical AI funding hit 47.4 billion dollars across 521 deals. That compares with 26.4 billion in the first half of 2025, and with 41.9 billion for the three years from 2022 through 2024 combined. One deal, Waymo's 16 billion dollar Series D in February at a 126 billion dollar valuation, accounted for nearly a third of the dollars. Anduril raised five billion in May at a 61 billion dollar valuation, double its valuation less than a year earlier. Shield AI raised two billion at 12.7 billion. Saronic raised 1.75 billion at 9.25 billion. And SpaceX raised 75 billion dollars in its June IPO at a valuation of 1.77 trillion. Those are not the numbers of a niche. They're the numbers of a new asset class getting priced for the first time.

Zach Martin

Now let me add one piece the funding trackers often miss: the funds behind the funds. Vestbee compiled the venture funds that raised capital in Q3 to invest in European startups, and the list tells you what the people who write checks to venture investors believe. Index Ventures raised two billion dollars across seed, venture, and growth vehicles. Highland Europe closed a 1.1 billion euro fund. Accel raised 800 million dollars for its ninth dedicated European fund, as part of a 3.5 billion dollar raise across four vehicles. And physical AI shows up directly in the mandates. White Star Capital closed a 250 million dollar Fund IV investing in foundational AI models, vertical AI applications, and physical AI. Uplift Ventures launched a 100 million euro deeptech fund that names physical AI, energy, enterprise AI, and logistics. Connect Ventures lists robotics among its areas. Defense is growing in parallel, with firms like BAE Systems and Keysight joining a fund as investors, not just customers.

Zach Martin

So what does all of this mean for you, whether you're an investor, an operator, or just someone trying to understand where the economy is heading? Let me be a coach about it for a minute. When capital is this abundant and this concentrated, two things tend to happen. First, the companies at the top get to hire the best people, buy the most compute, and wait out the slow parts. Second, everyone else gets measured against them, and starts to believe they need a billion dollar story to be taken seriously. Don't fall for that second trap. A business that doesn't have to raise a mega round is a business with options. Ask any founder who has watched a financing window close in a hurry. The discipline that survives every cycle is the same one: know your unit economics, know your runway, and let the capital be an accelerator, not a life support system. We'll apply that logic now to the companies building the robot brains, because that's where the distance between the headline and the revenue is widest.

Zach Martin

There's an architectural fight underway in robotics, and it's the key to understanding where the biggest private valuations come from. Some companies are building the brain, a general model that tells any robot how to see and move. Others are building the body, a complete machine, vertically integrated, with its own brain inside. The brain companies sell software and licensing. The body companies sell hardware and hope to own the whole experience. The research I'm drawing on frames it as two bets in one wave, and the smartest investors are tracking both. Let me walk through each side, with the numbers, and then we'll score them against the rubric. Start with the brains.

Zach Martin

Skild AI is the headline name. It was founded in 2023 by Deepak Pathak and Abhinav Gupta, two Carnegie Mellon professors who came out of Meta's research group. In January 2026, according to TSG Invest's research and the company's own announcement as reported elsewhere, Skild closed roughly a 1.4 billion dollar Series C at a valuation over fourteen billion dollars, led by SoftBank, with Nvidia, Bezos Expeditions, Samsung, LG, and others in the syndicate. Total funding is over two billion dollars. What's the product? Skild calls it the Skild Brain, a single model designed to be omni bodied. The same model is meant to control a humanoid, a four legged robot, or a robotic arm without being custom engineered for each one. Think of it like an operating system for robots. Robot makers upload their machine's specifications through a cloud platform and API and receive a working control system in return.

Zach Martin

Here's the claim that gets investors excited. TSG's report says Nvidia testing showed the Skild Brain can run on hardware costing between four thousand and fifteen thousand dollars, compared with 250 thousand dollars or more for traditional custom engineered systems. If that holds up in the field, you're looking at something like a tenfold drop in the cost of getting a robot to do useful work, and a tenfold drop in cost usually means a vastly larger market. Note the qualifier, though. That's testing, reported by a research firm with a stake in selling investment access, and the same report is explicit that production scale reliability across thousands of messy environments remains unproven. I'll treat the cost claim as an encouraging signal, not a settled fact.

