CHAPTERS
- 0:01 – 0:31
Figure AI’s live demo proves humanoids are real (and closer than people think)
The conversation opens with reactions to Figure AI’s long-form livestream, emphasizing it wasn’t cherry-picked footage. The demo highlighted both current capability and the physical toll on humans, reinforcing the case for automation and a near-term step-change in robot usefulness.
- •Live, multi-hour/multi-day stream suggests real capability vs curated highlights
- •Human competitor narrowly “won,” but the work was exhausting and unsustainable
- •Claim: humanoid/robot intelligence could be good enough for many daily tasks in ~2–3 years
- •Framing: this wave makes physical labor feel like an accessible “product”
- 0:31 – 1:02
RoboStrategy’s robotics-only investing thesis and portfolio
Andrew Kang introduces RoboStrategy and explains its singular focus on robotics and physical AI. He positions the fund as an early, concentrated bet across humanoids and task-specific industrial automation.
- •RoboStrategy: publicly listed venture fund focused exclusively on robotics/physical AI
- •Portfolio examples: Figure AI, Apptronik, Standard Bots, Path Robotics, others
- •Investment lens spans humanoids, industrial arms/cobots, and single-task automation
- 1:02 – 2:03
Why betting on Figure AI was non-consensus—and why they pivoted anyway
Kang recounts how investing in humanoids (specifically Figure) was widely viewed as too risky due to weak historical venture outcomes in robotics. RoboStrategy interpreted the moment differently, believing the pace of enabling tech had changed and warranted a full pivot.
- •Other investors doubted humanoids would work “anytime soon” and saw outsized risk
- •Robotics historically lacked venture-scale exits, shaping skepticism
- •RoboStrategy believed acceleration in robotics tech changed the context
- •Decision: pivot firm strategy to focus on robotics investing
- 2:03 – 3:33
Evaluating Figure: team quality, rare skill coverage, and competitive positioning
He explains why Figure stood out even before visiting the facility—based on founder history, execution evidence in videos, and coverage of hard-to-hire robotics specialties. The conclusion: Figure sits at the top tier alongside Tesla Optimus.
- •Due diligence relied heavily on public execution signals and founder/team track record
- •Figure’s advantage: hardware engineering + robot learning + controls + hands + fleet management
- •Robotics requires many niche disciplines; few teams cover them well
- •Current top-of-market view: Figure and Tesla Optimus lead
- 3:33 – 4:34
Why full-stack, vertically integrated humanoid companies have structural advantages
Kang argues the winners will co-optimize hardware, intelligence, deployment, and manufacturing rather than stitching together suppliers. Vertical integration improves training efficiency and real-world performance because the robot’s body and data pipeline are designed together.
- •Thesis: “vertically integrated robotics” is the most compelling model
- •Hardware/software co-optimization improves simulation fidelity and learning efficiency
- •Better sensing (e.g., torque) can translate into better modeling and control
- •Owning deployment + manufacturing compounds iteration speed and performance
- 4:34 – 6:05
The real bottleneck: embodiment data and robot manufacturing scale
Robot learning needs large volumes of embodiment-specific data, which requires fleets of physical robots. Even if models improve quickly, the industry can’t instantly scale robots like software instances; supply chain and factory throughput become gating factors.
- •Embodiment-specific data is essential; models learn best from the exact body they control
- •Collecting enough data requires many robots (teleop and autonomous runs)
- •Fleet availability is constrained—ordering 100–1,000 robots takes time
- •Manufacturing capacity and supply chain scale become the limiting step
- 6:05 – 8:36
Market size and the ‘labor-as-a-product’ economic unlock
He lays out why humanoids could become a tens-of-trillions market by tying pricing to human labor costs and scaling unit volumes. Cheaper, abundant labor expands what society can build—potentially unlocking new industries and accelerating infrastructure like data centers.
- •Top-down framing: global physical labor ~ $50T market
- •Bottom-up: ~$50k per robot (sale/lease) matches annual cost of many workers
- •Scaling to 100k units (~$5B) and 1M units (~$50B) shows plausible revenue trajectories
- •Abundant labor expands demand (space work, data center build-outs, new businesses)
- 8:36 – 10:07
AI acceleration helps robotics—but intelligence isn’t the only constraint
Kang connects LLM progress to faster physical-AI progress via automated research loops and transferable techniques (data, training, RL). However, he reiterates that physical deployment lags intelligence due to manufacturing realities.
