Farewell to Pure Software and Bubble Valuations: How Veteran Silicon Valley Figures Are Reshaping the Trillion-Dollar Industry with "Physical AI"

This material is very worth digging into because UP.Partners does not truly represent "a VC that dislikes software and specializes in hardware," but rather a deeper migration happening in Silicon Valley capital: For the past fifteen years, AI has primarily understood text, images, and code in the digital world; in the next phase, AI is beginning to enter airplanes, trucks, factories, energy systems, robotics, and defense equipment, starting to truly change the movement of atoms. The UP.Partners official website even directly states its positioning as: Intelligence Meets Infrastructure and clearly says: AI is moving from bits to atoms. So a more accurate understanding is not: "UP.Partners is going against the trend of AI." But rather: It is betting on AI moving from Software Intelligence to Physical Intelligence. ──────────────── 1. Change the title: "Three Major Hardcore Unicorns" cannot be written now This is the first important correction. The three companies in your video are actually: Skydio Autonomous drones. Range Energy Electric smart trailers. UnitX AI machine vision quality inspection in manufacturing. Among these, as of now, it can be clearly confirmed that the only one that has reached a unicorn-level valuation is: Skydio. In April 2026, Skydio completed a $110 million Series F, with a valuation of $4.4 billion. Moreover, the company disclosed that it now has hundreds of millions in annual revenue, strong unit economics, and claims that as its core business expands, its reliance on external capital is decreasing. Range Energy and UnitX do not have reliable public information proving they have reached a $1 billion valuation. So: "Three Major Hardcore Unicorns" should be deleted. ──────────────── 2. My most recommended title Precise and substantial Managing over $500 million in bets on the physical world: How UP.Partners is reconstructing drones, freight, and manufacturing with Physical AI Trend depth type AI moving from bits to atoms: Why UP.Partners is betting on drones, smart trailers, and the next generation of factories Investment type After software AI, where is the next trillion-dollar battlefield? Analyzing UP.Partners' Physical AI investment logic Entrepreneurial type Apple veteran Adam Grosser's hard tech investment method: Why the next round of AI giants may come from factories, trucks, and the sky I most recommend the second one. Because the real theme is: Bits → Atoms. ──────────────── 3. Change "$500 million fund" to "managing over $500 million" The first phase venture fund that UP.Partners officially completed in 2021 is: $230 million. At that time, it mainly invested in "people and goods movement," including transportation, logistics, aviation, etc. Recently, INSIDE Startups / UP.Partners described the company as: Over $500 million under management. Therefore: "$500 million Physical AI fund" is easily understood as a specific fund size of $500 million. More accurately: A Physical-World VC managing over $500 million. Moreover, in 2026, UP.Partners established a strategic capital cooperation with TWG Global, further expanding its investment scope to: physical AI, robotics, mobility, advanced manufacturing, engineering, logistics, supply chain. So it is much broader than the 2021 "mobility fund." ──────────────── 4. Do not write "Apple's 80th employee" for Adam Grosser at this time This is the second detail that needs to be deleted for dissemination. What can be confirmed is: Adam Grosser is indeed a very early engineering/management personnel at Apple, later worked at Lucasfilm and Sony, founded Catapult Entertainment, then spent about ten years as a GP at Foundation Capital, and then joined Silver Lake. SEC historical data can confirm his early engineering and management experience at Apple/Lucasfilm/Sony. But: "Apple's 80th employee" I have not found reliable first-hand information to confirm. Therefore, it is recommended to write: Early engineer/engineering manager at Apple. That is already impressive enough without needing to force an unverified number. ──────────────── 5. The section about Silver Lake also needs slight correction Adam Grosser indeed started leading: Silver Lake Kraftwerk around 2011. This is Silver Lake's growth-equity strategy focused on: energy, resource efficiency, automation, infrastructure technology. Public information describes Kraftwerk as approximately: $750 million growth-equity fund. And not "Adam managed $500 million at Silver Lake." So the "$500 million" in your material is actually more suitable to describe: The current asset scale managed by UP.Partners. ──────────────── 6. Why Adam Grosser is particularly suitable for investing in Physical AI? Because his life spans four worlds: Engineering He has really built things. Entrepreneurship He has started his own company. Venture Capital Foundation Capital. Growth / Private Equity Silver Lake. This is very rare. He is not: A VC who only knows how to look at SaaS dashboards. He is an ATP-level pilot and has long manufactured and repaired airplanes, boats, cars, and other mechanical systems. Public information states that he has built multiple airplanes, boats, and vehicles. This will change an investor's "technical intuition." ──────────────── 7. Because the most dangerous problems in Physical AI often do not exist in PowerPoint Software products: Bugs. Patch. Can be resolved in a few hours. Robots: Mechanical arms breaking. Drones: Crashing. Trucks: Brake failure. Batteries: Catching fire. Industrial quality inspection: Missing a critical defect. These are not: "User experience is slightly worse." But may be: Physical Failure. Therefore, someone who truly understands physical products will naturally