2026-06-18 👁️ — views

Physical AI Hardware Evolution — Lidar Cost Curves, Tesla HW4 to HW5, Waymo Gen 6, and the Race to Sub-$1,000 AV Bill of Materials

Lidar fell 99% from $75,000 to under $500. Tesla HW4 BOM is $300-700 vs Waymo Gen 6 at $5,000-15,000; Cybercab and Gen 7 are each company hardware cost gate.

Article 149 in the Physical AI Benchmark Series — Physical AI Hardware Evolution: Sensor Costs, Custom Silicon Roadmaps, and the Race to Sub-$1,000 AV Hardware Bill of Materials

Hardware cost is one of the most critical variables in autonomous vehicle economics. A lidar sensor that cost $75,000 (est.) in 2016 costs under $500 today. This decade-long cost collapse — approximately 99 percent over ten years — has transformed lidar from an economics-killing line item into a commodity sensor approaching camera-grade pricing. But lidar is only one part of the AV bill of materials. Compute, cameras, radar, and in-vehicle inference silicon together determine whether an autonomous vehicle can ever achieve positive unit economics at scale.

This article is Article 149 in the Physical AI Benchmark Series. It benchmarks the hardware BOM evolution for Waymo Gen 6 versus Tesla HW4/HW5, the custom silicon race between Tesla Dojo and Google TPU, and what the hardware cost curves mean for when AV unit economics turn positive. All figures labeled “(est.)” are derived from public disclosures, industry research, analyst estimates, and reported data rather than independently verified primary data. This article does not constitute investment advice.

Section 1 — Lidar Cost Curve: The Hardware Revolution That Made AV Economics Possible

What the Lidar Cost Curve Means for AV Economics

The 99 percent collapse in lidar pricing over ten years is not primarily a manufacturing breakthrough — it is a volume-driven commoditization curve playing out at the pace typical of semiconductor-adjacent hardware. The first Velodyne lidar sensors were assembled by hand in small quantities; the current generation of solid-state lidar sensors uses photonic integrated circuits and MEMS mirror arrays manufactured on semiconductor process lines, enabling mass production at automotive volumes.

The critical inflection point for AV economics was approximately 2022–2024, when solid-state lidar crossed below $500 per unit. Below that threshold, lidar stops being the primary objection in AV BOM conversations. A Waymo Gen 6 vehicle carrying three or four lidar sensors at $200–500 each (est.) is looking at $600–2,000 in lidar cost (est.) — meaningful but not the dominant line item. Compute, redundant camera arrays, high-accuracy GPS/IMU units, and the engineering integration overhead of a safety-certified sensor suite remain significant cost drivers.

Tesla’s camera-only strategy looked overwhelmingly cost-advantaged in 2016 when lidar was $75,000. It remains an advantage at $500 lidar. The question — answered below in the benchmark scorecard — is whether that advantage is decisive enough to determine the outcome of AV economics, or whether Waymo’s sensor redundancy provides enough safety-case advantage to justify the remaining BOM premium.

Section 2 — Tesla Hardware Evolution: HW3 to HW4 to HW5

The Tesla Hardware Philosophy: Camera-First, Compute-Scaling

Tesla’s hardware strategy rests on a core bet: that camera-based perception, trained at sufficient scale on sufficient real-world data, can match or exceed the safety performance of multi-sensor suites including lidar. Every hardware generation from HW3 onward has reinforced this bet — not by adding sensors, but by dramatically increasing inference compute.

HW3’s 144 TOPS (total) was state of the art for automotive inference in 2019. HW4’s estimated 500+ TOPS (est.) — a roughly 3x increase — enables higher resolution camera processing and more complex neural network architectures. HW5’s estimated 3,000–5,000+ TOPS (est.) would represent an additional 6–10x increase, designed to run the next generation of FSD models that Dojo’s D1 and D2 training infrastructure is producing.

The Dojo training cluster is not a vehicle component — it appears in no vehicle BOM. Its cost appears in Tesla’s capital expenditure, not in the marginal cost of each vehicle. This is a critical distinction for AV economics: Dojo’s benefit is amortized across every FSD-capable vehicle Tesla sells, making the marginal cost of better FSD capability approximately zero per incremental vehicle. If Dojo achieves its target of training cost reduction versus cloud alternatives, Tesla’s training cost per vehicle-improvement-cycle becomes a structural advantage.

Section 3 — Waymo Hardware Evolution: Gen 1 to Gen 6

Why Gen 6 Is Waymo’s Critical Economics Milestone

Waymo’s Gen 5 fleet — based on the Jaguar I-PACE — was a commercial validation platform, not an economics-optimized product. The I-PACE base vehicle cost approximately $70,000–80,000 (est.) before Waymo’s sensor suite was added. Gen 5 existed to prove that Waymo could operate a paid robotaxi service safely and reliably; it was never expected to generate positive unit economics.

