The AV Industry and the Data Bottleneck

Sandbox brief for the Applied Intuition Data Intelligence team. Mandate framing: "make high-quality labeled data easier to use for AVs and robotics." The angle taken here is: who builds AVs, what data do they consume, and where does labeled data become the rate-limiting reagent?

Last verified: 2026-05-08. Original draft was written from training knowledge alone; this revision adds inline citations from primary sources (company press releases, SEC filings, Reuters/Bloomberg/TechCrunch/CNBC). Items remaining as [unverified] could not be confirmed via public sources at verification time.

A. Player landscape

The industry is no longer a single race. It has split into four tracks with very different unit economics, regulatory exposure, and data needs.

A.1 Robotaxi / passenger AV

These are the players trying to remove the driver from a paid passenger ride. The bar is full SAE Level 4 in a defined ODD (operational design domain).

Waymo (Alphabet). The clear commercial leader. By late 2025, Waymo One ran driverless rides in Phoenix, SF, LA (launched Nov 12, 2024), Austin (via Uber, March 4, 2025), and Atlanta (via Uber, June 24, 2025), with Miami, DC, Dallas, Houston, San Antonio, Orlando, and Tokyo in rollout for 2026 (Waymo blog – five new cities, Nov 2025, TechCrunch on Austin/Atlanta partnership). Paid rides crossed 200,000/week in February 2025, 250,000/week in April 2025, and ~450,000/week by December 2025 (per a Tiger Global investor letter); Waymo completed ~14M trips in 2025, roughly 3x the ~4.5M done in 2024 (CNBC, Dec 8 2025). Cumulative rider-only miles reached 100M by July 15 2025 and ~170.7M by December 2025 (The Robot Report, Waymo Safety Impact). Stack: modular, multi-sensor (lidar + radar + camera), HD-map-anchored, mostly-classical planner with ML perception/prediction (MotionLM). Vehicles: Jaguar I-PACE (legacy), Zeekr-built purpose-built robotaxi, Hyundai IONIQ 5 partnership announced October 4, 2024 (reported plan for ~50,000 vehicles by 2028, ~$2.5B deal) (Waymo blog, TechCrunch). Toyota co-development of a new AV platform also confirmed in 2025. Data: own fleet logs + heavy internal sim ("Carcraft" / Simulation City). 2025 milestones: freeway and airport ODDs in select markets, Moove fleet-ops partnership in Phoenix and Miami. (Waymo, Safety Hub)
Tesla (FSD / Robotaxi / Cybercab). Largest data flywheel by fleet size — millions of cars reporting trigger-conditioned clips and shadow-mode disagreements. FSD v12 (early 2024) replaced ~300k lines of C++ planner with an end-to-end neural net. FSD v13 shipped to HW4 vehicles starting late November 2024 (v13.2.1 December 16, 2024), with v13 explicitly HW4-only (notateslaapp release notes). FSD v14 rollout began October 7, 2025 (v14.1.1 on HW4); v14.3.x shipping by April 2026, with an "FSD v14 Lite" planned for HW3 cars by end of June 2026 (Tesla Oracle). Robotaxi service launched in Austin June 22, 2025 as a small geofenced pilot (Model Y, safety monitor in front passenger seat) (TechCrunch); expanded to Dallas/Houston by April 2026. Early launch saw NHTSA scrutiny over wrong-way driving, phantom braking, and traffic violations. The Cybercab (two-seater, no steering wheel) was unveiled at "We, Robot" on October 10, 2024 (CNBC); the first production unit rolled off Giga Texas on February 17, 2026, with continuous production starting around April 2026 (InsideEVs, Wikipedia: Tesla Cybercab). The Dojo supercomputer project was shut down in August 2025, with Tesla pivoting to AI6/AI5 chip designs and external compute (NVIDIA H100/H200/B200) — a notable strategic retreat (TechCrunch, Sep 2025). Stack philosophy: vision-only, end-to-end, no HD maps, no lidar. (Tesla AI, Robotaxi)
