Robotics is not one market

Every technology cycle produces a shorthand that eventually gets in the way. In robotics, that shorthand is the humanoid.

The humanoid is easy to photograph, easy to imagine in a factory or kitchen, and easy to use as a symbol for the entire category. It is also a poor map of where the value is being created. Robotics is not one market. It is a stack of at least eight economically distinct layers, from rare-earth magnets and precision actuators to foundation models, fleet orchestration, and end applications. The margins, moats, capital needs, and geographic centers of gravity at each layer look nothing alike.

That distinction matters now because capital is arriving faster than deployment. Robotics and physical-AI companies raised record amounts in the first two quarters of 2026. Private valuations imply that some general-purpose robot companies have already won. Yet nobody ships a fully autonomous general home robot at consumer scale, and much of today's humanoid volume goes to entertainment, research, and data collection rather than useful work.

The thesis is simple: the robot will matter, but the most attractive early-stage investments may sit in the enabling layers that every robot needs.

Two markets moving at two speeds

Consumer robotics has split in two.

The first market is already real. Task-specific robots such as vacuums, lawn mowers, and pool cleaners make up a roughly $10 billion to $25 billion global category. These products solve frequent, legible problems. Their autonomy is narrow, but it works.

This market also offers a warning. The five largest robot-vacuum sellers are now Chinese. Chinese companies hold roughly 70% of the home-cleaning robot market, and Roborock alone has about 27%. In December 2025, iRobot, the company that created the Roomba category, filed for Chapter 11. Its assets went to Shenzhen Picea, its own contract manufacturer and lender.

Category creation did not create a durable moat. A supplier became a competitor, then an owner.

The second market is the one attracting the largest valuations: general-purpose home robots. 1X, Figure, Tesla, Sunday Robotics, and The Bot Company are pursuing versions of a machine that can navigate an unstructured home and manipulate the physical world.

That market remains pre-revenue at consumer scale. 1X offers NEO for $20,000 or $499 per month, but openly relies on remote human teleoperators for complex tasks. Tesla has committed more than $20 billion of capital expenditure to Optimus, but Elon Musk said in January 2026 that zero Optimus units were doing useful work. Figure's last private mark was $39 billion despite little meaningful revenue.

The valuation gap between these two markets is striking: the proven category is measured in product revenue, while the emerging category is priced on a future autonomy curve.

The autonomy gap is the diligence question

Every general-purpose robotics thesis reduces to one question: how does the company cross from a controlled demo or human-assisted workflow to reliable autonomy?

There are three main strategies.

Teleoperation first. A company deploys sooner with humans in the loop and turns those interventions into training data. This produces real-world trajectories, but adds privacy and support-cost risks.
Proxy data first. Companies use demonstrations captured from people, simulation, or video to train policies before broad deployment. This can produce a cleaner privacy story, but the transfer to real environments remains unproven.
Fleet data first. A vertically integrated company uses its own factory or robot fleet to create a proprietary data flywheel.

No approach has won. The right diligence questions are concrete: How many customer deliveries are verified? What percentage of tasks require a human? Is that ratio falling? Who owns the deployment data? What does support cost per robot per month? What happens to gross margin when the Chinese component cost curve moves again?

The stack under the robot

A useful map has eight layers.

Raw materials. Rare-earth magnets, specialty steels, battery cells, and structural materials.
Components. Actuators, reducers, roller screws, motors, encoders, tactile sensors, batteries, and wiring harnesses.
Compute. Training clusters and onboard inference silicon.
Models. Vision-language-action foundation models and locomotion policies.
Simulation and data. Physics engines, synthetic data, teleoperation pipelines, annotation, and evaluation.
Middleware and runtimes. Robot operating systems, real-time control, drivers, and hardware abstraction.
Orchestration and fleet operations. Mission assignment, monitoring, updates, failure handling, and security across mixed fleets.
Systems and applications. Robot makers, system integrators, and the end workflows customers buy.

The layers do not share one economic model. Precision components can carry 30% to 40% gross margins and multi-year qualification moats. Industrial robot makers such as FANUC and ABB can run above 20% operating margins. System integrators often take fixed-price project risk and earn thin margins. Foundation-model labs carry software upside but little current revenue and multi-billion-dollar entry prices. Fleet software can sell recurring subscriptions with far less capital.

Calling all of this "robotics" hides the allocation decision.

The physical supply chain is shifting east

Actuation, including reducers, screws, and motors, is roughly half of a humanoid's production cost. For decades, Japanese specialists held dominant positions in precision reducers. That structure is changing quickly.

