A 14-layer analysis of the entire robotics industry — rare-earth magnets to robot-delivered services — with competitive landscapes, market sizing, and strategic build opportunities across funding tiers. Humanoids, autonomous vehicles, drones and defence, surgical, agricultural, warehouse and industrial, plus every component and software layer beneath them. Compiled against publicly available information as of August 7, 2026.
Robotics has two industries wearing one name. The profitable one is purpose-built, works inside a structured environment and keeps a human in the loop: Intuitive Surgical booked $2.89B in Q2 2026 with zero autonomy, and Symbotic carries a ~$22.7B backlog. The funded one is general-purpose, works in unstructured environments and promises to remove the human. Almost none of it discloses a unit count.
Value concentrates at layer 14, not layer 12. Waymo is worth $126B selling rides. Figure is worth $39B selling robots it has never counted publicly. Uber has become the demand layer for six different autonomy stacks without building one. The companies that meter the output are worth more than the companies that build the machine.
The binding constraints are all below the software. China controls roughly 69% of rare-earth mining and 90% of magnet refining; gearboxes are 30–50% of actuator cost and the real bottleneck is qualifying robotics-grade parts, not raw capacity; dexterous hands have no volume supplier at all. Layers 01–05 are the least documented in this map and the most decisive.
The wall went up in a fortnight. On 28 July 2026 the FCC barred new foreign-made humanoids and quadrupeds from the US market, citing a Unitree vulnerability. On 30 July Beijing threatened countermeasures. On 6 August Unitree priced its Shanghai IPO at ~$9.04B. Meanwhile NVIDIA had picked Unitree, weeks earlier, to build the reference humanoid for its own platform.
Defence, not humanoids, is the most valuable private platform category. Anduril was valued at ~$61B in May 2026 on $2.2B of 2025 revenue, with reported talks near $100B; Shield AI reached $12.7B and Saronic $9.25B. Separately, SoftBank is acquiring ABB Robotics for $5.375B — the largest consolidation in industrial robotics history — while also leading Skild's round and backing Agility.
China ships; the West raises. AgiBot shipped an estimated 5,168 humanoids in 2025, more than every US humanoid company combined, and Chinese firms took the top three shipment positions. China's operational industrial robot base passed 2 million units, 54% of global installations, and domestic suppliers outsold foreign ones at home for the first time.
Report 01 runs silicon to verticals. Robotics runs raw material to service, because in a robot most of the cost and nearly all of the risk sit below the software. Value concentrates at the bottom, where China controls supply, and at the top, where the ride is metered — not in the middle, where the valuations are.
The mined and refined inputs a robot cannot be built without: rare-earth permanent magnets, specialty alloys, and structural composites. Invisible in every market map and decisive in every production plan.
Share of global rare-earth magnet processing and refining that sits in China, against roughly 69% of mining (McKinsey). A humanoid needs on the order of 3.5–4kg of NdFeB. There is no drop-in substitute at the same power density.
This layer has already been used as leverage. China's April 2025 export controls on seven rare earths and magnets hit Optimus directly, and Musk publicly attributed production impact to the magnet issue. Western capacity is being built — MP Materials is the only integrated US mine-to-magnet operator — but permitting and qualification run in years, not quarters. A tariff wall against finished Chinese robots does nothing about this layer.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | MP Materials | Mountain Pass mine, Fort Worth magnets | Only integrated US mine-to-magnet operator | ~$12B market cap | US strategic supplier of record |
| 2 | JL MAG Rare-Earth | Sintered NdFeB magnets | China's largest high-performance magnet maker | ~$9B market cap | Supplies the Chinese humanoid build-out |
| 3 | Lynas Rare Earths | Mt Weld, Malaysia, Texas | Largest separator outside China | ~$8B market cap | The non-China benchmark |
| 4 | Shenghe Resources | Mining and trading | Upstream of the Chinese magnet chain | ~$5B market cap | Controls feedstock flows |
| 5 | Proterial (ex-Hitachi Metals) | NeoMax magnets | Holds foundational NdFeB patents | Bain-owned | Japanese quality tier |
| 6 | Neo Performance Materials | Magnetics and advanced materials | Non-China magnetic powders and magnets | ~$0.5B market cap | Europe and North America focus |
The muscles. Gearboxes, motors and screws that turn electrical power into torque at the joint. Gearboxes alone are 30 to 50 percent of actuator cost, and a humanoid needs 14 to 28 of them.
