Stateful

Guests assess China's robotics manufacturing edge and US AI leadership amid deployment challenges

35 min Listen to the original
RoboticsChina ManufacturingAIInvestment ThesisSupply ChainAutomation

Summary

Cosmo Jiang and Cody Poh discuss their Shenzhen trip, revealing that many Chinese robots are pre-programmed rather than truly AI-driven. They argue China holds a supply chain advantage for data collection but faces profitability hurdles due to government intervention. The speakers contend that US firms lead in AI intelligence, though China may close the gap via model distillation. Value is expected to accrue at the deployment layer, where post-training data is critical. Finally, they project a long timeline for home robot adoption due to public trust issues, despite rapid AI progress.

Episode sections

  1. 0:00 – 0:26

    Testing Robot Intelligence and Pre-programming

    Cosmo Jiang describes testing a clothes-folding robot in a Shenzhen factory by throwing a shirt, revealing the robot could not handle the unstructured input. The guide admitted the robot was pre-programmed and lacked true AI generalization, a fact he was reluctant to share publicly but confirmed in private conversation.

  2. 0:26 – 4:29

    Rationale for Shenzhen Trip and Silicon Valley Assumptions

    Host Jonathan Gieg introduces guests Cosmo Jiang and Cody Poh. They explain their trip was driven by the 'Genshi Gembutsu' principle of seeing the real thing. They sought to validate the Silicon Valley assumption that hardware is impossible to compete on due to China's manufacturing dominance, questioning if US companies could onshore manufacturing.

  3. 4:29 – 10:40

    China's Supply Chain Advantage and Data Collection Lead

    Jiang argues that China's hardware advantage creates a three-to-six-month head start for Chinese companies in collecting tactile and vision data from US labs. This upstream manufacturing dominance in data gloves and headsets allows Chinese firms to access frontier lab insights earlier than competitors in Europe or the Middle East.

  4. 10:40 – 14:47

    AI Capabilities and Distillation Risks

    The speakers confirm US companies lead in AI intelligence, while Chinese robots observed were largely pre-programmed. However, Jiang predicts China could close the gap if LLMs become easier to distill, as Chinese labs excel at distillation techniques, potentially allowing them to replicate US model capabilities without equivalent hardware iteration.

  5. 14:47 – 16:39

    Value Accrual at the Deployment Layer

    The consensus is that value will accrue to the deployment layer rather than pure model companies. Tasks like folding napkins or cleaning floors require reliability and precision, not high IQ, and depend on post-training data. The bottleneck is not the model itself but the in-production data needed to refine robots for specific customer tastes.

  6. 16:39 – 20:17

    Economic Case for Robotics in Developed Markets

    The economic case for robotics is compelling in high-wage countries. A human worker costing $80k annually can be replaced by a robot working 24/7, potentially saving $240k in labor. With humanoid costs projected under $20k, leasing robots for $80k-$100k annually offers massive savings for employers, driving adoption in the US despite the need for post-training data.

  7. 20:17 – 27:33

    Near-Term Deployment Targets

    Deployment will likely start in high-turnover, dangerous, or repetitive jobs like package sorting or welding in tight spaces, rather than complex factory lines or homes due to safety and privacy concerns. Companies like Figure and Path are targeting these ergonomic challenges where labor shortages are acute.

  8. 27:33 – 30:38

    China's Manufacturing Speed and Profitability Challenges

    Shenzhen offers unmatched manufacturing speed, with a supplier able to customize and deliver a demo product within hours. However, profitability in China is extremely difficult due to intense competition. Even companies with superior technology, such as a robotic hand manufacturer with tactile sensors, compete primarily on price rather than technical differentiation.

  9. 30:38 – 31:38

    Profitability and Regulatory Environment in China

    Speaker_0 notes that extreme competition and government intervention make profitability in China difficult, citing flat performance in Chinese and Hong Kong stock indices over 15 years compared to US growth. This is attributed to a 'Chinese style capitalism' where the government steps in with price controls if companies become too profitable.

  10. 31:38 – 32:45

    Public Adoption Timeline for Humanoid Robots

    Cosmo Jiang compares current humanoid robot development to building cars when horses are still the primary transport, suggesting a 20-30 year education gap for public acceptance. He argues that while AI intelligence will improve significantly in 3-5 years, commercialization and home adoption will take much longer due to trust issues.

  11. 32:45 – 34:07

    AI Progress vs. Robotics Commercialization

    Speaker_0 contrasts the rapid progress in AI since the ChatGPT launch in late 2022 with the slower adoption of robotics. He suggests that while AI generalization is advancing quickly, the physical deployment of robots in homes may take 6+ years from the current date (2026), or potentially 4-6 years from the AI breakthrough if generalization accelerates.

  12. 34:07 – 34:33

    Podcast Conclusion and Blog Reference

    Host Jonathan Gieg concludes the episode and directs listeners to a blog post on the Pantera Capital website written by Cosmo Jiang and Cody Poh, which outlines their thesis and findings from the Shenzhen trip.

Claims worth verifying

Pantera Capital's first robotics investment was in OpenMind a couple of years ago.

Unidentified speaker (speaker_0) ·2:42

How to verify: Pantera Capital public disclosures or press releases regarding their portfolio history; SEC Form D filings for OpenMind. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Primary source

UniTree recently launched one of the largest robotics company IPOs.