Cruise (GM). Effectively shut down as a robotaxi business. After the October 2, 2023 SF pedestrian-drag incident (a woman struck by a human driver was thrown into a Cruise AV's path; the AV then dragged her ~20 feet during a pull-over maneuver) (CBS News), the CA DMV pulled the permit, the CEO departed, ~24% of staff were cut, and Cruise paid both a $500K DOJ fine and a $1.5M NHTSA penalty for filing an incomplete crash report. On December 10, 2024, GM announced it would stop funding Cruise robotaxi development and fold the team into GM's personal-vehicle ADAS effort, citing capital priorities and competitive intensity; GM had spent over $10B on Cruise (CNBC, GM News release). The Origin shuttle was abandoned; engineers dispersed to Waymo, Tesla, Nuro, Applied Intuition through 2025. With this decision, every US automaker except Tesla had exited the robotaxi race.
Zoox (Amazon). Purpose-built bidirectional vehicle (no front, no steering wheel). Public-road testing in SF, Las Vegas, Austin, and Miami in 2024–2025. Public rider service launched in Las Vegas on September 10, 2025 with free rides — a first in the US for a purpose-built robotaxi with no driver controls (CNBC, Bloomberg). SF early rider program followed; Austin and Miami announced as next markets. Vertically integrated; well-capitalized via Amazon. (Zoox)
Pony.ai. Nasdaq IPO November 27, 2024 (ticker PONY): $13/ADS, raised ~$413M total ($260M ADSs + ~$153M concurrent private placements), ~$4.55–5.25B initial valuation (TechCrunch, SiliconANGLE). Robotaxis in Beijing, Guangzhou, Shenzhen; among the few approved for fully driverless paid rides in multiple Chinese tier-1 cities. Also operates PonyTron robotrucks. Lidar-heavy modular stack.
Motional (Hyundai + Aptiv). In May 2024 Aptiv sold an additional 11% stake to Hyundai for ~$448M, reducing Aptiv's interest from ~~50% to 15%; Hyundai injected nearly $1B (~~$475M into Motional directly) (Aptiv press release, TechCrunch). Motional paused commercial launches and pushed its Ioniq 5 robotaxi commercialization to 2026. Restarted limited operations in Las Vegas / Pittsburgh in early 2026 (techbrew).
May Mobility. Low-speed shuttle / micro-transit (Ann Arbor, Sun City AZ). Toyota partnership; driverless ops in Sun City announced late 2024. MPDM (Multi-Policy Decision Making) planner. [unverified — exact 2024 driverless launch date]
WeRide. Nasdaq IPO October 25, 2024 (ticker WRD), $15.50/ADS, ~$120M raised; positioned as the world's first publicly listed universal autonomous-driving company (WeRide IR). Robotaxis, robobuses, robosweepers in China and the UAE; Uber partnership for the Middle East.
Baidu Apollo Go. Largest Chinese robotaxi by ride volume; Wuhan flagship. Apollo cited 8M+ cumulative rides by October 2024, 9M+ by January 2025, 11M+ by May 2025, and 14M+ by August 2025 (CarNewsChina, Gasgoo). Operates the RT6 6th-generation robotaxi at ~~200,000 yuan (~~$28k) per unit, a 60% cost reduction vs the prior generation (CnEVPost) — a real structural cost advantage vs Waymo's Zeekr or Cybercab.
Aurora Innovation. Public (Nasdaq: AUR). Trucking-first — passenger AV is not on the roadmap. Launched commercial driverless freight on Dallas–Houston I-45 in early May 2025 with launch customers Uber Freight and Hirschbach (Aurora IR).
Wayve. UK-based, end-to-end. Raised $1.05B Series C in May 2024 led by SoftBank with Nvidia and Microsoft participating — the largest UK AI raise to date (Wayve press, TechCrunch). Subsequently raised $1.2B Series D in February 2026 at an $8.6B post-money valuation, with Uber investing additional milestone-based capital (Wayve press, The Robot Report). Released GAIA-1 (2023), GAIA-2 (March 26, 2025) and GAIA-3 (2026) generative world models (Wayve GAIA-2 paper, Wayve press). Uber partnership announced June 2025; L4 trials in London targeting spring 2026 after the UK government fast-tracked AV legislation (TechCrunch). US testing presence (Bay Area) announced. Distinguishing claim: minimal HD maps, geography-portable.
Waabi. Toronto-based, founded by Raquel Urtasun (ex-Uber ATG). End-to-end "AI-first" with a Waabi Driver foundation model trained substantially in Waabi World, their high-fidelity neural simulator. Unveiled the Volvo VNL Autonomous truck with Volvo Autonomous Solutions at TechCrunch Disrupt 2025 (TechCrunch); planned end-of-2025 driverless launch was delayed to "the next few quarters" pending final validation. Uber Freight CEO Lior Ron joined Waabi in August 2025 (CNBC). Integrating NVIDIA DRIVE Thor into the Waabi Driver.

