2026-06-18 👁️ — views

Physical AI Supply Chain — Who Makes the Sensors, Chips and Actuators

Layer-by-layer map of the AV and humanoid robot supply chain: LIDAR, radar, cameras, AI compute, actuators, and geopolitical risks.

Article 67 in the Physical AI Benchmark Series — The Supply Chain Layer

Behind every autonomous vehicle and humanoid robot is a supply chain of specialized components that most investor analysis ignores entirely. The public narrative focuses on software stacks, miles-per-disengagement statistics, and deployment timelines. The harder question — and the one with the most concentrated geopolitical and competitive risk — is who makes the physical components that enable the AI to perceive, compute, and act.

This article maps that supply chain layer by layer: LIDAR sensors, automotive cameras and radar, AI compute silicon, and the actuators powering the new generation of humanoid robots. Each layer has its own competitive structure, its own cost trajectory, and its own exposure to the geopolitical pressures that increasingly shape technology supply chains.

The Physical AI supply chain is where the next wave of value creation is concentrated. Commoditization of LIDAR unlocks mass-market AV economics. 4D imaging radar closes the gap with LIDAR for object detection. NVIDIA DRIVE Thor ramps as the dominant compute platform. And humanoid actuator supply is the new frontier where competitive moats are being built from first principles.

Section 1 — LIDAR: The Critical Sensor Layer

LIDAR — Light Detection and Ranging — is the primary depth-sensing technology for autonomous vehicles operating at high speed in complex environments. Unlike cameras, which infer depth from image processing, LIDAR directly measures distance by timing laser pulses. Unlike radar, LIDAR produces high-resolution 3D point clouds that capture object geometry rather than just presence and velocity.

The cost trajectory is the critical variable. Spinning mechanical LIDAR (the Velodyne era) cost $75,000 or more in 2012. Solid-state LIDAR designs — no moving parts, manufactured using semiconductor processes — are targeting sub-$100 at automotive volume. The crossover to sub-$500 LIDAR at scale is the economic unlock for mass-market AV deployment: at that price point, LIDAR moves from a cost-prohibitive differentiator to a standard line item in the AV bill of materials.

Hesai Technology deserves particular attention because it represents both the competitive pressure and the geopolitical risk concentrated in this layer. Hesai is the dominant LIDAR supplier in China and has won reported design wins with international AV programs including Waymo. Its pricing is aggressive — enabled by Chinese manufacturing scale and vertical integration. That same profile has drawn US Congressional and BIS scrutiny, with proposals to restrict Chinese LIDAR vendors from US federal transportation programs on national security grounds. An AV company that built its sensor supply chain around Hesai faces a structural sourcing risk if those restrictions tighten.

Section 2 — Cameras and Radar: The Commodity Sensors

Cameras and radar are the commodity layer of the AV sensor stack — high volume, well-understood technology, multiple qualified suppliers. The interesting action here is not in the cameras themselves but in the image signal processors and the emergence of 4D imaging radar as a LIDAR complement.

The most significant development in this layer is 4D imaging radar. Traditional automotive radar provides range, azimuth angle, and velocity — three dimensions. 4D imaging radar adds elevation, producing a sparse but genuine 3D point cloud with velocity at each point. Continental’s ARS540 and Arbe Robotics’ chipset are the leading commercially available 4D radar products.

The competitive significance of 4D imaging radar is that it closes the object detection gap with LIDAR while retaining radar’s inherent all-weather performance advantage. LIDAR degrades in heavy rain, snow, and fog — the precipitation particles scatter laser pulses before they reach the target. Radar penetrates precipitation without degradation. A perception stack built on 4D imaging radar plus cameras offers robustness in adverse weather conditions that a LIDAR-plus-camera stack cannot match, at lower sensor cost.

This is why several Tier 1 automotive suppliers are investing heavily in 4D radar: it is the path to an all-weather, cost-competitive alternative to LIDAR for many AV applications. The technology is not yet at LIDAR resolution levels, but the resolution trajectory is improving rapidly.

