Embodied AI · Global Market Deep Dive
As of June 2026
The Humanoid
Ascendancy
The race won't be won by whoever builds the best robot — it'll be won by whoever controls its muscles. A value-chain analysis of the new industrial frontier.
The argument
Embodied AI is the next computing platform. But the decisive value won't sit in the robot — it will sit in the supply chain that builds its body.
Whoever controls the actuators, the sensors, and the foundation models — the picks and shovels of the humanoid gold rush — captures the decade. This deck follows that argument from market to machine to map.
What this deck answers
Three questions decide who wins the humanoid decade.
Q1
Is the market real — and how big?
Parts I. The inflection, the forecasts, the flywheel of capital. Today ~16,000 units; by 2050 a $5 trillion prize.
Q2
Where does the value actually concentrate?
Parts II–III. Brain, body, integrator — and why the body is the bottleneck and the beachhead is the factory floor.
Q3
Who controls it, and what could break it?
Parts IV–VI. The US–China–Korea race for the supply chain, the human element, and what investors, leaders and policymakers should do.
Executive summary · June 2026
By mid-2026, humanoids crossed from demo to deployment.
Installed base, 2025
~16k
Humanoids installed worldwide — still tiny, but real revenue.
Market leader value
$39B
Figure AI's valuation — up 15× in nineteen months.
Actuators / BOM
~53%
Share of a humanoid's hardware cost — the muscles.
2050 TAM (Morgan Stanley)
$5T
Revised up from $3T; over one billion machines.
I
Part One
The market is real, but the prize is the supply chain.
The dawn of the humanoid era — the inflection, the forecasts, and the flywheel of capital pulling the industry forward.
The inflection
In 2025 the conversation moved from "what if" to "how soon."
The catalyst was not one breakthrough but a convergence: mature robotic mechanics meeting multimodal AI. At CES 2025, NVIDIA's Jensen Huang spent nearly 40 minutes on physical AI, sharing the stage with 14 humanoids — a signal that re-rated the entire sector.
By CES 2026, that signal had become a product launch: Boston Dynamics unveiled the production Atlas, Figure was shipping Figure 03, and "embodied AI" had moved from an investment thesis to a deployment timeline.
The case for the human form
The humanoid wins because the world is already built for it.
Brownfield
No retrofit required
Specialized robots demand costly changes to the workplace. A humanoid uses the same doors, stairs, and tools we do — lowering the barrier to entry.
General-purpose
One platform, many tasks
The human hand is a universal tool. A machine that replicates it can do many jobs without a suite of bespoke end-effectors.
Addressable market
Built for human spaces
Factories, warehouses, hospitals, homes — all designed around the human body. The form factor is the addressable market.
Where we actually are
Today the market is tiny — about 16,000 robots were installed in 2025.
~16,000
units installed
worldwide, 2025
Real revenue, real deployments — but a rounding error against the long-run vision. The five largest suppliers accounted for roughly 73% of installations, and four of those five shipped from China.
Exhibit · Market forecasts compared
The forecasts disagree by 6× — and the disagreement is the real signal.
| Research firm | 2025 market | Forecast horizon | Forecast value | CAGR |
|---|
| MarketsandMarkets | $3.0B | 2030 | $15.3B | 39% |
| Goldman Sachs | — | 2035 | $38B | — |
| Research Nester | $4.4B | 2035 | $82B | 44% |
| Coherent Market Insights | $4.3B | 2032 | $70B | 49% |
| GlobeNewswire (aggressive) | $3.0B | 2035 | $243B | 49% |
The spread doesn't reflect doubt about the destination — it reflects uncertainty about timing. Aggressive forecasts assume non-linear breakthroughs in actuator cost, battery density, and AI reliability. The market's path is hostage to specific technical milestones.
The long-run prize
Morgan Stanley just doubled the prize — to a $5 trillion market by 2050.
$5T
Morgan Stanley's revised 2050 total addressable market — up from $3 trillion, driven by over one billion humanoids and a higher per-unit value.
The scale of the vision
The long-run vision spans one billion to twenty billion machines.
Morgan Stanley · 2050
1B+
Humanoids in operation — 302M in China, 78M in the US.
Elon Musk · vision
10–20B
Musk's projection — more humanoids than people.
