Transcriptly

What's Actually Holding Humanoid Robots Back Isn't the AI | Andrew Kang, RoboStrategy

EO21:10Added Aug 31, 2026

Andrew Kang, CEO of RoboStrategy and an early investor in Figure AI, breaks down why he's betting humanoid robots become a tens-of-trillions-dollar market, a...

Watch on YouTube →
Contributed by 刘嘉琪

Transcript

Transcript format

  1. Intro

  2. 00:01The amazing thing about the Figure AI
  3. 00:02live stream was it showed that this is
  4. 00:05real. It's not a video where they took a
  5. 00:07hundred attempts at [music] doing a task
  6. 00:09and they showed the best one. This was a
  7. 00:11live stream that went on for 8 or 10
  8. 00:13hours and it ended up going on for 8
  9. 00:15days. What was funny was that the human
  10. 00:17actually won. It won by a little bit,
  11. 00:19but you know, at the end of the day, the
  12. 00:20intern that was doing the challenge, his
  13. 00:22hands were blistered. He was not having
  14. 00:24a lot of fun. It was exhausting. It's
  15. 00:26not something that he'd probably want to
  16. 00:27do again. I actually think it's closer
  17. 00:29to something like two [music] to three
  18. 00:30years where humanoid intelligence, robot
  19. 00:32intelligence gets good enough to do most
  20. 00:34of the tasks that we need on on a daily
  21. 00:36basis. [music] This is a technological
  22. 00:38revolution that is different because it
  23. 00:40turns physical labor into in a product
  24. 00:42that almost anybody can access. Hey
  25. 00:45guys, I'm Andrew. I'm the CEO of Robo
  26. 00:47Strategy. Robo Strategy is one of the
  27. 00:49first publicly listed venture funds on
  28. 00:51NASDAQ and we are the only publicly
  29. 00:54listed venture fund that is exclusively
  30. 00:56focused on investing in robotics and
  31. 00:58physical AI. We're invested in quite a
  32. 01:00few robotics companies, Frager AI,
  33. 01:02Uptronic, Dino Robotics. We also have
  34. 01:04Centerbots in the portfolio. They build
  35. 01:06industrial arms, cobots, and also
  36. 01:08companies like Path Robotics that focus
  37. 01:10on specific tasks like welding.

