Building the SoftBank of Robotics | Andrew Kang

Summary

• Kang’s core thesis is that humanoid robotics today is where AI sat in 2022 and crypto sat in 2014–15 — a pre-ChatGPT-moment industry the general investment community still misprices. His unit math: at $50K per robot, $1B of revenue is only 20,000 units and $10B is 200,000 — trivial next to phones or cars made “in the hundreds of millions or billions” — and he expects industry revenue in the “tens of billions somewhere between the next two and four years.” His edge claim: “the more difficult an industry is to understand, the more alpha you can generate.”
• The TAM answer to humanoid skeptics is the general-purpose-device argument: “all human physical labor is your TAM,” and scale economics beat special-purpose machines the way smartphones beat cameras and GPUs beat ASICs. Proof by existence: BMW, Volkswagen and BYD still employ hundreds of thousands to million-plus workers despite decades of industrial automation, because most factory work changes too often to justify fixed automation below multi-million-unit SKUs.
• Robo Strategy (NASDAQ: BOT) exists because Tom Lee raised $18B in 8–9 months “for an asset that went down 50%… in a thesis that wasn’t that good” — so an asset class actually going up could raise “maybe an order of magnitude more.” Kang’s ambition is explicit: use public capital markets to reach beyond a16z-scale ($15B last fund series) and “become a new challenger to SoftBank,” applying a similar private-to-public valuation concept to the one TransDigm and Constellation Software use (buying at 3–8x earnings, marking at 15–40x).
• Post-ChatGPT power laws changed the venture math: you can now invest at multi-billion valuations and still expect “a 50x, 100x, 500x investing at a $2 billion valuation” — which is what makes deploying billions viable at growth stage. Portfolio construction is a barbell: 6–10 companies at ~70% of the fund, 20–30% in pre-seed-to-Series-B optionality, and “let winners ride.”
• He concentrates in vertically integrated players because embodiment-specific, on-robot data is “absolutely necessary” — a model trained in the wrong body is like learning basketball at 7 feet then playing at 5'10": “all wobbly.” The “Scale AI of robotics” cohort is the Buzz Lightyear meme — a thousand identical companies — and a “short-term cash cow” whose revenues might plummet “two, three, four, five years down the line” once robots self-improve; “it almost feels like trading as opposed to investing.”
• On US vs China he rejects the who-wins framing: both build great products, but American companies win for shareholders on 20–50%+ margins and stronger capital markets — “you can sell 10 times less product but be a bigger business.” Plus bipartisan bills restricting Chinese robots in the US are “very high probability,” potentially ring-fencing the biggest market (30–60% of any major product): “America’s the place you want to invest.”
• His philosophy is pricing risk/reward, not contrarianism: citing Coatue data that bigger public companies 10x more often than small ones, he backs the strongest founders at higher valuations — “winners win” — versus the Character AI/Inflection AI trap of buying “grittier and cheaper.” One admitted change of mind: more bearish on model-only companies, more bullish on open source — “Kimi getting as good as Claude on some benchmarks” will repeat in robotics — which he reads as bullish for hardware companies with secured supply chains, not for restricted Chinese robots.

Deep dive

1. From “cash is king” to all-in on humanoids

• José traces the arc from their chats: in June 2022 Kang said “Crypto’s done for a while. Cash is king”; by early 2024 he’d put $19M into Figure “before he’d even spoken to the founder and before the big markup.” Kang’s account of the pivot: robotics in 2022–23 looked like “AI 2022 or crypto back in 2014, 2015” — underappreciated, at an inflection in both acceptance and adoption.
• Why not AI labs, data centers, or drones? Not that it was too late — but “a lot of other people playing in that field already” meant no clear edge, whereas humanoids were “almost at a pre-ChatGPT phase.” The industry had never produced venture-scale winners — the biggest recent exit was Boston Dynamics to SoftBank around 2020 for roughly $1B — and its multidisciplinary difficulty (EE, ME, robot learning, deployments) was the attraction: “the more difficult an industry is to understand, the more alpha you can generate.”

