§1
Executive summary
Five findings from a fortnight that rearranged the sector.
The profitable robot industry is teleoperated, and nobody calls it robotics
Intuitive Surgical booked $2.89B in the second quarter of 2026, up 19%, and placed 468 da Vinci systems in three months. Symbotic carries roughly $22.7B of backlog. Amazon runs a million robots and says 75% of its deliveries are robot-assisted. None of this is autonomous in the sense the humanoid sector means: a surgeon drives every da Vinci motion. The machines that make money are purpose-built, work inside structured environments, and keep a human in the loop. The machines attracting capital are general-purpose, work in unstructured environments, and promise to remove the human.
Almost nobody in the West will tell you how many robots they have in the field
Figure is valued at $39B and discloses no unit count. Apptronik raised at more than $5B and its GXO engagement is still, in the company's own framing, a laboratory evaluation. Tesla targets 50,000 to 100,000 Optimus units in 2026 and Musk conceded in January that no Optimus robots are doing useful work at Tesla. The exception proves the rule: Agility Robotics, the lowest-valued of the leading Western humanoid companies at about $2.5B, discloses nine customer sites, four named customers and more than 65,000 operating hours — because it is going public and has to.
China ships; the West raises
AgiBot shipped an estimated 5,168 humanoids in 2025, more than any company in the world. Unitree was second and UBTech third; Chinese firms dominated global humanoid shipments outright. UBTech booked 13,361 orders for a consumer humanoid in under a month. Meanwhile China's industrial base already runs more than 2 million operational robots, 54% of global installations, and for the first time Chinese manufacturers outsold foreign suppliers at home. No Western humanoid company publishes a comparable shipment figure.
The United States closed its market nine days before China's champion went public
On 28 July 2026 the FCC added foreign-produced advanced robotic devices — humanoids and quadrupeds — to its Covered List, barring new imports on cybersecurity grounds, citing a critical vulnerability in Unitree hardware. On 30 July Beijing threatened countermeasures. On 6 August Unitree priced its Shanghai IPO at about $9.04B, raising roughly $904M. Unitree earned 13.3% of its 2025 revenue in the market that just shut. And NVIDIA had picked Unitree, weeks earlier, to build the reference humanoid for its own robotics platform.
The binding constraint is a supply chain the West does not control
High-torque actuators need NdFeB rare-earth magnets for power density. China controls roughly 69% of rare-earth mining and 90% of magnet processing and refining. Gearboxes are 30 to 50% of actuator cost. McKinsey's read is that aggregate gearbox capacity is adequate but the robotics-grade subset — compact, high-durability, low-clearance — is what has to be qualified and ramped. Musk has already said rare-earth export controls hit Optimus. A tariff wall against Chinese robots does not remove the dependence on Chinese inputs to build the alternatives.
How to read this report. Section 5 grades every deployment claim against five stated tests and sorts the sector into Shipped Piloted Demonstrated Announced. The rubric is printed in full so you can disagree with a placement rather than the whole table. Where sources conflict, section 7 records the conflict and the number chosen.
Eleven months, two industries
Capital events and evidence events, September 2025 to August 2026.
Sep – Dec 2025
Jan 2026
Feb – Apr 2026
Jun – Jul 2026
Aug 2026
Sources: Bloomberg, CNBC, TechCrunch, SCMP, The Robot Report, Electrek, TrendForce, FCC. Full list in §7.
§2
The stack: rare earth to deployment
Eleven layers. The interesting thing about a robot is that unlike a language model, most of its cost and nearly all of its risk sit below the software.
This is the briefing version. The full reference map — 14 layers, 181 companies, 84 sub-layers, a searchable company index and a force-directed knowledge graph, covering autonomous vehicles, drones and defence as well — is The Robotics Stack, report 05 in this series.
Rare earths and magnets
Permanent magnets are what make a joint strong for its weight. Every high-torque actuator in a humanoid depends on NdFeB magnets, and there is no substitute at the same power density.
Actuators and drivetrain
The muscles. Harmonic (strain-wave) drives dominate humanoid joints; planetary roller screws and frameless motors fill out the rest. Gearboxes alone are 30 to 50% of actuator cost.
Sensing
Force-torque sensing at the joint, depth and vision at the head, encoders everywhere. Tactile skin remains the least mature and the most often faked in demos.
Dexterous hands
Broken out from actuators 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.
Onboard compute
A robot has a power and thermal budget a data centre does not. NVIDIA's Jetson Thor pairs a Blackwell GPU with 64GB in a 40 to 70 watt envelope.