Share of actuator cost represented by the gearbox alone (McKinsey). A humanoid carries 14 to 28 actuated joints depending on configuration, which makes this the single largest addressable line in a robot's bill of materials after compute.
The popular claim that only a handful of firms can mass-produce harmonic drives overstates it. McKinsey's read is that aggregate gearbox capacity is adequate; the constraint is that humanoids need compact, high-durability, low-clearance configurations, so the effective supplier pool narrows and the real work is qualifying and ramping the robotics-grade subset. That is a multi-year exercise, not a purchase order. Chinese suppliers led by Leaderdrive are setting the cost floor, which is why Western humanoid BOMs still route through China.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | Harmonic Drive Systems | Strain-wave gears | Invented the category; reference precision joint | ~$3B market cap | Default for Western humanoid joints |
| 2 | Nabtesco | RV cycloidal reducers | ~60% share of precision cycloidal reducers | ~$3B market cap | Owns the industrial arm joint |
| 3 | Leaderdrive | Harmonic reducers | China's cost-competitive alternative | ~$4B market cap | Enables the Chinese humanoid price point |
| 4 | THK | Linear guides, ball screws | The rails robots move on | ~$3B market cap | Broad industrial base |
| 5 | Ewellix (Schaeffler) | Planetary roller screws | Linear actuation for humanoid legs | Schaeffler-owned | Named in Optimus supply chain |
| 6 | Sanhua / Tuopu | Actuator assemblies | Auto-parts scale pivoting to humanoids | ~$15B / ~$12B market cap | China's humanoid tier-one bench |
| 7 | Moog | Precision electric and hydraulic actuation | Aerospace-grade reliability | ~$7B market cap | Heavy and legged robotics |
How a robot measures the world and itself. Lidar and radar for vehicles and mobile robots, depth and vision for manipulation, force-torque and tactile at the contact point, encoders and IMUs everywhere.
Lidar units shipped by Hesai alone in FY2025, on revenue of RMB 3.03B (~$433M). Q1 2026 added another 471,723 units on RMB 681M, up 30% year on year, with 2026 guidance of 3–3.5M units and capacity doubling from 2M to 4M.
Lidar consolidated hard and the winners are Chinese. Hesai and RoboSense between them supply most of the world's automotive and robotics lidar, while Luminar's market capitalisation has collapsed to roughly $200M. NVIDIA named Hesai the primary lidar partner for DRIVE Hyperion 10 — and Hesai has since drawn the same US cyber-risk scrutiny that produced the FCC action against Chinese humanoids. Tactile sensing remains the immature end of the layer: Amazon's Vulcan made news in 2025 for being able to feel what it gripped.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | Hesai Group | AT128, robotics lidar | Volume leader; NVIDIA DRIVE Hyperion 10 partner | FY25 RMB 3.03B (~$433M) | Q1 2026 +30% YoY; US cyber scrutiny |
| 2 | RoboSense | RS lidar, robot platforms | Pivoting from sensor to full robot components | ~$3B market cap | Supplies WeRide |
| 3 | Bosch | Radar, MEMS, IMUs | Automotive scale; also a robotics investor | ~€90B group revenue | Backed NEURA's Series C |
| 4 | Ouster | Digital lidar | Industrial, infrastructure and robotics | ~$1B market cap | Non-automotive focus |
| 5 | Luminar | Automotive lidar | US challenger that could not scale | ~$0.2B market cap | Cautionary tale of the shakeout |
| 6 | ATI Industrial Automation | Force-torque, tool changers | The industrial force-sensing default | Novanta-owned | Ubiquitous in arm deployments |
| 7 | Orbbec | Depth cameras | Volume 3D vision for robots | ~$2B market cap | Supplies AgiBot |
| 8 | Meta / GelSight | Digit 360, optical tactile | The most serious tactile research, published | Meta research | No volume manufacturing yet |
The hand. Broken out from actuation because it gates every manipulation claim in the sector: a five-finger hand packs more actuators, sensors and failure modes into less volume than the rest of the robot combined.
No credible market number exists for dexterous hands as a component category, because almost nobody sells one at volume. That absence is the finding: a five-finger hand packs more actuators, sensors and failure modes into less volume than the rest of the robot combined.