ask: How is the material? What about vibrations? What about temperature? What about reliability? What about fault tolerance? What about maintenance? What about the supply chain? What about unit costs? This is a completely different investment discipline. ──────────────── 8. This is also why PE experience is actually very helpful for Physical AI Early software VCs can easily talk about: TAM. Narrative. User growth. Next round of financing. Physical AI must ultimately face: Economics. For example, a machine: How much to sell? What is the BOM? How much to install? How much maintenance per year? How long until the customer breaks even? What is the gross margin after scaling production? How much working capital is needed? How much investment is required for the factory? These questions are actually very close to PE thinking. ──────────────── 9. Therefore, Adam's mixed VC + PE thinking is very suitable for this era VC asks: "If successful, how big can it be?" PE asks: "Does the economic model actually hold?" Both questions must be answered for Physical AI. Only answering the first question: Will lead to failure. Only answering the second question: May miss companies like Tesla, SpaceX, and Anduril that truly create new industries. Truly excellent Physical AI investment is: Venture Upside + Industrial Discipline. ──────────────── 10. The data "the number of VC institutions has surged from hundreds to over ten thousand" cannot be written Adam's: "Now there are more companies than ideas." Is indeed his public statement and a very good point. But: "There are already over ten thousand VC firms in the U.S." Is not accurate data. The NVCA 2026 Yearbook shows that in 2025, there were approximately: 2,984 VC firms in the U.S. There was even a year-on-year decline for the first time. In 2024, there were still about 3,111 firms. So the best way to write it is: The number of VC firms in the U.S. has indeed increased significantly over the past few decades, with an explosion in capital and startups, while Adam believes that what is most scarce today is not the number of projects, but truly original and sufficiently large good ideas. This retains the thought without using false data. ──────────────── 11. Moreover, there is indeed a strong "AI capital concentration" in Silicon Valley today This background you wrote is correct. In 2025, the total VC investment in the U.S. was approximately: $320 billion. Among which: 65.4% of deal value is related to AI. This is a very astonishing capital concentration. So the environment Adam is in is indeed: A large amount of capital flowing into: foundation models, coding agents, enterprise AI, AI applications. ──────────────── 12. However, it cannot be written as "UP.Partners is against the AI tide" Because it itself is now an AI investor. It just asks: Besides helping you write emails and code, can AI also make an airplane fly by itself, make factories stop producing waste, reduce fuel consumption of trailers by 40%, and enable robots to understand the real environment? This is a different kind of AI. UP.Partners currently has clear investments in: sensing, perception, automation, manufacturing, energy systems, logistics, space infrastructure and other fields. So the real distinction is not: AI vs Non-AI. But rather: Digital AI vs Physical AI. ──────────────── Thirteen, what exactly is Physical AI? I would define it as: an intelligent system that can perceive the real world, understand the real world, make decisions, and ultimately change the state of the real world through machines. It typically requires four layers: Sense Camera, Radar, LiDAR, Sensors. Understand Computer Vision, World Model, AI. Decide Planning, Optimization, Control. Act Motor, Drone, Robot, Vehicle, Machine. This is completely different from a Chatbot. ──────────────── Fourteen, ChatGPT's errors usually occur in the "information world" A wrong answer in one sentence. You might: ask again. Physical AI's errors occur in: Reality. So the core metrics for Physical AI will not just be: benchmark score. It will also include: Reliability. Latency. Safety. Cost. Uptime. MTBF. False Positive. False Negative. Unit Economics. ──────────────── Fifteen, why is Physical AI suddenly starting to explode now? It's not because everyone suddenly likes robots in 2026. But because several cost curves have crossed critical points simultaneously. First: Compute. Computing power is getting stronger. Second: Sensors. The costs of cameras, radar, IMUs, etc. are decreasing. Third: Batteries. Energy density and electric drive have matured. Fourth: AI perception. Machines can now reliably understand complex visual environments for the first time. Fifth: Connectivity. 5G, edge compute, cloud. Sixth: Manufacturing software. Simulation, digital twin, generative engineering. After these curves overlap: Machines that were previously uneconomical are starting to become economical. ──────────────── Sixteen, this is very similar to the period before the iPhone appeared. The iPhone was not suddenly invented: screen. battery. ARM. wireless communication. GPS. camera. All these things existed. What really happened was: Cost / Performance Convergence. All technologies matured simultaneously to: be combined into an excellent product. Physical AI today is also very much like this stage. ──────────────── Seventeen, this is also why UP.Partners' website currently states very accurately: "AI is moving from bits to atoms." The capital significance behind this statement is very large. In the past, AI mainly competed for: advertising budgets. software budgets. employee knowledge labor. In the future, AI will start competing for: global industrial capital expenditures. factories. transportation. defense. energy. warehousing. construction. supply chains. This is much larger than SaaS budgets. ──────────────── Eighteen, now