The Race for General-Purpose Physical AI

wpnews.pro

Abstract #

This paper examines the race to build general-purpose physical AI, meaning learned systems that can do open-ended physical work in unseen places and on different robot bodies. Using filings, research, standards and market data available in September 2026, it proposes a six-level scale of physical generality, P0 to P5, and places the field at the boundary between P2 and P3. The best systems reach 56% zero-shot success on household tasks in unseen homes, and about 19,100 humanoids shipped worldwide in the first half of 2026, more than 97% of them from Chinese vendors.

The race is harder than the language model race because of a data gap of several orders of magnitude, compounding reliability demands, costly hardware and slow evaluation. It is also more geopolitical. The United States leads in models and capital, China in manufacturing, components and deployment volume, and decoupling began with the FCC’s July 2026 restrictions on foreign-made humanoids.

The paper finds strong support for the view that physical AI will matter more than today’s AI, though mainly in the 2030s and 2040s. It finds conditional support for the view that infrastructure will be rebuilt around it, mostly compute, power, data, networks and standards rather than buildings. It finds partial support for humanoids as the main interface, as the flagship general-purpose body rather than the majority of robots. Safety and security will set the pace. The paper documents standards gaps for non-industrial robots and exploitable flaws in shipping products, and recommends mandatory reporting of human support, a separation between learned control and certified safety layers, a cybersecurity baseline for robot fleets, and allied supply chain resilience.

1. Introduction #

The first wave of generative AI learned to manipulate symbols. The next wave has to learn to manipulate matter. This paper argues that the race for general-purpose physical AI, meaning systems that perceive, reason and act competently across many unstructured physical tasks, will be harder, slower and more consequential than the race for language models.

Three observations support that claim. Most economic activity still involves moving, assembling, cleaning, inspecting or caring for physical things, so the value at stake is larger than the market for digital cognition. The problem is also harder, because there is no internet-scale corpus of physical action, errors cause damage rather than a wrong sentence, and control must happen in real time under uncertainty. Finally, once a machine can act in the world, safety and security stop being questions of content moderation and become questions of bodily harm, critical infrastructure and national resilience.

The paper tests three propositions. First, that general-purpose physical AI will matter more than today’s AI advances. Second, that most infrastructure will be built around it. Third, that humanoid robots will be its main interface. The evidence gives strong support to the first, conditional support to the second and partial support to the third.

The analysis uses company filings, peer-reviewed and preprint research, standards documents, government policy and third-party market estimates available as of September 2026. It weights regulatory filings and independent deployment evidence above vendor claims, and labels forecasts as forecasts.

2. Defining general-purpose physical AI #

General-purpose physical AI is a learned system that can carry out a broad, open-ended range of physical tasks, in environments it has never seen and on different robot bodies, with little task-specific engineering and at a reliability and cost that make it economically useful. Today’s systems meet parts of this definition but not all of it.

Four properties separate it from earlier automation.

Task breadth. It learns many tasks from instruction or demonstration instead of running one programmed routine.
Environment transfer. It works in places it has not seen, without collecting data on site first.
Embodiment transfer. One model can drive different bodies, from arms to wheeled bases to humanoids.
Production reliability. It completes work at rates close to a trained person, with a small and falling need for human help.

The term is narrower than “embodied AGI”. A system can be broadly useful in physical work long before it matches human cognition. Self-driving offers the better analogy, where the SAE J3016 levels gave industry and regulators a shared vocabulary. Google DeepMind’s Levels of AGI framework did the same for digital AI by separating performance from generality. Table 1 adapts that approach to physical work.

Table 1. Proposed levels of general-purpose physical AI

Level	Name	What it can do	Human support needed	Representative evidence, 2025 to 2026
P0	Programmed automation	Fixed motions in a structured cell	Engineers reprogram for each change	Welding and painting arms
P1	Narrow learned skill	One task class with learned perception	Integration at every site	Bin picking, tote handling
P2	Multi-task, known site	Several tasks at a site it was trained on	Site data collection and tuning	Figure 02 at BMW ,Digit at GXO
P3	Zero-shot, known domain	Familiar tasks in unseen sites of the same type	Frequent supervision, success well below human	Helix 2.5 , 56% across 30 unseen homes
P4	Open-ended general worker	New tasks across domains from instruction	Rare remote help, human-level reliability	Not yet demonstrated
P5	Human-parity generality	Any physical task a trained adult can do, at equal or lower cost	Normal management only	Not yet demonstrated

The field sits at the P2 to P3 boundary. Commercial deployments are mostly P1 and P2, and the best research results reach early P3 on a narrow set of household tasks. The distance from P3 to P4 is the race this paper is about.