A.2 Trucking / freight

Long-haul highway autonomy is structurally easier (no pedestrians, simpler ODD, paying-customer freight) and the unit economics are clearer. Consolidation has been brutal.

Aurora Innovation. First commercial driverless freight in the US — Dallas–Houston, Peterbilt + Volvo trucks, early May 2025 (Aurora press release). Partners: Uber Freight, FedEx, Werner, Schneider, Hirschbach. Network expanded by tripling driverless lanes during 2025, including Fort Worth–El Paso (Aurora IR). Modular stack on FirstLight FMCW lidar. (Aurora)
Kodiak Robotics. Texas-based; commercial-scale highway autonomy plus US DoD autonomous-ground-vehicle work. Announced a SPAC merger with Ares Acquisition Corp II in April 2025; merger closed September 23, 2025 with the combined company trading on Nasdaq as KDK from September 25, 2025 at a pre-money equity value of $2.5B and >$275M raised (Kodiak press release, TechCrunch). Customer: Atlas Energy Solutions (initial 100-truck order, 8 driverless deployed); has also delivered loads for Maersk, IKEA, J.B. Hunt, Bridgestone.
Plus (formerly Plus.ai). Pivoted to L2+/L4 driver-out via OEM partnerships (Iveco, Hyundai). Less of an independent operator now. [unverified — current status]
Gatik. Middle-mile B2B (Walmart, Loblaws). Narrow ODD (short-haul, fixed-route) makes commercialization easier; driverless on select US/Canada routes. [unverified — current 2025 deployment status]
Torc Robotics. Daimler Truck subsidiary. Commercial launch with the Freightliner Cascadia targeted for 2027. [unverified]
Embark. Defunct since March 2023.
Stack AV. Founded by Argo AI alumni after Argo's late-2022 shutdown; backed by SoftBank. Lower profile through 2024–2025. [unverified — recent activity]
Bot Auto. Founded 2024 in Houston by Xiaodi Hou (TuSimple co-founder) after TuSimple's US wind-down; ~$20M pre-Series A and ~40 engineers (mostly ex-TuSimple) at debut (TechCrunch). Transformer-based stack.
TuSimple. Effectively wound down in the US in 2024, delisted, pivoted to AI-animation in China — a cautionary tale.

The trucking field roughly looks like: Aurora and Kodiak in the lead, Waabi/Torc as foundation-model and OEM-aligned challengers, Gatik in middle-mile, the rest consolidating.

A.3 ADAS / Tier-1 stack vendors

These companies sell driving software and silicon to OEMs rather than running fleets themselves. The volume is enormous (tens of millions of vehicles), the autonomy level lower (L2–L2++), and the relationship to data is mediated by the OEM customer.