Unidentified speaker (speaker_0) ·3:22

How to verify: SEC EDGAR filings for UniTree's IPO prospectus (S-1/F-1) and market cap data at the time of listing. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Primary source

Chinese companies have a three to six month head start in data collection due to manufacturing proximity to US labs.

Cosmo Jiang ·9:17

How to verify: Interviews with data collection vendors or US lab procurement records showing shipment dates of hardware from China vs. other regions. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Chinese AI is built on distillation.

Cosmo Jiang ·13:33

How to verify: Academic papers or technical blogs from Chinese AI labs detailing their model training methodologies, specifically regarding knowledge distillation. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

A human worker costs an employer $80,000 all-in (wages, insurance, turnover).

Unidentified speaker (speaker_0) ·18:55

How to verify: Bureau of Labor Statistics (BLS) data on average hourly wages and benefits for food caller workers, adjusted for turnover rates. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Tesla aims to get the cost of a humanoid down to less than $20,000.

Unidentified speaker (speaker_0) ·19:35

How to verify: Tesla Investor Day presentations or Elon Musk's public statements regarding Optimus production costs. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Figure targets package sorting as a primary use case.

Unidentified speaker (speaker_0) ·22:17

How to verify: Figure AI public case studies, press releases, or partnership announcements with logistics companies. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Path creates robot dogs for welding in tight spaces.

Unidentified speaker (speaker_0) ·22:57

How to verify: Path Robotics product documentation, website, or demo videos showing welding applications. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Shenzhen suppliers can customize and deliver a demo product within the same day.

Unidentified speaker (speaker_0) ·28:47

How to verify: Anecdotal verification through industry reports on Shenzhen supply chain logistics or direct interviews with hardware manufacturers. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

It is insanely hard to make money in China due to intense competition.

Unidentified speaker (speaker_0) ·29:33

How to verify: Financial analysis of Chinese robotics hardware companies' gross margins compared to US counterparts; industry reports on price wars in Shenzhen. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

The Chinese stock market has been flat for the last 15 years.

Unidentified speaker (speaker_0) ·30:38

How to verify: Check total return index data for CSI 300 or Shanghai Composite Index from 2026 to 2026 via Bloomberg or Yahoo Finance. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

The Hong Kong index has been flat for the last 15 years.

Unidentified speaker (speaker_0) ·30:58

How to verify: Check total return index data for Hang Seng Index from 2026 to 2026 via Bloomberg or Yahoo Finance. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

The US stock market is up many multiples over a 15-year time period.

Unidentified speaker (speaker_0) ·30:38

How to verify: Check total return index data for S&P 500 from 2026 to 2026 via Bloomberg or Yahoo Finance. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

The Chinese government creates price controls if companies make too much money.

Unidentified speaker (speaker_0) ·31:18

How to verify: Review Chinese regulatory filings and news reports regarding price control interventions in sectors like education, gaming, or tech from 2026-2026 via SEC EDGAR filings of Chinese ADRs or Reuters/Bloomberg news archives. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Primary source

Humanoid robots are currently like building cars when people ride horses, indicating a 20-30 year education gap for public acceptance.

Cosmo Jiang ·31:38

How to verify: This is a subjective analogy/claim; verify by checking public adoption surveys for robotics or autonomous vehicles from 2026-2026. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Robots are able to zero-shot simple tabletop tasks.

Cosmo Jiang ·32:18

How to verify: Review technical papers or demo videos from Chinese robotics companies released between 2026-2026 demonstrating zero-shot learning on tabletop tasks. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

The ChatGPT moment was at the end of 2022.

Unidentified speaker (UNKNOWN) ·33:00

How to verify: Check the public release date of ChatGPT (November 30, 2026) via official OpenAI announcements or major news archives. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Humans are no longer needed for math or coding.

Unidentified speaker (UNKNOWN) ·33:15

How to verify: Review industry reports on AI coding assistant adoption (e.g., GitHub Copilot usage stats) and AI performance on math benchmarks (e.g., GSM8K) from 2026-2026. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

AI labs like OpenAI and Anthropic suggest generalization is happening quickly.

Unidentified speaker (speaker_0) ·33:35

How to verify: Review press releases, blog posts, or technical reports from OpenAI and Anthropic between 2026-2026 discussing model generalization capabilities. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Pantera Capital published a blog post on robotics written by Cosmo Jiang and Cody Poh.

Jonathan Gieg ·34:07

How to verify: Search the Pantera Capital website for blog posts authored by Cosmo Jiang and Cody Poh on robotics released around November 2026. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Do your own research

  • Verify the 'pre-programmed' status of specific Chinese robot models by reviewing technical datasheets or independent lab tests.
  • Check the 'distillation' capabilities of Chinese AI labs by reviewing published papers on model compression and replication.
  • Analyze the cost-benefit analysis of robot leasing vs. human labor in specific high-wage industries using current wage data and robot pricing.
  • Review Pantera Capital's blog post by Cosmo Jiang and Cody Poh for detailed thesis notes and potential conflicts of interest.
  • Investigate government intervention policies in China's tech sector that may impact profitability, such as price controls.
  • Assess the timeline for public trust adoption by reviewing consumer sentiment surveys on humanoid robots.
  • Compare the speed of hardware iteration in Shenzhen against US supply chains using case studies of product development cycles.
Full transcriptexpand

0:00Unidentified speaker (speaker_0): We were walking through one of the factories and there's a station where one of the robots was folding clothes. A lot of people view like folding clothes as like one of the main eval benchmarks, and I decided to test it out, and so I grabbed a shirt and just threw it in front of him. As the person who was taking around was a second ago just telling us about how intelligent these robots were and how excited they were about their AI capabilities, so like he he was just like, whoa, whoa, whoa, what are you doing? That's obviously not gonna

0:23Unidentified speaker (UNKNOWN): work.