Two measures should anchor any claim of progress, because demos rarely report them. The first is task success in unseen settings, reported per task with trial counts. The second is human support per unit of accepted work, such as interventions per hour or teleoperation minutes per completed task. This mirrors the disengagement reports California requires from autonomous vehicle testers.

3. State of the field in 2026 #

In 2026 physical AI moved from lab demos to early commercial scale, but it is still a market of tens of thousands of humanoids, not millions. Four developments define the year.

3.1 Robot foundation models converged on one recipe

The leading labs now build vision-language-action (VLA) models on pretrained multimodal backbones, pair them with a fast low-level controller, and train on a mix of teleoperation, human video and simulation. Table 2 lists the main releases.

Table 2. Major physical AI model and platform releases, 2026

Date	Release	Main claim
Sep 2026	Figure Helix 2.5	56% zero-shot success on tidying, towel folding and bed making in 30 unseen homes, versus 9% without pretraining on Figure’s Index dataset of human behaviour
Aug 27, 2026	Anthropic Model Hardware Standard (research preview)	Shared drivers and device discovery so AI agents can run lab and factory hardware, with device-level safety limits and human approval for high-risk steps
Jul 30, 2026	Google DeepMind Gemini Robotics 2 , ER 2 and On-Device 2	Whole-body control on the Apptronik Apollo 2 humanoid, multi-robot coordination, and the ASIMOV-Agentic safety benchmark
Apr 16, 2026	Physical Intelligence π0.7	Compositional generalisation, plus laundry folding on a robot it had no training data for
Mar 16, 2026	NVIDIA Cosmos 3, GR00T N1.7 and GR00T N2 preview	An open world foundation model and humanoid models, with N2 due at the end of 2026

The common thread is broad pretraining followed by small amounts of task data. Figure says Helix 2.5 needed half the task-specific data of an earlier behaviour, and DeepMind says On-Device 2 adapts with fewer than 200 examples. This is the scaling logic that produced large language models, applied to action.

The gap to production is still wide. Gemini Robotics 2 reports 32% to 92% success on multi-finger dexterity tasks, and Helix 2.5 reaches 56%. By contrast, Figure’s own BMW deployment targeted over 99% placement success per shift.

3.2 Shipments are growing fast from a small base

Omdia estimates that about 13,000 humanoids shipped worldwide in 2025, led by AgiBot, Unitree and UBTECH. Smart Analytics Global (SAG) estimates 19,100 units in the first half of 2026 alone, up 272% year on year. SAG puts Chinese vendors at more than 97% of shipments and projects close to 60,000 units and about $1.6 billion of revenue for the full year.

Smart Analytics Global, 1H 2026 shipment estimates

Estimates disagree, which is normal in a young market. Omdia put Unitree at 4,200 units for 2025, while Unitree reported more than 5,500, and its prospectus data, as summarised by the Shanghai Stock Exchange , shows 5,632 humanoids sold from 2023 to 2025. Filings should outrank press releases. For scale, Merics notes that China made about 12,800 humanoids in 2025 against roughly 556,000 industrial robots a year.

3.3 Capital is concentrating in the robot brain

Crunchbase counts $15 billion of robotics venture funding in 2025 and $18.8 billion by late June 2026, already above the previous annual record of $14.1 billion set in 2021. The largest valuations sit with model builders and vertically integrated humanoid makers.

Figure closed a Series C of more than $1 billion at a$39 billion valuation- Skild AI raised$1.4 billion at over $14 billion- Physical Intelligence raised$1 billion at $11 billion- Agility Robotics agreed aSPAC merger- Unitree listed on Shanghai’s STAR Market in August 2026, raising about 6.1 billion yuan on 2025 revenue of 1.70 billion yuan and net profit of 278 million yuan.

The general-purpose AI labs are also moving in. In September 2026 Sam Altman said OpenAI “will definitely do a humanoid”, while Tesla is converting its Fremont Model S and X line for Optimus after Elon Musk conceded in January 2026 that no Optimus units were yet doing useful work in Tesla factories.

3.4 Deployments are still pilots, but they now publish output

Figure 02 at BMW Spartanburg ran for 11 months, logged more than 1,250 hours, loaded over 90,000 parts and contributed to more than 30,000 X3 vehicles. The forearm was the main failure point, which led Figure to redesign the wrist electronics for Figure 03(Figure ) .
Agility Digit moved more than 100,000 totes at a GXO site in 2025 and has logged over 65,000 operating hours across nine customer facilities, on a bill of materials of $125,000 for version 4(Agility ,The Robot Report ) .
1X NEO is one of the first humanoids marketed to households, at $20,000 or $499 a month, with a teleoperated “expert mode” for tasks it cannot yet do alone(TravTeks ) .