Hands are the component most likely to be quietly swapped for a two-finger gripper between the demo video and the deployment. Every serious humanoid OEM is building its own hand rather than buying a good one, which is a classic signal of a missing supplier. Meanwhile the profitable end of this layer is unglamorous: vacuum and suction gripping is how most warehouse robots actually pick, and Piab and Schunk sell into that all day.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | SCHUNK | Grippers, clamping | The broadest industrial catalogue | Private, family-owned | Industrial default |
| 2 | Shadow Robot | Shadow Dexterous Hand | The reference research hand for two decades | Private | Used by DeepMind and academia |
| 3 | Wonik Robotics | Allegro Hand | Made research dexterity affordable | Wonik Group | The academic workhorse |
| 4 | Sharpa | Dexterous hands with tactile | Integrated sensing for humanoids | Early stage | Targeting the OEM gap |
| 5 | OnRobot | Plug-and-play EOAT | Cobot-native tooling | Private | Universal Robots ecosystem |
| 6 | Piab | Vacuum and suction | How most warehouse robots actually pick | Patricia Industries | Profitable and boring |
| 7 | RightHand Robotics | Piece-picking systems | Gripper plus vision plus grasp planning | ~$100M raised | Sells outcomes, not hands |
Runtime is a product spec. Cells, power electronics, and the charging and docking infrastructure that decides whether a fleet works one shift or three.
Power is not a component choice, it is a product specification. Battery mass compounds into the whole mechanical design: heavier packs need stronger actuators, which need more power, which needs heavier packs. Robotics rides the automotive and consumer cell supply chain rather than commanding its own.
No robotics company is large enough to matter to CATL, which means this layer is effectively a price-taker relationship. The differentiated work sits in power electronics and charging infrastructure: fleet uptime is decided by how fast an AMR can dock and top up, not by cell chemistry. Safety MCUs from Infineon and NXP are the quiet gate on whether a robot can be certified to work near people at all.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | CATL | Battery cells and packs | World's largest cell maker | ~$180B market cap | Supplies the Chinese robot build-out |
| 2 | Samsung SDI | Cells for mobility and robotics | Korean quality tier | ~$25B market cap | Also an investor in Skild |
| 3 | Infineon | Power electronics, AURIX MCUs | Functional safety at industrial scale | ~$50B market cap | The certification enabler |
| 4 | NXP Semiconductors | Automotive and industrial MCUs | Safety-critical control | ~$55B market cap | Automotive-grade robotics |
| 5 | Bosch Rexroth | Hydraulics, drives, motion | Factory automation incumbent | Bosch division | Heavy industrial motion |
| 6 | Festo | Pneumatics, soft actuation | Long-running bionics research | Private | Soft robotics pioneer |
| 7 | Wiferion (PULS) | Wireless inductive charging | AMR fleet uptime as a product | Acquired by PULS | Deployed with Locus |
A robot has a power and thermal budget a data centre does not. Every joule spent on inference is a joule not spent on actuation, and battery mass compounds into the whole mechanical design.
The power envelope NVIDIA's Jetson Thor operates in, delivering a Blackwell GPU and 64GB of memory to a robot that also has to move. Compute is the largest single line in a humanoid bill of materials, and unlike a data centre it is bounded by watts, not by budget.
This is the layer where NVIDIA is executing the clearest platform strategy in robotics: compute, simulator and model sold as one stack, with 110 partners announced at GTC 2026 and an explicit ambition, in TechCrunch's phrasing, to be the Android of generalist robotics. The counterweights are vertical integrators (Tesla's AI5) and the automotive incumbents (Mobileye, Horizon Robotics) who already ship at scale into cars. NVIDIA choosing Unitree to build its GR00T reference humanoid, weeks before the FCC barred new Chinese robots, is the layer's defining irony.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | NVIDIA | Jetson Thor / Orin, Isaac, GR00T | Compute, simulator and model in one stack | ~$4.5T market cap | 110 partners; wants to be robotics' Android |
| 2 | Mobileye | EyeQ, Chauffeur | The incumbent ADAS silicon and stack | ~$14B market cap | Ships into most of the car industry |
| 3 | Horizon Robotics | Journey series | China's domestic Mobileye | ~$12B market cap | Supplies Baidu Apollo |
| 4 | Tesla | AI5 | Vertically integrated inference silicon | Tesla internal | Powers both FSD and Optimus |
| 5 | Qualcomm | Snapdragon Ride, Robotics | Mobile silicon scale applied to robots | ~$200B market cap | Investor in NEURA and Wayve |
| 6 | Ambarella | Vision SoCs | Low-power perception | ~$4B market cap | Cameras, drones, robots |
| 7 | Hailo | Edge AI accelerators | Perception at the robot's watt budget | ~$500M raised | Independent edge challenger |
The plumbing between sensors, planners and actuators. Message transport, real-time scheduling and the certified runtimes that let a robot work near a person without a cage.