  38. The Bet: All in on Figure AI and the full-stack future

  39. 01:16One of the largest investments that we
  40. 01:17had ever made into the company Figure
  41. 01:19AI, it wasn't a consensus investment
  42. 01:21because [music]
  43. 01:22everyone that we had asked, the other
  44. 01:23venture investors that were more
  45. 01:25familiar with investing in inferentiary
  46. 01:27technology, they didn't really believe
  47. 01:29that humanoid robotics was going to work
  48. 01:31[music] anytime soon or they perceived
  49. 01:32there was going to be a lot of risks.
  50. 01:34They saw that humanoids or uh companies
  51. 01:37building [music] the robotic space had
  52. 01:38never produced big venture scale
  53. 01:40outcomes as opposed to understanding the
  54. 01:44[music] context that things were
  55. 01:45changing and that technology for
  56. 01:47robotics was going to be accelerating
  57. 01:49and moving at a different pace than it
  58. 01:50[music] was before. And so that's why we
  59. 01:52really decided at that point to pivot,
  60. 01:54you know, our entire company into
  61. 01:56focusing on investing in robotics. It's
  62. 01:58funny because when I went and [music]
  63. 02:00invested in Figer for the first time, I
  64. 02:02had never actually been to their
  65. 02:03facility. I watched every single video I
  66. 02:05could of Brett of Figure um and you know
  67. 02:09all the work that they had done for
  68. 02:10previous companies as well. Just kind
  69. 02:12[music] of doing our research on the
  70. 02:13team, the founder, it was quite clear
  71. 02:16that this is [music] one of the few
  72. 02:17teams that were able to do it. They had
  73. 02:19the background in hardware engineering.
  74. 02:21They had the background in robot
  75. 02:22learning. They had the background in all
  76. 02:24these really niche fields like hand
  77. 02:26engineering or robot controls and
  78. 02:28[music] fleet management. Looking at all
  79. 02:30the competitors and all the other
  80. 02:31players in the space, it was pretty
  81. 02:32clear that they were one of the top
  82. 02:34teams to be able to to accomplish a
  83. 02:35task. We're not the type of investors to
  84. 02:37be very dogmatic. When we believe one
  85. 02:40thing, never change our minds. High
  86. 02:42conviction, strong beliefs loosely held.
  87. 02:45Um, but we're always trying to
  88. 02:47re-evaluate our beliefs of the world.
  89. 02:50And [music] if there's important
  90. 02:51information that comes up to lead us to
  91. 02:53believe we're wrong, then we're happy to
  92. 02:55change our minds. And it's important for
  93. 02:57us to always track the pace of
  94. 02:59development across [music] all robotics
  95. 03:00companies, not just the ones that we're
  96. 03:02invested in, so that we can understand
  97. 03:04how are the different companies stacking
  98. 03:05up against each other. How is the field
  99. 03:07developing across all the different
  100. 03:09characteristics that we look for
  101. 03:10robotics companies, right? How are
  102. 03:12different players scaling up their robot
  103. 03:14fleet? um how are they conducting robot
  104. 03:16learning research? You know, what are
  105. 03:18they doing on the hardware development
  106. 03:19side? And from all those kind of points
  107. 03:22of view, we still believe Figure is one
  108. 03:24of the top companies. Really, it's it's
  109. 03:26them in and Tesla Optimus at the top.
  110. 03:29We're really excited about the
  111. 03:30vertically integrated robotics
  112. 03:32companies. These are the companies that
  113. 03:34we're investing in the most. And these