The best-documented deployments measure output in thousands of hours, not millions. Most also rely on human support that vendors rarely quantify.

4. Why this race is harder #

Physical AI is harder than language AI for six connected reasons. Each one slows progress, and each one also creates advantages that are hard to copy once won.

Moravec’s paradox. In 1988 the roboticist Hans Moravec observed that computers find adult-level reasoning comparatively easy and a toddler’s perception and mobility very hard. Four decades later the pattern still holds. Models that pass professional exams still struggle to fold an unfamiliar shirt.

The data gap. Meta trained Llama 3 on over 15 trillion tokens of public text. Stanford’s 2026 emerging technology review notes that the largest robot datasets top out around 1 million episodes covering about 150,000 tasks. The Open X-Embodiment effort needed 21 institutions and 22 robot types to pool its data. Labs are closing the gap three ways, and each has a flaw.

Teleoperation fleets give clean action labels but cost human hours per robot hour.
Human video scales cheaply, and Figure says its Index dataset adds about 35 minutes of human experience every second, but video lacks force and joint data.
Simulation and world models such as NVIDIA’s Cosmos 3 multiply data, but still miss much of the physics of contact and soft materials.

Reliability compounds. A task made of many steps fails if any step fails. If each step succeeds with probability p, a task of n steps succeeds with the probability shown below.

P_{\text{task}} = p^{,n} At 99% per step, a 100-step task succeeds only 37% of the time. At 99.9% it succeeds 90% of the time, and reaching 99% per task needs about 99.99% per step. A chatbot’s error can be edited. A dropped pan or a missed pallet has a physical cost. This arithmetic is why the step from P3 to P4 in Table 1 is so steep.

Simulation and safe exploration. Stanford’s review warns that simulations often fail to capture the unpredictability of real environments and must be calibrated against real data. Robots also cannot explore freely the way software agents explore a codebase, because exploration breaks things and can hurt people. Physical Intelligence’s π*0.6, which its model card describes as π0.6 further improved through real-world reinforcement learning, is an early attempt to learn from deployment without unsafe trial and error.

Hardware, cost and energy. Stanford puts the average humanoid selling price at about $200,000 in 2024. Agility’s Digit v4 carries a $125,000 bill of materials. Merics finds Chinese humanoids average 300,000 to 500,000 yuan, against a commercial threshold of about 160,000 yuan, and that Schaeffler, THK and NSK supply 90% of the specialised ball screws. Hands and wrists fail first, as Figure’s BMW pilot showed. Battery life limits shifts, which is why Figure 03 charges wirelessly through its feet at 2 kW.

Evaluation is slow and costly. A language model can be scored on a benchmark in minutes. A robot policy needs robot-hours in varied real settings, so most results are self-reported and small. Figure’s 30-home test was large by robotics standards, yet it covered only three tasks.

The result is that progress will look slower and lumpier than in language models. But the same barriers create moats. Deployment data, manufacturing scale and safety records are much harder to copy than model weights.

5. The competitive landscape #

The race has two poles with opposite strengths. The United States leads in frontier models, chips and venture capital. China leads in manufacturing, component supply, deployment volume and state coordination. Other economies are positioning as component suppliers, niche specialists or rule-setters.

Table 3. Comparative position of the main players, September 2026

Region	Main strengths	Main weaknesses	Policy stance
United States	Frontier robot models (Google DeepMind, NVIDIA, Physical Intelligence, Skild AI, Figure, now OpenAI), AI chips, deep private capital	Adopts only about 70% of the robots its wage levels predict and ranks 13th on that measure, with no national robotics strategy (ITIF )	Security-led. New foreign-produced humanoids and quadrupeds need conditional approval since July 2026
China	More than 97% of 2026 humanoid shipments (SAG), 54% of world industrial robot installations in 2024, 63% of key humanoid component suppliers, about 70% of rare earth production	Reliance on the NVIDIA software ecosystem and on German and Japanese precision ball screws (Merics )	Industrial policy. Embodied AI is a priority of the 15th Five-Year Plan , alongside state and regional venture funds targeting 1 trillion yuan over 20 years
Europe	Precision components (Schaeffler), industrial robotics know-how, Neura Robotics’ Series C of up to $1.4 billion (Crunchbase )	Far less capital and scale than the US or China	Rule-setting through the Machinery Regulation and the AI Act
Japan and South Korea	Precision reducers and ball screws (THK, NSK), the world’s highest robot density in Korea at about 1,100 robots per 10,000 workers (Stanford )	Few frontier model builders	Korea’s K-Humanoid Alliance , launched April 2025, targets a shared robot AI model and a sub-60 kg commercial humanoid by 2028

Why this race differs from the language model race. Frontier language models needed compute, talent and web data, and those were concentrated in a handful of US labs. Physical AI also needs factories, component supply chains and fleets in the field, and China holds much of that base. Deployed robots generate the real-world data that improves the next model. A volume lead in hardware can therefore turn into a data lead in software, which makes China’s shipment share more important than its unit counts suggest.

