Physical AI: Deutsche Bank sees physical AI moving from digital models into robotics, autonomous transport and eventually off-world infrastructure

Overview

This thematic report examines how AI could extend from screens into machines that act in the physical world. Deutsche Bank connects the AI capital-spending cycle, early robotics deployment, demographic labor constraints, training-data limitations, component supply chains and the eventual use of autonomous systems in space.

Core views

Deutsche Bank frames physical AI as the next stage of the AI investment cycle: intelligence moves from software and foundation models into capital-intensive machines such as robotaxis, humanoids, warehouse robots, drones and surgical systems. The report notes that the $1 trillion market-capitalization club more than doubled over five years, with seven of the ten additions by September 2026 being semiconductor companies; Tesla and SpaceX are the other two additions identified as working on AI. It also highlights a combined 2026 capital-expenditure program by Amazon, Microsoft, Alphabet, Meta, Oracle and SpaceX, while cautioning through the prior US telecom buildout that heavy infrastructure spending can leave capacity underutilized after a boom. The report regards robotaxis as the first meaningful real-world deployment. 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 one million weekly rides. It also notes Tesla's Cybercab commercial service opening in Austin on September 3 without a steering wheel or pedals, and roughly 2,000 autonomous vehicles in paid service across Beijing, Shanghai, Guangzhou and Shenzhen. Deutsche Bank's illustrative Waymo cost stack totals $1.23 per vehicle-mile: $0.60 depreciation based on a $105,000 vehicle over 175,000 miles, $0.09 remote assistance, $0.26 depot, cleaning and maintenance, $0.23 insurance and liability, and $0.05 energy. The report's point is that the economics are already compelling before further optimization. Humanoids are positioned as a response to labor scarcity and resulting industrial bottlenecks. The report cites labor-force growth near zero in 2026, projected net migration to mid-2026 below the 2.7 million recorded in 2024, retirement of baby boomers, and 1.9 million manufacturing roles projected to be unfilled by 2033. It argues that reshoring requires workers as well as factories, with manufacturers ranking skilled workers above tariffs as a binding constraint. Its illustrative US 2026 humanoid operating-cost model reaches $5.86 per productive hour before the assumed two robots per worker, or $11.72 at that ratio, comprising depreciation, maintenance, supervision, fleet software and energy. Global humanoid shipments are running ahead of a prior roughly 50,000-unit forecast for 2026; China built more than 40,000 units in the first half alone and is projected to account for 88% of shipments. The report references a potential 100 million units by 2050 and a $1.5 trillion market. The central obstacles are training data quality and supply-chain concentration. Human video is abundant but does not contain robot joint angles, torques, gripper states, reliable scale, or failed attempts and recovery behavior. Teleoperation produces higher-fidelity demonstrations but costs $35–60 per usable hour and transfers imperfectly across robot bodies; simulation can produce millions of episodes overnight for under $1 per usable hour but needs real captures to build and validate the simulated world. Deutsche Bank describes a blended approach: pool demonstrations across robot types, infer latent actions from video and anchor them with a smaller labelled set, validate simulation by whether it ranks model checkpoints correctly, and deploy robots first on repeatable tasks so they generate data for later tasks. On components, the report argues that a US domestic humanoid supply chain remains incomplete relative to China and Asia. Every joint motor requires neodymium; China refines about 90% of global supply, while US refining capacity at scale is absent and new capacity would take years. Humanoids need 30–40 precision joints, with Harmonic Drive leading in Japan and Chinese challengers catching up, but no American supplier at scale. For batteries, CATL leads, followed by LG and Panasonic; US cell capacity has been added through EV investment, but the report expects humanoids to remain on an Asian cost curve for years. The report extends the physical-AI thesis to data centers and space. It argues that terrestrial data-center costs may rise as siting disputes, longer interconnection queues and community mitigation lift land and construction inflation to 6% annually versus a 2% base case. It assumes power and cooling costs rising 4% annually and servers 2%, versus base-case declines of 3% and 1%, respectively; electrical and mechanical trades are also a speed constraint, with contractors short about 349,000 workers entering 2026. Orbital compute could eventually become cost-competitive as launch reuse, solar and radiator scale reduce costs, initially serving terrestrial compute demand before enabling larger off-world intelligence systems. For the lunar economy, Deutsche Bank contrasts roughly $42.7 million per productive surface hour for an Artemis III-style human mission—based on $4.1 billion of SLS/Orion cost per launch and approximately 96 productive surface hours—with about $350 per hour for a $200,000 space-rated humanoid delivered at $50,000 per kilogram and operating 12,000 hours over five years. The Moon's 2.6-second round trip allows an Earth operator to remain in the loop, making existing imperfect autonomy potentially sufficient. Mars is more demanding: four to 24 minutes of one-way communication delay, launch windows every 26 months, long transit times and human health risks require onboard decision-making. The report therefore argues that Mars is a forcing function for autonomous control and data solutions, and that robots will necessarily form the first workforce before any human crew arrives.