Zach Martin

Now the number that should make every disciplined investor sit up straight. TSG reports Skild grew from zero to approximately thirty million dollars in revenue during 2025. That's fast. It's also thirty million dollars. At a fourteen billion dollar valuation, that works out to roughly 467 times revenue, which TSG itself flags as among the highest of any venture backed company in any sector. Compare that to what I walked through for Unitree. Unitree's revenue is about 235 million dollars with a profit margin attached. Skild's revenue is about thirty million, with a model that, by TSG's own account, carries heavy compute costs from training and running large models on GPU clusters. There's a legitimate bull case: the company is a horizontal intelligence layer, a picks and shovels play for the physical economy, with a data flywheel in which every deployment improves the model for every customer. But that is a case built on the future, not on the trailing numbers.

Zach Martin

Let me make the risk concrete, because sim to real sounds abstract. Skild trains largely in simulation, running billions of simulated scenarios on Nvidia's tools, and by learning from millions of online videos of humans doing everyday tasks. Then it transfers that learning to a physical robot. The risk is that simulated physics doesn't perfectly match real physics. A simulated floor is never slippery in quite the way a wet warehouse floor is. A simulated box never has a torn corner and a bulging side. When a chatbot gets something wrong, you get a bad paragraph. When a robot gets something wrong, you can get broken property or an injured worker, and a legal question about who is liable. That's why the reliability numbers matter more here than in any software category I've covered. A model that is right ninety nine percent of the time sounds great until you multiply by the number of motions a robot makes in a shift.

Zach Martin

The other brain companies fill in the map. Physical Intelligence, which pitched itself as ChatGPT for robots, released vision language action models called pi zero and pi zero point seven that can fold laundry, bus tables, and assemble boxes across different robot hardware. It closed a 600 million dollar Series B in November 2025 led by CapitalG at a 5.6 billion dollar valuation, after a 400 million dollar Series A backed by Jeff Bezos, OpenAI, and Thrive Capital. Bloomberg reported in March that it was in talks to raise about a billion more at an eleven billion dollar plus valuation. Generalist, founded in 2025 by former Google DeepMind researchers Pete Florence and Andy Zeng along with Andy Barry, formerly of Boston Dynamics, raised 400 million dollars in May 2026 at a two billion dollar valuation, led by Radical Ventures, to scale its general purpose model called GEN one. Fei Fei Li is among its investors. And Field AI went in a different direction, building autonomy software for places without GPS or pre mapping, like mines, construction sites, and energy facilities. It reached roughly a two billion dollar valuation with about 405 million dollars raised, and the coverage says it has production deployments on three continents.

Zach Martin

Notice a pattern in those cap tables. Nvidia's venture arm, Bezos Expeditions, and strategic names like Samsung, LG, and Salesforce show up across multiple companies. TSG makes an interesting observation about Nvidia. It has its own robotics foundation model, Isaac GR00T, which competes with Skild, and it still backs Skild. That looks like hedging. If you're the company selling the picks and shovels to the entire physical AI industry, you want a stake in whichever brain ends up on top. It's the same logic as a casino that owns a piece of several of the biggest gamblers. And the AMD and World Labs deal I mentioned earlier, plus the Hugging Face bid from Nvidia, suggests chip companies don't plan to leave the model layer to chance. For brain startups, that means the strategic buyers aren't just investors. They're potential acquirers, and potential competitors, at the same time.

Zach Martin

Now the bodies. Figure AI, founded in 2022 by Brett Adcock, built a humanoid called Figure 03 that's designed for homes and warehouses, powered by an in house AI system called Helix. In September 2025, per the company's release, Figure exceeded one billion dollars in committed Series C capital at a 39 billion dollar post money valuation, led by Parkway Venture Capital with Brookfield, Nvidia, Intel Capital, Salesforce, and Qualcomm Ventures participating. The company has talked about its BotQ production facility and a target of 100,000 humanoid shipments over four years. That's a company target, not a delivered result. The Unicorn Screener research describes a 39 billion dollar valuation for a pre mass production hardware company as the boldest bet on its list, and the wager is that vertical integration, building the robot and the brain together, beats the labs that only sell software.