Simulation and world models
Where policies are trained before they touch hardware. NVIDIA's Newton physics engine, built with Google DeepMind and Disney Research and hosted by the Linux Foundation, is the closest thing to a neutral standard the layer has.
Data engines
The scarcest input in physical AI. Unlike language, there is no internet-scale corpus of robot actions, so every lab is building a collection apparatus: teleoperation rigs, data gloves, instrumented grippers, egocentric video.
Robot foundation models
The VLA layer, and the only part of the stack where the software industry's economics apply. Gemini Robotics 1.5 drives ALOHA, bi-arm Franka and Apptronik's Apollo with no robot-specific post-training and transfers skills between bodies zero-shot.
Fleet operations, safety and teleoperation infrastructure
The unglamorous layer that decides whether a pilot converts. Remote operations, incident handling, over-the-air policy updates, and the certification paperwork.
Bodies and integrators
The layer that gets the coverage and the valuations. Also the layer with the worst gross margins and the longest qualification cycles.
Deployment and robots-as-a-service
Where robots meet a purchase order. This is the layer where the profitable robot industry actually lives, and almost none of it is humanoid.
Whitespace, by funding tier
Gaps that follow from the bottlenecks above rather than from a market map.
| Gap | Tier | Evidence of pain | Incumbent risk |
|---|---|---|---|
| Teleoperation operations tooling |
Bootstrappable |
1X sells scheduled teleoperator sessions into private homes. Nobody sells the back office for that. | Low. Too small and too operational for the model labs to want. |
| Manipulation evaluation and benchmarking | Bootstrappable |
No MMLU for manipulation. Every vendor grades its own homework on a task it picked. | Medium. Mecka AI now sells evaluation infrastructure alongside its data, but a vendor selling evaluation to its own data customers is not the neutral benchmark this gap describes. |
| Tactile skin at production cost | Venture |
Amazon's Vulcan made news for having a sense of touch. In 2025. | Medium. Meta is researching it; nobody is manufacturing it at volume. |
| Dexterous hand supply as a component business | Venture |
Every OEM is building its own hand badly rather than buying a good one. | High if the OEMs vertically integrate, which Tesla already has. |
| Humanoid safety certification and insurance underwriting | Venture |
ISO 10218:2025 admits humanoids but not their mobility hazards. Nobody can price the residual risk. | Low. Certification bodies move slowly; underwriters have no loss history. |
| Non-China NdFeB magnet supply | Deep-pocketed |
90% of refining in one jurisdiction, which has already used it as leverage against Tesla. | High. Capital-intensive, permit-bound, and competing against an incumbent with a decade of cost advantage. |
| Robotics-grade gearbox qualification capacity outside Asia | Deep-pocketed |
The constraint McKinsey names: not volume, but the compact high-durability subset humanoids need. | High. Harmonic Drive Systems and Nabtesco have decades of qualification data no entrant has. |
Tiers follow the convention in the companion Stack Landscape report. Opportunity assessments are editorial judgement, not investment advice.
§3
The atlas
Twenty companies that determine what physical AI does next, grouped by the position they have taken rather than by the market they claim. Every card carries its evidence grade and its sources.
§4
Fault lines
Six questions the sector cannot answer the same way twice, and the money riding on each answer. Placements are editorial judgement from public statements and shipped behaviour.
1. General-purpose humanoid, or purpose-built body?
The humanoid argument is that the world is built for human bodies, so a human-shaped robot needs no environment redesign. The counter-argument is that every profitable robot ever sold has been purpose-built, and that a warehouse is cheaper to redesign than a hand is to build.
Apptronik is the tell: it raised at more than $5B on a humanoid and then launched Elevate Robotics to build non-humanoid robots.
2. End-to-end VLA, or a modular stack?
The received wisdom in 2026 is that the architecture debate is settled and everyone is training a VLA. That is true of the model labs and much less true of anyone with a system in production, where classical planning, force control and safety interlocks still do the load-bearing work.
Notice that the left-hand end of this rail contains every company with material revenue.
3. Where does the training data come from?
There is no internet-scale corpus of robot actions, so this is the real competitive question. Teleoperation produces the best data at the worst unit cost. Simulation is free and slightly wrong. Video is abundant and lacks actions. Factory telemetry is the best of all, and you have to own a factory.
1X appears twice, at opposite ends. That is not an error: it teleoperates for delivery today and trains a world model on video for tomorrow. Mecka AI sits off this rail entirely — it captures human motion from body sensors and phones with no robot in the loop at all, which is the cheapest position on the board and the one nobody else occupies.