2. The timing question: how he underwrites a pre-ChatGPT moment

• José’s pushback is the graveyard of early: dot-com in the late 90s, autonomous vehicles “talked about for 20 years” — you could have waited for AI’s ChatGPT moment, bought the labs, and captured most of the IRR. Kang concedes you can’t call the month or year, but claims you can call the range: tens of billions of dollars of industry revenue “somewhere between the next two and four years.”
• The quantity math is the load-bearing step: “You sell one robot for $50,000. How many robots does it take to get to a billion dollars of revenue? 20,000 robots. It’s actually not that much.” $10B is 200,000 units — versus phones, cars, and PCs made “in the hundreds of millions or billions,” which “speaks to how underappreciated the scale that we could reach” is.

3. RT-1/RT-2 was the sign of life; Boston Dynamics backflips never were

• The 2022 evidence: Google DeepMind’s RT-1/RT-2 robot foundation models. The specimen he keeps returning to — a robot told to pick up a ball and place it on a picture of Taylor Swift, from among photos of celebrities, did it: picked from a random position, distinguished Swift from Kanye, dropped it correctly. “If it can pick and place and identify items… that’s what a lot of work is in factories today” — enough by itself for billions to tens of billions of revenue.
• His answer to “why not the Boston Dynamics demos ten years ago”: those “weren’t really signs of intelligence… they were demonstrations of the capability of the hardware itself.” A backflipping robot can’t identify an object and act on it under changed circumstances — the actual requirement.
• Mechanically, the first generation used vision-language models as a backbone, similar to the role of ChatGPT or Claude for language, with an action layer translating instructions into gripper joint angles, motor torques, and forces. Pre-LLM robotics was “for the most part very deterministic, which is why things would break”: move the object an inch, make the floor slippery, or hand it an empty versus full water bottle and the calibration collapses.

4. TAM: general purpose eats special purpose, and José’s confession

• José’s mea culpa, worth keeping: he mid-curved Figure by going looking at industrial robotics, finding deterministic robots already everywhere, and concluding the humanoid market was unclear — “which sounds incredibly dumb… in hindsight it definitely feels dumb.”
• Kang’s reframe: there was no humanoid market not because there was no demand but because the things didn’t work. If a $50K robot worked “24/7, 365” versus an eight-hour, five-day human, “every company would buy it.” If humanoids match human economics, “all human physical labor is your TAM.”
• The scale-economics argument: general-purpose devices dominate the way smartphones absorbed cameras, calculators and alarm clocks, and GPUs dominate despite ASICs and FPGAs — flexibility across billions of units beats per-task efficiency at 10,000 units a year. The naysayers “don’t appreciate fully” why BMW, Volkswagen and BYD still employ hundreds of thousands to million-plus workers: factory floors change quarterly or yearly, and “there are so many things in the world that exist at not multi-million-unit SKUs” that are still huge markets in aggregate.
• No single killer app needed: mixed-case palletizing, machine tending, shelf stocking across “hundreds of thousands of these types of stores… or millions,” the home (“also going to be tens of billions”), even space — “it’s going to make sense for robots to industrialize Mars or the moon first.”

5. Why VCs missed Figure — and why perception is flipping now

• Post-Figure-livestream (a month or two ago), Kang reports “a huge influx of VC interest” across his portfolio companies for follow-ons and new financings: “people just have to see it to believe it.” His team says they are among the most knowledgeable investors many founders and VCs meet — many conversations “are just education.”
• On skepticism of Brett as “a showman”: Kang says some of the skepticism may come from that. José responds with the example of VCs being skeptical of SpaceX and Elon Musk’s promotion of Tesla’s progress, arguing that showmanship is not disqualifying if investors can distinguish embellishment from reality.