0:26Jonathan Gieg: Welcome to another. episode of Stateful, a Pantera Capital podcast. I'm Jonathan D, your host today, and I'm joined by two of my colleagues, Cosmo Jong, General Partner, and Cody Poe on the investment team. Robotics is obviously, you know, a topic that uh has come up and I feel like everyone is talking about over the last year, and it's especially accelerated, you know, over the last six to nine months. Now when it comes to robotics, um, you know, there's really two pieces, right? There's the intelligence piece, which you can think of as the brain, and then there's the embodiment, you know, which is the actual physical robot. Um, the supply chain and the manufacturing chain that is involved in the embodiment side, um, right now, I believe is largely taking place in China, you know, and you guys just came back from a robotics trip to Shenzhen where you were able to meet with some of these uh these factories that are creating um these robotics components. So tell me a little bit about, you know, why you went to Shenzhen and maybe describe your first day on the ground there.

1:26Cosmo Jiang: Just to give everyone like a bit of background of myself. I actually grew up in Hong Kong. I served like pharmaceutical medical technology company in my past life as a consultant, but I never get to like experience the pace of innovation. that I see from the media. I come in as like a robotics investor trying to understand what's happening on the ground and what enables China or like Chinese companies to move like that quickly in such a fast growing industry. On the first day, we are trying to be like super open-minded, try to understand Then uh whatever that people are telling us about China. We want to not come in into this country with like a pre-established view. We actually wanted to um build our understanding on China from ground up.

2:13Unidentified speaker (speaker_0): You know, something that I have learned as an investor, um, you know, as a young investor, really carried with me. The reality is I think there's so much importance in this concept of of of Genshi Gembutsu or you know, the real real thing, thing, real play. real place. It comes from Toyota and this idea that you can't actually know something unless you actually see the real thing in the real location. And it's a really important investment value of mine that I've carried with me, and that we we now have here at Pantera. And you know, we've we've spent so much time over the last couple of years looking at robotics companies, starting with our first investment in OpenMine OpenMind a couple years ago. of to investing in earlier this year, uh, you know, a bunch of different robotics companies through a basket of them. But like all these companies that we talked to and visited have been in the US. And it was so consistent from all of our conversations, you know, obviously, that China is where the manufacturing happens. Like China is the manufacturing center of the world, right? And undoubtedly, anyone that's creating hardware has to rely on China in some form or fashion. And so we kept hearing this like, you know, all their parts are coming from China. There's all this uh things deck in China. Obviously, China

3:22just launched one of the largest robotics company, IPO, recently with UniTree. And so it just felt like, man, we can't possibly be uh like I think we're very we're we have a lot of knowledge base from talking to all these US companies, but man, it's it's gotta be so important to go to China. and see these companies where where those companies are and understand the whole vertical uh the whole geographic stack, not just the vertical stack. You know, we we're global investors, and so we do think about businesses in different areas. And China is a place that, as an investor, you really meet it's incumbent upon you to understand not just the companies that are coming out there, but really the the cultural mores that drive why those companies. succeed or not is so important, right? That's so important to American exceptionalism, and that's very important to why Chinese companies develop the way they are. I just felt it was you know critical for us to spend time there. And and that's why that's why we flew over and went to Sun Zen.

4:17Jonathan Gieg: So what were your goals going in? Were you trying to develop you know a thesis from the ground up? Were you trying to validate a thesis you already held about China? And you know, did it did the trip ultimately accomplish either of those things?

4:29Cosmo Jiang: I think for me it's just to relearn how we should think about the world. Like robotics is an industry that the US and China are kind of intertwined in the sense that

4:41Unidentified speaker (UNKNOWN): we

4:41Cosmo Jiang: all know that manufacturing is kind of like constrained domestically here. And a lot of companies that we talk to rely on like components or specific parts shipped from China. So we're trying to understand, okay, so knowing that China plays such an important role in the supply chain today, should we invest in China more because of that? Or what kind of advantage would that give Chinese company and how would that impact our thesis investing here in the US? I think that's kind of like the the the core thing I'm trying to answer. And I think after the trip we have like a pretty good sense of of that, which we can talk about is um later on. Uh but I would say that's like the core focus of myself going into Shenzhen.

5:24Unidentified speaker (speaker_0): I think there's this like very, very Silicon Valley thesis that um that we heard from the the the quote unquote you know classic Silicon Valley venture set that like you can't do hardware because the Chinese companies are doing it extremely well. And it just felt like this assumption that a lot of people have and they kind of throw out without actually having prosecuted that assumption. Uh and and uh and you know, we kept hearing this, but then there are also US companies like Tesla, like Figure, and others who are building their own hardware and saying they can do it in the US. Um, but those companies are getting, again, not a lot of attention from the classic Silicon Valley set that believes hardware is gonna gonna be done by Chinese. And we just felt like that was I don't know. I was I want to be open minded about whether that's true or not. Like maybe maybe the Chinese hardware uh advantage can be reproduced here. Obviously, we're trying to do that with the growth in uh what Tesla's re or Tesla and SpaceX both have done for onshoring manufacturing and bring hardcore like real-world physical manufacturing back to the US.

