Figure AI's Early Investor Predicts Human-Level Robots in 3 Years | Andrew Kang, RoboStrategy
In a Nutshell
Figure AI's early investor predicts human-level humanoid robots in 2-3 years, with the market reaching tens of trillions by converting physical labor into an accessible product at ~$2/hour. Vertically integrated companies like Figure, Tesla Optimus, and Apptronik lead because they control manufacturing for data collection, enabling faster AI development through embodiment-specific training. Model commoditization will occur in 3-5 years, shifting value to deployment execution, hardware production, and component innovation rather than intelligence itself.
These notes were generated by AI and may contain inaccuracies.
The Figure AI live stream demonstrated real-world capability rather than curated footage. The stream ran continuously for 8 to 10 hours and extended across 8 days total. The human intern ultimately won the challenge by a narrow margin, though the task proved physically exhausting, resulting in blistered hands and fatigue that made repetition unlikely.
Humanoid robot intelligence is projected to reach capability levels sufficient for most daily tasks within two to three years. This development represents a distinct technological revolution because it converts physical labor into a product accessible to nearly anyone.
Andrew Kang serves as CEO of Robo Strategy, a publicly listed venture fund on NASDAQ focused exclusively on robotics and physical AI. The fund maintains investments across multiple robotics companies including Figure AI, Apptronik, Sanctuary Robotics, Center Bots (industrial arms and cobots), and Path Robotics (welding automation).
The Figure AI investment was non-consensus at the time. Other venture investors familiar with frontier technology expressed skepticism about humanoid robotics feasibility and perceived significant risks. Previous humanoid companies had not produced venture-scale outcomes. The decision to pivot the entire fund toward robotics was based on recognizing accelerating technological change and different development pace compared to prior periods.
Initial Figure AI investment was made without visiting the facility. Research focused on founder Brett and prior work across previous companies. Background analysis covered hardware engineering, robot learning, hand engineering, robot controls, and fleet management expertise. Comparison across competitors indicated Figure represented one of the top teams capable of execution.
The approach emphasizes non-dogmatic evaluation with high conviction but loosely held beliefs. Regular re-evaluation occurs based on new information. Continuous tracking spans all robotics companies, not just portfolio holdings, to assess competitive positioning across fleet scaling, robot learning research, and hardware development. Current assessment maintains Figure and Tesla Optimus as the leading companies.
Investment priority focuses on vertically integrated robotics companies that build their own robot intelligence, hardware, conduct deployments, and scale manufacturing capabilities. Vertical integration enables co-optimization between hardware and intelligence systems. Hardware characteristics like torque sensing in joints can be incorporated into simulation models, improving training efficiency and research effectiveness, ultimately producing more performant robots.
Robot learning development requires embodiment-specific data from robots performing tasks or under teleoperation control. The analogy illustrates that a 7-foot-tall embodiment would create awkward world interaction compared to operating within familiar physical form. Embodiment-matched data produces more effective models. Large-scale data collection requires substantial robot quantities, creating current industry bottlenecks.
Companies like Figure, Tesla Optimus, and Apptronik operate their own manufacturing facilities, enabling dedicated robot allocation for data collection without external supply constraints. Supply chain vulnerabilities include inability to rapidly source 100 or 1,000 robots, requiring advance orders and production lead times. Historical parallels exist with GPU and commodity supply chain constraints during rapid demand scaling.
The humanoid robotics market is projected to reach tens of trillions of dollars. Top-down analysis starts with global physical labor market valued at $50 trillion. Bottom-up modeling assumes $50,000 per humanoid unit (sale or lease price), comparable to US physical labor annual costs including benefits. Scaling calculations show 100,000 units generating $5 billion revenue, while 1 million units annually would produce $50 billion. Historical parallels include billions of cellphones, hundreds of millions of cars, and PCs produced yearly, suggesting humanoid production volumes could exceed current estimates.
Abundant, affordable labor expands total addressable market beyond current physical labor applications. Potential applications include space deployment and data center construction, where current constraints center on labor availability for assembly, plumbing, and electrical work rather than component supply.