6. The Tom Lee revelation: why Robo Strategy went public

• The origin story, in substance: raising tens of billions privately requires an a16z-grade track record (“they raised what, 15 billion in their last series of funds”), but “you see Tom Lee raising 18 billion in the matter of eight to nine months… just him carrying all that weight himself, for an asset that went down 50%… in a thesis that wasn’t that good.” Kang’s inversion: with an asset class “actually going to go up substantially… you could potentially raise maybe an order of magnitude more.”
• The enabling shift: pre-ChatGPT, growth-stage meant 20–30% IRRs; post-ChatGPT power laws mean “you could still expect a 50x, 100x, 500x investing at a $2 billion valuation” — so billions can be deployed with seed-like outcomes, as happened with OpenAI. First-principles check on why nobody had done it: “is there a good reason why besides the fact that it’s hard? Not really.” Stated ambition: “establish a new capital markets model and become a new challenger to SoftBank.”

7. Premium to NAV is not premium to true value

• José notes BOT trades at what he calls a “3x premium” to NAV; Kang’s correction is the section’s spine: NAV marks the prior round, possibly 6–18 months stale — “a point in time in the past” — while public prices reflect a vastly larger participant set. Real-estate analogy: you don’t call a house’s value its transaction price from two years ago; you re-appraise against today’s buyers, and “if I’m going out to a larger group of buyers, more likely than not, I’m going to find a different price or a higher price.”
• The institutional precedent he’s replicating: TransDigm and Constellation Software acquire private companies at 3–8x earnings while the parent trades at 15–40x — spend $100M, book cash flows worth $400M, “$300 million almost immediate accretion” — scaled to tens of billions. “We’re applying a very similar concept to venture capital.”

8. Supply chain: every component can bottleneck, few capture value

• The nuance: robots contain PCBs, wires, batteries, torque sensors, actuators — and as with the iPhone, huge-volume suppliers can run “50 bps of margin, 1% margin” on commoditized parts, while a few GPU-style supply-chain players take “almost as much margin or more than NVIDIA” via genuine moats. “Made in America” actuators identical to everyone else’s is not a long-term moat; rethought motors with real IP, or unique magnet/battery sourcing, can be.
• On José’s memory-market analogy (demand inflects, supply lags, even commodities rip): Kang agrees “literally every single component might become a bottleneck” if robots go from hundreds of thousands to hundreds of millions or billions per year — margins jumping 1% to 10% is a 10x for incumbents. But the test is durability: “is that going to last two years, five years, ten years? …If it’s not clear it lasts more than five, I don’t think that’s an investment that’s interesting.”

9. Form factor: the welding specimen and the humanoid bear case

• His favorite application-specific bet: Path Robotics in welding — ~80% manual today, tens of billions paid to welders annually, a persistent labor shortage despite decent pay. The moat is failure data: capturing failed welds “means you have to scrap the pieces of metal… that is really expensive,” it lives siloed inside shipbuilders and plane manufacturers, and Path has collected “hundreds of thousands—maybe more—of welds,” though Kang says he cannot give the exact number.
• José puts the full bear case: humanoids are overkill for 95% of industrial tasks, an 80kg robot falling is a safety non-starter, and “the world is built for humans” is backwards — factories have always adapted to the machine. Kang’s rebuttal: “the goalposts are changing for the critics” — a wheeled mobile manipulator “is essentially a humanoid with wheels instead of legs,” a general-purpose robot doing hundreds of tasks. He concedes wheels make sense industrially (balance under load), but the home — “an even larger market or just as big” — needs legs for stairs and errands: “there’s no argument against having a humanoid within your home with legs.”