Each side holds a chokepoint over the other. China’s April 2025 export controls cover rare earth magnets containing dysprosium and terbium, which sit inside most high-performance actuators. A broader October 2025 expansion was suspended for one year after a US-China trade deal (China Briefing ). The US controls leading AI chips, while Germany and Japan control precision ball screws. No bloc can yet build a frontier humanoid alone.

Decoupling has started. On July 28, 2026 the FCC added networked humanoids and quadrupeds produced in any foreign country to its Covered List, so new models cannot be authorised for US sale without conditional approval from the Department of War (FCC fact sheet ). Previously authorised models are unaffected. The proposed GUARD Act would go further against robots from foreign adversaries. The likely outcome is two partly separate ecosystems with different standards, data rules and suppliers.

Bubble risk is real on both sides. China’s NDRC counts about 150 humanoid makers and has warned about repetitive product launches. Unitree’s offline IPO tranche was 2,760 times oversubscribed. In the US, Figure is valued at $39 billion on limited disclosed revenue. A shakeout would not end the race, but it would favour players with revenue, manufacturing and data rather than demos.

6. Humanoids as the main interface #

Humanoids are likely to become the main interface for general-purpose work in spaces built for people, but not the main form of physical AI by unit volume. The model, not the body, is the real platform, and it will run on many bodies.

6.1 The case for humanoids

The world is built for human bodies. Doors, stairs, tools, shelves and vehicle controls assume human reach and hands. As Sam Altman put it in September 2026, “we’ve kind of built this world for people”, so “matching that form factor seems good”(Humanoids Daily ) .
Human data transfers best to human-like bodies. The cheapest large source of physical data is video of people working. Figure’s Index dataset and its 6x zero-shot gain in Helix 2.5 depend on that match.
One platform spreads the cost. A body that can do many jobs lets makers amortise design, certification and manufacturing over one high-volume product, as Figure plans with aBotQ- People read human form. Gaze, gesture and posture signal intent, which matters in homes, hospitals and shops.

6.2 The case against

Most physical AI already runs on other bodies. Amazon passed1 million robots 500,000 paid rides a week- Specialised forms win structured jobs. The IFR expects industrial robots to remain “the backbone of high-speed, precision-driven manufacturing environments”(IFR ) . Stanford’s review notes simpler robots often beat humanoids on specific tasks.
Legs add cost and risk. A walking robot must actively balance, and a fall is a hazard in its own right. That is why ISO/TC 299 is draftingISO 25785-1- Portable brains weaken the lock-in. π0.7, Skild AI and Gemini Robotics 2 are built to run across bodies. If the model moves freely between forms, buyers will pick the cheapest body that does the job.

6.3 The likely mix

The most probable outcome is a layered market. Specialised machines, including arms, mobile bases, autonomous vehicles and drones, will dominate unit volume in structured settings. Wheeled humanoids will fill semi-structured sites such as warehouses, retail and labs. Legged humanoids will serve as the general-purpose interface where the environment cannot be redesigned, including homes, care settings, brownfield factories and disaster sites.

In that sense the humanoid is the flagship interface, not the whole fleet. Altman makes the same point, saying body design is “less important than really figuring out like the brain that makes the robot work”.

7. Infrastructure built around physical AI #

Physical AI will reshape the digital, energy and industrial layers of infrastructure far more than the built environment. The claim that nearly all infrastructure will be built around it holds well for compute, data, networks and standards. It holds only weakly for buildings, because humanoids are chosen precisely so that buildings do not have to change.