The de facto robotics middleware. Almost every robot in this report either ships with it or was explicitly built against it — and no company owns it. It is stewarded by the Open Source Robotics Foundation with Alphabet's Intrinsic among its backers.
An unowned standard at the centre of an industry is unusual and consequential: it means the integration layer cannot be captured the way CUDA captured AI compute. The commercial value migrates instead to the certified derivatives — Apex.AI for automotive-grade ROS, QNX and VxWorks for safety-critical real-time — and to transport, where Zenoh is displacing DDS inside ROS 2. This is the least glamorous layer in the stack and one of the two or three most strategically important.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | ROS 2 / Open Robotics | ROS 2 Jazzy and successors | The de facto standard; unowned | Open source | Everything integrates against it |
| 2 | Intrinsic (Alphabet) | Flowstate | Alphabet's industrial robotics software arm | Alphabet | Stewards parts of the ROS ecosystem |
| 3 | BlackBerry QNX | QNX RTOS | The safety-certified RTOS cars run on | BlackBerry | Automotive and medical incumbent |
| 4 | Wind River | VxWorks | Aerospace and defence RTOS | Aptiv-owned | Certified autonomy |
| 5 | Apex.AI | Apex.Grace | Automotive-grade certified ROS derivative | ~$100M raised | Production vehicle deployments |
| 6 | ZettaScale | Zenoh | Next-generation robot data transport | Private | Replacing DDS in ROS 2 |
| 7 | RTI | Connext DDS | Industrial-grade safety-critical messaging | Private | Defence and medical |
Where policies are trained before they touch hardware, and where autonomous vehicles are validated at a scale road miles cannot reach. Locomotion transfers well; contact-rich manipulation is where the sim-to-real gap still bites.
Simulated experience is effectively free and slightly wrong; real robot experience is accurate and enormously expensive. The entire layer exists to arbitrage that gap. For autonomous vehicles it is not optional — validation at the required coverage cannot be reached on public roads.
NVIDIA's Newton, built with Google DeepMind and Disney Research and hosted by the Linux Foundation, is the closest thing to a neutral standard the layer has, which is notable given NVIDIA's position everywhere else. The persistent technical problem is the sim-to-real gap on contact-rich manipulation: locomotion transfers well, grasping a deformable object does not. The AV branch of this layer is a real business with real customers — Applied Intuition is valued around $15B on it.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | NVIDIA Isaac Sim / Lab | Omniverse-based robot simulation | The dominant robot sim and synthetic data environment | NVIDIA | Bundled with the compute and the model |
| 2 | Newton | GPU physics engine | Built with DeepMind and Disney; Linux Foundation | Open source | The neutral standard attempt |
| 3 | MuJoCo | Contact physics | The research standard, open-sourced by DeepMind | Open source | Every manipulation paper uses it |
| 4 | NVIDIA Cosmos | World foundation models | Reasoning core inside GR00T | NVIDIA | World models as a platform product |
| 5 | Applied Intuition | AV development and validation | Toolchain plus, now, defence autonomy | ~$15B valuation | The layer's commercial proof point |
| 6 | Foretellix | Scenario verification | Coverage-driven AV validation | ~$130M raised | Measurable safety argument |
| 7 | Genesis | Generative physics engine | Fast open-source newcomer | Open source | Academic consortium |
| 8 | CARLA | Open AV simulator | Where driving research is benchmarked | Open source | Research standard |
The scarcest input in physical AI. There is no internet-scale corpus of robot actions, so every serious effort builds a collection apparatus: teleoperation rigs, data gloves, instrumented grippers, egocentric video, and the factory telemetry only an owner-operator has.
There is no internet-scale corpus of robot actions. Language models had the web; robotics has nothing equivalent, which makes data collection a capital expenditure rather than a download. Every serious effort builds an apparatus: teleoperation rigs, data gloves, instrumented grippers, egocentric video, factory telemetry.
