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

Overview

This thematic report examines how AI is moving from screen-based applications into machines that act in the physical world. Deutsche Bank links the development of robotaxis, humanoids, industrial robots and space infrastructure to a large AI investment cycle, labor constraints and the need to solve data, component and infrastructure bottlenecks.

Core views

Deutsche Bank frames physical AI as the next stage of the AI investment cycle: intelligence is moving from models and software into robotaxis, humanoids, warehouse robots, drones and surgical systems, supported by edge computing and foundation models. The report notes that the $1tn market-capitalization club has more than doubled over five years and says seven of the ten largest members are semiconductor companies; Tesla and SpaceX are cited among the remaining companies working on AI. It also cites combined 2026 capital expenditure of $865bn for Amazon, Microsoft, Alphabet, Meta, Oracle and SpaceX. While this spending supports the market and economy, the report recalls that much of US telecom investment in 1996–2000 remained unlit four years after the downturn, illustrating the execution risk of a capex supercycle. Robotaxis are presented as the first substantial real-world deployment. Waymo is cited as operating in 14 metros with 4,000 autonomous vehicles, more than 500,000 paid rides per week and 200m cumulative autonomous miles, targeting 1m rides a week. Tesla's Cybercab commercial service reportedly opened in Austin on September 3 without a steering wheel or pedals, while Chinese paid autonomous-vehicle services are cited at roughly 2,000 vehicles across Beijing, Shanghai, Guangzhou and Shenzhen. Deutsche Bank estimates Waymo's all-in cost per vehicle-mile at $1.23: $0.60 of depreciation based on a $105,000 vehicle over 175,000 miles, $0.09 for remote assistance, $0.26 for depot, cleaning and maintenance, $0.23 for insurance and liability, and $0.05 for energy. The report argues that this economics is compelling even before full optimization. The report sees labor scarcity as a powerful demand driver for automation. It cites projected net migration of 321,000 through mid-2026, down from 2.7m in 2024; near-zero labor-force growth; breakeven hiring below 10,000 jobs per month, the lowest in 65 years; fertility below replacement since 2008; and 1.9m manufacturing roles projected to be unfilled by 2033. Reshoring therefore requires workers as well as factories, with manufacturers said to rank skilled labor above tariffs as the binding constraint. Against this backdrop, Deutsche Bank estimates a humanoid's raw operating cost at $5.86 per productive hour, comprising $1.79 depreciation, $1.34 maintenance and spares, $1.52 supervision, $1.07 fleet software and $0.14 energy. At two robots per worker, that rises to $11.72 per productive hour. Humanoid shipments are described as ahead of forecast in 2026, with global shipments cited at about 50,000 versus a prior estimate of about 50,000 and China building more than 40,000 units in the first half, accounting for an indicated 88% share. The report identifies training data as the key technical obstacle. Text-based models can draw on vast human-written corpora, but robot learning needs actions, physical embodiment and failure data. Human video does not generally include joint angles, torques, gripper state, camera calibration or unsuccessful attempts. Deutsche Bank compares four data approaches: teleoperation offers 5–50 episodes per operator-hour and high fidelity but costs $35–60 per usable hour; simulation can produce millions of episodes overnight for under $1 but needs real-world captures to construct and validate the environment; human video is effectively unlimited and low cost but lacks core action data; and fleet learning scales with deployed robots but requires an installed fleet. The report’s proposed sequence is to pool demonstrations, validate simulations, infer actions from video, deploy initially on narrow repeatable tasks, and use operating units to both generate returns and collect data for future tasks. It adds that a validated simulation need only rank model checkpoints correctly rather than exactly reproduce real-world success rates. Physical-AI hardware also faces supply-chain concentration. Deutsche Bank states that every joint motor needs neodymium and that China refines about 90% of global supply, while the US lacks refining capacity at scale and new capacity takes years to build. Humanoids require 30–40 precision joints; Japan's Harmonic Drive is identified as the leader, Chinese challengers are closing the gap, and no US supplier is cited as operating at scale. For batteries, the report identifies CATL as the leader, with LG and Panasonic behind, and argues that even with some US cell capacity from electric-vehicle investment, humanoids are likely to remain on an Asian cost curve for years. The final thread extends physical AI to data centers and space. Deutsche Bank argues that terrestrial data-center costs could rise as siting disputes, longer interconnection queues and community mitigation lift land and construction cost growth to 6% annually from a 2% base case. It assumes power and cooling costs rise 4% annually and servers 2%, versus base-case declines of 3% and 1%, respectively. Electrical and mechanical trades are a further constraint, with contractors reportedly short about 349,000 workers entering 2026. The report sees orbital computing as a possible long-run response: dawn-dusk orbit offers near-24/7 sunlight with 40% higher intensity, and orbital data centers could initially serve terrestrial compute demand before enabling off-world scaling. Robots are central to the lunar and Mars cases because human labor remains exceptionally expensive off Earth. Deutsche Bank estimates an Artemis III-style SLS/Orion launch at $4.1bn for roughly 96 productive surface hours, or about $42.7m per human hour before lander and suit costs. By 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 an hour, five orders of magnitude less. The Moon's 2.6-second round-trip communication delay allows an Earth operator to remain in the loop, whereas Mars has a four- to 24-minute one-way delay, 26-month launch windows and a six- to nine-month transit. The report concludes that Mars rules out teleoperation and makes autonomous machines the initial workforce needed to build infrastructure before people arrive.