Table 4. Infrastructure layers and what physical AI demands of them

Layer	What physical AI needs	Evidence today	Pace of change
Training compute and power	Capacity for world models, simulation and VLA training	Data centres used 415 TWh (1.5% of world electricity) in 2024, heading for about 945 TWh by 2030 (IEA )	Fast
Onboard compute	Real-time inference inside every robot	NVIDIA Jetson Thor offers 2,070 FP4 teraflops in 130 W (DCD ) . π0.6 needs 63 ms per action chunk on an H100	Fast
Data pipelines	Teleoperation centres, human-video capture, simulation farms, fleet upload	Figure 03 offloads fleet data over 10 Gbps mmWave	Fast
Remote operations	Human operators who handle exceptions	1X NEO relies on a teleoperated expert mode in homes	Fast
Interoperability standards	Common interfaces between models, robots and machines	Anthropic’s MHS preview, VDA 5050 for mobile robot fleets, ROS 2	Medium
Site connectivity and charging	Low-latency private networks, charging without human help	Figure 03 charges inductively at 2 kW. Naver’s 1784 office links 110 robots to a cloud brain over local 5G (Korea Times )	Medium
Manufacturing base	Actuators, magnets, reducers and batteries at automotive scale	Figure’s BotQ line starts at 12,000 units a year. China hosts 63% of key component suppliers	Medium
Buildings	Flat floors, robot elevators, digital twins	Naver 1784 opened in 2022 with a robot-only lift, but about 80% of the buildings in use in 2050 already stand today (WorldGBC )	Slow

The humanoid paradox. The more successful humanoids become, the less the physical world needs to be rebuilt for machines. So the infrastructure built around physical AI will be mostly invisible, made of compute, power, data, networks, standards and remote operations centres. New sites such as warehouses, factories and offices will be designed robot-first, and there specialised robots will often beat humanoids.

Robots will build and maintain the infrastructure that trains them. Altman says OpenAI’s near-term robotics focus is industrial infrastructure and data centre automation. Inspection of grids, pipelines, rail and data halls is a natural early market, and it creates a loop in which robots expand the compute that improves the next robots.

Robots become critical infrastructure themselves. A fleet of thousands of machines tied to one cloud model shares one set of failure modes. An update, an outage or a compromise then touches every site at once, which is the subject of Section 9.

8. Safety #

Physical AI breaks the core assumption of machine safety, which is that a machine’s behaviour is fixed, known in advance and fenced off from people. General-purpose robots move among people, change behaviour with each model update and take instructions in natural language. Agility’s chief executive calls safety “the biggest blocker to adoption” (The Robot Report ).

8.1 Three layers of safety

Physical and functional safety. Limits on force, speed, energy and workspace, enforced by certified controllers and safe-stop functions. This is the layer existing standards know how to handle.
Behavioural safety of the learned policy. Whether the model does the right thing in a situation it has never seen, such as refusing to hand a knife blade-first or stopping when a child walks in. DeepMind’s ASIMOV-Agentic benchmark and the human-detection safe stop in Gemini Robotics 2 target this layer.
Operational safety. Deployment limits, supervision, incident reporting and control over model updates. This layer decides whether a fleet stays safe after the first day.

8.2 The standards gap

Table 5. Key safety instruments for general-purpose robots

Instrument	Scope	Status, September 2026	Gap for physical AI
ISO 10218-1 and –2:2025	Industrial robots and robot cells	Published February 2025. Absorbs the ISO/TS 15066 cobot guidance and adds cybersecurity where it affects safety	Written for fixed, programmed arms, not mobile learned systems
ISO/CD 25785-1	Dynamically stable industrial mobile robots, legged or balancing	Committee draft. Comment period closed July 8, 2026	Excludes non-industrial uses, so homes and public spaces are not covered
IEEE Humanoid Study Group pathway report	Classification, stability and fall response, human-robot interaction	Published September 2025	A roadmap, not a standard. Expects 18 to 36 months of standards work
EU Machinery Regulation 2023/1230	All machinery sold in the EU, including AI-driven safety functions	Applies from January 20, 2027	After the Digital Omnibus (Regulation 2026/1744), AI-specific obligations move into this regulation by delegated act by August 2, 2028

The pattern is clear. Industrial settings will have usable standards around 2027. Homes, care settings and public spaces, where the humanoid argument is strongest, have none in prospect.

8.3 Open problems

Contact force. A full-size humanoid near people carries real kinetic energy. In November 2025 a former Figure safety engineer sued, alleging the company’s robots could generate force sufficient to fracture a human skull, and the OECD AI Incidents Monitorlogged the case . The claims are allegations, not findings, but they show the stakes.
Falls. A robot that balances actively can fall onto a person or a child. The IEEE group asks for measurable stability metrics and fall-response tests that do not yet exist.
Verifying learned policies. A neural policy cannot be tested exhaustively. Assurance has to rest on runtime monitors and hard limits that do not depend on the model behaving well.
Updates. An over-the-air model update can change behaviour across a whole fleet overnight. Certification has to cover the update process, not just a frozen product.
Human fallback. Teleoperators are today’s safety net. Their workload, training and latency become safety-critical once one operator supervises many robots.

8.4 Separate the brain from the guardian

The most robust approach is architectural. A general-purpose model proposes actions. A simpler, certifiable safety layer, which is not learned, bounds force, speed, zones and energy whatever the model commands. Anthropic’s Model Hardware Standard applies the same idea at the interface, with device- level safety limits and human approval for high-risk steps. Regulators should require this separation for any robot that shares space with people.

