Physical AI: Deutsche Bank sees physical AI moving from digital models into labor-replacing machines and eventually off-world infrastructure

Overview

This thematic report examines how AI is extending from screen-based software into machines that act in the physical world. Deutsche Bank sees robotaxis as the first scaled deployment, humanoids as a potential response to labor shortages, and autonomy as a prerequisite for a longer-term space economy, while highlighting training-data and supply-chain obstacles.

Core views

Deutsche Bank frames physical AI as the next stage of the AI investment cycle: intelligence moves from tokens to torque through a stack of machines, edge compute and foundation models. The report points to an expanding capital base behind the shift. The $1 trillion market-capitalization club grew from six companies in September 2021 to 16 in September 2026; seven of the ten additions are semiconductor companies, while Tesla and SpaceX are also identified as working on AI. Combined 2026 capital expenditure by Amazon, Microsoft, Alphabet, Meta, Oracle and SpaceX is presented as $865 billion, versus $500 billion. The report nevertheless recalls that much US telecom capex in 1996–2000 went into long-haul fiber that remained 85–95% unlit four years after the bust, underscoring that large investment does not eliminate execution risk. Robotics capital is also accelerating, with cited large humanoid funding rounds including Figure at $1 billion, Skild AI at $1.4 billion, NEURA at up to $1.4 billion, and Apptronik at $935 million. The report identifies robotaxis as the first meaningful commercial deployment of physical AI. It cites Waymo at 14 metros, 4,000 autonomous vehicles, more than 500,000 paid rides per week, 200 million cumulative autonomous miles and a target of 1 million weekly rides. It also notes Tesla's Cybercab commercial service in Austin, opened on September 3 without a steering wheel or pedals, and approximately 2,000 autonomous vehicles in paid service across Beijing, Shanghai, Guangzhou and Shenzhen. Deutsche Bank's Waymo cost build totals $1.23 per vehicle-mile: $0.60 of depreciation based on $105,000 over 175,000 miles, $0.09 of remote assistance, $0.26 of depot, cleaning and maintenance, $0.23 of insurance and liability, and $0.05 of energy. The implication is that the economics can be compelling even before full optimization. For humanoids, the report argues that demographic and labor constraints create a practical demand driver. It cites projected net migration to mid-2026 of 321,000, down from 2.7 million in 2024; 11,200 baby boomers reaching retirement age every day; and 1.9 million manufacturing roles projected to be unfilled by 2033. Labor-force growth is described as near zero in 2026, while breakeven hiring is below 10,000 jobs per month, the lowest in 65 years. Fertility has been below replacement since 2008, at 1.6 births, and manufacturers reportedly rank skilled workers ahead of tariffs as their binding constraint. Against this backdrop, the report estimates a humanoid's raw operating cost at $5.86 per productive hour, consisting of $1.79 depreciation, $1.34 maintenance and spares, $1.52 supervision, $1.07 fleet software and $0.14 energy; at two robots per worker, the cost is $11.72. Global 2026E humanoid shipments are shown above a prior estimate of about 50,000, with China building more than 40,000 units in the first half of 2026 and holding an 88% share. The report also cites a 2050E market of 100 million units or $1.5 trillion. The central technical obstacle is not only model capability but the scarcity and fidelity of robot-action data. Text-scale learning is contrasted with the much thinner supply of data that records joint angles, torques, gripper state, failures and recovery behavior. Human video is effectively abundant and low-cost, but lacks commands, uses a different body, and usually omits failed attempts. Teleoperation produces higher-fidelity data but costs $35–60 per usable hour and yields only 5–50 episodes per operator-hour; simulation can generate millions of episodes overnight for under $1 per usable hour but needs real task capture and validation; fleet learning scales with deployed robots but cannot begin without a deployed fleet. Deutsche Bank describes a blended approach: share demonstrations across robot types, use latent-action learning to infer actions from video and anchor them with a smaller labeled set, build and validate simulations from real tasks, and deploy first on narrow, repeatable tasks that can both pay for themselves and create data for subsequent tasks. A validated simulation need not replicate real-world success rates exactly, the report argues, provided it ranks model checkpoints correctly. The report also identifies physical supply chains as a material constraint, particularly for a US-based humanoid buildout. Every joint motor needs neodymium, while China refines about 90% of global supply and the US lacks refining at scale; new capacity would take years to build. Humanoids require 30–40 precision joints, with Japan's Harmonic Drive leading and Chinese challengers closing the gap, while the report sees no American supplier at scale. On batteries, China’s CATL leads ahead of LG and Panasonic. Although EV investment has placed some cell capacity in the US, Deutsche Bank expects humanoids to remain on an Asian cost curve for years. The final section extends the thesis to off-world AI infrastructure. It argues that terrestrial data-center costs may rise as siting disputes, longer interconnection queues and community mitigation lift land and construction cost growth to 6% annually from a 2% base case. Power and cooling are projected to reprice at 4% annually and servers at 2%, versus base-case declines of 3% and 1%, respectively; electrical and mechanical trades are described as a build-speed constraint, with contractors short about 349,000 workers entering 2026. The report therefore considers orbital data centers, where dawn-dusk orbit offers near-continuous sunlight and 40% higher intensity, and expects initial orbital compute to serve terrestrial demand before enabling scaled intelligence off-world. For lunar and Martian activity, autonomy changes the economic and operational case. The report estimates that $4.1 billion of SLS/Orion cost per launch over roughly 96 productive surface hours on an Artemis III profile equates to about $42.7 million per astronaut-hour before lander and suit costs. In contrast, a $200,000 space-rated humanoid delivered at $50,000 per kg and working 12,000 hours over five years would cost about $350 per hour. The Moon's 2.6-second round-trip communications delay permits an Earth operator to remain in the loop, making imperfect autonomy usable today. Mars, however, has a four- to 24-minute one-way delay, a launch window every 26 months, six- to nine-month transit times and no practical complete shield from major radiation hazards. These conditions rule out teleoperation and make onboard autonomy, data and control systems essential; Deutsche Bank concludes that the first Mars workforce is robotic because a base must be built before crews arrive.