  114. 03:36are companies that are not just building
  115. 03:37their own robot intelligence, but
  116. 03:39they're building the hardware. uh
  117. 03:41they're doing the deployments and
  118. 03:42they're also uh scaling up their own
  119. 03:44manufacturing capabilities. When we
  120. 03:46think about why these [music] companies
  121. 03:48exist in the first place is because when
  122. 03:51you're training the robots, it also
  123. 03:52makes sense to be able to, you know,
  124. 03:54have built the robot hardware yourself
  125. 03:55[music]
  126. 03:56so that they're co-optimized for each
  127. 03:58other. Maybe a robot that has better
  128. 04:00torque sensing within its joints uh is
  129. 04:03able to be better modeled in simulation
  130. 04:05[music]
  131. 04:06or you can build a model that
  132. 04:08incorporates that type of data that
  133. 04:10you're capturing. So there's a lot of
  134. 04:11advantages in [music] building these
  135. 04:13systems in parallel with each other um
  136. 04:15because it makes the training more
  137. 04:16efficient, research more efficient. At
  138. 04:18the end of the day, the robots are going
  139. 04:20to be more performant [music] as well.
  140. 04:21One of the key data pieces that are
  141. 04:24required for robot learning development
  142. 04:26is the actual robot data itself. Robots
  143. 04:29that are either doing a specific task
  144. 04:31running a model or robots that are
  145. 04:33controlled using teleaoperation to
  146. 04:35collect the data. One of the ways that
  147. 04:37you can think of this is if you were
  148. 04:39transformed into the body of somebody
  149. 04:41that was seven foot tall, you probably
  150. 04:43would have be a little bit awkward in
  151. 04:45interacting with the world around you as
  152. 04:47opposed to you continuing to interact
  153. 04:49with the world around you in your
  154. 04:51current body in your current physical
  155. 04:52form because that's the body that you're
  156. 04:54used to. And so having that embodiment
  157. 04:57specific data is going to create more
  158. 05:00effective models. and to be able to
  159. 05:02collect a lot of embodiment specific
  160. 05:03data, [music]
  161. 05:04you're also going to need need a lot of
  162. 05:05robots. That is one of the bottlenecks
  163. 05:07that the industry is currently working
  164. 05:08through right now is if you're trying to
  165. 05:10buy 100 robots or a thousand robots.
  166. 05:13That's going to be pretty tough. You
  167. 05:14[music] can't get that in a day. You
  168. 05:16need to make those orders ahead of time
  169. 05:17and it's going to take time to produce
  170. 05:19those robots. And so, if I have my own
  171. 05:22manufacturing facility, I can earmark
  172. 05:24all those robots just for the sole
  173. 05:25purpose of collecting data myself. And
  174. 05:28that's what companies like Figure
  175. 05:29[music] are doing, what companies like
  176. 05:30Tesla Optimist are doing, Electronic as
  177. 05:33well. So I'm not going to have a
  178. 05:34bottleneck because I don't have to worry
  179. 05:36about say a robot company supplier in
  180. 05:38China where I'm getting my [music]
  181. 05:40robots from just not having enough
  182. 05:42available because demand has
  183. 05:44skyrocketed. And that's what you've seen
  184. 05:45with GPUs or you know other components
  185. 05:48of the supply [music] chain is that
  186. 05:49things demand for a lot of these items
  187. 05:51are scaling up really really quickly and
  188. 05:53it's hard for these supply chain vendors
  189. 05:55to be able to produce them enough to
  190. 05:57fill that demand.