6:24Uh and maybe there's a shot at that. And so we really want to test this thesis of is there something special about hardware that has to remain, or is something special about China, such that their hardware advantage is insurmountable? And the corollary to that is that we've seen a lot of our

6:40Unidentified speaker (UNKNOWN): uh

6:41Unidentified speaker (speaker_0): a lot of venture dollars go into just pure AI model companies, right? All Uh including in robotics. And uh, you know, there are large companies like physical intelligence and generalists Generalist and other others, and they're doing incredible work. But the reason why venture investors are so crowded into those names is because they believe they've already made the assumption that hardware's not worth doing. Maybe hardware is worth doing, all right. And and and so that's what that's what we really wanted to figure out.

7:07Jonathan Gieg: And what did you walk away

7:10Unidentified speaker (UNKNOWN): uh

7:10Jonathan Gieg: with an opinion um you know of of that specifically? More

7:13Cosmo Jiang: specifically,

7:14Unidentified speaker (UNKNOWN): is hardware. is

7:15Jonathan Gieg: worth you know pursuing

7:17Unidentified speaker (UNKNOWN): i

7:17Cosmo Jiang: think

7:18Unidentified speaker (speaker_0): we got even more certainty that the verticalization and like having a deep understanding of each piece of the stack is incredibly important there's definitely a world where like

7:28Unidentified speaker (UNKNOWN): you

7:28Unidentified speaker (speaker_0): know as we break down robotics there's you know there's hardware there's software or the model there's the um there's the uh there's the data collection that drives the training there's the actual go-to-market of getting these robots into people's hands and deploying them. And as we think about each of those layers, like there's definitely a real reason why each of those layers could succeed or not succeed, and you know, we want to remain open-minded and that we're very early. But uh we started off with this

7:56Unidentified speaker (UNKNOWN): um

7:57Unidentified speaker (speaker_0): inkling assumption that like actually verticalizing and owning each piece of the stack is so important because of the speed of innovation, uh uh and the speed of being able to iterate on things is is faster if you own every piece of the stack. I think we came away a lot more confident that that is true, but also that wow, the advantage in hardware in China is massive. And so to think that you could totally ignore China supply chain is yeah, that is

8:23Unidentified speaker (UNKNOWN): you

8:23Unidentified speaker (speaker_0): definitely cannot. So like there's some truth to that.

8:26Unidentified speaker (UNKNOWN): I think

8:26Cosmo Jiang: I asking like whether hardware is worth pursuing is might not be the right question. I think the the extension to that is like, oh, knowing that this is China's advantage, what kind of competitive edge would that give Chinese company? And my read is that hey, like a bunch of like US model companies are like experimenting, trying out different ways to collect alternative data to make it more tactile, better in forced sensing.

8:57Unidentified speaker (UNKNOWN): They

8:57Cosmo Jiang: want to do like vision capture with more cameras just to make sure that their model can like reconstruct it well with a comprehensive view. But the upstream of everything that I just mentioned comes from China, because Chinese company or Chinese supply chain manufacture manufactures all these like data glove, vision capture. headset, hardware that enables large scale data collection. If you are a Chinese company and you talk to these US labs, you basically have like a three to six month head start against everyone else in Europe, in Middle East, in in in other parts of the world, and you get a glimpse of like what are Frontier Labs actually thinking about the next generation model. And I think this actually gives like China a like massive lead to everyone else that if they manage to capitalize on

9:46Unidentified speaker (UNKNOWN): it

9:46Cosmo Jiang: properly, over time it will be like a very big advantage. So far, we haven't seen that playing out in China. Um maybe what Cosmo mentioned, like the cultural reason investor asking for like uh um return in the short run. that kind of like make them rethink their priority. But I I I I do think that this is something super underrated and if used properly, uh can reshape what we're seeing in robotics industry today.

10:18Unidentified speaker (UNKNOWN): Yeah,

10:18Jonathan Gieg: I think it's you know widely understood, um, or at least posted a lot online that uh China has a pretty significant advantage in manufacturing and the US has a significant advantage in the intelligence side. Um, after your visit, do you think that is still true? Do does the US have um a significant advantage on the intelligence side, or is China catching up um there?

