Where the New Humanoid Robotics Race Is Playing Out (And No, It’s Not Just About Chips)
We have not yet reached a “ChatGPT‑level” moment in humanoid robotics, but we are fast approaching it, and the scaling laws that have driven progress in large‑language models (LLMs) appear to apply to this sector as well.
Companies that are amassing the largest volumes of training data through already‑operational fleets enjoy self‑reinforcing advantages that are set to expand.
Historically, robotics applications have relied on specialized software trained for narrow, task‑specific functions. Tomorrow’s market leaders, by contrast, will differentiate themselves through general‑purpose AI models that enable a robot to learn a new job in hours rather than months. It is precisely this race to develop versatile models that forms the foundation of a long‑term investment thesis.
A pivotal architectural breakthrough powering next‑generation humanoid intelligence is the Vision‑Language‑Action (VLA) model. A robot equipped with a VLA receives a simple language instruction, interprets its surrounding environment, and executes the task—without having been pre‑programmed for that specific scenario.
The industry is closely watching a series of general‑purpose model releases from Physical Intelligence (π), such as its pi‑0.7 model announced in April 2026, which performed tasks it had never been trained on. Physical Intelligence is also reportedly negotiating a new financing round that could push its valuation to roughly $11 billion.
In parallel with VLA development, we are witnessing the emergence of world models: AI systems that simulate physical environments and predict how objects and forces interact over time.
The strategic importance of this development is clear. Unlike LLMs, which are trained on the entire textual corpus of the Internet, physical AI requires concrete data that does not exist online in a usable format. World models provide a way to generate such data synthetically at scale—by describing a scenario, simulating it with physics‑aware models, and creating training sessions without a single physical robot present in the room. In the first quarter of 2026 alone, about $6 billion flowed into six or seven companies focused on world models. Whoever controls the simulation infrastructure will control the ability to scale robot training without proportionally expanding physical hardware fleets.
Near‑Term Catalysts for H2 2026
Tesla’s Reallocation of the Model S/X Assembly Line – Tesla’s decision to repurpose its Fremont Model S/X assembly line for the first‑generation Optimus humanoid robot—designed for a capacity of up to one million units per year—maps the sector’s trajectory, irrespective of exact delivery timelines. Moreover, Tesla’s vertical integration of actuators, compute, and AI training infrastructure could yield cost advantages of 30 %–40 % versus competitors that rely on third‑party suppliers.
Unitree Robotics IPO on the STAR Market – The upcoming IPO of Unitree Robotics will be the most significant short‑term liquidity event in humanoid robotics. Financials may surprise those who view the segment merely as a cash‑burning cost center. Unitree’s gross margins in both its humanoid and quadruped lines reached roughly 60 % in 2025. Revenue from humanoid robots surpassed quadruped revenue for the first time last year, accounting for over 51 % of total sales. The company delivered more than 5,500 humanoid robots in 2025—outpacing the combined production of all U.S. rivals (including Tesla, Figure AI, and Agility Robotics)—and aims to ship 20,000 units in 2026.
China’s Advantage
Many of the most critical humanoid subsystems—motors, harmonic gearboxes, power‑management electronics, battery systems, and sensors—benefit from proximity to China’s advanced electric‑vehicle value chain. This enables supplier reuse, process transfer, and faster scalability. On the innovation front, China has filed roughly 7,700 humanoid‑related patents over the past five years, compared with about 1,560 in the United States.
Rather than a single global winner, the humanoid supply chain is likely to bifurcate:
China will lead on hardware cost compression thanks to manufacturing efficiency and economies of scale.
U.S. and European ecosystems will differentiate on frontier AI sophistication, system architecture, and high‑reliability, safety‑certified implementations.
Portfolio Implications
In this environment, an equally weighted exposure to the entire humanoid robotics and physical‑AI ecosystem is preferable to an overconcentration in large‑cap U.S. tech stocks. A truly global representation that includes the United States, China, Japan, and Europe offers the best chance to capture opportunities across the full value chain.
By Derek Yan, Senior Investment Strategist, KraneShares