Mobileye (Intel-controlled, public). The dominant ADAS chip+stack vendor — EyeQ SoCs at >100M cumulative volume. 2024 was a difficult year: revenue fell ~20% YoY to $1.65B (from $2.08B in 2023) on customer inventory destocking; ~200 employees laid off (mostly Israel) and the LiDAR development unit was shut down (Times of Israel, Calcalist). Bright spot: a CES 2024 announcement of design wins with a "large global Western automaker" for SuperVision, Chauffeur, and Drive across 17 ICE/EV models starting 2026 (later identified as VW Group, with VW intensifying its Mobileye partnership in March 2024 across Audi, Bentley, Lamborghini, Porsche) (Mobileye press). On the BMW side: BMW Chauffeur was historically a Mobileye partnership but Qualcomm has been winning new BMW ADAS contracts; the company re-emphasized "Mobileye Drive" hands-off L2+/L3 across more OEMs. Initial Mobileye Chauffeur production planned on Polestar 4. REM (Road Experience Management) — crowd-sourced HD maps from customer cars — remains the distinctive data asset. [unverified — exact mechanics of any "BMW SuperVision loss" to Qualcomm]
Nvidia DRIVE. DRIVE Thor (~2,000 TOPS) ramping 2025–2026 across Mercedes, Volvo, JLR, BYD, Xiaomi, Lucid, Lotus, and others. Nvidia bundles silicon + DriveOS + DRIVE Sim (Omniverse) + the Cosmos World Foundation Model platform, announced at CES on January 6, 2025, openly released for physical AI / AV / robotics; first AV adopters include Waabi, Wayve, Foretellix, and Uber (NVIDIA newsroom, arXiv: Cosmos).
Qualcomm Snapdragon Ride. Cockpit-plus-ADAS bundling is Qualcomm's wedge. Wins at BMW (from Mobileye), Mercedes, Stellantis. More chip + middleware than end-to-end stack.
Bosch, Continental, Magna, Aptiv. Classical Tier-1s; mostly L2/L2+. Bosch and Continental cut significant ADAS staff in 2024 in an industry-wide L3 reset. [unverified — exact headcount figures]
Horizon Robotics. Chinese ADAS chip+stack vendor. Hong Kong IPO October 24, 2024, raising ~$696M (HK$5.4B) at the top of the price range — the largest Hong Kong IPO of 2024 (Bloomberg, Horizon press). Strong domestic penetration (BYD, Li Auto, etc.).
Huawei ADS. "ADS 3.0+" is widely considered the top Chinese consumer ADAS in independent comparisons; ships in HarmonyOS-Smart-Alliance brands (Aito, Luxeed, Stelato, Maextro). End-to-end neural planner since 2024.
BYD "God's Eye" (天神之眼). Released February 10, 2025 across BYD's full lineup down to ~100,000-yuan models like the Seagull. Three tiers (A/B/C) tied to DiPilot 600/300/100 compute and varying lidar counts. By November 2025 ~2.3M BYD vehicles in China shipped with God's Eye (CarNewsChina, CleanTechnica). The democratization-of-ADAS event of the cycle, and a structural threat to anyone selling ADAS into low-end Chinese segments.

A.4 "Foundation-model" / end-to-end approach

Crosscutting category — these companies share a technical philosophy rather than a vertical: drive policy is a single (or near-single) learned model, with HD maps de-emphasized.

Tesla — vision-only, end-to-end, HW4/AI5 (above).
Wayve — embodied-AI / "AV2.0", GAIA-2 (above).
Waabi — Waabi Driver + Waabi World neural sim (above).
Pi. [unverified — no widely-known AV company by this name through May 2026; possibly a confusion with another player. Removing or verifying recommended before customer-facing use.]
Comma.ai. George Hotz's open-source-leaning ADAS — openpilot runs on the comma 3X retrofit device across 250+ supported car models. End-to-end-ish (perception + lateral/longitudinal control) with a community data fleet. Not L4, but a uniquely crowd-driven flywheel.
Helm.ai. California. "Deep Teaching" — large-scale self-supervised learning from unlabeled video to reduce manual labeling. Multi-year ADAS joint development with Honda confirmed (initial partnership announced 2024, ADAS mass-production agreement signed August 2025; mass production targeted post-2027) (Helm.ai press, Honda Xcelerator). Honda made an additional Helm.ai investment in October 2025 (Honda Global). Generative-sim products: VidGen-2, GenSim-1/2, WorldGen-1.
Nuro. Announced its pivot September 11, 2024 from operating delivery bots to licensing the Nuro Driver stack (L4) and Nuro Driver Assist (L2++) to OEMs and AV operators; Nuro Driver runs on NVIDIA DRIVE Thor with Arm Neoverse (TechCrunch, Nuro press). Subsequently raised $106M to support the licensing model (The Robot Report).

The end-to-end bet: as compute and data scale, hand-coded planners and HD-map dependencies become liabilities. The modular bet: safety cases and regulators demand interpretable intermediate representations. Both camps still need vast labeled data — just for different supervision targets.

B. Data strategies and pipelines

The AV data lifecycle, end-to-end, looks roughly like this:

[Collection] → [Ingest/triage] → [Curation/mining] → [Labeling] → [Training] →
[Eval (sim + closed-loop + replay)] → [Deployment (shadow → active)] → [Re-collection]

Every loop iteration is a flywheel turn, and the rate-limiter shifts depending on the team's stack philosophy.

