2026-06-18 👁️ — views

Physical AI Supply Chain — Lidar, Compute, Actuators, and the Hidden Ramp Bottlenecks

Hardware, not software, is the hidden constraint on Physical AI: lidar lead times, NVIDIA Orin allocations, harmonic drives, and Waymo Zeekr dependency mapped.

Article 122 in the Physical AI Benchmark Series — Physical AI Supply Chain: Lidar, Radar, Camera, and Compute Bottlenecks; What Waymo’s Gen 6 Ramp Depends On; and Why Semiconductor Supply Is the Hidden Constraint on AV Fleet Expansion

The Physical AI Benchmark Series has spent 121 articles mapping technology readiness, operational metrics, safety records, regulatory frameworks, and market valuations across autonomous vehicles and humanoid robotics. Article 122 introduces a dimension that has been implicit throughout but never mapped in its own right: the hardware supply chain. Every Waymo Gen 6 vehicle requires lidar units, cameras, radar modules, and a compute platform. Every Tesla with HW4 requires proprietary silicon manufactured at TSMC. Every humanoid robot requires servo actuators, joint motors, harmonic drives, and batteries. These components have their own supply chains, lead times, cost curves, and supplier concentrations — and these supply chains, not software maturity or regulatory approvals, may be the binding constraint on how fast Physical AI fleets can grow in 2026 to 2030.

This article maps the hardware supply chain as a Physical AI benchmark dimension, with data labeled “(est.)” throughout where figures are derived from public market information, analyst estimates, and industry reporting rather than primary supplier disclosures.

Section 1 — AV Sensor Supply Chain: Lidar, Radar, and Camera

The sensor stack that enables autonomous vehicles to perceive the world consists of three primary modalities: lidar (light detection and ranging), radar, and cameras. Each has a distinct supply chain maturity profile, cost trajectory, and strategic positioning across the leading AV companies.

The lidar supply chain has undergone substantial consolidation since 2020. The near-monopoly that Velodyne held over spinning mechanical lidar — and which created genuine AV supply chain risk in 2018 to 2021 — has dissolved through consolidation (Velodyne merging with Ouster), new entrant scale-up (Luminar, Innoviz, Cepton), and the technology shift toward solid-state MEMS-based lidar that enables fundamentally different manufacturing. Solid-state lidar has no moving parts, which allows it to be manufactured using existing semiconductor-adjacent production lines rather than precision mechanical assembly. This shift is reducing both cost and supply constraint simultaneously.

The compute platform is the sensor supply chain segment with the most active constraint in 2025 to 2026. NVIDIA DRIVE Orin — the leading third-party automotive compute platform — is integrated across a wide range of AV and ADAS platforms from BYD, Xpeng, Volvo, and others. Demand has strained TSMC production capacity at the relevant nodes, with lead times reaching 12 to 18 months at peak (est.) in 2023 to 2024. Supply is normalizing in 2026 (est.) as TSMC expands capacity, but the structural lesson is clear: automotive-grade compute silicon is subject to the same supply shocks as consumer electronics semiconductors during demand spikes.

The radar and camera supply chains are effectively commoditized. Both components leverage decades of automotive-grade manufacturing investment, and neither is a constraint on AV ramp timelines.

Section 2 — Waymo Gen 6 Vehicle Supply Chain

The Waymo Gen 6 vehicle is the most strategically important hardware platform in the AV industry today. It represents Waymo’s attempt to close the unit economics gap that has prevented AV from reaching commercial viability at scale. The Gen 6 platform is built on the Zeekr RT base vehicle (a product of Geely’s Zeekr subsidiary, manufactured in China) rather than the modified consumer vehicles (Chrysler Pacifica minivans, Jaguar I-PACE SUVs) used in prior Waymo fleet generations. This manufacturing decision has significant supply chain implications.

The Zeekr dependency is the most consequential supply chain decision in Waymo’s Gen 6 strategy. The manufacturing cost reduction it enables — moving from $100,000-plus per vehicle (est.) in prior generations toward a target significantly closer to Zeekr’s RT6 manufacturing cost of approximately $37,000 (est.) — is fundamental to the unit economics improvement that Waymo needs to reach marketplace-level EBITDA margins. However, manufacturing in China introduces a supply chain risk category that is external to Waymo’s operational control: US-China trade policy. A 25%-plus tariff scenario, analogous to the automotive tariff actions taken in 2025, would add $9,000 to $15,000 (est.) to each vehicle’s cost. At meaningful fleet scale, this could materially impair the Gen 6 unit economics thesis.

Waymo’s custom compute platform — a TPU-derived chip built on Google’s silicon design heritage — represents the opposite supply chain logic: vertical integration at the compute layer (following the same strategy as Tesla’s HW4) while accepting dependency at the vehicle platform layer. This tradeoff reflects Waymo’s position as a software and AI company operating a fleet rather than a vehicle manufacturer.

Section 3 — Tesla’s Vertically Integrated Supply Chain Advantage

Tesla’s supply chain strategy for Physical AI hardware stands in deliberate contrast to Waymo’s. Where Waymo sources its base vehicle from a Chinese OEM and its lidar from custom manufacturing partnerships, Tesla has systematically vertically integrated the most strategically valuable components — compute and battery — while accepting commodity supplier dependency for less differentiated components like cameras.

Tesla’s camera-only sensor philosophy (no lidar) is not merely a philosophical stance — it is a supply chain decision with compounding strategic implications. By eliminating lidar from the vehicle bill of materials entirely, Tesla removes a $100 to $5,000 per vehicle cost item (est.) and a 6 to 18 month lead time component from its manufacturing critical path. Whether the perception quality tradeoff is acceptable remains actively debated across the AV industry, but from a pure supply chain perspective, Tesla’s camera-only approach produces the leanest and most predictable sensor supply chain of any major AV company.

