Driver Assistance·May 23, 2026·7 min read·6 sources

Tesla FSD vs Waymo Driver: architecture and fleet economics

Tesla bets on end-to-end vision under a $1–2k BOM ceiling. Waymo bets on multi-modal fusion plus HD maps with hardware near $20k per vehicle.

#Tesla#Waymo#Autonomy#ADAS#LiDAR#Sensor Fusion#Neural Networks#Fleet Economics
Key takeaways
  • Tesla FSD v14.3.2 (April 2026) unifies one end-to-end multimodal network across Robotaxi, FSD, and Summon, with an MLIR-rewritten runtime delivering ~20% faster reaction time.
  • Waymo went fully driverless on the 6th-gen Driver in Feb 2026 — 13 cameras, 4 LiDAR, 6 imaging-radar, external audio receivers, ~42% fewer sensors than 5th-gen.
  • Hardware BOM differs by ~10×: ~$1–2k per Tesla vs sub-$20k per Waymo suite (50% lower than 5th-gen) — the central constraint on each fleet model.
  • Tesla scales globally under stochastic safety bounded by compute and regulator acceptance; Waymo scales city-by-city under verified safety bounded by capex, targeting ~1M paid rides per week across 15 metros.
  • Over 12–36 months, watch Tesla critical-disengagement curves and Waymo per-city cost-per-mile — those two numbers decide the race.

Tesla and Waymo run incompatible playbooks, and 2026 has sharpened the gap rather than closed it. Tesla shipped FSD v14.3.2 in April 2026, collapsing Robotaxi, customer FSD, and Summon into one multimodal end-to-end network running on ~$1–2k of in-car compute. Waymo dropped the safety driver from its 6th-gen Driver in February 2026 and is scaling toward ~1M paid rides per week on a sub-$20k sensor suite tied to centimeter-accurate prior maps. The architectures, the BOMs, and the deployment models all follow from that split.


Two scaling theses

The question is not which stack drives better in a demo. It is which one converges first on two numbers: critical disengagements per million miles, and fully-loaded cost per autonomous mile. Tesla minimizes BOM and bets that the multi-million-car HW4 fleet uploading video closes the long tail. Waymo minimizes uncertainty per mile and bets that paid rides across its current six metros — plus the nine expansion cities planned for 2026 (Washington, Detroit, Las Vegas, San Diego, Denver, Dallas, Houston, San Antonio, Orlando) — reach unit-economic positive before BOM falls far enough to threaten the moat.


System architecture

DimensionTesla FSD v14Waymo 6th-gen Driver
Primary perception8× cameras + audio, monocular passive video13 cameras (17 MP) + 4 LiDAR + 6 imaging-radar + EARs
Network topologyEnd-to-end multimodal MoE, unified across productsModular perception + fusion + planner
Map dependencyLow — real-time topology from visionHigh — centimeter HD prior maps required
Inference computeHW4 in-car NPU, MLIR runtime (~20% faster v14.3)Redundant onboard compute, kW-class draw
LocalizationOccupancy networks + GPS3D LiDAR registration against HD base map
Safety envelopeDriver-supervised L2; supervised Robotaxi pilotDriverless L4 live in 6 metros (Feb 2026)
Hardware BOM / car~$1,000–$2,000~$15,000–$20,000 (50% under 5th-gen)
Operational domainGlobal, unbounded, driver-on-loopBounded geofences, fully verified per city

Tesla: one multimodal network across all products

FSD v14 replaces hand-written C++ perception and planning with a single end-to-end network. v14.3.2 (April 2026) extends that further: the same model now drives Robotaxi, customer FSD, and Actually Smart Summon. Inputs include video, navigation, vehicle state, and cabin/exterior audio. A Mixture-of-Experts (MoE) architecture activates specialized sub-networks per driving regime. The v14.3 MLIR compiler rewrite cut runtime overhead enough to lift reaction-time benchmarks ~20%.