10:40Unidentified speaker (speaker_0): Yeah, it's so funny. It reminds me, like, yeah, we did go in to try to understand the the AI capabilities of all these companies. You know, we we went in with this belief that we wanted to we went with a number of beliefs, some of them that we wanted to either overturn or confirm, right? And they definitely had the belief of exactly what you said that the US is far ahead on AI model capabilities and the US that China is far ahead on hardware. And we went there, I mean we saw robots doing pretty incredible things, right? If you go online and go to like the But look at like the robotics Olympics that that happened happen in China or the you know annual. Chinese New Year show with a bunch of robots. You would think that wow, these robots in China are incredible. It turns out though that these robots are being pre-programmed to do these things. And so we really understand like where where things are being pre-programmed and where things are actually like truly AI-driven. And it just reminded me, we were walking through one of the

11:33Unidentified speaker (UNKNOWN): uh

11:34Unidentified speaker (speaker_0): one of the factories, and there's a station where one of the robots was folding clothes. And I was like, oh, well, that's pretty cool. Like most robots, a lot of people view like folding clothes as like one of the main eval benchmarks uh for like how advanced their robot intelligence is. And I just thought, wow, that's actually like pretty impressive. I mean, it's sacking them in a very particular way, but that that's fine. And I decided to test it out, and so I grabbed a shirt The and just threw it in front of them. him. And uh as the as the person who was taking around was

12:03Unidentified speaker (UNKNOWN): a

12:03Unidentified speaker (speaker_0): second ago just telling us about how intelligent these robots were and how excited they were about their AI capabilities. So he was just like, Whoa, whoa, whoa, what are you doing? That's obviously not gonna work, right? As like uh in other words, like these robots obviously have no intelligence, they're only doing pre-program work. Um, and I was just I you know I tried to ask him about generalization and all that, and almost amongst a big group, like when he was answering us, he was kind. of like, Oh, well, you know, it's still early and all these things, but we believe in all these things. Uh, but it was really clear that if you even move the shirt slightly ajar from where it was placed, like the robot was just completely unable to do it. And, you know, as we walked away, y and this is part of why being there in person is really helpful. As we walk away and you kind of pull on him the aside, side, he's like, Well, yeah, I mean, you know, it's still early, there are all these things we're doing, like, you know, there's nowhere close. it's all pre-programmed. And so you get all these small uh pieces of tidbits that people are uncomfortable saying over Zoom, right? But maybe they would mutter under a breath to you in in quiet conversation. And so we were pretty uh like it was we felt that was pretty confirmatory that you know this idea that

13:10Unidentified speaker (UNKNOWN): the

13:11Unidentified speaker (speaker_0): AI in China just especially for robotics just isn't there. It's

13:14Cosmo Jiang: pretty clear that. US companies lead in terms of intelligence, but I have kind of like a prediction that hey if if if LLMs are bet getting better in like controlling robots, it went from like the system two reasoning to like system one rope motion controlling. And like what you said just now, Cosmo, that Chinese AI is basically built on distillation. Like, what if LLM gets to a point that it's so good at controlling robots and it's so easy to distill? Like it's reasonably hard to distill VLAs or role action model because you actually need the hardware to do it and there's a lot of investment involved. But what if LLM can actually dominate? That can actually shorten the path of these Chinese lab distilling state-of-the-art models. And that can kind of like close the gap. So this is kind of like an interesting force that is like that can play out in the near term, depending on how Frontier Labs are thinking about incorporating robotics capability in DLLM. I think it's still very early to say that US will can will it's clear that US is leading, but it's kind of unclear And to clear me that whether they will continue to lead because China is so good at distillation. And the development of foundation. foundational model is actually going to a place where it's becoming easier to distill.

14:31Jonathan Gieg: Interesting. So what does that mean for companies that are developing their own in-house robotics models today? Do you think they will survive into the future? Do you think everything will uh sort of converge to the cost of compute and tokens as LLMs really take over on the intelligence side? Um what are your your thoughts there? I

14:47Cosmo Jiang: think Valley ultimately accrues to the deployment layer.

14:53Unidentified speaker (UNKNOWN): I

14:53Cosmo Jiang: actually don't think that it requires true intelligence to perform tasks that can replace labor today, making beds or cleaning the floor.

15:05Unidentified speaker (UNKNOWN): You

15:05Cosmo Jiang: don't need like true generalization or like 200 IQ to do that. But what you need is like precision, accuracy and reliability. And that comes from the post-training post of training things. Well

15:15Unidentified speaker (speaker_0): I think ultimately like who who controls the customer usually captures the most value, right? Uh and the end customer that is.

15:23Unidentified speaker (UNKNOWN): Um

15:23Unidentified speaker (speaker_0): that said, you can build massive businesses servicing a lot of these these people. And so I think there can be a lot of I think it's it's probably too uh broad A of a rush to say that you know the only company companies is worth in i investing. There there's only one area where companies are worth investing. But certainly our our working thesis is exactly what Cody said, which is that because these production sites where robots are working are so customized, and that there is an element of taste uh whenever there some a task is being done, that getting robots into people's hands and being able to get real post-training data at the site is probably the most important thing to becoming um to becoming a sticky customer to becoming a sticky vendor to one of your customers. So

16:11Unidentified speaker (UNKNOWN): it

16:11Unidentified speaker (speaker_0): really important is getting robots into people's hands so that they can actually start using them, so that robots can use that data to post-train, which is refine on the actual task specific task tasks based on that specific customer's taste. And so getting that flywheel going of just getting actual re-sur. real data, actual real product data in production, we think is probably the biggest bottleneck right now. And we hear from everyone that data is the biggest bottleneck, but really specifically, it's like in production data that's the biggest bottleneck.