Analysis framework

The report moves from market-scale evidence and capital expenditure to the physical-AI technology stack, then tests the thesis through robotaxi and humanoid unit economics. It assesses adoption drivers through labor-market constraints, identifies training-data and component-supply bottlenecks, and finally compares terrestrial and off-world operating costs and communication constraints to explain why autonomy becomes more valuable in space.

Methodology notes

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  Physical-AI component supply-chain analysis

  The report traces humanoid deployment constraints from neodymium magnets and precision reducers to battery cells, showing how concentrated Asian supply can affect US production costs and capacity.

• Industry AnalysisVolume-price decomposition

  Unit-cost and productive-hour economics

  The report decomposes robotaxi vehicle-mile cost, humanoid productive-hour cost and lunar labor cost into operating components to compare automation with human labor.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Waymo

  Example of scaled robotaxi deployment and cost economics

  Strengths

  More than 500,000 paid rides per week, 4,000 AVs, 14 metros and 200 million cumulative autonomous miles.

  Comparison

  Presented alongside Tesla and Pony.ai as robotaxi deployments.

  Risks

  Insurance and liability account for $0.23 of the modeled $1.23 per vehicle-mile cost.

• Tesla

  Example of physical-AI and robotaxi deployment

  Strengths

  Cybercab commercial service opened in Austin on September 3 without a steering wheel or pedals.

  Comparison

  Discussed alongside Waymo and Pony.ai in robotaxis.

• SpaceX

  Example of AI-related capital investment and space-economy infrastructure

  Strengths

  Included in the cited combined 2026 capex group and used in the report's launch-cost and lunar-labor analysis.

  Comparison

  Its launch economics are contrasted with SLS/Orion costs in the lunar discussion.

Key data

• $1 trillion club6 companies in September 2021; 16 in September 2026Seven of the ten additions are semiconductor companies.
• Combined 2026 capex$865 billionFor Amazon, Microsoft, Alphabet, Meta, Oracle and SpaceX, versus $500 billion.
• Waymo operating scale>500,000 paid rides per weekAcross 14 metros with 4,000 AVs and 200 million cumulative autonomous miles.
• Waymo cost per vehicle-mile$1.23All-in modeled delivery cost.
• Humanoid raw operating cost$5.86 per productive hour$11.72 at two robots per worker.
• China humanoid production>40,000 units in 1H26The report cites an 88% China share.
• Lunar productive-hour cost~$42.7 million for an astronaut-hour versus ~$350 for a robot-hourBased on the report's Artemis III and space-rated humanoid assumptions.

Impact & implications

Deutsche Bank's thesis is that physical AI can become economically viable first in constrained, repeatable tasks where it offsets labor shortages and accumulates operating data. Scaling beyond those settings depends on solving action-data limitations and reducing dependence on concentrated Asian component supply; deeper autonomy would be especially consequential for Mars, where real-time human control is not possible.

Risks

• Robot-action training data remain scarce and often lack joint commands, force information and examples of failure recovery.
• Teleoperation improves data fidelity but is costly, while simulation requires real-world task capture and validation.
• Humanoid supply chains face concentrated exposure to China for neodymium refining, Japan and China for precision reducers, and Asia for battery cost curves.
• Terrestrial data-center construction may be constrained by siting opposition, grid interconnection queues, equipment repricing and skilled-trade shortages.
• Mars operations require onboard autonomy because communications delays make teleoperation impractical.