9. Security #

Security is where physical AI differs most from digital AI. A compromised chatbot leaks data. A compromised robot fleet can see, listen and move inside factories, homes and critical sites. Serious flaws have already been found in products on sale today.

9.1 Attack surfaces and documented cases

Table 6. Security threats to general-purpose robots

Attack surface	What can go wrong	Documented example
Wireless setup and firmware	Takeover with root access from nearby	Every Unitree G1 shared the same hard-coded AES key, letting an attacker in Bluetooth range inject commands as root (Help Net Security , October 2025)
Telemetry and data residency	Covert surveillance of a site or home	The G1 sent battery state, joint torque, camera video, microphone audio, lidar and GPS data to servers in China every five minutes without notice to users (same source)
The model itself	Jailbreaks and visual prompt injection that make the robot do harm	Penn’s RoboPAIR often reached 100% attack success on three LLM-controlled robots, including the first jailbreak of a commercial robot, a Unitree Go2
Fleet and cloud	One exploit or bad update spreads across many robots	Recorded Future warns wireless exploits allow lateral movement between nearby robots, a “physical botnet”
Teleoperation channel	A remote control path that insiders or intruders can abuse	Home robots such as 1X NEO rely on remote operators looking through the robot’s cameras
Supply chain	Hidden access built in by a vendor or state	The FCC’s July 2026 action cites risks of data collection and remote commandeering

9.2 Fleet scale turns flaws into systemic risk

In digital AI, a jailbreak usually affects one conversation. In physical AI, a policy flaw or a poisoned update replicates across thousands of bodies at once, in many locations. That makes failures correlated rather than independent, which is the defining property of systemic risk. Insurers, grid operators and hospitals will treat robot fleets the way they treat industrial control systems, and they will be right to.

Physical AI is also dual-use. Recorded Future notes a reported robot-only assault on a Russian position by Ukrainian forces in April 2026. The same advances in locomotion and manipulation that serve warehouses serve battlefields.

9.3 Security is now geopolitics

The FCC order and the proposed GUARD Act show that robot security is following the path of telecom security. Buyers will increasingly ask where a robot was made, where its data goes and who can push updates to it. Trusted supply chains, local data residency and verifiable firmware will become purchasing criteria. This will fragment the market, raise costs outside China, and make security a competitive feature rather than a compliance cost.

9.4 Baseline controls every deployment should require

Unique keys per device, secure boot, and signed firmware and model weights, with no hard-coded credentials.
Disclosed, minimal telemetry with user control and a choice of data residency.
Staged, signed model updates with canary fleets and rollback.
A physical emergency stop and local safe mode that work without any network.
Authenticated, logged and consented teleoperation.
Regular red-teaming of robot models for jailbreaks and physical prompt injection.
Network segmentation and monitoring of outbound telemetry, as Recorded Future recommends.

10. Economic and labour implications #

In the long run general-purpose physical AI should change labour markets more than digital AI, because it reaches the work that digital AI cannot. In the near term the reverse is true, and the published forecasts differ by an order of magnitude.

10.1 Digital AI captures more value first

The McKinsey Global Institute estimates that current technology could in theory automate about 57% of US work hours, 44% through software agents and only 13% through robots. Activities that need both physical and cognitive skills make up about 35% of US work hours. In its midpoint scenario, automation yields about $2.9 trillion a year of US value by 2030, with robots contributing 23%.

The gap between the 35% and the 13% is the prize. It is locked behind the dexterity and situational awareness that separate P3 from P4 in Table 1. The claim that physical AI will matter more than today’s AI is therefore a claim about the 2030s and 2040s, not about this decade.

The World Economic Forum’s Future of Jobs Report 2025 makes the same point from the other side. It expects 170 million jobs created and 92 million displaced by 2030, with the largest growth in frontline roles such as farmworkers, delivery drivers and construction workers, plus care. Those are exactly the jobs general-purpose physical AI would eventually reach.

10.2 Forecasts diverge widely

Table 7. Selected humanoid market forecasts

Source	Published	Forecast
Goldman Sachs	September 15, 2026	75,000 units in 2026, 890,000 in 2030 and 6.5 million in 2035, a $138 billion market. Up from 1.38 million units and $38 billion previously
Smart Analytics Global	2026	Close to 60,000 shipments and about $1.6 billion of revenue in 2026
Morgan Stanley	May 14, 2025	Over 1 billion humanoids in use by 2050 and a $5 trillion market including supply chain and services. 90% industrial and commercial. Prices fall from about $200,000 in 2024 to $50,000 by 2050, with slow adoption until the mid-2030s

These are scenarios, not order books. Goldman’s own note stresses that robots must still show the reliability, operating cost and task coverage needed for millions of annual sales. The more useful question for decision-makers is which leading indicators would confirm the steep path (Section 11).