Section 3 — Compute: The AI Silicon Layer

The AI compute chip is the most value-dense component in the AV stack. Every sensor modality — camera, LIDAR, radar, thermal — produces data that must be processed in real time at automotive-grade reliability. The compute platform determines what perception, prediction, and planning algorithms are possible, and it is the component where NVIDIA has built the most durable competitive moat in the AV supply chain.

NVIDIA DRIVE Thor is the most significant product ramp in this layer for 2026. At 2,000 TOPS, Thor combines AV compute and cockpit AI into a single system-on-chip — reducing the number of electronic control units a vehicle manufacturer must integrate. The design win list (BYD, NIO, Volvo under Geely, Foxconn automotive) confirms that Thor is the platform on which the next generation of high-performance AV systems will be built.

TSMC dependency is the systemic risk. Tesla HW4, NVIDIA DRIVE Thor and Orin, and Qualcomm Snapdragon Ride are all manufactured at TSMC fabs. This creates a structural single-point dependency for the entire AV compute supply chain. TSMC’s Arizona expansion provides some geographic diversification, but Arizona volume remains limited relative to Taiwan capacity. Intel Foundry and Samsung offer alternative manufacturing options in principle; neither is at parity with TSMC on the leading-edge nodes these chips require. Taiwan Strait geopolitical tension is therefore not just a geopolitical news item — it is a direct AV supply chain risk that every investor in the sector must model explicitly.

Section 4 — Humanoid Robot Actuators: The New Frontier

The humanoid robot supply chain is less mature than the AV supply chain and represents the highest-growth opportunity within Physical AI hardware over the next decade. The key insight from observing the leading humanoid programs is that the critical scarcity is not at the AI model layer — it is in the physical actuation hardware that enables robots to move with the precision and force necessary for real-world manipulation.

The vertical integration pattern visible in Tesla Optimus is the defining strategic choice in humanoid hardware. Tesla is designing custom BLDC motors and actuators specifically for Optimus — trading short-term sourcing flexibility for long-term cost structure and performance density advantages. The same pattern is emerging at Figure AI and 1X Technologies, though they are less vertically integrated than Tesla at this stage.

Harmonic drive gearboxes are the component most likely to become a supply constraint as humanoid production volumes ramp. Harmonic Drive AG — the German-Japanese company that has dominated this category for decades — produces gearboxes with the ultra-low backlash and high torque density required for precise robot joint actuation. Nidec acquired a harmonic drive capability through M&A. But at the production volumes implied by Tesla’s stated Optimus production targets (tens of thousands to hundreds of thousands of units), harmonic drive supply would need to scale by an order of magnitude beyond current capacity. This is the kind of supply constraint that does not appear in software-layer analysis of humanoid robotics but will be a real binding constraint on production ramp timelines.

Section 5 — Geopolitical Risks in the Physical AI Supply Chain

The geopolitical risk map for Physical AI hardware is more concentrated than for software. Software can be replicated, retrained, and redistributed across borders with minimal friction. Physical components cannot. A TSMC disruption does not have a 90-day workaround — it takes years to qualify alternative semiconductor fabs at leading-edge nodes.

The rare earth magnet dependency deserves more investor attention than it currently receives. China controls an estimated 85% of global rare earth processing capacity (est.), and rare earth permanent magnets are the enabling technology for the high-performance BLDC motors in both EV traction systems and humanoid robot actuators. The US, Australia, and Canada have rare earth mining programs underway, but processing capacity outside China remains limited. Every humanoid robot and every EV produced globally today depends on this supply chain.

The Chinese LIDAR scrutiny risk has a clearer near-term resolution path. If Hesai and Robosense are restricted from US programs, there are qualified alternatives: Luminar for long-range highway applications, Innoviz and Cepton/Ouster for solid-state short-to-medium range, and Aeva for the FMCW velocity-sensing use case. The switch is disruptive and carries qualification cost, but it is executable.