The current robotics and AI wave represents the fourth industrial revolution. Previous industrial advances created machines for production and convenience, but this development enables machines and intelligence capable of any human task. Labor cost reduction to approximately $2 per hour transforms physical labor into an accessible product. Applications include personal assistants comparable to iPhone ubiquity, and enabling entrepreneurship through affordable robot labor previously inaccessible due to cost or availability.
AI research acceleration creates feedback loops where improved models automate more research processes, enabling 24/7 operation and faster information processing. These efficiency gains extend to robot AI research across multiple dimensions including research speed, data annotation methodologies, data collection infrastructure, mid-training structures, and reinforcement learning approaches. LLM concepts transfer to physical AI models, accelerating development timelines beyond conventional expectations.
Intelligence capability reaching human-level within 2-3 years does not guarantee equivalent deployment timelines due to manufacturing constraints. Unlike instant chatbot instance scaling, physical robot production requires factory and supply chain expansion for all components, adding substantial time requirements.
Open-source model market share increased from less than a few percent of tokens 2-3 years ago to 25-30% or more currently. Benchmark saturation is occurring with the gap between open-source and frontier models compressing from approximately 2 years to 6 months. Future trajectory suggests open-source models will saturate benchmarks, though gaps may persist without practical significance for routine tasks like shelf restocking or computer mouse assembly that do not require high-level intelligence.
Physical AI model layer commoditization is projected within 3-5 years. At this point, intelligence becomes inexpensive, shifting value toward deployment capabilities, hardware production, and component innovation. Most valuable companies will likely be those executing deployments, manufacturing hardware, or developing improved designs and components.
Nvidia produces open-source models across multiple domains including Nemotron for general LLM and software engineering applications, autonomous vehicle models, and physical AI/robot intelligence models including Root, Cosmos, and Dream Zero. The company maintains significant resources and research talent. Development of open-source alternatives serves strategic interests by reducing dependency risks if closed-source labs optimize for alternative hardware platforms like Google TPUs.
Both US and Chinese robotics industries are expected to achieve massive scale with independent development of hardware and robot intelligence. Robots deployed in America will likely originate from American companies, while Chinese-purchased robots will come from Chinese manufacturers. Government interest in domestic production capability exists in both countries, with China having invested billions directly and indirectly through government funds and municipalities. US domestic robotics support is anticipated to follow patterns established with rare earths processing and semiconductor companies like Intel.
US leads in physical intelligence models while China maintains competitive research groups, including Alibaba-affiliated teams producing near-frontier robot models. Both countries possess strong research talent. Future development will likely occur through independent capability building combined with collaboration via open-source research publication that benefits both nations. The dynamic is characterized as short-term positioning rather than permanent competitive gaps, with both countries expected to achieve independent capability within 5-10 years.
Hardware platform availability reduces development barriers for application-specific builders. Robotics development parallels smartphone ecosystems where platform creators (like Apple iPhone) enable external developer communities building applications for time management, note-taking, and specialized functions. Robot applications could address cooking skills, elder care, farming efficiency, agricultural techniques, and any physical labor domain.
Unitree represents a watched but non-portfolio company employing different market approach than US companies. US companies typically delay release until products achieve perfect performance for home or factory environments. Chinese companies release hardware as research and entertainment platforms for external development. This strategy provides developer access, deployment familiarity, and Unitree-specific data collection that could create model compatibility advantages for Unitree robots versus competitors.
Dexma portfolio company employs similar platform-release strategy, enabling external application development by researchers and startups focusing on AI rather than hardware and manufacturing challenges.
Current robotics industry stage remains early compared to LLM deployment where models like Opus 4.8 and ChatGPT execute substantial human work across companies globally. Robotics has not reached equivalent capability levels, creating significant opportunity for company participation, investment, and growth trajectories.
Recommended initial steps include research and networking with industry professionals. Robotics company building requires substantial accumulated knowledge across mechanical design, electrical engineering, high-rate manufacturing, and real-world deployment. Unlike software companies, successful humanoid or robotics ventures require expertise from individuals with years or decades of manufacturing and robotics environment experience. While investment and startup activity will increase, appreciation for execution difficulty and need for genuine experts is essential for success.
Keep EO in your library
Save the videos and channels worth coming back to, and find them again in one place.