B.1 Collection: fleet logs vs simulation vs customer-shadow

Fleet logs. The traditional Waymo-style approach: instrument a small fleet (hundreds to low thousands of vehicles) with the full sensor suite (lidar, multiple cameras, radar, GNSS/IMU), drive a lot of miles, and store the raw drives. Cost is high per mile (~$10s per mile loaded) but signal quality is maximal because you control sensor placement and synchronization.
Customer-shadow data. Tesla's flywheel: deploy on millions of consumer vehicles, only upload trigger-conditioned clips when (a) human disengages FSD, (b) the policy disagrees with the human, (c) a rare-event detector fires. This is bandwidth-efficient and biases toward long-tail scenarios. Mobileye's REM is a similar idea for HD-map crowd-sourcing.
Simulation. Two flavors: (1) Replay-and-perturb — take real logs and edit actor trajectories (Waymo's Simulation City, Cruise's now-defunct system, Aurora's Virtual Testing Suite). (2) Generative — world models that synthesize entirely new driving video conditioned on actions or scenario parameters (Wayve GAIA-2, Nvidia Cosmos, Waabi World, Tesla's neural sim). Generative sim is the area that most rapidly closed the realism gap in 2024–2025.
Crowd / open data. nuScenes, Waymo Open Dataset, KITTI, Argoverse 2, ZOD (Zenseact), Lyft L5 (legacy) — useful for benchmarking but too small for production training.

B.2 Auto-labeling and human-in-the-loop

This is the crux of the Data Intelligence pitch.

Auto-labeling pipelines typically use a much larger / slower / multi-pass / multi-modal "teacher" model that has access to the full log (including future frames) to produce 3D boxes, lane geometry, semantic seg masks, and trajectory tracks at quality levels that approach human labels. The fleet/online "student" model is then distilled from this teacher.
Waymo published their "Block-NeRF" and offboard 3D auto-labeling pipelines. Tesla disclosed its "occupancy networks" and 4D-BEV auto-labeling in 2022–2023 AI Day talks. Most serious AV teams now treat human labeling as a last resort for the long tail and edge cases — not a default.
Human-in-the-loop is concentrated in (a) correcting auto-label failures on rare classes (pedestrians in unusual poses, construction zones, emergency vehicles, debris, animals), (b) behavioral-intent labels (is that pedestrian going to cross?), (c) safety-critical edge case adjudication, and (d) reward/preference labeling for end-to-end policy fine-tuning (the AV analogue of RLHF).
The mature labeling-tool stack (Scale AI, Labelbox, Encord, Voxel51, V7, plus in-house tools at every major lab) is increasingly an orchestration layer over auto-labelers, not a labelers-of-pixels layer. Applied Intuition's data tooling sits in roughly this orchestration / curation / eval band.

B.3 Scenario mining and long-tail discovery

The expensive miles are the rare ones. Productive teams have:

Embedding-based mining. Embed every clip (CLIP-style or domain-specific), then nearest-neighbor query around known failure cases, or cluster to find under-represented modes. This is now table stakes.
Tag/rule-based slicing. "All night-time left turns at unprotected intersections in rain on streets with streetcar tracks." Tesla and Waymo both maintain hundreds-to-thousands of named slices for regression eval.
Disagreement mining. Compare student model output to teacher (auto-label) or to human driver — disagreements are training gold.
Counterfactual / scenario perturbation. Take a real scenario, sweep parameters (other-actor speed, cut-in aggressiveness, weather), and measure failure surface. This is Applied Intuition's classic wheelhouse.

B.4 Data flywheel concepts

Tesla flywheel. Millions of cars → trigger-conditioned uploads → auto-label cluster (Dojo, then H100/B200 fleets after Dojo's apparent de-prioritization in 2024–2025) → train next FSD net → ship via OTA → measure new disengagement rate → repeat. Theoretically the most data-rich flywheel in the world; the ongoing question is whether vision-only and trigger-conditioned sampling are enough without lidar ground-truth.
Mobileye REM. Customer cars upload sub-1KB-per-km landmark and lane observations; aggregated into a continuously updated HD map ("AV Map"). Used as both a perception prior and a localization input.
Waymo flywheel. Smaller fleet, but each mile is fully sensed and fully owned. Heavier reliance on simulation-replay-with-perturbation to scale beyond the real-mile bottleneck. The bottleneck is fleet size, partially offset by sim.
Wayve / Waabi flywheel. Flywheel is centered on the world model — collect real driving, train the generative world model, then sample synthetic-but-realistic scenarios from it for downstream policy training. The bet is that the world model is a compressed and infinitely sample-able substitute for re-collection.