  191. The Future: How far humanoids actually go

  192. 06:03I think the market for humanoid robotics
  193. 06:04is going to be in the tens of trillions.
  194. 06:06It's a crazy number. I think the way
  195. 06:08that you can get there is you can take
  196. 06:10two views. You can take the top down
  197. 06:12view, which is you just look at all of
  198. 06:14the market for physical labor in the
  199. 06:16world, and that's a $50 trillion market,
  200. 06:18but it's it's a little bit hard to
  201. 06:19conceptualize. And so the way that we
  202. 06:21thought about it was imagine one
  203. 06:23humanoid. It might be sold or it might
  204. 06:25be leased for $50,000. That's that's a
  205. 06:27pretty good price because a laborer in
  206. 06:29the US or physical work in the US, you
  207. 06:32have to pay maybe $50,000 a year when
  208. 06:34you're considering all of the benefits
  209. 06:36and and all- in costs or sometimes more
  210. 06:38than that. And then you take that
  211. 06:39$50,000 and you multiply it by h 100,000
  212. 06:42just as a starting point. That number is
  213. 06:44already $5 billion. Company that's
  214. 06:46making $5 billion a year is is a pretty
  215. 06:48sizable company. But then right you just
  216. 06:50scale it but up by 10 [music] and you
  217. 06:52say what if I have a company that sells
  218. 06:54a million humanoids per year. It's $50
  219. 06:56billion. We make billions of cell phones
  220. 06:59per year. We [music] make hundreds of
  221. 07:00millions of cars and PCs. And so I think
  222. 07:03we're probably going to make a lot more
  223. 07:04humanoids. And so you can really clearly
  224. 07:06see that there's a [music] there's a
  225. 07:08trajectory for this industry for
  226. 07:10humanoid robots to get to trillions of
  227. 07:11dollars of revenue. And that would imply
  228. 07:14tens of trillions of market cap. And
  229. 07:16that's almost an underestimate because
  230. 07:18when we start making labor more abundant
  231. 07:21uh more affordable then it expands the
  232. 07:24market as well. We can start sending
  233. 07:25robots to space. We can start sending
  234. 07:27robots to build more data centers.
  235. 07:28Right? That is kind of a key constraint
  236. 07:30for the data center buildout right now.
  237. 07:32It's not the things that go into making
  238. 07:34them. It's the the labor. It's the
  239. 07:36people that are actually putting things
  240. 07:37together, doing the plumbing,
  241. 07:38electricity. I actually think it's
  242. 07:40closer to something like 2 to 3 years
  243. 07:42where humanoid intelligence, robot
  244. 07:44intelligence gets good enough to do most
  245. 07:46of the tasks that we need on on a daily
  246. 07:48basis. So, I think you could almost
  247. 07:49characterize this new wave of robotics
  248. 07:52as almost the fourth industrial
  249. 07:53revolution. This wave of robotics and
  250. 07:55AI. We've created machines that [music]
  251. 07:57allow us to produce many different
  252. 07:59things and to make the everyday life
  253. 08:01easier. But this one is really different
  254. 08:03because this is the first time that
  255. 08:05we've been able to create machines in
  256. 08:07intelligence that can really do anything
  257. 08:09a human can do. And that opens the door
  258. 08:11for a lot of different things that
  259. 08:13weren't possible before. If labor gets
  260. 08:16as [music] cheap as say $2 an hour or it
  261. 08:20just becomes a product that we can buy.
  262. 08:21So every single person in the world they
  263. 08:23can have a personal assistant like
  264. 08:25everyone has their own iPhone. People
  265. 08:27can also buy robots or rent robots to
  266. 08:30maybe even produce things or to build
  267. 08:32companies that previously maybe they
  268. 08:34couldn't afford or maybe they couldn't
  269. 08:36find the right people to do. This is a
  270. 08:38technological revolution that is
  271. 08:40different because it turns labor
  272. 08:42physical labor [music] into a product
  273. 08:44that almost anybody can access. AI
  274. 08:46research has really been accelerating.
  275. 08:48When you think about research, right, AI
  276. 08:51development, it's not something that is
  277. 08:52on a on a slope that is completely
  278. 08:54[music] flat. It's something that
  279. 08:55changes and it feeds back on itself
  280. 08:57because the better AI models get. The
  281. 09:00more of AI research can be automated,
  282. 09:03the faster it can be done. Loops that
  283. 09:04were previously required a lot of humans
  284. 09:06can now be running right 24/7 365. And
  285. 09:10they're also able to process a lot of
  286. 09:12information a lot faster. And so a lot
  287. 09:14of that uh what you consider efficiency
  288. 09:16gains is also going to be applied to
  289. 09:19robot AI research. And that can exist
  290. 09:21across multiple dimensions, right? It
  291. 09:23helps with the actual speeding up of the
  292. 09:25research. But there's also a lot of
  293. 09:27innovation and learnings from AI
  294. 09:30research that can be applied for
  295. 09:32physical AI research. Learnings in how
  296. 09:34to best do data annotation
  297. 09:36infrastructure around collecting data
  298. 09:38and annotating data. Learnings and
  299. 09:40innovations on how to structure
  300. 09:42mid-training on how to do reinforcement
  301. 09:44learning. A lot of the same concepts
  302. 09:46from LLMs can also be applied to [music]
  303. 09:49physical AI models. And so that's why I
  304. 09:51think the amount of time for these
  305. 09:53models to get really good is is probably
  306. 09:55a lot faster than people think. But at
  307. 09:57the same time, the models are going to
  308. 09:59get really good, but that doesn't mean
  309. 10:00we're going to have robots [music] doing
  310. 10:01all of that work in the next 2 to 3
  311. 10:03years. Because even though the
  312. 10:04intelligence can get there, [music]
  313. 10:06we're still going to have a bottleneck
  314. 10:07with manufacturing. I can spin up a
  315. 10:09million instances of a chatbot
  316. 10:11instantly, but I can't do that for
  317. 10:13robots. I can't produce them out of thin
  318. 10:14air. [music] And so we're going to need
  319. 10:16to scale up all the factories. We're
  320. 10:18going to have to scale up the supply
  321. 10:20chain for all the components that go
  322. 10:21into a robot and that's going to take
  323. 10:23some additional time.