10. Vertical integration wins; the “Scale AI of robotics” is a trade, not an investment

• Why he concentrates in vertically integrated players, including Figure and Dyna Robotics: more shots on goal (“if the model layer becomes monetized, at least you have the hardware layer”), and embodiment-specific on-robot data is “absolutely necessary.” The analogy as told: “if I were learning how to play basketball in a 7-foot-tall body and then immediately transformed into a 5'10” basketball player, I’d probably be all wobbly and uncoordinated." Relying on third-party robots means overheating, actuator failures, uncertain joint-torque data, supply shortages — and the model-only companies now trying to build hardware face iteration cycles the integrated players already put in the time to develop.
• On the data-layer pattern-match (Mercor/Scale-shaped bets): “there’s just a thousand of these companies trying to be the Scale AI of robotics” — the Buzz Lightyear meme. Pressed on market size, on how many hours of egocentric video are needed (millions? billions?), and on what happens if continual learning and self-play become dominant, “they don’t really have good answers.” His view: this is a “short-term cash cow” that might produce a short-term spike in revenue and valuations, then start plummeting “maybe two, three, four, five years down the line.”
• José’s counter: Scale and Mercor were still massive outcomes — Mercor at “a couple billion dollars of ARR,” every round looking expensive then growing through it. Kang’s reply is about exit, not revenue: Scale’s acquisition was more around talent and know-how than the business itself, and the risk is no liquidity when growth stalls. “It almost feels like trading as opposed to investing… I want to invest in businesses that are durable over decades.”

11. China: both countries win the technology; America wins the equity

• An honest update: he’s “been impressed” by Robbyant, an Ant Group subsidiary, releasing open-source models at the frontier, among the first built on a video-model backbone rather than a VLM — where he says the field has transitioned and where robot foundation models are headed. But he rejects the country-versus-country framing: as with EVs and phones, Chinese firms can make more impressive products while “the biggest companies in the world are the American companies — the Apples, Teslas,” on stronger capital markets and 20–50%+ margins. “You can sell 10 times less product but be a bigger business,” versus Chinese firms “operating to the bone” — great for consumers, “not great from a shareholder perspective.”
• The policy kicker: a few bills in Congress to restrict Chinese robots — national security (robots “that can see and hear everything” in American homes) plus domestic-industry promotion — have a lot of bipartisan support and are “very high probability” to pass, potentially protecting a market that’s “usually 30 to 60% of any major product.” He looked at Chinese companies, including Galvi AI, but did not invest: it was not compelling enough to offset dual-use risk for a US-listed vehicle.

12. Philosophy: price risk/reward, back winners, verify everything

• On contrarianism: “I don’t really think about contrarian versus not contrarian” — only mispriced risk/reward. He cites Coatue data showing bigger public companies have a higher probability of 10x-ing than small ones — “there’s a reason why these companies are bigger” — so within a contrarian field he pays up for the strongest teams. The cautionary pair: three years ago you could buy Anthropic/OpenAI at tens of billions, or go “grittier and cheaper” into Character AI or Inflection AI — “most people probably don’t even remember these names.”
• On crypto habits: early-stage crypto “instills really bad habits” because less-than-best founders could get rewarded by rising tides and liquidity; venture-scale wins require founders who recruit and inspire — validated by reference-checking down to the engineering level, not just executives (“this person worked at a brand-name company, but there were all these issues… he wasn’t delivering”).
• Process: gut-instinct yeses are “very rare” outside a Bezos/Musk-grade founder (“Prometheus… okay”). For foundation-model companies, test the model yourself — move the objects, change the table height. For a novel-actuator bet, he leaned on Scott, “a foremost expert on actuators… 40 years in the space, built and sold two robotics companies,” plus validation from a leading actuator manufacturer and humanoid customers.
• Sizing: “let winners ride,” subject to regulatory limits if a company approaches 80% of the portfolio; the barbell is 6–10 companies at ~70%, with 20–30% in pre-seed-to-Series-B to track execution and earn follow-on allocation. What he got wrong: gone “a little bit more bearish on model-only companies” because open source will commoditize closed models — “Kimi getting as good as Claude on some benchmarks” — which, with Chinese robots restricted in America, he reads as “bullish hardware companies… and companies that can do their own manufacturing.” One or two companies were not on his radar because they were in stealth; José says it is really one, and names will be withheld “until we’ve built positions.”