Per-unit cost · 2024
~$200k
Falling fast toward a mass-market target.
The innovation flywheel
Capital is compounding into an innovation flywheel.
01
Capital funds aggressive R&D and talent
Mega-rounds let pure-plays out-hire and out-spend, compressing development timelines.
02
R&D produces visible breakthroughs
New VLA models, in-house actuators, and live factory pilots de-risk the thesis.
03
Breakthroughs attract still more capital
Each milestone re-rates valuations — and the cycle accelerates.
Exhibit · Figure AI valuation
Figure's valuation rose 15× in nineteen months.
Series BFeb 2024 · $675M raised
$2.6B
Series CSep 2025 · $1B+ raised
$39B
Led by Parkway Venture Capital with Brookfield, NVIDIA, Intel, Qualcomm, Salesforce and T-Mobile. Total capital raised: ~$1.9B — the best-funded pure-play in the field.
Exhibit · The 2025–26 funding wave
The mega-rounds put real money behind the thesis.
| Company | Round | Amount | Lead / notable backers | Valuation |
|---|
| Figure AI US | Series C · Sep 2025 | $1B+ | Parkway, Brookfield, NVIDIA, Intel | $39B |
| Apptronik US | Series A-X · Feb 2026 | $520M | Google, Mercedes-Benz, Jabil | ~$5B |
| UBTech China | Financing · Sep 2025 | $1B | State-linked & strategic | — |
| Agility Robotics US | Growth · 2024 | $400M | Existing & strategic | — |
| LimX · Neura · Leju CN / DE | 2025–26 | $200M+ | Regional & strategic | — |
Symbiotic ecosystems
The tech giants are funding robots to own the next computing platform.
NVIDIA
Sells the brains
Chips and the Isaac simulation stack that train and run the robots — exposure to every integrator at once.
Microsoft · Google
Sell the cloud & models
Azure compute, and Google DeepMind's Gemini Robotics foundation models powering multiple platforms.
Amazon · Mercedes
Buy the first-mover edge
Strategic adopters get a head start deploying automation against their own labor and efficiency problems.
II
Part Two
A humanoid is a brain, a body, and an integrator.
Anatomy of the machine — and why value concentrates unevenly across the three layers of the value chain.
The value chain
Value splits three ways — and it concentrates unevenly.
~20%
The Brain
Software & chips
Foundation models, perception, planning, AI accelerators. High margin, few winners.
~50%
The Body — the muscles
Actuators & sensors
The mechanical bottleneck. Half the cost, the hardest to source, the contested ground.
assembly
The Integrator
The robot maker
Tesla, Figure, Agility, Boston Dynamics. Brand and data, but thin without the chain.
The Brain · foundation models
The brain evolved from language models to vision-language-action models.
An LLM reads and writes text. A Vision-Language-Action (VLA) model interprets multimodal input — sight, language, touch — and turns it into motor commands. "Put the apples in the basket" decomposes into perception, planning, and control.
This is the difference between a chatbot and a worker. The VLA is what lets a humanoid generalize to tasks it was never explicitly programmed to do — the capability unlock behind the 2025–26 step-change.
The strategic divergence
Two philosophies are fighting for the brain.
Horizontal platform
NVIDIA — the open OS
Project GR00T is a general foundation model for many robots, trained in Isaac Sim. NVIDIA aims to be the AI operating system for the whole industry — Agility, Boston Dynamics and others as partners.
Vertical integration
Tesla & Figure — the closed stack
Peak performance comes from a bespoke AI tightly coupled to bespoke hardware. Figure exited its OpenAI partnership in Feb 2025 to build Helix in-house; Tesla extends its FSD stack to Optimus.
A re-run of PC (Microsoft/Intel platform) vs. Apple (integrated). The winner decides where the value pools.
The new third force
A third force arrived: Google's Gemini Robotics now powers two leading robots.
Google DeepMind's Gemini Robotics foundation models are integrated into Boston Dynamics' Atlas and Apptronik's Apollo.
It reshapes the platform-vs-vertical map: a second horizontal brain provider, but one tied to the world's deepest AI research and willing to embed directly in partners' hardware. Atlas can reportedly learn a new industrial task in under a day.