4. Own the body, or own the brain?
Vertical integration buys margin and data; the horizontal play buys reach. NVIDIA has said the quiet part out loud — it wants to be the Android of generalist robotics, which casts every body maker as a handset vendor.
Gemini Robotics 1.5 driving Apptronik's Apollo with no robot-specific post-training is the strongest public evidence yet for the right-hand side of this rail.
5. Sell the robot, or sell the hours?
Selling a robot is a capital sale into a procurement cycle. Selling hours is a labour substitute with a comparison price everyone already knows. The second is easier to buy and much harder to deliver, because the vendor absorbs every failure.
Intuitive's instruments-and-accessories annuity is the model everyone else is trying to reinvent: sell the machine once, sell the consumable forever.
6. Can the West build a robot without China?
This stopped being rhetorical on 28 July. The FCC barred new foreign-produced humanoids and quadrupeds from the US market. It did not, and could not, bar the magnets, gearboxes and sub-assemblies that Western robots are built from.
The insulated end of this rail is occupied by a medical device company and by the Chinese incumbent. Neither is a humanoid startup.
Three ironies worth naming
NVIDIA picked a Chinese body for an American platform, weeks before the ban
The Isaac GR00T Reference Humanoid — the physical embodiment of NVIDIA's robotics platform — is to be built by Unitree and shipped in late 2026. That was announced on 1 June. On 28 July the FCC barred new Unitree-class devices from the United States, citing a critical vulnerability in Unitree hardware.
The most valuable robot company by revenue has zero autonomy
Intuitive Surgical earns more in a quarter than the Western humanoid sector has earned in total, and a human surgeon drives every single motion. The sector's central premise is that removing the human is where the value is. The evidence so far says the value is in making the human better.
The company with the best deployment evidence has the lowest valuation
Agility Robotics discloses nine customer sites, four named customers and more than 65,000 operating hours, and is going public at roughly $2.5B. Figure discloses none of those things and is valued at $39B. The difference is not performance. It is that one of them has to file.
§5
The reality ledger
Every deployment claim in this report, graded against five tests. The rubric is printed in full so you can disagree with a placement rather than with the table.
The five tests
- Units in the field. A verified count of robots doing work. Orders, pre-orders, letters of intent, reservations and production targets do not count.
- A named paying customer, with a duration attached.
- Autonomy share. What proportion of task time runs without a human teleoperating. Undisclosed is treated as unfavourable, because the companies that have good numbers publish them.
- Third-party verification. A customer statement, a regulatory filing, or on-site reporting. A vendor-produced video is not evidence.
- Repeatability. The same task across sites or shifts, with cycle time and accuracy disclosed.
Grades. Shipped passes four or five. Piloted passes two or three: real units doing real work, but bounded. Demonstrated passes one: capability shown outside a demo reel, no named customer in production. Announced passes none: orders, targets, reveals.
| Company | Grade | Units in the field | Named customers | Autonomy share | Valuation |
|---|
What is disclosed, and what is not
Autonomy share, by company. Hatched bars are undisclosed.
The pattern is the finding: the two companies that disclose an autonomy figure both disclose a low one, and both are commercially successful. Six of nine disclose nothing.
Humanoid shipments, 2025
Estimated units shipped. Hatched bars are companies that did not publish a figure.
Bloomberg reported that Chinese firms dominated global humanoid shipments in 2025, with AgiBot first, Unitree second and UBTech third. Bar widths for Unitree and UBTech reflect rank order, not measured volume. Since then AgiBot has reported producing its 15,000th unit and UBTech has booked 13,361 consumer orders — both are production and order figures, not shipments, and are kept separate for that reason.
Programmes that did not survive contact
Amazon Blue Jay — launched October 2025, shut down January 2026
A multi-armed sortation robot for same-day delivery facilities, developed in about a year, which is fast by Amazon's standards. It was cancelled inside four months of launch; the core technology was redirected and the team reassigned. Amazon has more robots deployed than anyone on earth, the best internal data, and complete control of the environment. It still killed the programme. Any model of the sector that cannot accommodate that is too optimistic.
Tesla Optimus — the gap between the target and the admission
Tesla is converting Model S and X production lines at Fremont to build Optimus, targets 50,000 to 100,000 units in 2026, and describes a path to 10 million a year at Giga Texas. In January 2026 Musk acknowledged that no Optimus robots were doing useful work at Tesla, having previously indicated otherwise. Both statements are on the record, three months apart.
