Detailed content
📈 Size and projections: humanoids and AI robotics
Projections vary enormously because the market is still taking shape. Goldman Sachs estimated ~US$ 38B for humanoids alone by 2035; broader estimates for AI robotics reach US$ 111B+ by 2030. The exact number isn't what matters—the order of magnitude and the direction are: tens to hundreds of billions over a decade.
📊 Read projections with healthy skepticism
- US$ 111B+ — broad estimate of AI robotics through 2030 (varies widely by methodology).
- ~US$ 38B — Goldman for humanoids by 2035, in the base case.
- Real TAM — if the target is global physical work, it’s trillions; the question is timing.
TAM
Total addressable market.
CAGR
Growth rate.
Goldman/BCG
Projection sources.
Timing
The most uncertain factor.
🔗 Value chain: chips, actuators, models, data
The value chain has layers: silicon (NVIDIA, edge), actuators and hands, foundation models, integrators e data. As on any platform, the margin tends to concentrate where there is defensible differentiation—and the consensus in 2026 is that operational data e models form the moat, while hardware becomes a commodity.
💡 Where the moat is
Humanoid hardware is getting cheaper (Unitree showed this). What isn't easily commoditized is data flywheel: those with robots in operation collect data that improves the model, which sells more robots. The data/model layer is the one that looks most like software in terms of margins.
Value chain
Value layers.
Layers
Chip→actuator→model→data.
Integrators
Captures the end customer.
Data moat
Defensible flywheel.
💸 Players and funding: PI, Figure, Skild, 1X, Covariant
Capital shows where the market is betting. Physical Intelligence (π0/π0.5) raised billion-dollar funding rounds for a robotics foundation model; Figure e 1X in humanoids; Skild AI in a general-purpose "brain"; Covariant in warehouse manipulation. Billion-dollar valuations for pre-revenue companies signal conviction—and risk.
| Player | Focus | Bet |
|---|---|---|
| Physical Intelligence | Foundation VLA | π0/π0.5, flow matching, generalist |
| Figure | Humanoid | Helix System 1/2, industrial focus |
| 1X | Humanoid | NEO for home use, teleop→autonomy |
| Skild AI | Generalist brain | Body-agnostic foundation model |
| Covariant | Logistics manipulation | RFM-1, warehouse picking |
Physical Intelligence
Robotics foundation model.
Figure / 1X
Humanoids leading.
Skild
Agnostic brain.
Valuation
Billions before revenue.
🏷️ Business models: RaaS, licensing, data
How do you monetize? Three models dominate. RaaS (robot-as-a-service) charges by the hour/task, removing the client's CapEx and aligning revenue with the value delivered. Model licensing sells the brain to hardware OEMs. And data monetization turns operations into an asset that feeds back into the model.
RaaS — robot-as-a-service
The customer pays per use instead of buying the robot. This lowers the adoption barrier and creates recurring revenue.
Model licensing
The NVIDIA-style model: sell the brain + tools to whoever builds the body.
Data flywheel
Every robot in the field generates data that improves the model—a compounding advantage over time.
RaaS
Pay per use.
Licensing
Brain for OEMs.
Data flywheel
Data becomes an asset.
Recurrence
Predictable recipe.
💼 Where the jobs/skills are
The sector needs profiles that this track maps out. Research in VLA (architectures, training), Robotics MLOps (deploy, fleet monitoring), integration (ROS2, hardware, safety) and teleoperation data engineering (data collection pipeline) are the hottest areas. The competitive advantage lies in combining ML with physical engineering.
✓ In-demand skills
- ✓PyTorch + VLA training, fine-tuning, sim-to-real.
- ✓Robotics MLOps: edge deployment, fleet monitoring.
- ✓ROS2 + hardware integration and safety.
- ✓Teleoperation data engineering and curation.
✗ Career pitfalls
- ✗Only ML theory, without ever touching real hardware.
- ✗Ignore safety/certification — what blocks deployments.
- ✗Betting on a single company in a high-churn sector.
Robot MLOps
Deployment and monitoring.
Data engineering
Teleoperation pipeline.
Integration
ROS2, hardware, safety.
VLA Research
Architectures and training.
⚠️ Risks: hype cycle, sim-to-real gap, regulation
Reading risks is as strategic as reading opportunity. The sector may enter the trough of disillusionment of the hype cycle if viral demos don't convert into profitable deployments. The residual sim-to-real gap, immature regulation, and the adoption timing are the main constraints. Anyone investing or building a career needs to model the pessimistic scenario.
⚡ The lesson of the hype cycle
Every transformative technology goes through a peak of inflated expectations, a trough of disillusionment, and only then a plateau of productivity. General-purpose robotics is near the peak in 2026. The winners will be those who survive the trough—usually those with real revenue, not just demos.
📊 The 4 risks to model
- Hype cycle — delivery expectations pulled forward; valuation correction.
- Residual sim-to-real gap — the last 10% of reliability is the most expensive.
- Regulation — still immature; certification may delay deployment (see M3.6).
- Timing — Enter too early and you burn cash; too late and you miss the window.
Hype cycle
Peak → valley → plateau.
Trough of disillusionment
Filter for serious cases.
Timing
Too early vs. too late.
Regulation
Adoption barrier.
✅ Module summary
Next module
3.6 — The 2026–2030 Frontier: Scaling Laws, World Models, Certification. What determines which scenario comes to pass.