looking at the first case: Skydio Here, let’s correct the name: The founder is: Adam Bry not Adam Brie. Skydio was founded in 2014 by three engineers with MIT backgrounds: Adam Bry, Abe Bachrach, Matt Donahoe. Bry and Bachrach previously participated in Google Project Wing. ──────────────── Nineteen, the story of "initially just wanting to make a skiing selfie drone" is also overly simplified. Skydio indeed started from consumer products. But it is not: "Someone just wanted to ski and take selfies, so they casually made a drone." Its true long-term technology thesis from the very beginning was: Autonomous Flight. The consumer-level follow-camera drone was just the first market validation. Skydio later clarified: The reason for starting with consumer products was that the consumer electronics supply chain iterates quickly, and autonomous follow-and-film can help them build the technological foundation needed for broader autonomous flight. This is a very important distinction. ──────────────── Twenty, Skydio R1 also did not "only sell a few units." The R1 launched in 2018 indeed belongs to: high-end autonomous selfie/follow drone. But early reports even claimed it sold out quickly. So: "The product is good but only sold a few units" this narrative is inaccurate. What really happened was: The Consumer Market ultimately did not become Skydio's best business opportunity. ──────────────── Twenty-one, this is its truly impressive pivot. Skydio did not remain in the consumer space just because: Consumer products could sell. It later discovered that the same set of: computer vision, 3D mapping, autonomous navigation, obstacle avoidance had much higher economic value in: public safety, defense, infrastructure inspection. Thus: the same core technology, but a different TAM. ──────────────── Twenty-two, in 2023, Skydio officially exited the consumer drone market. This is the key historical milestone. The company clearly stopped the consumer drone business and focused its efforts on: Enterprise. Public Safety. Government. Defense. This is an extremely classic entrepreneurial case: Do not mistake your first product for your final market. ──────────────── Twenty-three, true PMF sometimes is not about technological changes, but about "who is willing to pay the most for it." A skier: is willing to pay: $1,000. A police department: if the drone can: arrive at the shooting scene 30 seconds early, reduce unnecessary police dispatches, lower the risk to law enforcement personnel, the entire DFR network's value could be: tens of thousands, hundreds of thousands, even higher. The same technology: Customer Value is completely different. ──────────────── Twenty-four, this is called: Value Migration. The technology did not change. The value pool changed. This is also one of the most important lessons in entrepreneurship: Do not just ask "who likes my product?" Ask: "Who suffers the most from not having my product?" The latter type of customer usually has greater purchasing power. ──────────────── Twenty-five, by 2026, Skydio will no longer be a "small drone company." In April, it completed: $110M Series F. Valuation: $4.4B. The company claims: its core business has reached hundreds of millions in annual revenue, and now has about 4,000 customers, including military, public safety, and infrastructure users. Therefore: "hundreds of millions public safety empire" as a narrative is not particularly exaggerated. But it is not just Public Safety. ──────────────── Twenty-six, Skydio's real product is not actually the Drone. If it only sells aircraft: it is easy to be compared: how long does it fly? how much does it cost? camera pixels? DJI can compete very strongly. What Skydio really wants to sell is: Autonomous Drone Infrastructure. Drone Dock Autonomy Connectivity DFR Command Data Software. This is an evolution from: hardware company to: Physical Intelligence Platform. ──────────────── Twenty-seven, the Las Vegas case is strong, but needs to be accurate. As of May 2026, in the public activities between Skydio and LVMPD, Las Vegas already has: 12 dock locations. 38 dock drones. Additionally, there were previously 16 mobile DFR drones. Covering about 130 square miles. So: "nearly 40 automated pods" is close to the direction. But more accurately: 12 sites, 38 dock-based drones. ──────────────── Twenty-eight, and these drones are currently not completely "unmanned autonomous police." This boundary is very important. The Las Vegas project indeed supports: automatic launch, automatic route calculation, dock automatic management. However, LVMPD has clearly stated: Each flight is still managed by trained operators. Pathfinder is still an auxiliary autonomous navigation capability, not "AI deciding where to catch people on its own." So: Autonomous needs to be broken down into: navigation autonomy and: decision authority. They cannot be mixed together. ──────────────── 29. "Only responding to 911, not conducting random patrol surveillance" is currently also the clear stance of LVMPD. Officers responsible for flight publicly stated: No random surveillance, Flight is used for: calls for service. LVMPD also stated that each flight is recorded, audited, and linked to active incidents. This is very important. Because: Once Physical AI enters public spaces, Social License and technical capability are equally important. ──────────────── 30. The biggest bottleneck for drones in the future may not be flight capability, but public trust. Technically: They can fly. The question is: To what extent does society allow you to fly? Who can see the data? How long is it stored? Can facial recognition be used? Who can access it? Can continuous tracking be done? This will determine: public-safety Physical AI whether it upgrades from: tool to: infrastructure. ──────────────── 31. The data on 