Johnson & Johnson Ottava — what a regulated timeline actually costs
Delayed by two years, granted an investigational device exemption in late 2024, clinical study completed and presented in 2026, now filed for De Novo classification. This is what it takes to put a robot near a person under a regulator's eye. Humanoid pilots announced and converted inside twelve months are not operating under the same constraints — yet.
§6
The money
Robotics startups raised more in the first seven months of 2026 than in the whole of 2025, and more than in 2021, the previous peak. Where it went, and what it bought.
Robotics startup funding, global
Crunchbase category, which is wider than this report's scope: it includes autonomous vessels and defence.
Source: Crunchbase News. Note the discrepancy in §7, conflict 6: a widely repeated figure of $40.7B for 2025 does not reconcile with Crunchbase's $15B, and this report uses Crunchbase.
The ledger
Disclosed rounds of roughly $500M or more into physical AI since September 2025, plus the two public-market events. Rounds reported as under negotiation are listed but excluded from totals.
| Date | Company | Round | Amount | Valuation | Status |
|---|
Valuation, against what is publicly verifiable
Bars are valuation. The label is the strongest deployment evidence in the public record.
Valuations are post-money or implied by the transaction. The company with the most disclosed evidence is valued at one-fifteenth of the company with the least.
Revenue, on a single axis
Most recent disclosed period. Not annualised, not adjusted.
Unitree converted at approximately 6.75 CNY/USD. Intuitive's single quarter is roughly eleven times Unitree's entire 2025.
Where China sits in the bill of materials
Share of global capacity, by input.
Sources: McKinsey (magnets, gearbox cost), IFR World Robotics (installations, domestic share). The last bar is a cost share, not a country share, and is shown for scale.
§7
Method, conflicts & sources
Scope
Physical AI, drawn to exclude three adjacent industries that each deserve their own report:
- Autonomous vehicles — a separate regulatory regime, a separate capital structure, and a decade-longer history. Saronic appears once in the funding ledger, marked out of scope, only so the totals reconcile with Crunchbase's wider robotics category.
- Defence drones and autonomous weapons — procurement dynamics that have almost nothing in common with commercial robotics.
- Maritime and space robotics — different physics, different customers.
What is in: humanoids, mobile manipulation, warehouse and logistics automation, surgical, agricultural and industrial robotics, and the model, simulation, component and supply-chain layers beneath them.
Sourcing
This report was compiled on 7 August 2026 against a deliberate tier-1 restriction: company filings and earnings releases, exchange disclosures, FCC and ISO primary documents, and reporting from Bloomberg, CNBC, TechCrunch, South China Morning Post, Nikkei, Rest of World, The Robot Report, IEEE Spectrum, Crunchbase News, IFR, McKinsey and TrendForce.
Robotics has an unusually dense layer of search-optimised aggregator content that recycles and inflates figures. Several numbers that appear widely online — a $40.7B 2025 robotics funding total, a "$25 per operating hour" robots-as-a-service benchmark, teleoperation data costs falling from $340 to $118 an hour, and a set of precise Figure-at-BMW performance metrics — could not be traced to a tier-1 source and have been left out rather than repeated. Their absence from this page is deliberate.
Conflicts log
Where sources disagree, the disagreement is recorded rather than smoothed.
| # | Claim | The disagreement | What this report uses, and why |
|---|---|---|---|
| 1 | 2025 global robotics venture funding | $40.7B, up 74% year on year (widely repeated, tier 2) against $15.0B (Crunchbase News) | $15.0B. Crunchbase is tier 1 and internally consistent with its own 2021 and 2026 figures. The larger number most likely bundles autonomous vehicles, defence and late-stage corporate deals. |
| 2 | Agility deployment scale | "More than 100,000 totes at GXO" (tier 2) against "more than 65,000 operating hours across nine sites" (SEC filing and joint press release) | 65,000 hours. Tier 1, and hours across named sites is a harder measure than totes at one. |
| 3 | Physical Intelligence valuation | $5.6B closed (Nov 2025), ~$11B reported in talks (Mar 2026), "$11B target" (Jul 2026) | $5.6B as the last confirmed close, with the $11B flagged as reported-in-talks and excluded from totals. The July item recycles the March report; it is one round, not two. |
| 4 | Apptronik valuation | $5B (CNBC) against "more than $5.5B" (Bloomberg), same round, same day | $5B, the more conservative figure, with the conflict noted. Both outlets are tier 1; the difference is probably pre- versus post-money. |