The HW4 custom silicon strategy provides Tesla with a structural cost advantage that compounds over time. TSMC’s 7nm node, while not the cutting edge of semiconductor manufacturing, is a mature and well-supplied process that provides predictable unit economics at Tesla’s volume. As Tesla’s vehicle volumes scale — approximately 2 million per year in recent production (est.) — the per-unit cost of HW4 falls on a learning curve that is internal to Tesla, not subject to third-party supplier pricing decisions. This is the textbook definition of a supply chain moat.

Section 4 — Humanoid Robot Supply Chain: The New Frontier

The humanoid robot supply chain is the newest and least mature of the Physical AI supply chains. Where automotive component supply chains have decades of development behind them and lidar supply chains have had five to ten years of AV-specific investment, humanoid robot components — particularly the high-torque precision actuators that enable human-like movement — are manufactured by a small number of specialists with limited production capacity. This creates a supply chain constraint that is arguably more binding on the near-term humanoid ramp than any software or regulation variable.

The harmonic drive situation illustrates a supply chain dynamic that is familiar from the early days of lidar: a near-monopoly supplier with limited production capacity serving a rapidly accelerating demand curve from a new product category. Harmonic Drive AG has been the leading supplier of precision strain wave gearing — the mechanism that enables the precise, backdrivable joint control required for humanoid robots — for decades. Its production capacity was dimensioned for industrial robot arms and aerospace applications, not the tens of thousands of units per year that humanoid companies are beginning to project. This mismatch is producing lead times and allocation constraints that directly limit how fast humanoid robot manufacturers can ramp production.

The Unitree G1 at $16,000 per unit (est.) demonstrates an important supply chain insight: achieving low unit costs in humanoid robots is partly a bill-of-materials decision about actuator specifications. A humanoid that uses lighter-duty actuators — with less torque density, less precision, and simpler gearboxes — can be assembled from a broader supplier base at lower cost. The tradeoff is task capability: lighter-duty actuators limit the weight the robot can lift, the speed at which joints can move, and the precision of manipulation. Tesla Optimus and Boston Dynamics Atlas target the high-capability end of the actuator spectrum, which is why their cost structures remain significantly above Unitree’s — and why harmonic drive supply is their binding constraint.

Section 5 — Semiconductor Supply as the Cross-Cutting Constraint

Across all three Physical AI hardware categories — AV sensors, AV compute, and humanoid robot compute — the semiconductor supply chain is the cross-cutting constraint. The observation is simple but important: Physical AI is a semiconductor-intensive industry that competes for the same TSMC and Samsung wafer capacity as consumer electronics, data center AI accelerators, and defense electronics.

The data center AI buildout — NVIDIA H100, H200, and GB200 GPU demand from hyperscalers — has created a supply environment in which TSMC’s most advanced nodes (3nm, 4nm, 5nm) are heavily committed to the highest-revenue-per-wafer customers. Automotive and robotics applications, which require automotive-grade qualification (AEC-Q100 for automotive, AEC-Q200 for passives) in addition to the standard semiconductor process, compete at a disadvantage for wafer allocation during periods of peak demand. This is the structural semiconductor supply risk for Physical AI: when data center AI demand spikes, Physical AI hardware companies face extended lead times because their per-wafer revenue cannot compete with H100-class GPU economics.

The mitigation strategies vary by company. Tesla’s vertical integration and multi-year TSMC supply agreements provide allocation priority. Waymo’s status as part of Alphabet provides leverage through Google’s deep TSMC relationship. Independent AV companies and humanoid robot startups face the greatest exposure to semiconductor supply volatility because they lack the procurement scale to secure priority allocations.

Section 6 — Supply Chain Risk Matrix: What Could Slow the Physical AI Ramp

The supply chain risk matrix reveals an asymmetry between the AV and humanoid robot categories. The AV supply chain — after years of investment, consolidation, and cost reduction — is relatively mature except for the Waymo-specific geopolitical risk of its Zeekr dependency. The humanoid robot supply chain, by contrast, faces genuine near-monopoly constraints at the actuator layer that have no near-term resolution short of new entrants scaling production. This asymmetry suggests that AV fleet expansion faces primarily financial and regulatory constraints in 2026 to 2028, while humanoid robot scale faces a physical supply constraint that is harder to resolve with money alone.

Section 7 — Supply Chain Benchmark Summary

Mapping the hardware supply chain as a Physical AI benchmark dimension produces a differentiated picture of ramp constraint by company and product category.

The Physical AI ramp is not just software, regulation, and capital — it is also hardware. The supply chains analyzed in this article will shape how fast the Physical AI fleet can actually grow. Software can be updated remotely at negligible cost. Regulatory approvals can accelerate. Capital can be deployed at scale. Hardware supply chains require physical manufacturing capacity that takes years to build. The companies that recognized this earliest and invested in vertical integration — Tesla being the clearest example — have secured a supply chain advantage that is not easily replicated by competitors in the 2026 to 2030 window.

Note: All figures labeled “(est.)” are derived from public market information, analyst estimates, industry reporting, and company investor relations materials as of mid-2026. Component costs are indicative estimates that vary significantly by volume tier, specification, and supplier contract terms. Supply constraint ratings reflect assessment of publicly available information and are subject to change as supply chains evolve. This article does not constitute investment advice.

Sources

• Waymo Gen 6 Zeekr vehicle announcement — Waymo ↗
• Luminar lidar technology — Luminar Technologies ↗
• Tesla HW4 full self-driving computer — Tesla ↗
• NVIDIA DRIVE Orin automotive SoC — NVIDIA ↗
• Harmonic Drive robotics gearboxes — Harmonic Drive ↗

Tags