128x zoom is basically true, but it also needs explanation. The X10 VT300-Z sensor package supports: about 128x system zoom. But the actual lossless resolution is about: 16x (1080p mode). The 128x includes subsequent digital magnification. So readers should not be misled to think: it has 128x pure optical zoom. Professional expression: Maximum about 128x system zoom. ──────────────── 32. The most valuable lesson for VCs from the Skydio case is not that "the drone is great" but: Technology Reuse. Consumer: Training autonomy. Enterprise: Validating inspection. Public Safety: Validating DFR. Defense: Expanding national security. The same technology stack continuously enters: higher value, higher barriers, stronger payment capability markets. This is one of the most beautiful deep-tech scaling paths. ──────────────── 33. The second company, Range Energy, is more interesting because it discovered something that hasn't been seriously changed for 100 years: Trailers. The entire trucking industry focuses on: Tractors. Diesel engines. Electric trucks. Charging stations. Hydrogen energy. But trailers have always just been: Passive Assets. Range's question is: Why must trailers always be dead weight? ──────────────── 34. Here, I will first correct the 72% of U.S. freight. The accurate data is: In 2024, U.S. truck transportation will account for: 72.7% of domestic freight tonnage. At the same time, it will obtain about: 76.9% of transportation revenue. In 2024, U.S. trucking revenue is about: $906 billion. So: "72% of U.S. goods rely on trucks" is generally correct. But the professional statement is: 72.7% of domestic freight tonnage. ──────────────── 35. The truly clever aspect of Range's technology is not "turning trailers into electric vehicles" but: Turning originally completely passive trailers into active machines. It adds: Battery. Electric Axle. Sensors. Control Software. Regenerative Braking. Refrigeration power. Thus, trailers can assist tractors in: Accelerating. Cruising. Braking. ──────────────── 36. "A 6000-pound trailer feels weightless" is a very good demo, but it shouldn't be interpreted literally in physics. It certainly does not: Actually lose mass. What it really does is: The motor actively compensates for rolling, acceleration, and other loads based on traction and motion state. Thus, the driver/tractor feels: Effective Load decrease. So in the video: A person can push a several-thousand-pound trailer, showing: motor assistance. Not: Gravity disappearing. ──────────────── 37. "Reducing fuel consumption by 40%" should also add two words: At most. Range currently states on its official website: reduce fuel consumption by up to ~40%. Not: All fleets, all road conditions will definitely decrease by 40%. The actual results will be affected by: Route, Gradient, Load, Diesel prices, Electricity prices, Speed, Refrigeration demand impacts. ──────────────── 38. Moreover, Range has another often-overlooked value: cold chain. The trailer itself has: Up to about 300kWh onboard energy. Not only can it: Assist propulsion, But also power the refrigeration unit of the refrigerated trailer. This is very important. Because traditional refrigerated trailers may still operate separately: Diesel refrigeration units. Range can simultaneously address: Tractor Fuel + Refrigeration Fuel two cost pools. ──────────────── 39. This is a typical characteristic of excellent industrial entrepreneurship: ROI does not rely on carbon credits to tell a story. Customers may like it because of: Emission reduction. But what really determines large-scale procurement is: Payback. How much can be saved in a year? How much less maintenance? What about speed performance? How is safety? Is there no need to change routes? Range has even provided a fleet savings calculator, directly breaking down fuel, maintenance, TRU fuel, etc. into economic accounts. This is what industrial products should really sell. ──────────────── 40. The best Physical AI projects usually share a common feature: Customer saves money on Day 1. Not: The world will be better in ten years. But: Install today, This month's P&L will improve. Such products will have much lower adoption friction. ──────────────── 41. Range also has a very advanced strategy: Backward Compatibility. It does not require the 3 million heavy trucks in the U.S.: To all switch to a brand new tractor. But rather: Hook to existing equipment. Existing tractors. Existing operating routes. Existing fleets. Only: Trailer upgrades. This is a very advanced industrial product strategy: Don't replace the system. Upgrade the bottleneck. ──────────────── 42. This is similar to the PC expansion cards, local enterprise SaaS, and even the early internet. The easiest new technology to popularize is usually not: Asking customers to throw away all old assets. But rather: Layer on top. Range puts electrification capabilities onto: Trailers. This bypasses: A large fleet replacement cycle. ──────────────── 43. Moreover, Range is now advancing towards supply chain scaling. Range completed a $23.5 million financing in 2024, and subsequently deepened cooperation with global large automotive parts supplier ZF, utilizing mature components like ZF AxTrax 2 e-axle to push towards mass production. This is also the most difficult step for Physical AI from demo to company: Industrialization. It's not about whether the prototype works. It's about: Does it work with 10,000 units too? ──────────────── 44. Many hard tech companies ultimately do not fail due to technology, but due to "mass production." Prototype: 10 units. Nice. Series A. Production: 10,000 units. Suddenly discover: Suppliers are unstable. Tolerances are inconsistent. Yield is poor. Maintenance is expensive. Warranty issues explode. Working capital is insufficient. This is: The Valley of Death. So hard tech VCs