  324. The Shift: Why Open Source wins

  325. 10:29One of my views is that open source
  326. 10:31models are going to get really good. 2 3
  327. 10:33years ago, open source models were
  328. 10:36probably less than a few percentage of
  329. 10:38all tokens that were produced. Nowadays,
  330. 10:40open- source models produce something
  331. 10:42like 25 30% maybe even more of all
  332. 10:46tokens that are produced. They're
  333. 10:48getting really good and they're also
  334. 10:50saturating benchmarks. And so the gap
  335. 10:52between open source and frontier models,
  336. 10:54it used to be around 2 years. That was a
  337. 10:56few years ago. Now it looks something
  338. 10:58more like 6 months. We're going to get
  339. 11:00to a point where the open source models
  340. 11:03start to saturate the benchmarks. Even
  341. 11:05though there might be a gap between open
  342. 11:06source and Frontier, that gap may not
  343. 11:09matter for a lot of tasks in the world.
  344. 11:11Because if I'm doing a simple task like
  345. 11:14for example restocking shelves or uh you
  346. 11:17know assembling a computer mouse I don't
  347. 11:20need a really high level intelligence to
  348. 11:21do that. I don't need an Einstein
  349. 11:23[music]
  350. 11:24to be able to do these tasks. And so as
  351. 11:26long as these open source models get to
  352. 11:28that level which I believe they will the
  353. 11:31model layer were almost commoditized for
  354. 11:33physical [music] AI. We're not going to
  355. 11:34be there yet but I think that's
  356. 11:36somewhere something that we're going to
  357. 11:37get to in somewhere maybe the next 3 to
  358. 11:395 years. And so I think at that point
  359. 11:42intelligence [music] it becomes really
  360. 11:44cheap. What I consider you know the most
  361. 11:46valuable companies or the most important
  362. 11:48companies are probably going to be the
  363. 11:49ones that are doing deployments. [music]
  364. 11:51They're they're producing the hardware
  365. 11:53uh or you know they're innovating on new
  366. 11:56designs or components to make these
  367. 11:57robots even better. Nvidia is also a
  368. 11:59very big player in open source model
  369. 12:01development. NVIDIA if you look at them
  370. 12:03uh they're producing open source models
  371. 12:04for just general um you know LLM
  372. 12:07software engineering. Neatron. [music]
  373. 12:09They're really climbing the benchmarks.
  374. 12:10They're producing open source models for
  375. 12:12autonomous vehicles and they're also
  376. 12:14producing open source models for
  377. 12:15physical AI and and robot intelligence.
  378. 12:17And you know, some of the best
  379. 12:18researchers in the field are yes,
  380. 12:21they're across some of these closed
  381. 12:22source labs, but they also are at
  382. 12:24companies like Nvidia. I think everyone
  383. 12:26needs to keep in mind that for Nvidia,
  384. 12:29they're one of the most powerful
  385. 12:30companies in the AI space. They have a
  386. 12:32lot of resources. They have a lot of
  387. 12:34really smart people. It is an almost an
  388. 12:36existential threat for closed source
  389. 12:39models to win because as you saw with
  390. 12:42Anthropic starting to train on Google
  391. 12:45TPUs, if companies decide to optimize
  392. 12:48for and train on other hardware, Nvidia
  393. 12:51starts to lose their business. It
  394. 12:52becomes a bit of a threat to them. And
  395. 12:54so that's why they're putting so much
  396. 12:55effort into developing their own open-
  397. 12:57source models like Neatron, like the
  398. 13:00autonomous vehicle models, like you
  399. 13:02know, all the different physical AI
  400. 13:04models that they're developing, Brute,
  401. 13:06Cosmos, Dream Zero, etc. That is
  402. 13:09something that I I feel like can't be
  403. 13:11understated because it's if you're
  404. 13:12building just, you know, physical AI
  405. 13:14models, you have to think about I'm
  406. 13:16competing with one of the best AI
  407. 13:17companies in in the world.