[1]8× cameras @ 36 Hz + audio + nav + vehicle state

[2]Temporal transformer (MoE)

[3]3D occupancy + vector space

[4]Planner head

[5]Steering / throttle / brake actuation

The cost of collapsing the stack is interpretability. The system encodes scene understanding as learned weights, so unexpected behavior at a novel construction zone or a stalled emergency vehicle cannot be debugged by inspecting an object tracker — only by mining the failure clip and retraining.

Waymo: deterministic fusion, now fully driverless

The 6th-gen Driver fuses active and passive sensing against a pre-built 3D map. LiDAR returns centimeter point clouds, imaging-radar tracks velocity through fog and precipitation, and 17 MP cameras carry classification. External audio receivers (EARs) localize emergency-vehicle sirens before they enter visual range. As of February 2026 the suite is running without safety drivers in commercial service, validating the BOM reduction without falling back to the 5th-gen rig.

[1]LiDAR + imaging-radar + 17 MP video

[2]Sensor fusion layer

[3]Localization on HD prior map

[4]Constrained planner

Scaling and unit economics

The 10× BOM gap is not cosmetic. At ~$1.5k per car amortized into the vehicle sale, Tesla’s capex per autonomous seat is paid by the owner. Waymo carries the ~$15–20k suite on its own balance sheet for every vehicle in service, on top of the ~$50k base platform (Zeekr RT, Hyundai IONIQ 5 next-gen). That cost has to be recovered over the vehicle’s revenue miles before depreciation and remote-ops overhead.

  • Tesla — asset-light, supervision-gated. Network value compounds only if v14.x achieves an unsupervised safety case regulators accept. The Austin Robotaxi pilot still runs with in-car safety monitors as of mid-2026. Until that gate clears, revenue is software-attach on owner-driven cars.
  • Waymo — asset-heavy, ODD-gated. Each new metro costs survey, mapping, depot, and remote-ops capex before the first paid ride. Scaling is linear in cities, not in software releases — but the nine new 2026 launches are the largest single-year expansion in the program’s history.

Operational reality

Each architecture fails along a predictable axis. Tesla’s monocular geometry degrades when temporal cues thin out (parked-car occlusion at night, oncoming high-beam glare, snow erasing lane markings). Without LiDAR ground truth, the recovery path is fleet data plus the next training run. Waymo’s active sensing holds in those conditions, but the system stops outside the map — unsurveyed construction reroutes and ODD edges generate remote-assistance calls that cap throughput per operator.

The published numbers worth tracking are NHTSA standing-general-order incident reports, California DMV disengagement filings, and Waymo’s own quarterly safety disclosures. Anything else — demo videos, owner anecdotes, leaked builds — does not move the safety case.


Strategic outlook (12–36 months)

  • Tesla. The binary question is whether v14 (or v15) produces a defensible miles-between-critical-disengagement curve before the AI4/HW4 cohort ages out of warranty. If yes, robotaxi scales over-the-air across a multi-million-car installed base. If no, FSD remains a software-attach product.
  • Waymo. The binary question is per-city payback. Sub-$20k BOM, partnerships with Zeekr and Hyundai, and rideshare integrations compress the cost curve, but each new metro still requires a multi-quarter ramp before unit economics turn.
  • Convergence risk. A vision-only stack adding cheap automotive LiDAR (sub-$500 class) and a fusion stack stripping a LiDAR ring would narrow the BOM gap inside 36 months. Neither company has signaled that move; both have committed cycles to the current bet.

Related vehicles

Vehicles referenced in this analysis.

Tesla Model Y 2026Tesla Model 3 2026
About this analysis

MotiveGrid Engineering Team

Written and reviewed by engineers with production experience across powertrain, battery, ADAS / autonomy, functional safety, and large-scale consumer hardware. Every analysis follows the published MotiveGrid methodology — primary sources, transparent assumptions, explicit confidence levels.