must not only ask: "Can it be made?" But also ask: "Can it be made continuously 100,000 times?" ──────────────── 45. The third company is actually not "the underlying brain of industrial robots." The third company in the video is: UnitX. What it does is: AI-powered machine vision for manufacturing quality inspection. Not a general-purpose robot foundation model. Not a humanoid robot "brain." UP.Partners' description of it is also: Using AI to detect manufacturing defects. ──────────────── 46. What UnitX truly addresses is the most invisible yet extremely costly problem in industrial manufacturing: Defects. Producing 1 million parts. As long as: 0.5% errors. That’s: 5,000. If these errors are discovered only at the end of the production line: Waste has already occurred. Worse: If discovered after selling to customers. It could lead to: Recalls. Warranties. Brand damage. Safety issues. ──────────────── 47. Therefore, the value of industrial AI is not "AI looks smart." The real value is: Yield. If a factory: Produces tens of billions of dollars worth of products in a year, Improving yield: By 1 percentage point, The economic value could be enormous. This is why manufacturing AI may be easier to charge than many consumer AIs. Its ROI: Is very clear. Forty-eight, UnitX's system actually has three layers. It is not just: Uploading photos to a visual model. It has: OptiX Imaging and visual acquisition. CorteX AI defect detection / classification. DeteX Intelligent visual inspection system. It also needs to connect: PLC, MES, Production lines. This is: Physical AI Stack. ──────────────── Forty-nine, why might "imaging" be more important than the AI model? An engineer would tell you: Garbage In, Garbage Out. If: The light is unstable. Severe reflections. The angle is wrong. Motion blur. Poor camera calibration. No matter how good the foundation model: It is useless. So the moat of Physical AI is often not a single model. But rather: Sensor + Optics + Data + Model + Integration. ──────────────── Fifty, this also explains why "model commercialization" may not necessarily kill Physical AI companies. Assuming: OpenAI, Google, Meta make the vision model ten times stronger tomorrow. UnitX still needs to: Put the camera into the production line. Control lighting. Calibrate equipment. Connect PLC. Design rejection mechanisms. Conduct reliable validation for each manufacturing defect. Provide uptime. This is: Last-Mile Industrialization. Model companies may not be willing to do this. ──────────────── Fifty-one, the data that UnitX currently discloses is already not very early. The company claims: The system has accumulated over: 5.7 million hours of operation. Deployed: 100+ production lines. Served: 190+ customers. Every year, it detects product value exceeding: $6.1 billion. These are self-reported by the company, not independently audited data. But at least it shows: It has clearly moved beyond the pure laboratory stage. ──────────────── Fifty-two, so Skydio, Range, and UnitX actually represent three levels of Physical AI. Skydio Machine autonomy. Machines move by themselves. Range Machine augmentation. Making a previously dumb mechanical asset smart. UnitX Machine perception. Letting factories truly "see". Putting these three companies together is much more advanced than classifying them as "three hardware companies". ──────────────── Fifty-three, true Physical AI can be understood with a closed loop: Sense → Think → Act → Learn UnitX: Leans more towards Sense + Think. Range: Think + Act. Skydio: Covers all four layers. Future robots: Integrate all four layers. This is what is called: Embodied / Physical Intelligence. ──────────────── Fifty-four, this is also why the next batch of truly huge AI companies may not look like OpenAI. They may look like: Drone companies. Robot companies. Energy companies. Industrial automation companies. Logistics companies. Defense companies. But at the core, they are all: AI Companies. Because what truly determines product capability is: software + intelligence. ──────────────── Fifty-five, from a historical perspective, this is not the first time Silicon Valley has returned to hardware. Silicon Valley was originally: Hardware Valley. Fairchild. Intel. HP. Apple. Semiconductor. Later, in the internet era: Capital became lighter. 2000s: Web. 2010s: Cloud + SaaS. The biggest advantage of software is: Zero marginal replication cost. Thus, capital naturally preferred: bits. ──────────────── Fifty-six, however, after SaaS reached extreme success, it produced another problem: Software Abundance. Every CRM: Dozens of companies. Every Marketing Tool: Dozens of companies. Every AI wrapper: Hundreds of companies. Coding: Many companies. Enterprise AI: Many companies. So Adam's statement: "More companies than ideas." Truly expresses: Originality Scarcity. Not a problem of the number of startups. ──────────────── Fifty-seven, why might Physical AI have a deeper moat? Because it is difficult. It requires: Mechanical Engineering. Electrical. Firmware. AI. Manufacturing. Regulatory. Supply Chain. Customer Deployment. Capital. A new competitor cannot: Copy in two months. This is: Complexity Moat. ──────────────── Fifty-eight, but do not draw the wrong conclusion that "hardware is inherently better than software." Atoms are harder Also means: Atoms are more expensive. Hardware failures are very severe. Software startups: $5M May validate PMF. Some robotics companies: After $100M Still do not know if they can scale. So the biggest risk of Physical AI is also: Capital Intensity. ──────────────── Fifty-nine, the investment formula for Physical AI is completely different from SaaS. SaaS investors often look at: ARR. NRR. CAC. Gross Margin. Burn Multiple. Physical AI must also add: BOM. Manufacturing Yield. Warranty Rate. Deployment Time. Utilization. Hardware