Analysis framework

The report moves from the scale of the AI capital-spending cycle to real-world deployment economics, then tests the thesis against labor shortages, robot-training data constraints, component supply chains and infrastructure costs. It uses cost-per-unit and cost-per-productive-hour comparisons to assess automation economics, and extends the same logic to terrestrial data centers, lunar operations and Mars.

Methodology notes

• Industry AnalysisSupply-demand framework

  Labor-supply constraints and automation demand

  The report links weak labor-force growth, lower migration, retirements and unfilled manufacturing jobs to rising demand for physical automation.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  Physical-AI component supply chain

  The analysis traces robotics deployment back to inputs including rare-earth magnets, precision reducers and battery cells, highlighting regional concentration and capacity constraints.

• Other

  Unit-economics comparison

  The report compares the cost per autonomous vehicle-mile, productive humanoid hour and lunar worker-hour to explain when robotic deployment could become economic.

Key data

• Combined 2026 capital expenditure$865bnAmazon, Microsoft, Alphabet, Meta, Oracle and SpaceX combined
• Waymo operations14 metros; 4,000 AVs; >500,000 paid rides/week; 200m cumulative autonomous milesThe company is targeting 1m rides per week
• Waymo all-in cost per vehicle-mile$1.23Includes depreciation, remote assistance, depot operations, insurance and energy
• Humanoid raw operating cost$5.86 per productive hourEstimated 2026 US cost; $11.72 at two robots per worker
• Manufacturing roles projected unfilled1.9mProjected by 2033
• China rare-earth refining shareAbout 90%Relevant to neodymium requirements for robot joint motors
• Construction-worker shortageAbout 349,000 workersContractor shortfall entering 2026
• Lunar productive-hour cost comparisonAbout $42.7m per astronaut-hour versus about $350 per robot-hourBased on the report's Artemis III and space-rated humanoid assumptions

Impact & implications

Deutsche Bank's central implication is that physical AI could broaden the economic effects of AI beyond software by replacing or augmenting scarce labor in transport, manufacturing and other real-world tasks. Progress depends on deployment economics and on resolving action-data, hardware-component and infrastructure constraints; in space, the same economics favor machines before human activity can scale.

Risks

• A capital-expenditure supercycle can leave infrastructure underutilized, as illustrated by the report's comparison with unlit US telecom fiber after the 1996–2000 buildout.
• Robot training remains constrained by limited high-fidelity action and failure data.
• US physical-AI supply chains remain dependent on concentrated Asian supply for rare-earth materials, precision reducers and battery cells.
• Rising land, construction, power, cooling and skilled-trade costs could impede terrestrial data-center expansion.
• Simulation requires real-world capture and validation, while fleet learning cannot scale before a deployed fleet exists.