| 5 | Harmonic drive supply | "Fewer than five companies can mass-produce harmonic drives" (tier 2) against McKinsey's finding that aggregate capacity is sufficient and the constraint is qualifying robotics-grade configurations | McKinsey's version. Tier 1, and it identifies a more precise mechanism. The popular framing overstates the scarcity and understates the qualification problem. |
| 6 | AgiBot cumulative units | 5,168 shipped in 2025 (Bloomberg) against a 10,000th and then 15,000th unit produced (company milestones) | Both, labelled. Produced is not shipped and shipped is not deployed. Collapsing the three is exactly the error §5 exists to prevent. |
| 7 | NVIDIA Jetson Thor compute | 1,200 TFLOPS against 2,000 TFLOPS in different NVIDIA-derived materials | Neither quoted as a headline figure. The discrepancy is almost certainly dense versus sparse or reduced-precision throughput. The 40 to 70 watt envelope is the number that matters for robot design and is consistent across sources. |
| 8 | Figure at BMW Spartanburg | Precise performance metrics (operating hours, parts handled, cycle time, placement accuracy) circulate widely but appear only in tier-2 outlets | Omitted. Tier-1 coverage confirms the deployment and the progression to Figure 03 at Leipzig, but not the metrics. |
Volatility tiers
High churn, verified first and expected to move within weeks: valuations, unit counts, funding status, the FCC exemption process, Unitree's post-listing price. Medium churn, months: deployment counts, model releases, production targets. Low churn, quarters to years: supply-chain concentration, safety standards, installed industrial base.
Watchlist — eight things that would change this page
- Unitree's post-IPO disclosure. A listed company must publish unit economics. It will be the first real gross-margin data anyone has for a humanoid.
- The Agility SPAC closing. A second set of audited humanoid financials, and the first for a US company.
- FCC exemptions. The rule allows them on a national-security finding. Who gets one, and on what basis, defines how real the wall is.
- Chinese countermeasures. Beijing has threatened them and has used rare-earth licensing before. Magnets are the obvious lever.
- Optimus at Fremont. Production was to begin in late July or August 2026. An actual unit count would be the single most informative number in the sector.
- UBTech U1 deliveries from 16 September. 13,361 orders convert, or they do not.
- Figure disclosing anything. A unit count, an autonomy rate, or a customer statement would move it up the ledger immediately.
- The first humanoid-specific safety standard. The IEEE study group has framed the gap; someone has to close it before humanoids work near people at scale.
Caveats
Valuations, unit counts, shipment estimates and revenue figures are press-reported or company-stated unless drawn from a filing, and are not independently verified. Evidence grades, fault-line placements and whitespace assessments are editorial judgement. August is a partial month. Nothing here is investment advice. This report is not affiliated with, endorsed by, or published by any company named.
Principal sources
| Source | Used for |
|---|---|
| FCC — Covered List fact sheet, national security determination, FAQ | The 28 July import restriction and its stated basis |
| SEC — Intuitive Surgical 8-K and 10-Q, Symbotic 10-Q, Churchill Capital Corp XI 8-K and 425 | Revenue, backlog, the Agility transaction |
| Bloomberg | Unitree IPO pricing, Figure and Skild valuations, 2025 humanoid shipment rankings, Hyundai |
| CNBC | Unitree pricing, Apptronik, NEURA, Amazon Vulcan, the FCC ban and China's response, Tesla |
| TechCrunch | Agility SPAC, Apptronik, Skild, Physical Intelligence, Amazon Blue Jay, NVIDIA platform strategy |
| South China Morning Post | Unitree prospectus and Q1 profits, UBTech U1 orders, US-China robotics rivalry |
| The Robot Report | Deployment detail across Agility, Figure, 1X, Apptronik, Boston Dynamics, NVIDIA, AgiBot; ISO and IEEE standards |
| IEEE Spectrum | FCC Covered List analysis, supply-chain consequences |
| Electrek | Optimus production timeline and the January admission |
| Crunchbase News | Robotics venture totals for 2021, 2025 and 2026 |
| McKinsey | Humanoid supply-chain constraints, magnet concentration, gearbox cost share |
| IFR World Robotics | Installed base, installations, Chinese domestic supplier share |
| TrendForce | Chinese humanoid output growth and shipment concentration |
| Google DeepMind | Gemini Robotics 1.5, ER 1.5 and On-Device capabilities |
| Rest of World, Nikkei Asia | Unitree filing, UBTech consumer launch |
| ISO | ISO 10218-1:2025 and 10218-2:2025 scope |
Per-claim source links are attached to each company card in §3 and to each round in the §6 ledger. Underlying data ships with this page and is also published as data/companies.json and data/rounds.json in the repository.