Case study · Figure's Helix
Figure bet the company on building its brain in-house.
Architecture
One net, three layers
Helix 02 unifies walking, balance and manipulation. Layers S0 (control), S1 (visuomotor), S2 (reasoning) run on low-power onboard GPUs — no cloud required.
Project Go-Big
Trained on human video
First-person video from Brookfield's 100,000 homes. Demonstrated zero-shot transfer: a robot navigating a home from "go to the fridge" with no robot-specific data.
Thesis
"Vertically integrate"
Adcock: to solve embodied AI at scale, you must own the AI and the body. The bet that off-the-shelf models won't be good enough.
The power-latency dilemma
Power, not intelligence, limits how long a humanoid can work.
Real-time autonomy runs on power-hungry GPUs, throttling battery life and uptime — a top complaint in early industrial pilots. The fix may be architectural, not chemical.
Neuromorphic computing — Intel's Loihi 2, IBM's NorthPole — fuses memory and compute in brain-inspired, event-driven chips. Orders of magnitude more efficient for the sparse, asynchronous data of robotics. Still research-stage, but decisive for long-run viability.
The semiconductor backbone
Every robot brain still runs on the same short list of chips.
Compute
NVIDIA · Qualcomm · Intel
High-performance GPUs and edge AI accelerators for training and onboard inference.
Vision
Ambarella · Mobileye
Specialized vision processors interpreting the robot's camera streams.
Memory
Micron · SK Hynix · Samsung
High-bandwidth memory feeding the models.
Make
TSMC · Synopsys · Cadence · Arm
Foundry, EDA software, and the licensed architectures underneath it all.
Exhibit · Bill of materials, Tesla Optimus Gen 2
The body, not the brain, is the bottleneck — actuators are half the cost.
0%SHARE OF BILL OF MATERIALS100%
A1
Frameless torque motors
A2
Planetary roller screws
—
Everything else: brain, sensors, battery, structure, hands
53%
Of a humanoid's hardware cost is its actuators. Whoever drives that cost down — or controls its supply — sets the pace of the entire industry.
Inside the muscle
An actuator is three hard parts: a screw, a reducer, and a motor.
| Component | Does what | Leaders | The catch |
|---|
| Planetary roller screw linear | Pushes & pulls (legs, torso) | SKF, NSK, Hengli, Beite | $1,350–2,700 each; supply-constrained |
| Harmonic drive rotary | Zero-backlash joint rotation | Harmonic Drive Systems, LeaderDrive | Lower shock tolerance |
| Cycloidal drive rotary | High-torque, shock-resistant | Nabtesco, Sumitomo | Heavier, more backlash |
| Frameless torque motor core | Converts current to motion | Kollmorgen, TQ, Inovance | Needs thermal management |
The strategic vulnerability
The hardest part to buy is also the one the West doesn't make.
The best high-load screws are expensive, scarce, and increasingly made in China.
Harmonic reducers are led by Japan; routing components through China can trigger tariffs. Western developers face a direct dependency on a strategic competitor for the single most critical subsystem — a structural drag on their cost and pace of innovation.
Perception
To act safely, a robot must first perceive — vision, force, and touch.
Vision · the debate
LiDAR + camera vs. camera-only
Most developers fuse LiDAR and cameras for robust depth. Tesla bets on camera-only — cheaper hardware, far heavier compute, weaker in poor conditions.
Force & torque
6-axis sensing at wrists & ankles
Enables compliant control — adjusting grip on a fragile object, balancing on uneven ground. Led by ATI (Novanta) and Hypersen.
Touch · the frontier
Robotic skin
High-resolution tactile arrays sensing pressure, temperature and slip — the final gap between robotic and human dexterity.
III
Part Three
Adoption is pragmatic: factories first, homes last.
Humanoids at work — the phased rollout through structured environments where the business case is clearest and the risk is lowest.
The adoption pattern
Adoption follows a "crawl, walk, run" path through structured work.
Crawl
One high-pain, repetitive task
Move empty totes; insert a sheet-metal part. Prove reliability and safety in a low-risk role first.
Walk
Integrate into existing systems
Connect to the warehouse or manufacturing execution system. The hard part is the workflow, not the robot.
Run
Expand to adjacent tasks & fleets
Use the learnings to broaden scope — and let robots dispatch other robots.