Zach Martin

Apptronik, the Austin company behind the Apollo humanoid, raised 520 million dollars in February 2026 at roughly a five billion dollar valuation, bringing its Series A to 935 million dollars in total, according to CNBC. Apollo is designed for case picking, palletizing, and machine tending, and it's piloting with Mercedes Benz and GXO Logistics. Apptronik also has a partnership with Google DeepMind to power Apollo with Gemini Robotics, which is interesting because it's buying a brain while building a body. Then there's 1X Technologies, backed by OpenAI, which is chasing the home with its NEO humanoid. Reports say its entire first year of production sold out within days of preorders opening in October 2025, and TechCrunch and Sifted reported it was targeting about a ten billion dollar valuation in a round of up to a billion dollars. Selling the home is the hardest problem in the category, a messier environment and a higher trust bar. It's also the biggest prize if it works.

Zach Martin

To put the scale in perspective, TSG notes that global humanoid shipments reached only about 13,300 units in 2025, with forecasts of 250,000 by 2030. Goldman Sachs projects the humanoid market could reach 38 billion dollars by 2035, and a forecast the research says it revised upward roughly sixfold. Morgan Stanley's number is a lot bigger, five trillion by 2050. Forecasts that long are more of a compass than a map. What matters for today is the gap between the current unit volume and the valuations. Thirteen thousand robots shipped worldwide in a year, and one private company alone, Figure, is valued at 39 billion dollars. Agility Robotics, by contrast, is valued north of 1.75 billion after a 400 million dollar Series C, and its Digit humanoid is described as the first humanoid in commercial deployment, with customers including Amazon, GXO, and Schaeffler. That's a lot less capital per unit of proven deployment.

Zach Martin

So let me run the rubric across this field and give you my reading. On purchase orders, the strongest evidence I can point to is thin everywhere: GXO's deployment of Agility's Digit, Apptronik's pilots with Mercedes Benz and GXO, 1X's preorders, and Skild's thirty million dollars of revenue. These are signals, not backlogs. On deployed unit economics, the Skild claim of cheaper hardware is promising but unproven in production, and I can't tell you what a Figure robot costs per working hour because the sources don't say. On gross margins, Unitree is the one name in this universe with a disclosed, strong number. And on balance sheet runway, the brain and body companies are well capitalized, which is both the good news and the trap. Plenty of cash lets you survive long enough to find out whether the economics work, and it also lets you avoid asking the question for longer than is healthy.

Zach Martin

So who wins, brains or bodies? The honest answer from the research is that nobody knows yet, and the open question is stated plainly: whether horizontal, body agnostic models like Skild and Physical Intelligence can out execute vertically integrated machines like Tesla's Optimus and Figure in real production environments. Tesla and Figure's advantage is that they control the whole loop, so when something breaks, there's one company to call. The horizontal model's advantage is that it doesn't have to build every body, and it gets data from every robot that runs it. There's an analogy I like from the phone world, with the caveat that analogies are not predictions. Android spread across many hardware makers, while Apple built the whole stack, and both made enormous businesses. The question is whether robotics has room for both, or whether one approach bleeds the other. For now the capital markets are paying for both, which tells you investors are hedging their own uncertainty.

Zach Martin

Now let's leave the demo stage and go to the warehouse. I want to talk about where this technology is most likely to earn its keep first, because the answer isn't a glamorous humanoid in a living room. It's a bounded, instrumented space with a labor problem. Think of a warehouse as a sports field for robots. The lines are painted, the rules are fixed, the weather doesn't change, and nobody jaywalks across the outfield. That's exactly why CXTMS, a logistics software firm, argues investors are choosing warehouses over highways. Their piece notes that robots in a warehouse don't have to handle unpredictable weather, pedestrians, or construction zones, and the regulatory burden is far lower than for autonomous vehicles on public roads. Note the source: it's a vendor with a commercial interest in this narrative, so I'm using it for the numbers it cites and the logic, not as a neutral judge.

Zach Martin

Here are those numbers. The global warehouse automation market is valued at nearly thirty billion dollars in 2026, and CXTMS cites a projection to reach 59.5 billion by 2030, which implies an 18.7 percent compound annual growth rate. The labor side is where the pressure comes from. Warehouse wages are up more than 20 percent since 2020, and turnover exceeds 40 percent annually at many facilities. Think about what a 40 percent turnover rate means to a manager. Four out of every ten people you train this year may be gone next year, so you're running a school as much as a warehouse. And CXTMS cites McKinsey research saying logistics firms plan to put 25 percent of capital spending into automation, with fulfillment operations targeting more than a third. If those intentions hold, you have a budget line, not a science project, which is exactly what a venture investor wants to see.