Gross Margin. Service Margin. Inventory Turns. Working Capital. CapEx. Payback Period. This is another set of financial language. ──────────────── Sixty, especially must look at Payback Period. Assuming a robot: $200K. Saves: $20K a year. Customers will not be particularly excited. Payback: 10 years. If a machine: $200K. Saves: $150K a year. Then: Less than two years. The procurement logic is completely different. So a very nice benchmark for Physical AI is: Can the machine pay for itself? ──────────────── Sixty-one, the second metric is Utilization. A robot works: 22 hours a day. Very good. Works: 2 hours a day. Asset returns may be very poor. So the biggest problem for Physical AI is not: "Can it perform this action?" But rather: "Can it reliably perform it 10,000 times a day?" ──────────────── Sixty-two, the third metric is Human Replacement versus Human Leverage. This is a very critical distinction in AI investment. Replacing a: $50K worker with a machine. The price ceiling is limited. Allowing one: Person to manage 10 devices, to manage 100, the value may be higher. What Skydio DFR is doing has a bit of this nature: In the future, a few operators can manage a larger distributed drone fleet. This is: Labor Leverage. ──────────────── Sixty-three, the fourth metric: Revenue Mix. The most attractive Physical AI companies are often not: Selling machines once. But rather: Hardware Software Subscription Service Data Maintenance. Why? Hardware revenue: One-time. Software: Ongoing. So the most attractive model may be: Hardware acquires the customer; software compounds the margin. ──────────────── Sixty-four, Skydio has a particularly good opportunity to do this. Selling X10: One-time revenue. But: DFR Command. Cloud. Dock management. Connectivity. Fleet operations. Software. Theoretically can generate more sustainable recurring economics. This will directly affect the company's future: valuation multiple. ──────────────── Sixty-five, the fifth metric: Regulatory Moat. Many entrepreneurs hate regulation. In Physical AI: Regulation may even become a moat. Drone BVLOS. Aviation FAA. Medical FDA. Autonomous driving permits. Defense procurement qualifications. If you spend five years clearing the regulatory path: Latecomers cannot replicate in a day. This is called: Regulatory Capital. ──────────────── 66. The sixth metric: Installed Base Once the factory installs: 200 sets of UnitX. Processes, Data, Employees, MES all depend on it. The replacement cost becomes increasingly high. Once Las Vegas builds dozens of Skydio docks: It is no longer just: buying drones. But in establishing: Infrastructure. The Installed Base will form: Switching Cost. ──────────────── 67. This is the most beautiful business model of Physical AI Software has: Network Effects. Physical AI can have: Installed-Base Effects. More machines deployed: ↓ More real-world data. ↓ Better models. ↓ Higher reliability. ↓ More customer deployments. ↓ More data. Thus: Physical Data Flywheel. ──────────────── 68. This type of data is very difficult to replicate Web data: Many people can access it. Real drones: Millions of flight missions. Factories: Millions of hours of defect data. Trucks: Different routes, weights, gradients, battery, braking data. These belong to: Proprietary Reality Data. As AI progresses, this type of data may become increasingly valuable. ──────────────── 69. This also explains why the next round of AI moats may shift from "internet data" to "real-world data" First-generation foundation models: trained on the internet. Next-generation robotics/physical models: need: how machines grasp. how they move. how they fly. how they fail. how they recover. Real-world data is much harder to obtain than web data. Therefore: Deployment becomes a data advantage. ──────────────── 70. This is precisely what makes UP.Partners interesting: it not only provides money but also tries to control industry entry UP.Partners has: UP.Ventures investments. UP.Labs directly collaborates with large enterprises to create startups. UP.Summit brings together industry, capital, and founders. And UP.Labs has collaborated with companies like Porsche, Alaska Airlines, J.B. Hunt to build vertical AI startups. This adds a very important capability compared to ordinary VCs: Distribution. ──────────────── 71. Why is this extremely important for Physical AI? A SaaS: sells directly online. An industrial startup: the biggest problem is often not financing. But rather: Who lets me into the factory? Who lets me test in the fleet? Who lets me connect to the production line? Who provides me with real data? Who is willing to take the risk of the first deployment? This is the value of large corporate partners. ──────────────── 72. So the real VC moat that UP.Partners can build is not "we understand robots better" but rather: Capital + Customer Access + Industrial Network. If you are an industrial founder: An ordinary VC: gives $5 million. UP.Partners: provides money. Then: helps you meet J.B. Hunt, Porsche, airlines, industrial companies. The latter type of value is completely different. ──────────────── 73. This is why VCs in the Physical AI era also need to change Software VCs can provide: Recruitment. Financing. GTM. Physical VCs also need: Manufacturing. Supply Chain. Certification. Government. Corporate Pilots. Factory Deployment. Engineering Network. It increasingly resembles: Industrial Company-Building Platform. ──────────────── 74. And Adam Grosser's engineering background creates a real edge here When a founder presents a complex machine to an ordinary VC: VC: "How big is the market?" Adam might first ask: "Why is this actuator designed this way?" This kind of founder conversation is completely different. Truly top-tier hard-tech founders can easily judge: this investor: really understands or not. This determines