16:39Unidentified speaker (UNKNOWN): You

16:39Jonathan Gieg: know, if I'm a business, right, and I'm purchasing a lot of robots to come, you know, work in my restaurant, work in my factory, work in my hotel. Um,

16:46Unidentified speaker (UNKNOWN): I

16:46Jonathan Gieg: think it's reasonable to assume that a lot of them would expect these robots to have some degree of functional. functionality from day one, from minute you know zero of deploying. Um, if each robot has to take in post-training data from specific job sites they're that they're deployed to, isn't that a blocker to scaling and a blocker to businesses wanting to adopt robotics?

17:07Cosmo Jiang: I think it comes to the point where like, how will adopting robotics technology create cost saving for Discopp? these corporates? And that's like a lot of that's basically the reason why US corporates are willing to experiment with um robots in their production line or even in like uh for for tasks that is like less significant today, just to try out the things and generate the data required to make

17:33Unidentified speaker (UNKNOWN): it

17:33Cosmo Jiang: um deployment ready. If you're running a restaurant today, you probably need to have like two percent for um eight hours uh a day each of them costs like eighty to ninety thousand dollars not to mention there's like insurance the four oh one K and stuff like that but if you were to like rent a robot from one of the human art or verticalized companies today it costs like far less than that. So I think it's completely understandable that corporates are willing to experiment experiment with that and see if it can drive cost saving over time. What we're seeing by interacting with a lot of companies here in the US, a lot of the robots are like performing jobs in a fairly autonomous way that requires like very little human intervention. And that's like It's the proof point of their capability. And on the other side of the world, like um the labor costs in China or in emerging markets is like significantly low. So it's understandable why even for the leading Chinese human arm manufacturer, most of them robot is are are still being sold to like to labs, to universities,

18:43to to perform like Kung Fu Food performance at at at at some shows, because the cost saving that's going to bring to corporates is not super clear to them. And that's why they do not want to experiment with them.

18:55Unidentified speaker (speaker_0): Yeah, I think just to be clear, like the math is extremely compelling, and I just want to make sure we laid out the math for everyone. A an example, food caller worker in a factory, maybe maybe costs an employer eighty thousand dollars all in. That's the cost of pay wages, but also insurance, you know, payroll tax, uh turnover, which is very high for these jobs, and then need to continue. continually hire. You know, a human works eight hours a day. A robot, in theory, can work 24 hours a day, seven days a week. Uh, and so it can take over three human shifts. Uh, and so you know, it can replace $240,000 of waiver. And we're getting to a point where many humanoids, uh, you know, the Tesla has has been out there saying that they think they can get the cost of a humanoid down to less than $20,000. Um, you know, we've heard similar things from other people. And for non-humanoid form factors, which are even uh less expensive, meaningfully lower than $20,000 of uh for the bomb or the bill of materials, which means if you can go to these factories and say, hey, I can sell you this robot, I can lease you this robot

19:55for let's just call it 80k a year, 100k a year, and replace three shifts of humans for 240k. a year, that's a massive cost savings for the uh for the employer or the customer, but also huge margin for the robotics company. And so that's why this is the the economic use case is so compelling and so obvious, especially in developed countries.

20:17Jonathan Gieg: Great. So where do you think the most deployment will start for robotics? Do you think you know it'll be in hospitality and sort of like these repetitive skills? Do you think maybe it will move towards the the more dangerous tasks, maybe like underwater um welding or or those sorts of um jobs where where you know danger is a factor. I'm I'm curious where you think the most deployment will happen in the earliest time.

20:42Cosmo Jiang: I believe that most of the near-term deployment is still going to primarily replace simple tasks, um maybe like making bats. in in in in in hotels. Um and the reason for that is there is additional work to be done if you were to deploy a robot in a factory or even homes. Let's say factories, right? Like we're asking we we're talking about like ninety nine percent safety accuracy because if something goes wrong in a factory line someone could get hurt, someone could die. Uh if you were to deploy in homes, I'm not sure if like everyone that purchased a robot is happy to have their household footage is being leaked to whatever server that is running the inference. The extension to that is like the privacy has to be done in a proper way.

21:31Unidentified speaker (UNKNOWN): Uh

21:32Cosmo Jiang: it requires time to build trust, even if you can afford that that that robot. robots.

21:37Unidentified speaker (speaker_0): The way that I frame this is where is there the most acute need? Uh that's one vector, and the other vector is how capable are companies at being able to address that. I think from the most acute need, it is areas where there's high turnover, uh, where it's really hard to find labor, uh, and maybe it's uh because it's a task that maybe people don't enjoy doing. And so those tend to be highly repetitive, maybe brain-numbing. brain numbing, mind-numbing tasks. Those tend to be very dangerous or physically demanding tasks. And so you know when we talk to someone like Figure, uh, and you know, this is public and so it's out there, Figure talks about how one of our tasks is literally just sorting packages. Sorting packages is very, very simple, but ergonomically challenging for a human to do. Like imagine you're pushing boxes eight hours a day nonstop. It's first of all, you get blisters on your hands, and so that's why they all wear gloves. Uh, but also your shoulders hurt and all this thing, and it's um and so you face really high turnover in that kind of job. Well, getting a robot to do that makes a ton of sense. Or another example, or the other example is you know, dangerous jobs in tight spaces. You know, one of the companies out there, this company called Path, they're creating these uh robot dogs that effectively go into tight spaces in like in in factories or s or uh uh where uh in order to weld

22:57together pieces of steel. And the reason why that's important is because it it's kind of dangerous to send a human out there into these very hot areas. It's a boiler room, or it's you know, uh or it's an underwater place that's nearly impossible to get to. But you can send a robot there. You know, people have talked about all these defense use cases, I mean drones. are the most obvious robot that we don't talk about. And drones are, you know, effectively the battlefields in Ukraine are all drones, right? Uh accompanied by some humans. And so these highly repetitive, dangerous, high turnover tasks where it's hard to find humans, is where we're gonna see the fastest product market fit where customers are saying, yes, we really want that. And then from a capability perspective, um, it also tends to match reasonably well. Like, where are there things where we have a ton of training data, first of all, so that we can actually go and do that thing? Where there can be a lot of training data uh can be the first places where robot robotics companies feel like they're very comfortable doing that.