10.3 Who gains and who loses

Workers in physical roles face exposure later than office workers but on a larger scale, starting with logistics, manufacturing, cleaning and agriculture, then care and construction.
New jobs appear in the human support layer. Teleoperators, robot trainers, fleet technicians, safety assessors and data curators already exist, and early deployments depend on them.
Remote physical work becomes tradable. Teleoperation lets a person in one country do physical work in another. That opens a new export for lower-wage economies, and new questions about labour standards and data rights.
Low-wage manufacturing economies lose part of their cost advantage if robot labour costs converge worldwide. Countries that build robots, components or models capture the value. Those that only buy them capture productivity but not the industry.
Ageing societies gain most. China, Japan, South Korea and much of Europe face shrinking workforces, which turns robots from a threat to jobs into a response to labour shortages.

10.4 Where the value will settle

Two models compete. If robot brains become portable across bodies, hardware commoditises as PCs and Android phones did, and value settles in models, data and fleet operations. If tight coupling of body and brain matters more, vertically integrated makers such as Tesla, Figure and Unitree capture it, as Apple did. The evidence so far, including cross-embodiment models and Anthropic’s hardware standard on one side and OpenAI’s move into its own humanoid on the other, supports both. A likely outcome is an integrated premium tier plus an open, commoditised volume tier.

11. Scenarios for 2026 to 2040 #

The most likely path is a long plateau, in which physical AI spreads quickly through structured work with heavy human support while general-purpose household robots take until the late 2030s. Three capability scenarios and two overlays capture the range. The probabilities are the author’s judgement, not model outputs.

Table 8. Capability scenarios

Scenario	Probability	What happens	Signature by 2030
Steep curve	25%	Action scaling works like language scaling. Deployment data compounds, and P4 arrives in logistics and manufacturing around 2030	Shipments near Goldman’s 890,000 a year. Published intervention rates fall year after year
Long plateau	55%	P3 systems spread through warehouses, factories and labs with remote support. Homes wait for another hardware and safety cycle. A funding shakeout around 2027 to 2028 consolidates the field	Hundreds of thousands of units a year, mostly industrial, consistent with Morgan Stanley’s slow adoption until the mid-2030s
Stall	20%	Reliability and data limits prove harder than expected. The humanoid boom deflates, while specialised robots keep growing	Humanoid shipments flatten. Capital shifts back to arms, mobile bases and vehicles

Two overlays can occur in any scenario.

Split world. US and Chinese ecosystems separate, with different standards, data rules and suppliers, and other countries choose between them. The FCC’s July 2026 action makes some version of this likely.
Safety or security shock. A serious injury in a home or a fleet-wide compromise triggers moratoria or licensing, delaying adoption by years. Robotaxi programmes went through the same thing after a fatal crash and later permit suspensions.

Table 9. Leading indicators to watch

Indicator	Reading today	What would confirm the steep curve
Zero-shot success in unseen sites, with trial counts	56% on three tasks in 30 homes (Helix 2.5)	Above 90% on broad task suites across hundreds of sites
Human support per accepted task	Rarely disclosed	Audited, falling intervention and teleoperation rates
Unit cost	About $200,000 average price in 2024. Digit v4 bill of materials $125,000	Volume prices near $50,000 for industrial units
Paying shipments and repeat orders	About 19,100 units in 1H 2026	Repeat fleet orders in the thousands from non-affiliated customers
Hand and wrist reliability	The forearm was the main failure point at BMW	Published mean time between failures in the thousands of hours
Standards	ISO 25785-1 at committee draft	Published industrial standard and a started home-robot standard
Supply chain	China’s broader rare earth controls suspended for one year from November 2025	Stable magnet licensing or credible non-Chinese capacity

Dates to watch. China’s one-year suspension of its October 2025 rare earth rules runs out in November 2026. NVIDIA expects GR00T N2 at the end of 2026. The EU Machinery Regulation applies from January 20, 2027, and AI-specific obligations are due to be folded into it by August 2, 2028.

12. Recommendations #

The priority for every actor is the same. Measure real autonomy, separate learned intelligence from certified safety, and treat robot fleets as critical infrastructure before they become it.