deal flow. ──────────────── 75. The three founder criteria you wrote—authenticity, resilience, contrarianism—are all very correct But a fourth should be added: Engineering Honesty. A truly hard-tech founder can be in a very dangerous state: Vision too strong, but does not respect physics. ──────────────── 76. The world does not care about your pitch Batteries: just have energy density. Materials: just fatigue. Aircraft: just subject to aerodynamic constraints. Factories: just have tolerances. So Physical founders must possess: Extreme Vision + Extreme Respect for Reality. Both are essential. ──────────────── 77. The fifth criterion should be: Capital Discipline. Because hard-tech startups can burn money endlessly. It is very easy to say: "Just give me another $500 million, and we can solve it." A truly good founder will constantly ask: What is the cheapest experiment? How to prove the most critical technology risk? How to get customers to pay in advance? How to leverage existing supply chains? This determines whether the company can survive. ──────────────── 78. The sixth item: Wedge Market. Range did not say: "First redo the entire American heavy truck." It cut in from: Trailer. Skydio: Consumer autonomy cut in, then into public safety. UnitX: Defect inspection cut in. Great grand visions all need: Small Commercial Wedge. Otherwise, it is just a research project. ──────────────── 79. This is the most important rule of Deep Tech entrepreneurship Vision can be huge; first invoice must be specific. You can think: Change global manufacturing. But the first customer must be willing to: pay for defect detection of this battery production line. This is when it starts to become a company. ──────────────── 80. From the perspective of capital allocation, I also would not simply believe that "Physical AI valuations are cheaper" This judgment is already starting to become outdated. Robotics, Defense technology, Drones, Fusion, AI infrastructure are all very hot today. Skydio: $4.4 billion. Anduril: $61 billion. Many humanoid robotics companies have also achieved extremely high valuations. So: Physical ≠ Cheap. Price is still important. ──────────────── 81. The correct investment judgment is not: "Software bubble, so buy hardware." But rather: Where does technology create measurable economic value before valuation prices it all in? If a robotics company: $20B valuation, revenue: $10M, it can still be very expensive. So always remember: Great Technology ≠ Great Investment. ──────────────── 82. True Physical AI investments may have a very beautiful "valuation migration opportunity" Initially: The market sees the company as: Hardware Company. Giving: low multiple. Later: It is discovered that it also possesses: software, data, recurring revenue, network effects. The valuation system may gradually migrate from: Industrial Multiple to: Technology Platform Multiple. This is a very beautiful capital opportunity. ──────────────── 83. Skydio has this potential If it ultimately only: sells drones. It is like: hardware OEM. If it becomes: an autonomous aerial infrastructure for public safety, it is much closer to: Platform. The two types of companies: even if the revenue is the same, the multiples given by the capital market may be different. ──────────────── 84. Range is the same If it is just: selling an electric trailer. Industrial equipment. If in the future: every trailer is connected, software-defined, producing telematics, battery optimization, energy management, fleet intelligence, it may gradually evolve into: Software-Defined Freight Asset. The valuation logic may also upgrade. ──────────────── Eighty-five, UnitX is also If it is just: selling cameras. Machine Vision Vendor. If it controls: factory vision data, defect intelligence, quality analytics, process optimization, it could become: Manufacturing Intelligence Layer. This is what investors are really looking for: Platform Expansion. ──────────────── Eighty-six, so when studying Physical AI, I will particularly look at one question: Does intelligence grow faster than hardware? If in the future, every time a company sells a machine: the software value increases, the data increases, the margin increases, that is a good company. If it is always just: selling one more piece of hardware, earning one more hardware margin, the ceiling is much lower. ──────────────── Eighty-seven, from the perspective of American history, this wave has another force that cannot be ignored: Reindustrialization. In the past thirty years, American tech capital has become increasingly software-oriented. Manufacturing: mass globalization. Now: US-China competition, supply chain security, national defense, semiconductors, energy security, AI data centers are all driving the US to refocus on: Industrial Capacity. This has brought a very strong era Beta to funds like UP.Partners. ──────────────── Eighty-eight, therefore Physical AI is not only a tech trend but also a geopolitical trend. Skydio is a very typical case. The US is highly focused on: non-Chinese drone supply chains. Skydio emphasizes domestic design and manufacturing as well as supply chain security, and is rapidly expanding in public safety and US defense sectors. Thus, its growth is driven by: technology, business, national security three forces. This is called: Geopolitical Tailwind. ──────────────── Eighty-nine, the strongest companies in the next decade are likely to be at the intersection of several trends: AI × Robotics × Energy × Defense × Manufacturing × Supply Chain Resilience. This is why Physical AI is so worth watching. Not because: robots are cool. But because: multiple huge capital cycles are converging simultaneously. ──────────────── Ninety, but the most dangerous part is right here. When capital discovers: Physical AI is very sexy, there will immediately