Chinese companies' share of the harmonic-reducer market rose from roughly 15% in 2022 to about 38% by early 2024. Leaderdrive, Shuanghuan, and Zhongda have qualified into major robotics programs. Roughly 70% of the components in Tesla's third-generation Optimus are reported to be Chinese-sourced. The five largest humanoid shippers in the first half of 2026 were all Chinese and accounted for 86% of shipments.

Unitree shows what that cost base can do. The company cut average selling price by roughly 72% in two years, sells its R1 humanoid for $4,900, and is profitable. Western robotics companies should assume that component prices will fall and that Chinese competitors will move up the stack.

The deepest supply-chain chokepoint sits even lower. High-temperature rare-earth permanent magnets are essential to robot servo motors. China still handles roughly 80% to 90% of processing. Since April 2025, qualifying shipments have required export licenses. The broader control suspension is scheduled to expire on November 10, 2026. Western contingency supply from MP Materials and Lynas remains limited.

Every hardware investment now needs a magnet, battery-cell, and precision-machining supply audit.

The integration market is bigger than the robot market

Industrial robotics offers another useful correction. The robot arm is only about 25% to 35% of the cost of a working automation cell. A $65,000 arm can become a $220,000 installed system after end-effectors, safety equipment, and hundreds or thousands of engineering hours.

That gap is the system-integration market. At roughly $55 billion in 2025, it was more than three times the size of the industrial robot market itself. It is fragmented across thousands of regional specialists and full of fixed-price contracts, component shortages, delays, and liquidated damages.

This is a difficult services business and a strong software opportunity. The winning product may not be a new robot. It may be the layer that makes robots faster to deploy, easier to operate, safer to certify, and cheaper to maintain.

Where value can accrue

We see four broad margin pools.

The toll pool includes precision components and edge compute. These suppliers get paid regardless of which robot maker wins, but they face vertical-integration and commoditization risk.

The brain pool includes foundation models and synthetic-data platforms. This layer may produce software margins and a small number of large winners, but current entry prices already assume much of that outcome.

The body pool includes robot makers, integrators, and commodity sensing. It can create generational winners, but it looks more like the early automobile industry: many entrants, capital-intensive production, compressed margins, and eventual consolidation.

The operations pool includes middleware, orchestration, deployment tooling, observability, safety, and compliance. Nobody photographs it. It is also the layer that can produce recurring-revenue, capital-efficient companies on top of every deployment the other pools generate.

Five areas we are watching

1. Tactile and force sensing

Cameras and LiDAR benefit from smartphone and automotive scale. Tactile and six-axis force/torque sensors do not. They remain low-volume precision instruments with high margins, and no technology has won at robot scale. Dexterous manipulation depends on solving this bottleneck.

2. Physical-AI data engines and evaluation

Every model lab needs teleoperation pipelines, egocentric data, annotation, and benchmarks. The model race may concentrate, but the need to produce and evaluate data spans every contender. We are especially interested in teams with real fleet access rather than a synthetic-data-only story.

3. Vendor-neutral fleet orchestration and robot operations

Mixed-vendor fleets are becoming normal. Factories and warehouses need software that can connect, monitor, secure, update, and command heterogeneous robots across ROS 2, OPC-UA, MQTT, and legacy industrial systems. This is recurring software revenue tied to deployment volume rather than one robot maker's survival.

4. Safety, compliance, and insurance

ISO 10218:2025 moved more of the certification burden from the robot to the application. Each deployment now needs application-level risk assessment and ongoing documentation. Walking robots add fall zones and new forms of physical liability. This is required spend, and no dominant software or underwriting platform exists.

5. Vertical robotics where autonomy works today

Welding, machine tending, inspection, agriculture, and construction offer narrow tasks in structured environments with measurable labor pain. The opportunity is to turn custom integration into a repeatable product: a standardized cell, software, and outcome-based pricing with unit economics that work now.

What we are not underwriting

We are cautious on consumer humanoid rounds priced above $10 billion, foundation-model financings where a small fund cannot buy meaningful ownership, commodity perception components, and reducer manufacturing caught in a Japan-China capital-expenditure race. We are also skeptical of any plan that requires full home autonomy before roughly 2028 to hit its numbers.

The technology can be real while the price is wrong.

The robot is the visible layer. The stack is the market.

The central mistake in robotics investing is to confuse the object with the industry.

The robot is the interface the customer sees. Under it sits a long chain of materials, precision components, compute, models, data systems, runtimes, fleet software, compliance, and integration. Some of those layers will commoditize. Others will become toll booths. A few will determine whether the machine can do useful work at all.

The best opportunities for an early-stage investor may be the least cinematic ones: a sensor, a data pipeline, an orchestration layer, a compliance product, or a narrow application that works every day.

That is where we are looking.