Gen 6 is different in intent. The Zeekr RT is a purpose-built AV platform designed from the vehicle structure up to accommodate Waymo’s sensor suite at lower integration cost. Waymo has publicly stated that Gen 6 is significantly cheaper to manufacture than Gen 5. Analyst estimates of 40–60% BOM reduction (est.) imply a sensor suite cost in the range of $5,000–15,000 per vehicle (est.), down from $10,000–20,000+ (est.) for Gen 5.

That range still implies negative unit economics at current Waymo pricing. At $30–50 per ride (est.) and 3–5 rides per vehicle per hour (est.) during peak periods, even a continuously running Gen 6 vehicle generates $90–250 per vehicle-hour (est.) in gross revenue. A vehicle that costs $40,000–60,000 (est.) all-in, with 18–24 month depreciation cycles and high maintenance costs for sensor suites, is unlikely to achieve positive unit economics below a Gen 7 cost profile. The Gen 7 target of $3,000–8,000 (est.) total sensor BOM would be the critical threshold.

Section 4 — Custom Silicon Race: Dojo vs Google TPU

The Significance of Vertical Silicon Integration

The custom silicon race in AV is not primarily about peak TOPS per chip. It is about the end-to-end cost of training a model, running inference at vehicle scale, and continuously improving the model over the vehicle lifetime via OTA updates. Both Tesla and Waymo have concluded that general-purpose silicon from NVIDIA is too expensive at the scale they need, and have invested in custom silicon accordingly — but via different strategies.

Tesla’s bet on Dojo is a vertical integration of training infrastructure: design the chip, design the interconnect, design the software stack, own the complete training pipeline. If Dojo achieves its stated cost target of approximately $1 per FLOP versus cloud renting alternatives, Tesla would have a structural training cost advantage that compounds with fleet scale — more vehicles generate more shadow mode data, which requires more training compute, which costs Tesla less per FLOP than any competitor relying on NVIDIA or cloud providers.

Waymo’s bet on Google TPU is horizontal integration: rely on Alphabet’s world-class AI infrastructure, benefit from Google Brain and DeepMind’s advances in TPU architecture, and access the most advanced training hardware available without the capital expenditure of building a custom training cluster. The trade-off is that Waymo’s training cost is subject to Alphabet’s internal pricing decisions, and Waymo does not control the training silicon roadmap.

The in-vehicle inference dimension is where Tesla has the clearer advantage today. Tesla HW4, deployed across millions of vehicles in production, is a proven at-scale custom inference chip. HW5 is designed to be manufacturable at similar cost with dramatically higher compute. Waymo’s in-vehicle compute is purpose-built for a fleet of approximately 2,500 vehicles (est.) — it has never been subjected to the cost pressure of mass automotive manufacturing.

Section 5 — Hardware Cost Benchmark Scorecard

Overall Verdict

Tesla’s camera-only approach maintains a decisive sensor BOM advantage today — roughly 10 to 20 times lower sensor cost per vehicle (est.) compared to Waymo Gen 6. That gap is real and meaningful. But the strategic question is not who has the lower BOM today; it is which company’s hardware cost trajectory leads to positive unit economics first, and at what fleet scale.

The gap narrows as lidar falls below $500, and continues narrowing toward sub-$200 by 2030 (est.). A Waymo Gen 7 vehicle with $3,000–8,000 (est.) in total sensor cost — carrying lidar that costs $100–300 (est.) per unit — would have a safety case argument that camera-only HW5 Tesla vehicles cannot match, at a BOM premium that might be justified by higher rider trust and regulatory approval in more jurisdictions.

Tesla’s counter-argument is Cybercab: if Tesla can mass-produce a purpose-built robotaxi below $30,000 (Tesla stated target), the total cost of fielding an AV fleet per vehicle would be dramatically lower than Waymo’s Gen 6 or Gen 7. The company that crosses its hardware cost gate first — Waymo’s Gen 7 sensor BOM, or Tesla’s Cybercab production cost — will be the first to demonstrate positive-margin AV economics at scale. Both gates are expected to open in the 2027–2029 window (est.).

Note: All figures labeled “(est.)” are derived from public disclosures, industry research, analyst estimates, and reported data as of mid-2026. This article does not constitute investment advice.

Sources

• Lidar cost trends — Luminar Technologies ↗
• Tesla HW4 and FSD chip — Tesla AI Day 2022 ↗
• Waymo Gen 6 vehicle — Waymo blog ↗
• Tesla Dojo supercomputer — Tesla ↗
• AV hardware cost analysis — McKinsey Center for Future Mobility ↗

Tags