appear: Physical AI washing. In the past: every company called itself AI. In the future: every motor company will say: Physical AI. Every sensor company will say: Embodied Intelligence. Every robotic arm will say: Foundation Robotics. So VCs return to Adam's question: Signal vs Noise. ──────────────── Ninety-one, how to judge whether a Physical AI company is real? I would ask five questions. First: If the word AI is removed from the Pitch Deck, would customers still be willing to buy it? If the answer is No: danger. Second: How much money does the customer save in a year? It must be calculable. Third: How many hours has the machine operated in a real environment? Demo is meaningless. Fourth: What happens when it fails? Physical systems will definitely fail. Fifth: Is the economics of 10,000 units the same as that of 10 prototypes? This is a matter of life and death. ──────────────── Ninety-two, and the biggest investment opportunity in Physical AI may not be "humanoid robots." The media loves: Humanoid. But the truly profitable things may be more boring: factory vision. power grids. warehouses. mines. trailers. agricultural machines. Inspection. Maintenance. These do not have: cool demos. But: the ROI is huge. Historically, the most profitable infrastructures are often very boring. ──────────────── Ninety-three, so I particularly like the cases of Range and UnitX. Skydio is very sexy. Drones are easy to understand. Range: modifying trailers. UnitX: checking defects. Sounds boring. But real capital allocators should ask: "How many hundreds of millions of dollars are wasted on this boring problem every year?" Instead of: "Is this demo visually appealing on Twitter?" ──────────────── Ninety-four, this may be one of the biggest cognitive arbitrages for VCs in the next decade. Consumer AI: high attention. extremely high competition. extremely high valuations. Industrial AI: huge problems. ugly products. slow sales. But: once solved, customers are very hard to switch. This may produce: Low Hype / High Economic Value. This is the combination that investors like the most. ──────────────── Ninety-five, from the perspective of billionaires and capital allocators, I think what UP.Partners is most worth learning is not "investing in hardware" but rather: going to places where capital density is lower than economic value density. If: a market has huge value, but because: it is too difficult, too slow, too complex, capital is unwilling to go, then there may be: Mispricing. This is the eternal logic of investment. ──────────────── Ninety-six, it does not conflict with Warren Buffett's thinking. Buffett looks for: long-term cash flows that others underestimate due to short-term issues. Hard-Tech VCs look for: futures industries that others underestimate due to: technical difficulties, long cycles, lack of understanding. Both are: Non-consensus + Correct. Only Non-consensus: is useless. It must ultimately be: Correct. ──────────────── Ninety-seven, from the perspective of entrepreneurs, these three companies also collectively prove one thing: Skydio: did not stubbornly stick to the consumer market. Range: did not try to completely reinvent the entire tractor. UnitX: did not attempt to manufacture general-purpose robots from the start. They all did: Constraint Selection. Found: the smallest entry point, and then expanded. This is the true strategy of entrepreneurship. ──────────────── Ninety-eight, this is also where entrepreneurs are most likely to misunderstand Physical AI. After watching the video, they say: "I want to make robots." This is meaningless. The real question should be: Which very expensive actions in the real world still have to be performed by humans because machines are not smart enough? Then ask: Are AI, sensors, robotics just making this action: Economically Automatable? This is the opportunity. ──────────────── Ninety-nine, from an investment perspective, I would compress UP.Partners' thesis into a very simple formula: Huge Physical Spend × New Intelligence × Measurable ROI × Hard-to-Copy Deployment First: the market is already huge. Second: AI can now solve problems that could not be solved in the past. Third: customers can directly calculate the economic account. Fourth: real-world deployment forms a moat. Only when all four are together, can truly huge Physical AI companies emerge. ──────────────── One hundred, finally, compressing the entire UP.Partners case into one sentence: The biggest opportunity for the last generation of software companies was to digitize information in the real world; the biggest opportunity for the next generation of Physical AI companies is to reinject intelligence from the digital world back into the real world. Skydio: gives machines in the sky eyes and autonomous navigation. Range: gives a trailer that has been almost completely passive for a century power, sensing, and software. UnitX: allows factories for the first time to automatically identify defects that only human eyes could judge at machine speed. So what UP.Partners is really investing in is not: Hardware. It is investing in: Intelligence becoming physical. And this is also what I believe is one of the most important migrations of AI in the next decade: 2022—2025: AI learns to speak. 2025—2028: AI learns to see, reason, use tools. Next: AI begins to learn to move, manufacture, transport, inspect, drive, and operate in the real world. If this round really happens, the largest value pool will not only exist in: chatbots, Coding Agents, Enterprise SaaS. But will begin to enter: tens of trillions of dollars in global industrial, energy, transportation, defense, and manufacturing assets. This is the true high-level significance of UP.Partners. It is really betting that: after software eats the world, intelligence begins to take over the machines in the world.