Inside the Machine: A Buy-Side Deep Dive on Unitree G1
We tore apart a Unitree G1. Working alongside hardware engineers here in Hong Kong with direct visibility into the Chinese supply chain, we mapped the G1's architecture joint by joint in a way sell-side teardown notes rarely do with sufficient discipline. What we found challenges several assumptions driving meaningful capital into the U.S. public-market humanoid trade.
The core conclusion: the G1 is not an industrial robot in humanoid form. It is a highly cost-engineered, mobility-optimised development platform. Understanding that distinction is the single most important analytical step U.S. investors have failed to take.
What Is Actually Inside
At roughly 35 kilograms, the G1 is strikingly light. Aluminium alloy dominates the frame; plastics appear wherever tolerances permit; steel is reserved only for critical stress-bearing junctions. Topology optimisation and hollow routing have already been pushed aggressively — Unitree's mechanical team has likely harvested most available lightweighting gains within the current architecture. Future performance improvements will require better materials, higher-efficiency actuation, or architectural changes rather than straightforward mechanical trimming.
The joint modules are the most telling design choice: each integrates motor, gearbox, encoder, and drive electronics into a single compact assembly. Complexity is absorbed inside modular subassemblies rather than distributed across the robot. The downstream effects are real — improved assembly efficiency, faster iteration cycles, simpler serviceability, and a cost-down curve that steepens with volume.
Sensors and compute follow the same philosophy: commercially proven components, established supply chains, no bespoke silicon. This is a company optimising for deployment velocity and cost structure, not engineering prestige.
The Thermal Constraint Analysts Are Missing
The G1 appears thermally conservative. In demonstrations, this is invisible. In sustained industrial operation, it becomes definitional. There is a categorical difference between dynamic locomotion across a five-minute showcase and consistent torque output across an eight-hour duty cycle. Repetitive lifting, continuous warehouse shifts, sustained high-load actuation — these are thermal problems as much as mechanical ones.
The G1's effective high-load operating window, under real industrial conditions, is materially limited relative to what demonstration footage implies. Thermal robustness is not a technical footnote. It is a market segmentation variable. Investors extrapolating from "walks impressively" to "replaces industrial labour economically" are making a leap the evidence does not yet support.
The Strategic Logic Being Underpriced
Unitree's limitations are real: low arm payload, constrained thermal headroom, an architecture unsuited to heavy-duty industrial applications. And yet Unitree may be more strategically dangerous than the market appreciates — precisely because of its price point.
A humanoid affordable enough to seed universities, research labs, AI developer environments, and commercial pilot programmes at scale is not merely a cheaper robot. It is a data-flywheel asset. Every deployed unit generates locomotion data, edge cases, and software iteration cycles that compound into algorithmic advantage. In platform markets — and humanoid robotics is fundamentally a platform market — early volume is often more strategically decisive than early capability. Unitree appears to understand this. Much of the U.S. investment community, focused on industrial TAM slides and payload specs, does not.
What the Supply Chain Actually Tells Us
Presence in a teardown is not investability. Public-market humanoid supply-chain names fall into three buckets:
Fundamental beneficiaries — validated production relevance, meaningful content-per-unit economics, earnings leverage that matters at listed-company scale. This group is small.
Strategic watchlist — technically relevant, potential design wins, but sourcing share duration, ASP durability, gross-margin contribution, and internal-substitution risk remain unresolved. Most serious candidates sit here.
Thematic proxies — adjacent exposure, no credible earnings bridge, valuations reflecting narrative over fundamentals. This is the largest cohort in the current U.S. humanoid basket.
The layers with genuine investable substance: system integration and control software, where the real moat lives; integrated actuation and power management at the subsystem level; thermal and power-cycle management, the primary barrier between current capability and industrial duty-cycle requirements; and embodied-AI data infrastructure, the longest-duration and potentially highest-reward layer.
Our View
We are constructive on humanoid robotics as a multi-year industry formation story. We are not constructive on the current structure of the U.S. public-market trade.
The near-term market will be won by affordable, iterable, widely-deployed platforms — not the most industrially ambitious ones. Monetisation will lag adoption. Industrial validation will take longer than consensus assumes.
Do not buy the theme. Buy the specific layers where volume, differentiation, and earnings leverage can simultaneously coexist. That list is shorter than the current humanoid basket implies. For the names that qualify, the opportunity is real.
not investment advice