Zach Martin

The business model changing underneath this market matters as much as the robots. Robotics as a Service, where a customer pays a monthly fee per robot rather than buying the hardware outright, is becoming the default, per CXTMS. It converts a capital expense into an operating expense, and it lowers the barrier for a mid sized shipper who can't write a seven figure check. But look at what it does to the vendor. The vendor now owns the robots on the customer's floor, which means the vendor carries the hardware cost, the maintenance burden, and the replacement risk, while collecting revenue slowly over years. That's a model with a cash flow shape like a lease, which is why the balance sheet runway part of my rubric becomes the most important question. A RaaS company with beautiful demos and a thin balance sheet is like a landlord who has borrowed against buildings that haven't filled up yet.

Zach Martin

Here's a practical way to think about it, if you run a facility or advise one. When a vendor offers you RaaS, the monthly fee is the headline, and the real cost of ownership hides in the fine print: integration with your warehouse management system, training, maintenance, and the cost of switching if the vendor disappears. CXTMS makes the point that capital abundance cuts both ways. More startups survive long enough to compete, which gives you better pricing and more flexible contracts. But consolidation follows. Startups get acquired, pivot, or fail. So their advice to shippers, which I think is sound, is to assess a vendor's financial health and funding runway alongside the technology. In a market where 27 companies raised a billion dollars in a single quarter, the vendors who are not in that club need to be checked more carefully, not less.

Zach Martin

One category CXTMS calls the most overlooked is orchestration software. Picture a warehouse in a few years. You've got autonomous mobile robots moving totes, a robotic arm picking items, maybe a humanoid unloading trucks, and human workers walking the aisles. They come from different vendors. Each has its own software, its own data, its own idea of what is happening. Somebody has to coordinate all of it, and tell the warehouse management system what is true. That integration layer is where silos form and where the whole investment can quietly fail, because a robot that can't share data with your logistics stack is an island. I'd put it this way: buying robots is like buying musicians, and orchestration is the conductor. A hundred excellent players with no conductor is noise. The CXTMS examples of enterprise scale validation are GXO deploying Agility's Digit commercially and DHL Supply Chain partnering with Robust.AI, which tell you the large logistics operators are moving from pilot to deployment, at least in some sites.

Zach Martin

Now let me show you where the heavy money is flowing outside the warehouse, because it contradicts the humanoid headlines. AlleyWatch's July report highlights Atoms, a Los Angeles robotics company that raised 1.7 billion dollars, the second largest U.S. round of the month, backed by Andreessen Horowitz, Bain Capital Ventures, and Craft Ventures. What does Atoms build? Physical AI systems for mining, food infrastructure, and transportation. And AlleyWatch emphasizes the approach: task specific industrial machines rather than general purpose robots. In the same month, Walden Robotics in Cambridge, Massachusetts raised 300 million dollars in what AlleyWatch called one of the largest early stage rounds in recent memory for a hardware company. Industrial robotics, the report says, is emerging as a distinct, capital intensive category in its own right.

Zach Martin

Why would a smart investor prefer a machine that does one hard thing over a robot that promises to do everything? Because specificity is how you earn a return. If you build a machine for a mine, you know what the ore looks like, how heavy the load is, what the operator is paid, and what an hour of downtime costs. You can quote a payback period. A general purpose humanoid has to be good enough at everything to replace someone, and that's a much taller order. This is the same logic Ryan Ziegler of Edison Partners gave Crunchbase. He said the appeal is in high value, traditionally analog industries, including manufacturing, supply chain, utilities, agriculture, and transportation, where return on investment is measurable. He said many of these companies look like vertical software businesses, with attractive unit economics, large deal values, and multi year deployments, and that hardware becomes the distribution model for a data intelligence flywheel. That's an investor's view, and it's a good one: sell the machine, collect the data, keep the customer.