This is the scarcest input in physical AI and the layer where a genuinely new business model is forming. Mecka AI collects human motion from body sensors and iPhones and sells it, and projects a $100M run rate on signed contracts — though it names no customer and discloses no valuation, so treat the projection as a projection. The structural winners may be neither startups nor labs but owner-operators: Hyundai's Robot Metaplant Application Center exists to turn factory operations into Atlas training data, an advantage no one can buy.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | Mecka AI | Human motion capture from sensors and iPhones | Sells training data as a product | $60M raised; $100M run rate projected | No customer named, no valuation disclosed |
| 2 | Scale AI | Data labelling and annotation | The AV industry's annotation backbone | ~$1.6B raised | Pivoted toward LLM data |
| 3 | MANUS | Data gloves | Teach by demonstration | Private | Used in humanoid imitation learning |
| 4 | Xsens (Movella) | Inertial motion capture | Industrial standard for human movement | Movella | Long-established |
| 5 | UMI / handheld rigs | Universal Manipulation Interface | Collect manipulation data without a robot | Open research | Radically cheaper per hour |
| 6 | Encord | Multimodal data engine | Training and evaluating perception | ~$50M raised | Perception-focused |
| 7 | Hyundai RMAC | Factory telemetry as training data | Owner-operator advantage | Hyundai | Opens 2026, feeds Atlas |
The brain. Vision-language-action models for manipulation, and the very different, far more mature driving and flight stacks that got there first under a regulator's eye.
Waymo's post-money valuation after a $16B round in February 2026 — more than three times the combined valuation of every humanoid foundation-model company in this report. The driving branch of this layer is a mature industry; the manipulation branch is a research programme with a very large cheque attached.
The received wisdom is that the architecture debate is settled and everyone is training a vision-language-action model. That is true of the labs and much less true of anyone with a system in production, where classical planning and safety interlocks still do the load-bearing work. The strongest published evidence for the generalist thesis is Gemini Robotics 1.5 driving ALOHA, bi-arm Franka and Apptronik's Apollo with no robot-specific post-training, transferring skills zero-shot. The strongest evidence against it is that every deployed, revenue-generating autonomy stack on this list is narrow.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | Waymo Driver | L4 driving stack | Most validated by paid driverless miles | Waymo, $126B | 500,000 paid rides a week |
| 2 | Skild AI | Skild Brain | Omni-bodied; any body, no prior knowledge | $1.4B at >$14B | ~$30M revenue disclosed |
| 3 | Physical Intelligence | π0 → π0.7, RECAP | Open-sourced π0; coach-through-corrections | $600M at $5.6B confirmed | ~$1B at >$11B reported in talks |
| 4 | Google DeepMind | Gemini Robotics 1.5 / ER 1.5 | Zero-shot transfer across embodiments | Alphabet | Drives ALOHA, Franka and Apollo |
| 5 | NVIDIA Isaac GR00T | GR00T N1.6 | Humanoid VLA with whole-body control | NVIDIA | Reasons through Cosmos |
| 6 | Tesla FSD | Camera-only end-to-end | Largest consumer fleet as a data engine | Tesla | Shares silicon with Optimus |
| 7 | Wayve | GAIA, end-to-end driving | Sells to carmakers rather than running a fleet | ~$2.8B raised, ~$8.6B | NVIDIA, Microsoft, Uber, Qualcomm backed |
| 8 | Shield AI | Hivemind | Flies without GPS or comms | $1.5B at $12.7B | Pentagon low-cost drone programme |
The unglamorous layer that decides whether a pilot converts. Remote operations, incident handling, over-the-air policy updates, and the certification paperwork that has not yet caught up with a robot that can fall over.
ISO 10218 was overhauled in 2025 and its definition now admits robots fixed to a mobile platform, so humanoids can fall under the industrial robot safety standard. But the standard explicitly does not govern their mobility hazards. ISO/TS 15066 on collaborative operation was folded into 10218-2:2025.
The IEEE's Humanoid Study Group put the gap plainly: a humanoid that loses power falls over, and fixed-base rules never contemplated that. Until someone closes it, humanoids working unsupervised near people at scale is a legal question before it is a technical one. Commercially, this is the layer that decides whether a pilot converts — remote operations, incident handling, over-the-air policy updates and the certification paperwork — and it is chronically underinvested relative to its leverage. Nobody can price the residual risk, so nobody underwrites it.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | UL Solutions | UL 3300 | Writes the service-robot safety standard | ~$15B market cap | The North American gate |
| 2 | TÜV SÜD | Functional safety assessment | European certification body | Private | AV and robot assessment |
| 3 | Formant | Fleet ops and remote intervention | Observability plus teleoperation | ~$50M raised | Multi-vendor fleets |
| 4 | Foxglove | Robotics data visualisation | The ROS developer's console | ~$30M raised | Debugging standard |
| 5 | InOrbit | Vendor-neutral fleet management | Mixed robot estates | ~$20M raised | Interoperability play |
| 6 | Ottopia | Teleoperation software | Vehicles and robots over public networks | ~$30M raised | Powers 1X-style remote assist |
| 7 | Rerun | Multimodal time-series visualisation | Modern robotics data tooling | ~$20M raised | Developer traction |
The layer that gets the coverage and the valuations, and the layer with the worst gross margins and the longest qualification cycles. It inherits every constraint from the eleven layers beneath it.