Analysis framework

The report links market and capital-spending evidence to deployment case studies and operating-cost models, then tests the adoption thesis against labor availability, data collection constraints and component supply chains. It finally applies the same cost and autonomy logic to rising terrestrial data-center costs and off-world operations.

Methodology notes

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  Physical-AI supply-chain analysis

  The report traces how robot deployment depends on upstream rare-earth magnets, precision reducers and battery cells, and explains how regional supply concentration can constrain downstream humanoid production.

• Industry AnalysisVolume-price decomposition

  Unit-cost and operating-cost analysis

  Deutsche Bank breaks robotaxi vehicle-mile costs and humanoid productive-hour costs into depreciation, labor support, maintenance, software and energy to assess economic viability.

• Other

  Robot training-data sourcing and simulation validation

  The report compares teleoperation, simulation, human video and fleet learning by throughput, cost and fidelity, and describes validating simulation by its ability to rank model checkpoints.

Asset mapping & comparison

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

• Waymo

  Example of scaled robotaxi deployment

  Strengths

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

  Comparison

  Presented alongside Tesla and Pony.ai as robotaxi deployments.

  Risks

  Insurance, liability, remote assistance and vehicle depreciation remain material elements of the cost stack.

• Tesla

  Example of physical-AI deployment through autonomous driving

  Strengths

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

  Comparison

  Discussed alongside Waymo and Pony.ai in robotaxis.

• Pony.ai

  Example of robotaxi deployment in China

  Strengths

  Paid autonomous-vehicle service is referenced across Beijing, Shanghai, Guangzhou and Shenzhen.

  Comparison

  Discussed alongside Waymo and Tesla.

• CATL

  Battery-cell supply-chain example for humanoids

  Strengths

  Identified as the battery-cell leader.

  Comparison

  Discussed ahead of LG and Panasonic.

  Risks

  Humanoid production is expected to remain on an Asian cost curve for years.

• Harmonic Drive

  Precision-reducer supply-chain example

  Strengths

  Identified as a leader in precision reducers.

  Comparison

  Chinese challengers are closing the gap.

  Risks

  No American source is identified at scale.

• SpaceX

  Example of launch-cost and orbital-compute enabling infrastructure

  Strengths

  Rapid reusable Starship launches are presented as part of the path to lower orbital costs.

  Comparison

  Falcon 9 is cited at $1,400 per kilogram today.

  Risks

  Launch-cost improvements alone do not solve the need for productive off-world labor.

Key data

• Waymo paid rides>500K per weekAcross 14 metros with 4,000 autonomous vehicles; the company targets 1M weekly rides.
• Waymo all-in cost per vehicle-mile$1.23Illustrative cost stack before full optimization.
• US manufacturing roles projected unfilled1.9M by 2033Used to support the labor-shortage case for humanoids.
• Humanoid operating cost$5.86 per productive hour; $11.72 at 2 robots per workerIllustrative US 2026 fully loaded cost calculation.
• China humanoid production40K+ units in 1H26China is projected to account for 88% of 2026 global shipments.
• China rare-earth refining share~90%A supply-chain constraint for neodymium-dependent robot motors.
• Lunar productive-hour cost~$42.7M for an astronaut-hour versus ~$350 for a robot-hourIllustrative comparison based on an Artemis III profile and a space-rated humanoid.

Impact & implications

The report argues that physical AI can turn AI spending into demand for autonomous vehicles, robots, chips, components, power infrastructure and data systems. Near-term deployment is supported by labor shortages and improving unit economics, but scaling depends on solving embodied-data problems and reducing dependence on concentrated Asian supply chains.

Risks

• Physical-AI scaling is constrained by limited high-fidelity robot action data, especially data on failures and recovery behavior.
• US production faces multi-year gaps in rare-earth refining, precision reducers and competitively priced battery supply.
• Capital-intensive infrastructure investment may risk overcapacity, as illustrated by the earlier US telecom-fiber buildout.
• Terrestrial data-center expansion may face higher land, construction, power, cooling and skilled-labor costs.

What to watch

• Robotaxi paid-ride growth, fleet scale and per-mile cost reductions.
• Humanoid shipment growth, particularly China's production share and progress toward repeatable commercial tasks.
• Progress in teleoperation, simulation validation, latent-action learning and fleet-generated training data.
• Development of US capacity for rare-earth refining, precision reducers and battery cells.
• Data-center construction costs, interconnection queues and the evolution of reusable-launch economics.