Beachhead one · manufacturing
The factory floor is the first beachhead.
Automotive plants are structured, repetitive, and acutely short of labor — the clearest business case in the industry.
Every major integrator is targeting the same ground: BMW, Mercedes-Benz, Hyundai/Kia, Tesla, and a wave of Chinese OEMs. The task list is deliberately narrow — part sequencing, material handling, body-shop insertion — chosen to prove ROI before broadening scope.
Exhibit · Figure 02 × BMW Spartanburg
Figure's BMW pilot ran for eleven months and built 30,000 cars.
Duration
11 mo
Continuous run on the X3 line, completed Nov 2025.
Vehicles supported
30k+
BMW X3 vehicles produced with robot assistance.
Parts handled
90k+
Sheet-metal parts placed into fixtures.
Operating hours
1,250
Logged before the generation was retired for Figure 03.
Titan play · Boston Dynamics Atlas
Hyundai committed 25,000 Atlas robots to its own plants.
25,000
Atlas units committed across
Hyundai & Kia — 83% of capacity
The production Atlas debuted at CES 2026 — 56 degrees of freedom, 50 kg payload, autonomous battery swap, Gemini Robotics inside. Hyundai is investing $26B in US operations and a 30,000-unit/year factory by 2028. Crucially, Hyundai Mobis makes the actuators: the first fully vertically integrated humanoid supply chain.
Vertical play · Tesla Optimus
Tesla is converting a car factory into a robot factory.
In early 2026 Tesla moved to wind down Model S/X at Fremont and repurpose the line for Optimus.
Roughly 800–1,000 units already work inside Tesla's own factories — a closed loop for data and iteration. The bet is that controlling FSD-derived AI, chip design, and manufacturing together drives Optimus toward a $20–30k mass-market price. External deployment is slated for later in 2026.
The China lane
China is deploying humanoids on its lines faster than anyone.
UBTech · Walker S
Multi-OEM pilots
Deployed across BYD, Dongfeng and Nio factory logistics. $1B raised in Sep 2025 to scale.
BYD
In-house ambition
Aggressive internal deployment targets across its enormous manufacturing base.
AgiBot · Unitree
Volume & price
AgiBot shipped 10,000+ units by Mar 2026; Unitree's G1 starts near $13,500 — undercutting the West.
Beachhead two · logistics
The warehouse is the second beachhead — Digit has moved 100,000 totes.
100k+
Totes moved by Agility's Digit at GXO Logistics — widely considered the first commercial humanoid deployment, now joined by Mercado Libre in Texas.
Agility's model differs from the vertical players: a Robots-as-a-Service subscription and a partner ecosystem. Its Arc cloud platform now lets Digit dispatch autonomous mobile robots — robots coordinating robots.
The human-centric frontier
Healthcare and the home are the prize, and the hardest to reach.
Hospital logistics
Giving time back
Diligent Robotics' Moxi fetches supplies and samples — returning 600+ hours to care teams in months.
Rehabilitation
Patient, consistent
Robots guide repetitive therapy exercises and keep patients engaged, freeing therapists for higher-level care.
Elder care
Demographic pull
Aging populations in Japan, China and the West create powerful demand for companionship and monitoring.
The trillion-dollar question
The home is the trillion-dollar question no one has answered yet.
Figure is training Helix on human video from 100,000 homes; 1X offers its Neo home robot for $499/month. But unstructured domestic environments demand a reliability no one has shown. The home is where the largest market and the hardest problem meet.
Exhibit · Deployment scorecard, mid-2026
A scorecard of who is deployed, where, and doing what.
| Robot | Maker | Customer | Task | Status |
|---|
| Figure 03 | Figure AI | BMW; UPS (reported) | Body-shop, commercial | Scaling — ~240/mo |
| Atlas | Boston Dynamics | Hyundai; Google DeepMind | Part sequencing | Production committed |
| Optimus | Tesla | Tesla (internal) | Material handling | Internal · ext. 2026 |
| Digit | Agility | GXO; Mercado Libre | Tote handling | Commercial |
| Apollo | Apptronik | Mercedes-Benz; GXO; Jabil | Logistics, assembly | Pilots |
| Walker S | UBTech | BYD; Dongfeng; Nio | Factory logistics | Pilots |
IV
Part Four
The real contest is geopolitical.