Anduril's May 2026 valuation, with reported talks near $100B — making defence autonomy, not humanoids, the most valuable private category in robot platforms. Against that: Figure at $39B with no disclosed unit count, and Unitree listed at ~$9B having actually shipped.
This is the layer that gets the coverage and the valuations, and it has the worst gross margins and the longest qualification cycles in the stack. It inherits the magnet problem, the gearbox qualification problem and the hand durability problem, and is valued as though it had solved them. Two structural facts dominate 2026: SoftBank is buying ABB Robotics for $5.375B, the largest consolidation in industrial robotics history, and Chinese firms shipped more humanoids than everyone else combined while the FCC closed the US market to them.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | Anduril Industries | Lattice, Ghost, Fury | Rewrote how the Pentagon buys autonomy | $61B; 2025 revenue $2.2B | Reported talks near $100B |
| 2 | Waymo | Waymo One robotaxi | Most validated driving in the world | $126B valuation | 500,000 paid rides a week |
| 3 | Figure AI | Figure 02 / 03, BotQ | Highest humanoid valuation | $39B valuation | BMW Spartanburg then Leipzig; no units disclosed |
| 4 | Unitree Robotics | G1, H1, quadrupeds | China's listed humanoid champion | ~$9.04B; FY25 RMB 1.7B | IPO 6 Aug 2026; barred from US 9 days earlier |
| 5 | AgiBot (Zhiyuan) | Genie, GO-1 | World's largest humanoid shipper | ~5,168 units in 2025 | 15,000th unit produced July 2026 |
| 6 | Intuitive Surgical | da Vinci 5, Ion | Most profitable robotics company on earth | Q2 2026 $2.89B, +19% | 468 systems placed in one quarter |
| 7 | ABB Robotics | Industrial arms | Being acquired by SoftBank | $5.375B enterprise value | Replaced a planned spin-off |
| 8 | FANUC | Yellow industrial arms | Largest installed base in the world | ~$30B market cap | Industrial robot sales fell 16% in the cycle |
| 9 | Agility Robotics | Digit | Best disclosed evidence in Western humanoids | ~$2.5B via SPAC | 9 sites, 65,000+ operating hours |
| 10 | Boston Dynamics | Atlas, Spot, Stretch | Owned by its own biggest customer | Hyundai | Tens of thousands of robots committed |
Where a robot meets a purchase order. Integrators design the cell, finance the deployment and carry the risk, which is why the most profitable robotics businesses on earth sit here rather than at the OEM layer.
Symbotic's backlog as of 28 March 2026, the vast majority of it from Walmart and the Exol joint venture. A single quarter's guided revenue — $700–720M — exceeds what most humanoid companies have ever raised.
Integration is where a robot meets a purchase order, and it is where the money in robotics actually is. Integrators design the cell, finance the deployment and carry the delivery risk, which is exactly why they capture the margin the OEM layer does not. The pattern repeats across geographies: AutoStore and Exotec in goods-to-person, Dematic and Honeywell in conveyance, Comau on automotive lines. The strategic risk is customer concentration — Symbotic's backlog is essentially one retailer.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | Symbotic | Case and each handling systems | Largest pure-play deployment business | ~$30B market cap | Acquiring Walmart's robotics unit |
| 2 | Dematic (KION) | Warehouse automation | Global integrator at industrial scale | KION Group | Broad customer base |
| 3 | AutoStore | Cube storage robots | Densest goods-to-person system | ~$3B market cap | Licensed worldwide |
| 4 | Ocado Group | Ocado Smart Platform | Grocery robotics licensed as a platform | ~$4B market cap | Kroger is the US operator |
| 5 | Exotec | Skypod climbing robots | European ASRS scale story | $335M Series D at ~$2B | Fast international growth |
| 6 | Honeywell Intelligrated | Conveyance and robotics | Industrial integration incumbent | Honeywell | Deep parcel presence |
| 7 | Schaeffler | Tier-one supplier and integrator | Supplier, customer and investor at once | ~$5B market cap | Buys Agility, backs NEURA |
| 8 | Formic | Robots-as-a-service financing | Priced per hour for small manufacturers | ~$60M raised | Removes the capex barrier |
The demand side: companies that sell the robot's output rather than the robot. A ride, a delivery, a picked order, a completed procedure. This is where robotics revenue is actually metered, and it is almost entirely a different set of companies from layer 12.