- •AI research accelerates via feedback loops: better models automate more research
- •Robotics benefits from LLM-era advances: data annotation, infrastructure, RL, mid-training
- •Prediction: models get good faster than most expect
- •Key caveat: robots can’t be instantiated instantly—factories and components must scale
- 10:07 – 11:38
Open source closes the gap and commoditizes the model layer
He predicts open source models will become “good enough” for many real-world robotic tasks, shrinking the advantage of frontier closed models. As intelligence cheapens, differentiation shifts toward hardware, deployment, and manufacturing excellence.
- •Open source share of usage has risen dramatically; quality gap shrinking (~2 years → ~6 months)
- •Many robotic tasks don’t need frontier-level intelligence to be valuable
- •Expectation: model layer commoditizes for physical AI in ~3–5 years
- •Future value concentrates in deployments, hardware design, and scaling production
- 11:38 – 13:09
NVIDIA’s open-source push and the strategic implications for robotics builders
NVIDIA is framed as a major force advancing open models across domains, including physical AI. Kang argues this is partially defensive—ensuring AI workloads remain on NVIDIA hardware—meaning robotics companies may be competing with an exceptionally resourced player.
- •NVIDIA is releasing strong open models across LLMs, AV, and robot intelligence
- •Named efforts include Nemotron and multiple physical-AI initiatives
- •Strategic driver: protect GPU demand as some labs shift training to alternatives (e.g., TPUs)
- •Implication: physical-AI builders must account for NVIDIA as a formidable competitor
- 13:09 – 15:40
Why ‘US vs China’ is the wrong framing: parallel ecosystems and shared research
Kang argues robotics will develop in largely separate national ecosystems due to security and industrial-policy priorities. While the US may lead in some model work, China has strong labs too, and open research creates cross-pollination—making “the race” a short-term narrative.
- •Robots used in the US likely come from US companies; China similarly favors domestic suppliers
- •Governments seek independence and resilience, accelerating local robotics capability
- •China has invested heavily; US support is likely to grow (analogous to chips/rare earths)
- •Open research and timelines imply both countries converge over 5–10 years
- 15:40 – 16:40
Where to build now: massive ‘white space’ for robotics applications and platforms
He describes an app-store-like future where hardware platforms enable a broader developer ecosystem building task-specific capabilities (cooking, elder care, agriculture, etc.). This creates opportunities for startups that focus on software, data, or vertical applications rather than full humanoid stacks.
- •Robotics platforms reduce barriers for application-layer builders
- •Analogy: iPhone platform + third-party apps → similar ecosystem for robots
- •Potential verticals: home skills, elder care, agriculture, many forms of physical work
- •Opportunity spans software, tooling, data pipelines, and specialized applications
- 16:40 – 18:12
China’s platform-first hardware strategy (Unitree) vs US perfection-first go-to-market
Kang contrasts Chinese firms that ship earlier as research/entertainment platforms with US firms that wait for near-perfect products for factories/homes. Early platform distribution can seed developer tools, data moats, and ecosystem lock-in effects.
- •Unitree approach: ship broadly as a platform, not a polished end-user worker
- •Pros: developer adoption, comfort/familiarity, tooling ecosystems, and data collection in the wild
- •Potential downside: capability may not meet mainstream task expectations yet
- •Similar strategy noted in portfolio example DexMate
- 18:12 – 21:10
Getting involved: it’s early, but robotics is much harder than software
Kang closes by emphasizing the industry is still early with ample opportunity to join, start, or invest—while warning that robotics requires deep, multidisciplinary expertise. Success demands seasoned specialists across mechanical, electrical, manufacturing, deployment, and operations.
- •Robotics maturity lags LLM adoption—so the window isn’t “missed”
- •Ways to engage: research, network with insiders, join/start/invest
- •Caution: robotics is exceptionally hard and knowledge-intensive
- •Winning requires experienced experts across design, engineering, manufacturing, and deployment