The competitive and geopolitical arena — a commercial race among integrators, set against a US–China–Korea battle for the supply chain.
The integrators
The field divides into incumbent titans and venture-backed pure-plays.
The titans
Scale, supply, a captive use case
Tesla bends its manufacturing and FSD data toward Optimus. Hyundai pairs Boston Dynamics' locomotion with its own factories and actuator supply. They start with distribution others must buy.
The pure-plays
Speed, focus, fresh capital
Figure, Agility, Apptronik and Sanctuary move fast on a single mission. They win on agility and partnerships — and live or die by access to capital and components.
Exhibit · The leading platforms, head-to-head
Head-to-head, the leading robots are converging on the same spec.
| Platform | Payload | Runtime | Vision | AI brain | Target price |
|---|
| Tesla Optimus | ~20 kg | ~2 h | Cameras only | FSD / in-house | $20–30k |
| Figure 03 | ~20 kg | ~5 h | RGB cameras | Helix (in-house) | $130k+ / RaaS |
| Atlas | 50 kg | batt. swap | LiDAR + depth | Gemini Robotics | $150k+ (est.) |
| Agility Digit | ~16 kg | ~4 h | 2D/3D + LiDAR | Agility Arc | RaaS |
| Apptronik Apollo | ~25 kg | swap packs | Cameras | Gemini Robotics | sub-$50k |
| Unitree G1 | light | ~2 h | LiDAR + cam | In-house | $13.5–16k |
Where value pools
The platform-vs-integrated war will decide where the value pools.
If the open platform wins, NVIDIA and Google capture value as the brains for everyone. If vertical integration wins, the market consolidates around a few players — Tesla, Figure, Hyundai — who own their entire stack. Either way, the integrator alone, without a brain or a body it controls, is the most exposed position on the board.
Exhibit · The geopolitical fact
China already controls the humanoid value chain.
56%
of value-chain companies are based in China
77%
of integrators are Chinese
This is the structural fact behind the whole thesis. Western developers can design the best robot and still depend on a strategic competitor for the components that make it move.
Exhibit · 2025 installations by supplier origin
Four of the world's five top suppliers shipped from China in 2025.
Top-5 suppliersshare of all installations
73%
Chinese among top-5four of the five leaders
4 / 5
Concentration plus geography: the installed base is dominated by a handful of suppliers, and the center of gravity sits in China — across both robots and the components inside them.
Two strategies
Two strategies collide: China's state plan vs. America's market.
China · top-down
State-directed mobilization
Five-Year Plans, state-backed funds, a "1 Million AI Robots" target, and provincial subsidies. Risks misallocation — but enables rapid scaling and a deep component base.
United States · market-led
Decentralized innovation
The CHIPS Act revitalizes semiconductors; robotics funding is diffuse across NSF and DoD. Fosters breakthroughs and agility — but leaves mid-stream hardware exposed.
The wildcard
Korea is the wildcard — it owns its robot's supply chain end-to-end.
Hyundai controls the platform (Boston Dynamics), the actuators (Mobis), and the deployment site — a model no rival can match.
With Hyundai Mobis supplying actuators that are 60%+ of material cost, Hyundai sidesteps the dependency that constrains every other Western player — and Korea already has the world's highest robot density. Vertical integration as national strategy.
The takeaway
The supply chain, not the showroom, is the real battlefield.
The captivating race between Tesla, Figure and Boston Dynamics plays out on stage. The decisive contest happens upstream — in actuators, reducers, rare-earth magnets, and the foundation models. Whoever secures those secures the decade.
V
Part Five
The human element decides acceptance.
Labor, trust, and ethics — the factors that will determine whether the technology is welcomed or resisted.
The future of labor
Automation will displace some work and create work that doesn't exist yet.
~13%of US jobs face high
displacement risk near-term
The roles most at risk are routinized and physical — exactly what the first wave targets. But history says technology destroys and creates: the challenge is managing the transition through reskilling, not pretending it won't happen.
The work that appears
The robot economy invents new jobs: fleet managers, pilots, trainers.
01
Robot fleet managers & remote "pilots"
Supervising and teleoperating fleets across sites — the air-traffic control of embodied labor.