Paid Waymo rides per week across ten US cities as of late March 2026, up from 400,000 across six metros at the start of the year. This is the layer where robotics revenue is actually metered, and it is very largely a different set of companies from the OEM layer.
The most important structural fact in this report is that value is concentrating at layer 14, not layer 12. Waymo is worth $126B selling rides; Figure is worth $39B selling robots it has not counted publicly. Uber has become the demand layer for everyone else's autonomy without owning a stack at all, aggregating Waymo, WeRide, Nuro, Pony, Zoox and Wayve. Intuitive has run this model for two decades: sell the system once, sell the instruments forever. Every profitable robotics business in this report either controls the environment or keeps a human in the loop.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | Waymo One | Robotaxi service | The metered end of the most validated stack | $126B valuation | 500,000 paid rides a week |
| 2 | Uber | AV demand aggregation | Becomes the demand layer without a stack | ~$180B market cap | Waymo, WeRide, Nuro, Pony, Zoox, Wayve |
| 3 | Baidu Apollo Go | Robotaxi service | China's largest operator, now exporting | Baidu ~$40B | Dubai: 100 by end 2026, 1,000 by 2028 |
| 4 | Intuitive Surgical | da Vinci programmes | Sell the system once, the instruments forever | Q2 2026 $2.89B | The model everyone else is reinventing |
| 5 | GXO Logistics | Contract logistics | Robotics' most important test bed | ~$6B market cap | Runs Agility, Locus and Apptronik trials |
| 6 | Zipline | Autonomous delivery network | Most-flown delivery autonomy in the world | $7.6B valuation | $800M Series H |
| 7 | Exol (ex-GreenBox) | Warehouse-as-a-service | Symbotic systems sold as a service | JV | Underwrites much of the backlog |
| 8 | Pony.ai / WeRide | Robotaxi services | Fastest fleet growth; Gulf-first expansion | ~$6B / ~$4B market cap | Uber partnership to 15 more cities |
Gaps that follow from the bottlenecks in the layer map rather than from a market-size spreadsheet. Every one of them sits below or beside layer 12, because that is where the constraints are and where nobody is competing for attention.
Gap: There is no MMLU for manipulation. Every vendor grades its own homework on a task it selected, which makes capability claims across Physical Intelligence, Skild, GR00T and Helix formally incomparable.
Pain: Buyers cannot run a bake-off. Investors cannot diligence a model claim. The absence is structural: no vendor will fund the benchmark that could rank it last.
Two or three people with a robotics PhD and a lab. Publish a hard, reproducible task suite with public leaderboards; monetise through certification and private evaluation contracts. The precedent is MLPerf, which became unavoidable.
Gap: 1X sells scheduled teleoperator sessions into private homes and Ottopia sells the transport layer, but nobody sells the back office: operator scheduling, hand-off protocols, session audit, quality scoring, and the compliance trail for someone looking through a robot's eyes inside a house.
Pain: Teleoperation is not a temporary crutch — The Robot Report's read is that it is 1X's path, not its stopgap. Every humanoid deployment in this map either uses it or hides it. Nobody has productised the operations around it.
Three-person team building the Zendesk of robot teleoperation. Sell per operator seat to every humanoid company running a pilot. Natural acquirer is any fleet-ops vendor in layer 11.
Gap: McKinsey's finding is that the gearbox constraint is qualification, not capacity. There is no independent lab that certifies a component as robotics-grade for humanoid duty cycles.
Pain: Every OEM re-runs the same qualification internally and slowly; suppliers cannot prove readiness without a customer. A shared testbed removes months from both sides.
Test rigs, a durability protocol and a published standard. Charge suppliers for certification and OEMs for access. Capital-light, credibility-heavy, defensible through accumulated failure data.
Gap: Amazon's Vulcan made news in 2025 for being able to feel what it gripped. Meta and GelSight do the best research; nobody manufactures tactile skin at volume and automotive-grade reliability.