Section 6 — Investor Signal

The Physical AI supply chain is where the next wave of value creation is happening in the AV and robotics sector — not just the software layer. Four investment theses follow from this supply chain map.

LIDAR commoditization unlocks AV unit economics. Sub-$500 solid-state LIDAR at automotive volume changes the cost structure of AV deployment fundamentally. Luminar, Innoviz, and Ouster/Cepton are the western candidates for that commoditization ramp. The company that achieves automotive-volume pricing first while maintaining reliability will be the Bosch of AV sensors.

4D imaging radar is an underappreciated disruptive force. Continental’s ARS540 and Arbe Robotics are building toward a perception capability that challenges LIDAR’s necessity for many AV applications — at lower cost, with superior all-weather performance. Continental is a public company; Arbe is publicly listed on NASDAQ. Both deserve positions in any Physical AI component portfolio.

NVIDIA DRIVE Thor is the platform bet. NVIDIA’s design win concentration in AV compute is the most durable moat in Physical AI hardware. Thor’s combination of AV and cockpit AI compute reduces OEM integration complexity while locking in NVIDIA’s position at the center of the vehicle software-defined platform. The risk is TSMC dependency — a Taiwan disruption hits NVIDIA as hard as anyone.

Humanoid actuator supply is the 2026–2030 constraint. Harmonic drive gearboxes, precision BLDC motors, and force/torque sensors will be the binding constraints on humanoid production ramp. The companies supplying these components — Harmonic Drive AG, Maxon, ATI Industrial Automation — are the picks-and-shovels play in humanoid robotics. They are less visible than the humanoid platform companies but will benefit from every unit of production ramp regardless of which humanoid platform wins.

The TSMC dependency is the systemic risk that every Physical AI investment thesis must model. It is not a remote tail risk — it is the central geopolitical variable for the entire sector’s hardware supply chain.

Section 7 — About This Series

This is article 67 in the Physical AI Benchmark Series. Previous articles have covered the ramp index, the humanoid race, unit economics, global competition, HD mapping, fleet operations, software and OTA, insurance and liability, consumer demand, competitive moats, Cybercab versus Model Y, safety data, Waymo Gen 6, Optimus manufacturing, scorecard snapshots, 2030 forecast scenarios, the investor framework, city expansion pipelines, Tesla FSD state approval maps, AV weather and climate constraints, the talent war, regulatory calendars, robotaxi fare pricing, humanoid deployment trackers, supply chain analysis, consumer adoption demand index, valuation and IPO analysis, the Physical AI 2026 mid-year roundup, AV unit economics cost-per-mile breakdown, the AV data flywheel comparison, and AV cybersecurity attack surfaces.

This article adds the supply chain dimension: LIDAR layer (Luminar, Hesai, Innoviz, Aeva, Waymo custom, Tesla camera-only), camera and radar layer (Sony, Onsemi, Continental, Arbe, 4D imaging radar emergence), AI compute layer (NVIDIA DRIVE Thor/Orin, Qualcomm Ride Elite, Mobileye EyeQ Ultra, Tesla FSD HW4, Horizon Robotics), humanoid actuator layer (harmonic drives, BLDC motors, force/torque sensors, battery packs), and the geopolitical risk map (TSMC dependency, Chinese LIDAR scrutiny, rare earth processing concentration, HBM supply, BIS export controls).

Note: Price estimates for sensors and chips are labeled “(est.)” and reflect publicly reported ranges or analyst estimates; actual contract pricing between OEMs and suppliers is not publicly disclosed. Design win information is based on public company disclosures, earnings calls, and industry reporting. Geopolitical risk assessments reflect publicly available information as of the article date. This article does not constitute investment advice.

Sources

• Luminar Technologies investor relations — Luminar ↗
• NVIDIA DRIVE Thor — NVIDIA automotive ↗
• Hesai Technology LIDAR products — Hesai ↗
• Continental ARS540 4D radar — Continental Automotive ↗
• TSMC geopolitical risk — TSMC annual report ↗

Tags