23:52Jonathan Gieg: And let's talk about the economics, you know, of robot labor. I know Cosmo mentioned um it may cost eighty $80,000, thousand dollars right per year to replace a blue-collar worker. Um that might not be the case in countries where wages are are lower, right? There it might cost 20,000 or 25,000. So I'm curious how you see robot deployments into countries where wages are naturally lower and there's a surplus of labor willing to do those jobs. Will robot deployments remain um a a function of higher wage countries, or will you see proliferation around the world?

24:26Unidentified speaker (UNKNOWN): It

24:26Unidentified speaker (speaker_0): reminds me of a point. of that technology, very you can think about technology from a high level, and in theory, this is how it should play out. But the reality is that technology plays out in a very stepwise function, and and that path determin at least very path deterministic. Um and the I think

24:45Unidentified speaker (UNKNOWN): all

24:45Unidentified speaker (speaker_0): that is to say is your question around how uh what you're alluding to, which is right, is that it's more obvious that the economic case makes a lot more sense here in the US than it does in like Bangladesh. Um, and why we're gonna see is why we're gonna see robots first here in the US and not in Bangladesh or India or or something like that, because if the economic case is more useful. But because that is true, that means the companies that are able to get that, what I just described as the most important. unlock, which is this training data, this actual in live production data, uh, and being with the customers, building with them, it may be limited to the US companies or these companies in the in the in these developed markets. Cause they're the they're the ones that are there to service these companies. And then they're able to be able to get all this data that helps them inform how to make their next robot more efficient, how to make their next model more efficient. And maybe that actually means that you have to be able to get to that point of building a robot, developing a robot, and getting into the customers' hands. But actually, despite China being the best at manufacturing and despite them being able to manufacture as many robots as they want, if they can't get their robots in their hands of customers, but the US companies can, the US companies are going to have the key bottleneck, which is the training data, to really succeed and grow. And I think that is a very likely

26:05uh you know path for how this might turn out. And

26:08Jonathan Gieg: Cody, we've taken a look at you know quite a few data vendors, right, in the data space. Obviously, it's been huge for training robotics models. Uh, you know, Frontier labs are buying tons of data right now. What has changed over the last you know three to six months uh on the data set? Uh, where do you see that market going? We

26:23Cosmo Jiang: have full hop on calls with a bunch of data collection companies. And the one question that I'm trying to ask them is like how are you positioning yourself when the best model architecture has not been determined by the market? How are these companies going to position themselves? And interestingly a lot of them told us that oh the type of uh data or footages that they are requiring from us is still the same. And the reason for that is like they are still trying to experiment and understand like what works and what does not. So it speaks to how early we are still in the entire like pre-training phase. But my understanding is if you are run, if if you are a data vendor, if you you want to build like a successful business, being able to understand the needs, being able to verticalize the hardware that helps the Frontier Labs to capture the data, be able to build a really robust evaluation framework, um, and consistently uh refine the type of data that you guys are collecting so that it can actually drive real model incremental improvement is is is very important. So

27:33Jonathan Gieg: I'm curious what your guys' biggest wow moment was in in China, specifically with with manufacturing. and these different um, you know, plant owners that you're meeting with. Was there a a single point where you said, wow, China really does have a significant advantage uh maybe in speed or quality of manufacturing components?

27:49Unidentified speaker (UNKNOWN): Oh

27:50Unidentified speaker (speaker_0): my god, yeah, there are so many. I mean, one very specific example is, you know, people talk about this fast iteration cycle and how it's so important, uh, and how that you know, because Shenzhen is this manufacturing hub, you have all these suppliers that are existing, coexisting, very close to each other, and how the US lacks something like that, and why that's a big disadvantage. We really saw that come into play when we were meeting with, I remember it was around noon, we were meeting with a company that builds um motion capture headsets or envisioned visual capture headsets, right? The idea is to collect egocentric data through these headsets um and gloves. And um and he would he had just rushed in, he was a little late to our meeting, and he was explaining how he was actually just at one of his customers earlier that morning who had who was testing out their product and mentioned that they wanted to make some changes and that he had on his way texted like his had texted his factory to make those changes to create a demo product. And he expected, he anticipated that he had to rush out at our meeting right on time so that he could then get the actual demo product the new product uh that just been customized based on the customer specifications that morning and delivered to the customer that afternoon and I was just like wow that is that's speed right like the idea that you can ask for something to change and actually get the thing in your

29:07hands that same day like you just don't have that here in the US right um uh and so I thought that was that was incredible.