  408. US vs China: Why it's not a race

  409. 13:24I think some people like to frame this
  410. 13:26as US versus China. I think both
  411. 13:28industries are going to be massive in
  412. 13:30the future and I think they're both
  413. 13:31independently going to build really
  414. 13:33great hardware and and robot
  415. 13:35intelligence industries are going to
  416. 13:36develop a little bit independently in
  417. 13:38the sense that the robots that are sold
  418. 13:41and they're used in America are probably
  419. 13:43going to come from American companies
  420. 13:44and the robots that are bought and used
  421. 13:46in China, they're going to come from
  422. 13:47Chinese companies. the world is kind of
  423. 13:49coming to a place where a lot of
  424. 13:51countries they're interested in
  425. 13:53independence. They want to produce
  426. 13:54things in their own country. They don't
  427. 13:56want to be dependent on another country.
  428. 13:58They want to make sure that on their own
  429. 14:00they can survive and they can thrive.
  430. 14:02And so there's a lot of interest right
  431. 14:03now in the governments from both China
  432. 14:06[music] and America to really accelerate
  433. 14:08the development of robotics in in those
  434. 14:10individual countries. Some of them uh
  435. 14:12like China they've invested many
  436. 14:15billions of dollars either directly or
  437. 14:17indirectly through government funds in
  438. 14:19municipalities and [music] in the US
  439. 14:22that hasn't exactly happened yet but I
  440. 14:24leave believe we're going to get to
  441. 14:25there in [music] the future. The US has
  442. 14:27already shown that they're interested in
  443. 14:29funding domestic companies. They've
  444. 14:31funded and provided financing to rare
  445. 14:34earths processing companies directly
  446. 14:36invested in semiconductor companies like
  447. 14:38Intel. And I think it's pretty clear
  448. 14:40that there's a similar amount of support
  449. 14:41that's going to come to the domestic
  450. 14:43robotics industry in America as well. It
  451. 14:45is true that the US is somewhat ahead on
  452. 14:48the physical intelligence models. At the
  453. 14:50same time, there are some really great
  454. 14:52research groups in in China. Some that
  455. 14:54are associated with Alibaba, for
  456. 14:56example, that are building robot models
  457. 14:58that are pretty close to the frontier.
  458. 15:00They have really smart researchers
  459. 15:02there. And there's also really smart
  460. 15:03researchers in America as well.
  461. 15:05Eventually, both countries are going to
  462. 15:07get there. probably independently, but
  463. 15:09also they're going to collaborate in
  464. 15:11doing so because there's a lot of open-
  465. 15:13source research that's published,
  466. 15:14research that helps both countries.
  467. 15:17[music] And so, I wouldn't really think
  468. 15:18about it as as a race or, you know,
  469. 15:20one's a little bit ahead and one's a
  470. 15:22little bit behind. I I think it's really
  471. 15:24kind of short term because I think at
  472. 15:26the end of the day, in 5 years from now,
  473. 15:2810 years from now, both countries are
  474. 15:29going to be able to get there
  475. 15:31themselves.