02
Maintenance technicians & integrators
Keeping hardware running and wiring robots into enterprise systems.
03
AI trainers & data analysts
Curating demonstrations and interpreting the data streams fleets generate.
The backlash, already here
The labor backlash has already begun — a union blocked Atlas.
In January 2026 Korea's Metal Workers' Union declared Atlas would not enter Hyundai plants without a labor agreement.
The union cited Atlas's ~$145k unit cost as evidence management sees it as a labor-cost tool, and flagged the summer 2026 contract talks as the escalation point. Change management is not a footnote — it is a gating risk.
Building trust
Trust is the gating asset, and safety standards are still being written.
Standards in flux
ISO 25785-1 is still in draft
The first safety standard for dynamically-stable walking robots — written by Agility, Boston Dynamics and A3 — isn't expected until 2026–27. Until then, OSHA has no humanoid-specific rules.
Accountability
A fragile asset
A 2025 lawsuit alleged a Figure robot was strong enough to fracture a skull. One high-profile failure could set the industry back years — making safety the foremost imperative.
The quieter risks
Bias, privacy, and over-trust are the quieter risks to manage.
Bias
Models inherit our data
A robot trained on narrow data performs unevenly across people. Diverse data and audits are not optional.
Privacy
Mobile sensor platforms
Cameras and microphones on legs collect vast personal data. Encryption, consent and transparency are obligations.
Over-trust
Automation bias
People defer to machines even when wrong. A robot should always be legible as a machine, with honest limits.
VI
Part Six
What to do.
Strategic outlook — concrete recommendations for investors, corporate leaders, and policymakers navigating the ascendancy.
For investors
Buy the picks and shovels, not just the robots.
01
Own the enabling technology
Actuators (Harmonic Drive, Nabtesco, SKF), sensors (Novanta/ATI), and the AI platform layer (NVIDIA) — the de-risked exposure every integrator must buy.
02
Judge integrators on ecosystem, not specs
Access to capital, anchor customers (BMW, Hyundai), and a viable model — direct sales vs. Robots-as-a-Service.
03
Watch the milestones, not the hype
Actuator cost, battery density, and AI uptime are the variables that move the whole curve.
For corporate leaders
Start small, prove ROI, then scale.
Start small
Pick one painful task
Don't wait for a general-purpose robot. Deploy against one repetitive, ergonomically hard bottleneck and build internal expertise.
Integrate
The robot is the easy part
The real work is wiring it into your WMS/MES and data flows. Partner with experienced systems integrators.
Prepare people
Augment, don't blindside
Communicate early, train for new roles, and create pathways into fleet management and maintenance.
For policymakers
Secure the domestic supply chain.
01
Fund the components, not just the chips
Extend CHIPS-style incentives to motors, reducers and sensors — the mid-stream vulnerability.
02
Coordinate national R&D
Re-energize a cross-agency robotics strategy from basic research to applied deployment.
03
Invest in people and write agile rules
Fund reskilling at scale, and co-develop safety and ethics frameworks before a vacuum erodes public trust.
What could derail it
Three risks could still derail the ascendancy.
Geopolitical
A supply-chain shock
A US–China rupture in actuators or rare-earth magnets could stall Western integrators overnight.
Technical
A wall on cost or power
If actuator cost or battery uptime fails to improve, the optimistic forecasts simply don't arrive.
Regulatory
A trust-breaking incident
One high-profile safety failure could trigger restrictive rules and a public backlash.
The bottom line
The robots will get the headlines. The supply chain will get the returns.
Embodied AI is real, and the deployment has begun. But the durable advantage belongs to whoever owns the muscles — the actuators, the sensors, and the models that move them. Own the muscles, own the decade.
The Humanoid Ascendancy
June 2026
Explore the value
chain yourself.
An interactive companion lets you click through the anatomy of a humanoid — brain, body, and joints — to see component costs, suppliers, and bottlenecks. Open the Value-Chain Explorer →
Sources — Morgan Stanley "The Humanoid 100"; Counterpoint Research; Goldman Sachs; company announcements (Figure AI, Boston Dynamics, Tesla, Apptronik, Agility, UBTech); CES 2026. Figures actualized to June 2026.