Zach Martin

Joe Fath of Eclipse Capital added a line I keep coming back to. He said that funding is shifting away from experimentation and toward companies that can hit production milestones, land customers, and scale efficiently, and that customers value operational efficiency, reliability, and revenue, not technical sophistication alone. Eclipse describes its approach as investing on the shoulders rather than the head, meaning it backs the infrastructure and the companies applying AI to the physical world, while generally avoiding standalone large language model providers. And Fath expects the strongest moats to belong to companies that vertically integrate and own multiple layers. Put that together with Atoms and Walden, and you can see a version of the future that is less about a single magical brain and more about a thousand well engineered machines, each paying for itself in a specific job.

Zach Martin

Let me bring this home with a scorecard for the industrial and logistics layer, using the rubric one more time. Purchase orders: this is where the logistics operators are ahead of the humanoid crowd. The planned spending is documented, with that 25 percent of capital spending figure for automation, and named deployments like GXO with Digit and DHL with Robust.AI exist. Deployed unit economics: a warehouse can calculate it, because wages are up and turnover is brutal, so the comparison against human labor is quantifiable. Gross margins: not disclosed in my sources for most of these private names, so I'd ask for them. Balance sheet runway: the RaaS model makes this the question to press. A vendor who owns the robots needs deep pockets, and one who doesn't have them can leave you with stranded equipment. If you're a buyer, those four questions are a sturdy checklist. Ask them in the first meeting, not after the pilot.

Zach Martin

And to the investors listening, here's the lesson from this part of the market. The easiest dollars to underwrite in physical AI are not the ones with the best videos. They are the ones attached to a customer who is already losing money on a human labor problem, in a controlled environment, with a measurable payback. That is a duller story than a humanoid in a kitchen, and it's likely a safer one. Dull, repeatable, and measurable is the sound of compounding. Small, steady wins that stack up over years are how real businesses get built, and the same principle applies to how you evaluate them: small checks against clear milestones, not a single leap of faith on a demo.

Zach Martin

I want to end by talking about something that doesn't have a face or an arm: power. Every robot needs a battery, and every model needs a data center, and every data center needs electricity that doesn't flicker. So the last piece of capital flow I want you to see this week is the money chasing energy. AlleyWatch's July report shows energy and power infrastructure took three of the top five U.S. rounds that month. Houston based Joulent raised 1.75 billion dollars, the largest single round of the month, for power systems that combine gas turbines, long duration battery storage, and renewable generation to supply reliable, large scale power to AI data centers and industrial loads. AlleyWatch's reading was that investors are betting on the energy layer underneath the AI economy as heavily as on the models themselves.

Zach Martin

The other energy deals fill out the picture. Commonwealth Fusion Systems in Cambridge, Massachusetts raised a billion dollars to advance its SPARC demonstration device toward net energy gain. Antora Energy in San Jose raised 550 million for thermal energy storage systems that convert renewable power into industrial heat and electricity on demand. Antares, in Los Angeles, raised 370 million dollars for nuclear microreactor technology. So fusion, thermal storage, and fission, three very different approaches, all commanded nine figure or larger rounds in a single month. My interpretation, and I'll mark it as such, is that investors have understood a simple constraint. You can fund the smartest robot on earth, but if the grid can't give a data center the firm power it needs to train and run the models, the robot is a very expensive statue. Energy is the bottleneck behind the bottleneck.

Zach Martin

Now consider the silicon side, which the InvestorPlace analysis lays out as a layered stack. Their argument is that physical AI extends beyond humanoids to autonomous vehicles, smart wearables, and industrial machines, and that investors can get exposure through six hardware layers: edge chips, sensors, optics, robotics, memory and power, and connectivity infrastructure. The examples they point to include Qualcomm's Snapdragon X2 chips shipping in new Microsoft Surface AI laptops for on device inference, Nvidia's Jetson modules, and Arm's embedded architectures for low latency work. On sensing, names like Cognex, onsemi, and Ambarella supply vision hardware. Micron makes low power LPCAMM memory, and companies like Broadcom, Marvell, and Arista provide the high throughput optical networking. I'm sharing that as a map of where the physical AI economy touches public markets, not as a recommendation to buy anything. The takeaway is that the picks and shovels already trade publicly, and they have for years.

Zach Martin

Why does that matter? Because it shows you a second way to think about risk. If you believe physical AI is real but you're unsure which robot company wins, the chip and component layer is a way to be exposed to the trend without picking the winner. Of course, it comes with its own risks, since those stocks move with the broader AI cycle, and I'm not your financial adviser. But conceptually, it's the same logic that makes the foundation model labs attractive: be the layer that gets paid no matter which body wins. The difference is that those component companies already have revenue, margins, and customers, which is exactly what the rubric asks for. Unitree's IPO was interesting for the same reason. For the first time, a pure play humanoid manufacturer, with profits, is a public stock.