Pain: Every manipulation roadmap in layer 10 assumes contact sensing the supply chain cannot deliver. Hands fail on durability under load, which is why demos quietly swap to grippers. The gate on dexterity is a component, not an algorithm.
$20–40M to industrialise a published sensing modality: yield, calibration, and a connector that survives a million cycles. Customers are every OEM in layer 12. Picks-and-shovels with a single-supplier endgame.
Gap: Every serious humanoid OEM builds its own hand rather than buying one — the classic signature of a missing supplier. Shadow and Wonik serve research; Sharpa is early. There is no Bosch of robot hands.
Pain: Hands are the highest-failure, highest-complexity subsystem, and duplicating them across a dozen OEMs is enormous waste. Vertical integration here is a symptom, not a strategy.
$30–50M to build a hand a third party will actually specify: published MTBF, standard mounting and bus, serviceable fingertips. Tesla and Figure will not buy — the next forty OEMs will.
Gap: ISO 10218:2025 admits humanoids into the industrial standard but explicitly excludes their mobility hazards. The IEEE study group has framed the gap and nobody has closed it. Insurers have no loss history to price against.
Pain: A humanoid that loses power falls over, and no standard governs that. Deployments stall at legal review, not technical review. Certification is the rate limiter on layer 12's entire thesis.
Assemble the actuarial dataset nobody has by instrumenting early fleets, then sell certification services and later the underwriting itself. Slow, unglamorous, structurally impossible to disintermediate.
Gap: Roughly 90% of magnet refining sits in one jurisdiction that has already used it as leverage — April 2025 export controls hit Optimus by Musk's own account. MP Materials is the only integrated Western mine-to-magnet operator.
Pain: Every Western robot programme, including defence, runs on this dependency, and there is no substitute at the same power density. Policy cannot tariff its way out of a materials position.
$500M+ and a decade, with permitting and offtake risk throughout. The only credible sponsors are sovereign-backed or defence-anchored. Anduril, Shield AI and the humanoid OEMs are natural offtake partners.
Gap: The constraint McKinsey names is the compact, high-durability, low-clearance subset humanoids need. Harmonic Drive Systems and Nabtesco hold decades of qualification data no entrant can shortcut.
Pain: If humanoid volumes reach even the low hundreds of thousands, the Western supply base cannot serve them. Leaderdrive is already setting the price, and it is not in the West.
$200M+ for tooling, metallurgy and a qualification programme measured in years. The buyer of last resort is a defence department that does not want its robots gated on Chinese reducers.
Gap: Warehouses already run Locus, Zebra, AutoStore and soon humanoids side by side. Formant and InOrbit sell into that, but nobody owns the interoperable data and safety layer across vendors.
Pain: Operators are accumulating incompatible fleets with no common telemetry, incident model or safety envelope. The integrator layer captures the margin precisely because this is unsolved.
$100M+ to build the standard and buy enough distribution to make it stick. A land grab: the winner becomes the control plane for physical labour the way Kubernetes did for compute.
| # | Opportunity | Tier | Rating | Key Rationale |
|---|---|---|---|---|
| 1 | Tactile Skin at Production Cost | Venture ($5–50M) | HIGH | Gates every manipulation roadmap in layer 10; the research exists, the manufacturing does not |
| 2 | Humanoid Safety Certification & Underwriting | Venture ($5–50M) | HIGH | ISO 10218:2025 excludes mobility hazards; deployments stall at legal, not technical, review |
| 3 | Manipulation Evaluation & Benchmarking | Bootstrappable (<$5M) | HIGH | No vendor will fund the benchmark that could rank it last; the MLPerf precedent |
| 4 | Non-China NdFeB Magnet Supply | Deep-Pocketed ($100M+) | HIGH | ~90% of refining in one jurisdiction that has already used it against Optimus |
| 5 | Dexterous Hands as a Merchant Component | Venture ($5–50M) | HIGH | Every OEM building its own badly is the signature of a missing supplier |
Both are horizontal, vendor-neutral and reach revenue without hardware risk. Benchmarking has the greater strategic leverage: whoever defines the measure defines the market.
Tactile skin and dexterous hands are both being vertically integrated out of necessity, not preference. The moat is manufacturing yield and accumulated failure data, not algorithms — and it compounds across every OEM in layer 12.
Magnets and robotics-grade reducers decide whether Western robotics scales at all. Slow, capital-intensive, permit-bound — and the only opportunities here that a competitor cannot replicate with software.