  476. Where to build now: The white space of robotics

  477. 15:37in terms of where people might want to
  478. 15:39build. I think there's so much white
  479. 15:40space because there are hardware
  480. 15:42platforms that exist. It makes the
  481. 15:44development a lot easier for someone
  482. 15:46that wants to build for a specific
  483. 15:47application. And so I think you can
  484. 15:49really think about robots as kind of
  485. 15:51like the smartphone, Apple, right? They
  486. 15:54make the iPhone, but there's this whole
  487. 15:56developer community that exists outside
  488. 15:58of people that are just building
  489. 15:59applications. people that were building
  490. 16:01applications for time management, for
  491. 16:05taking notes, etc. That can also happen
  492. 16:07for robotics where maybe I want to build
  493. 16:10uh you know robot applications to teach
  494. 16:13robots or the skills on how to cook
  495. 16:16really well or maybe how to do elder
  496. 16:18care or maybe I want to build a robot
  497. 16:20application to help uh increase uh you
  498. 16:23know the efficiency of certain [music]
  499. 16:25farming standards or uh agriculture
  500. 16:28techniques. I mean you can really think
  501. 16:29of anything where physical labor is
  502. 16:31involved as a potential robot
  503. 16:33application that can be built that that
  504. 16:35is you know such a large white space.
  505. 16:38A company that we haven't invested in
  506. 16:40but we're watching quite closely is is
  507. 16:42Unatree and also you know other Chinese
  508. 16:44companies. A lot of these companies,
  509. 16:46what they do is they haven't they don't
  510. 16:48take the same approach as the US
  511. 16:49companies or a lot of the US companies,
  512. 16:51they wait until they have the product
  513. 16:52that's perfect that they're ready to
  514. 16:55basically sell into, you know, the home
  515. 16:57or factory environment and and it works
  516. 16:59absolutely perfectly. The approach that
  517. 17:01the Chinese companies are taking is
  518. 17:03they're releasing their hardware, their
  519. 17:06robots as more of a platform for
  520. 17:08research [music]
  521. 17:09or entertainment for people to build on.
  522. 17:11It's not necessarily a case where I can
  523. 17:13buy a Unistry robot and then it it'll
  524. 17:16immediately be able to do everything I
  525. 17:17wanted to do. But it's also a pretty
  526. 17:20interesting business or commercial
  527. 17:22strategy because now that the robots are
  528. 17:25out in the world, you get a little bit
  529. 17:27of of a of a developer mode. You get a
  530. 17:29little bit of a deployment mode because
  531. 17:31people then become comfortable with
  532. 17:33using those Unity robots, right? There
  533. 17:35could be developer tools. There could be
  534. 17:37data collection platforms that are
  535. 17:39specific to the Unitary robots. If
  536. 17:41people are doing a lot of robot-based
  537. 17:43data collection, that might be now
  538. 17:45unitry specific. And so if you're
  539. 17:48building models that use a lot of this
  540. 17:51unitry [music] specific data, those
  541. 17:53models might run better on Unitry robots
  542. 17:54as opposed to other robots. And so I
  543. 17:56think that's a pretty interesting
  544. 17:58strategy that we're not seeing [music]
  545. 17:59too many companies in the US take.
  546. 18:01There's one company in our portfolio
  547. 18:02called Dexmate that is selling robots
  548. 18:04and you can consider them using a
  549. 18:06similar strategy. But that I think is a
  550. 18:10pretty interesting approach that can
  551. 18:12result in a lot of other maybe
  552. 18:15downstream effects because [music] now
  553. 18:17that everyone has access to these
  554. 18:19robots, other people can build
  555. 18:20applications on them. [music] And so
  556. 18:22those applications don't necessarily
  557. 18:24need to come from the the company that's
  558. 18:26building the robots. It can come from
  559. 18:27outside researchers or other startups
  560. 18:30that just want to focus on the AI side
  561. 18:32of things as opposed to building the
  562. 18:35hardware and figuring out the
  563. 18:37manufacturing component themselves. I
  564. 18:39actually think the industry is really
  565. 18:40early. So I don't think anybody's missed
  566. 18:42anything yet because you look at the
  567. 18:44robots today and they're getting a lot
  568. 18:46better, but they're nowhere near if for
  569. 18:48example Opus 4.8 8 or chat GPT or any of
  570. 18:52the LLM models, right, where they're
  571. 18:54actually doing a lot of the work that
  572. 18:56humans would do and they're being used
  573. 18:58across almost every company in in the
  574. 19:00world today. We're not there for
  575. 19:02robotics and that's what makes it really
  576. 19:04exciting time as well because there is
  577. 19:05still still a lot of opportunity for
  578. 19:07people to join really exciting companies
  579. 19:09that have a great growth trajectory or
  580. 19:12start making investments in the space
  581. 19:14themselves. I I think the first thing to
  582. 19:15do is is really just start doing more
  583. 19:18research, talking to friends that might
  584. 19:19be working in the industry. If you
  585. 19:21really want to, right, be involved in
  586. 19:22some way, either by joining a company,
  587. 19:25starting your own company, or investing.
  588. 19:28It's pretty underappreciated how hard
  589. 19:30building a robotics company actually is.
  590. 19:33I'm seeing online these days a lot of
  591. 19:36people from different industries saying,
  592. 19:37"Hey, look, I'm going to go out and
  593. 19:39start a robotics company." We're really
  594. 19:41excited about the industry and we think
  595. 19:43there's going to be a lot of great
  596. 19:44companies that come out of this, a lot
  597. 19:45of great technology that comes out of
  598. 19:46it. But it is also really hard the
  599. 19:48amount of kind of uh knowledge that you
  600. 19:50need to kind of accumulate over over
  601. 19:52many many years, amount of experience
  602. 19:53you need to have, right? Like this is
  603. 19:55not building a software company. To be
  604. 19:58able to understand uh you know all the
  605. 20:01components needed to build a successful
  606. 20:02humanoid company or robotics company
  607. 20:04from you know mechanical design to
  608. 20:06electrical engineering, high rate
  609. 20:08manufacturing, how to actually deploy
  610. 20:10the robots in the real world. It's a
  611. 20:12little bit maybe uh underappreciated.
  612. 20:14There's going to be a lot of maybe
  613. 20:16investment that goes into the space.
  614. 20:17There's going to be a lot of startups
  615. 20:18that go out. But I would maybe caution
  616. 20:20people to um just appreciate a little
  617. 20:23bit more how how difficult it is. And
  618. 20:25you're you're going to need a lot of
  619. 20:26real experts. People that have years,
  620. 20:29decades of experience in the space,
  621. 20:30people that have had a significant
  622. 20:32amount of experience working at other
  623. 20:34real, you know, manufacturing or
  624. 20:36robotics environments before to be able
  625. 20:38to really build a successful company.