Zach Martin

Let me gather the gaps, because the most valuable thing I can do for you is tell you what's still unknown. Number one: the valuation gap. U.S. brain companies trade at several hundred times revenue on early revenue, with Skild at roughly 467 times on about thirty million dollars, while Unitree trades on real revenue and a controlled supply chain. Number two: the architecture gap. We don't yet know whether horizontal, body agnostic models can out execute vertically integrated machines in production environments. Number three: the operations gaps that investors have to watch, which are long term fleet maintenance costs, liability in mixed human and robot work cells, and battery constraints. Think about that last set. A robot that works eight hours then needs a charge is a different economic asset from one that runs a full shift, and a robot that needs a technician every week is a different cost center from one that runs for months. None of those details show up in a funding announcement.

Zach Martin

There's also a gap in how we should read liquidity. It's tempting to say the Unitree debut proves the whole industry is real. I'd say it proves a narrower thing, that public markets are willing to pay a very high price for a robotics company with revenue, margins, and a vertically controlled supply chain. It doesn't prove the same price is available for a pre revenue American humanoid maker. Likewise, AMD's planned 8.2 billion dollar acquisition of World Labs proves a chip company will pay a lot for spatial intelligence, but the deal was announced as an intent, and the price reflects what AMD needs strategically, not necessarily what the standalone business would be worth. Strategic buyers pay for strategy. Every liquidity event is a data point, but a single data point isn't a trend line.

Zach Martin

So let's turn this into a routine you can actually use, the way a coach would turn a game film into a practice plan. First, for any physical AI company you evaluate, write down the four rubric questions and answer them with evidence, not with adjectives. Purchase orders, meaning signed demand, not pilots and not preorders unless you understand the refund terms. Deployed unit economics, meaning what one robot earns per hour compared with what it costs per hour, fully loaded. Gross margins, and watch whether they hold when the company moves from demos to volume. And balance sheet runway, which is how many months of operations the company can fund if the next round arrives late. If you can't answer one of those four, that's not a reason to panic. It's the next question to ask.

Zach Martin

Second, separate the headline from the base rate. The headline this week is that venture funding is 679 billion dollars through three quarters, that 27 companies raised a billion or more in Q3, and that robotics finally has a public exit. The base rate is that the median U.S. deal in July was six million dollars, and that most of the capital was in sixty five deals. If you're a founder, aim at the base rate for planning and the headline for inspiration. If you're a buyer, remember that the vendors with the biggest rounds are not automatically the ones with the best service. And if you're an investor, remember that concentration is a risk in its own right. When a third of venture capital rides on 27 companies, a stumble by a few of them can ripple across the whole asset class.

Zach Martin

Third, respect the slow work. Physical AI is a field where progress compounds like a muscle, not like a lottery ticket. Every deployment teaches the model something, every month of uptime builds a customer's trust, and every improvement in a motor, a battery, or a supply chain shaves a little cost. The companies that win won't be the ones with the best launch video. They'll be the ones whose robots are still working, still paid for, and still improving two years later. That's the thing I'd tell anyone building a business around this wave: don't use the technology everywhere, use it where the return is clear, and make a small improvement every week. Small gains, stacked consistently, beat heroic bets nearly every time.

Zach Martin

Let me close with where we started. A robot company that was worth nine billion dollars at the opening bell was worth fifty billion at the close, and InvestorPlace put it plainly: physical AI has moved from the demo reel into the public markets. Last episode, I said robotics was waiting for its liquidity moment. It has arrived, twice over, with a Shanghai listing and a set of chip company acquisitions. But liquidity is only the starting gun. The race is now about whether the machines earn their keep, whether the power is there to run them, and whether the balance sheets can carry them to the finish. Keep asking the four questions. Keep your own balance sheet strong. And as always, this is educational commentary, not investment advice, so do your own homework or talk to a qualified professional before you act. I'm Zach Martin, and this has been Robot Capital on KMKT IR Hub Radio. Thanks for spending the hour with me. I'll see you next time.