12.1 For policymakers

Require disclosure of human support. Robots deployed in commercial or public settings should report interventions and teleoperation time per task, as California requires disengagement reports from autonomous vehicle testers.
Mandate a brain and guardian architecture. Any robot sharing space with people should have a certified, non-learned safety layer that bounds force, speed, zones and energy regardless of model output.
Close the home and public-space standards gap. Fund test beds, start a standard for non- industrial humanoids now, and create an incident reporting channel modelled on the OECD AI Incidents Monitor.
Set a cybersecurity baseline. Unique device keys, signed firmware and models, a software bill of materials, disclosed telemetry and a local emergency stop should be conditions of sale and public procurement.
Secure the supply chain with allies. Build allied capacity in magnets, actuators and precision components. Target controls at identified risks, since blanket rules on all foreign-made robots also block trusted partners.
Prepare the workforce. Fund training for robot technicians, trainers and remote operators, and update labour and data law for teleoperated work that crosses borders.
For smaller and emerging economies, pick a niche. Data collection, teleoperation services, component manufacturing and sector-specific deployment, such as agriculture and mining, are realistic entry points. Adopting international standards is cheaper than writing new ones.

12.2 For industry

Publish per-task success rates with trial counts and unseen-site conditions, not just videos.
Report intervention rates and support cost per accepted task to customers, and let them audit it.
Build security in from the first prototype, and run a coordinated vulnerability disclosure programme.
Ship model updates in stages, with canary fleets and rollback.
Invest in hand and wrist durability, and publish mean time between failures.
Adopt open interfaces such as ROS 2, VDA 5050 and the Model Hardware Standard to cut integration time.

12.3 For investors

Weight third-party deployment evidence and repeat orders above demos and vendor claims.
Model unit economics fully, including bill of materials, price, uptime and human support cost per hour.
Price in regulatory and geopolitical risk, from the FCC Covered List to rare earth licensing.
Favour companies that own a data loop from paying deployments, since that is the hardest asset to copy.

13. Conclusion #

The race for general-purpose physical AI is real, it is harder than the language model race, and it will be won on trust as much as on capability. Each of the paper’s three propositions holds, with limits.

It will matter more than today’s AI, but later. Physical and mixed physical-cognitive work is about 35% of US work hours, and the fastest-growing jobs to 2030 are frontline and care roles. Digital agents still capture most automation value this decade, so physical AI’s larger impact belongs to the 2030s and 2040s.
Infrastructure will be rebuilt around it, mostly out of sight. Compute, power, data pipelines, networks, standards and remote operations will be reshaped. Buildings will change slowly, because most of the 2050 building stock already exists and humanoids are designed to fit it.
Humanoids will be the flagship interface, not the whole fleet. They are the best general-purpose body for spaces built for people. Specialised robots will still outnumber them, and portable robot brains will run on both.

Safety and security are not side issues in this race. They decide its pace. The documented flaws in shipping robots, the absence of standards for homes, and the correlated risk of cloud-connected fleets mean that one serious incident could set the field back years.

The winner will not be the first company to show a robot doing everything once. It will be the first to show robots doing useful work safely, securely and reliably, thousands of times, in places they have never seen.

14. References #

Sources are listed alphabetically by publisher and graded by evidence type. Vendor claims are self-reported and have not been independently verified.

Source	Publisher	Date	Evidence type
Previewing the Model Hardware Standard	Anthropic	Aug 27, 2026	Vendor announcement
Digit moves over 100,000 totes	Agility Robotics	Nov 20, 2025	Vendor claim
Amazon deploys over 1 million robots	Amazon	Jul 1, 2025	Vendor claim
Autonomous vehicle disengagement reports	California DMV	Ongoing	Regulator
China’s rare earth export controls	China Briefing	2025 to 2026	Third-party analysis
Robotics venture funding surges to record in 2026	Crunchbase News	Jun 22, 2026	Third-party data
Nvidia launches Jetson Thor for humanoid robots	DatacenterDynamics	Aug 2025	Media report
Tesla pushes Optimus V3 reveal later this year	Electrek	Apr 22, 2026	Media report
Machinery Regulation 2023/1230 and the Digital Omnibus on AI	Eurogip	2026	Third-party legal analysis
FCC updates Covered List to include foreign-produced advanced robotic devices	FCC	Jul 28, 2026	Regulator
F.02 contributed to the production of 30,000 cars at BMW	Figure	Nov 19, 2025	Vendor claim

| [Figure Series C](https://www.figure.ai/news/series-c) | Figure | Sep 16, 2025 | Vendor announcement | 
| [Introducing Figure 03](https://www.figure.ai/news/introducing-figure-03) | Figure | 2025 | Vendor claim |

source & further reading

humanoidanalytics.com — original article General-Purpose Physical AI May Need More Data and Compute Than LLMs XPENG IRON’s Catwalk Gait Points to a Fashion-Show Use Case Humanoid Robotics Has a Training Data Market, but Not Yet a Commodity Market