29:15Jonathan Gieg: With such a density of you know manufacturers, um, there there's obviously a lot of choices for for companies to go to. So I'm I'm curious how do these manufacturers make money? And in general, how do businesses in China make money? It seems like there's a lot of undercutting and and um pushing prices lower. So uh yeah, I'm curious your thoughts there, Cosmo.

29:33Unidentified speaker (speaker_0): That was also one of our biggest learnings about China. And you know, I you kind of see this as an investor, but you don't really understand until you get there, which is that it is insanely hard to make money in China because everyone is so goddamn competitive and there's so many people. I that sounds kind of obvious in the context of yeah, there's a 1.4 billion people in China. So obviously there's a lot of people. The implications of that are so profound. found Taking it back to robotics, um, we met with there was this hand manufacturer that we thought was doing. this incredible thing with their hands I don't want to give away the name of the company but uh you know they were talking about how their hands had all this all this range of dexterity had tactile sensors um that they had that they had these incredible technology they put into the actuators

30:18Unidentified speaker (UNKNOWN): and

30:18Unidentified speaker (speaker_0): I was just like wow that sounds like you know real IP and as I asked them you know okay well when you talk to your customers like how do you even think the sales processes go? How do you compete against, how do you compete against other vendors? And they're like, Well, of course, it's just about cost. I was like, Oh, wait, what about like that last hour that we spent talking about your technology? Like, okay, so it turns out like it's all about cost. And so the the it's it's sort of practical and like very comp hyper competitive. And so it just made us realize that wow, it's insanely hard to make money. No wonder the Chinese stock market has been flat for the last 15 years, right? Like, isn't that incredible? The US stock market is up many multiples over a 15-year time period. The Chinese index is flat. The Hong Kong index is flat, right? And it's because this, I think it comes down to this Chinese style capitalism, which is that we they want to encourage everyone to be entrepreneurial and take risks and create companies. If you get too big, there's first of all, there's gonna be a lot of competitors that can be could you. pee down. And then if everyone makes too much money, the Chinese government then feels like they step in and kind of create these price controls. And so uh it's just like, man, it's insanely hard to make money in China.

31:26Jonathan Gieg: I'm curious to your personal thoughts on you know, how many years out are we from having, you know, humanoid robotic robots in our own homes, doing things like washing the dishes, folding clothes, making the bed, etc. et cetera.

31:38Cosmo Jiang: One interesting conversation that we we have with like a Chinese entrepreneur that is building a human art company which focuses on like elderly care use cases. Basically told us human arts today are like building cars when like someone can basically like ride horses to get to other places. Like there there there is at least like a twenty, thirty year education gap for the general public to fully embrace the idea of a new technology and there's even a longer period of time for them to be willing to pay for that. The intelligence part is like getting there. Like we we we see it in real person. Robots are able to like zero shot some simple tabletop task and if scaling law applies, like their intelligence is going to improve. like significantly in the coming three to five years. But I think the commercialization it's going to take a much longer. And be able and in in order for people to like actually trust a robot and buy one robot and put it at their home, it's going to take much longer.

32:45Unidentified speaker (UNKNOWN): Yeah,

32:45Unidentified speaker (speaker_0): my view is that things have progressed we're in an age where things can progress a lot faster than we might think. There are all these reasons that Cody said why there are real obstacles, especially in the home, to getting robots there.

33:00Unidentified speaker (UNKNOWN): But

33:00Unidentified speaker (speaker_0): man, it it's like truly incredible what's been done to why call of work over the last three short year or r y I guess coming on four years, right? So the chat ChatGPT GPT moment was at the end of twenty 2022. twenty two. And we're sitting here in twenty 2026, twenty six and

33:15Unidentified speaker (UNKNOWN): I

33:15Unidentified speaker (speaker_0): don't know if humans are needed for math anymore. I don't know if humans are needed for coding anymore. I don't know that we're gonna see this ChatGBT moment for robotics, although many would argue that there are these signs of generalization we're seeing coming out of large AI labs like a PI, like a generalist, or from OpenAir and Netropic themselves, suggest that we are getting to generalization very quickly. And if that's the case, then man, it feels like maybe it's only four years out until this is really, really uh this has penetrated a lot of industrial use cases. And if that's the case, I don't know, five, six years out, just like I feel like personal agents are probably a couple years out of being used in a large way.

33:54Unidentified speaker (UNKNOWN): So

33:54Unidentified speaker (speaker_0): six years out from ChatGPT 2022, maybe six years out from now we're actually see start to see humans or start to see robots in the home. If not humanoids, which may not be the ultimate form factor, really proliferate in a big way.

34:07Jonathan Gieg: Well, guys, it was a lot of fun. Uh thank you for sharing your thoughts. Um if you guys haven't checked out our new blog post on robotics, you can check that out on our website written by Cosmo and Cody, where they outline their their trip to Shenzhen and a little bit about what our thesis looks like moving forward. So thank you for joining me guys. It was it was a lot of fun again. and um we'll see you in the next episode.

This page was produced automatically: the audio was transcribed with nvidia/parakeet-tdt-0.6b-v3+sortformer-v2.1 and interpreted by Qwen3.6-35B-A3B-UD-IQ3_S.gguf. Transcription and interpretation can contain errors; always confirm against the original episode. Nothing here is investment advice.
All insights