B.5 Where the bottlenecks actually are (in practice)

Compute is rarely the limit at the labeling stage — H100/H200/B200 is plentiful for the labelers. The bottleneck is throughput of the human-in-the-loop and the trust placed in auto-labels for safety-critical classes.
Scenario coverage is the limit at the eval stage. Nobody has a defensible answer to "have you covered enough rare events to argue safety to a regulator?" except by running enormous simulation suites and arguing about coverage metrics.
Cross-team data interop is the limit at the org stage. Sensor configurations differ, label schemas differ, coordinate frames differ. A perception team and a planning team inside the same company often can't share data without a translation layer. This is exactly the gap a tooling vendor sells into.

C. Key inflection points 2024–2026

C.1 Cruise wind-down (Dec 2024)

GM's December 10, 2024 decision to halt Cruise robotaxi funding and absorb the team into personal-vehicle ADAS removed one of the three large US robotaxi efforts overnight (CNBC, GM news). The downstream effects: Waymo became effectively unchallenged in US commercial robotaxi for 12+ months (until Tesla's Austin pilot in June 2025 and Zoox's Las Vegas launch in September 2025); talent dispersed broadly (visible flow into Applied Intuition, Nuro, Waymo, and several end-to-end startups). [unverified — GM's specific 2028 personal-vehicle stack details]

C.2 Waymo geographic expansion

Through 2024–2025, Waymo went from a two-city service (Phoenix + SF) to live service in LA (Nov 12, 2024), Austin (via Uber, March 4, 2025), Atlanta (via Uber, June 24, 2025) (Uber IR, Engadget), and announced rollouts in Miami (announced Dec 2024 via Moove for 2026), Washington DC (March 2025, for 2026 launch) (Waymo blog), plus Dallas, Houston, San Antonio, Orlando (Waymo Nov 2025 blog). Tokyo testing began April 2025 with Nihon Kotsu and GO partnerships (Waymo blog); Toyota AV-platform co-development confirmed. The pattern matters: Waymo started running its commercial product on partner fleets (Uber app + Moove fleet ops) rather than vertically operating, which reduces the capex curve and accelerates city onboarding.

C.3 Tesla End-to-End FSD v12/v13/v14 and Robotaxi

v12 (Q1 2024): end-to-end neural-net planner replaces hand-coded code. Step-change in smoothness; modest step in safety metrics.
v13 (Q4 2024): HW4-native (HW3 stayed on v12), longer context, better night driving, materially improved unprotected-left and roundabout behavior per Tesla's own and third-party trackers (notateslaapp release notes).
v14 (October 7, 2025 onwards): 14.1 began rollout to HW4 in October 2025; 14.3.x by April 2026; "FSD v14 Lite" planned for HW3 by end-June 2026 (Tesla Oracle).
Robotaxi service: Austin launch June 22, 2025, with safety monitors in the front passenger seat (TechCrunch). The strategic significance: Tesla committed publicly to the L4 robotaxi market on a vision-only stack in the same year that Waymo grew from ~200k to ~450k weekly rides on a lidar-heavy stack — directly testing two opposing hypotheses about what data architecture can solve driving.
Dojo shutdown (August 2025): Tesla disbanded the Dojo team and pivoted to AI6/AI5 chip designs plus external H100/H200/B200 compute, reflecting that the bottleneck for FSD scaling is increasingly not training compute but data quality, sim, and policy supervision (TechCrunch). About 20 ex-Dojo engineers spun out into DensityAI.

C.4 Wayve $1.05B, $1.2B Series D, and GAIA-2/3

May 2024 Series C ($1.05B led by SoftBank with Nvidia and Microsoft) established Wayve as the best-capitalized end-to-end AV pure-play outside the US/China (Wayve press). GAIA-2 (March 26, 2025) demonstrated multi-camera, action-conditioned generative driving video at higher fidelity than GAIA-1, with cross-country transfer (UK/US/Germany) (GAIA-2 paper). GAIA-3 (2026) extended world models from simulation to evaluation. Wayve raised an additional $1.2B Series D in February 2026 at an $8.6B post-money valuation, with milestone-based capital from Uber and a partnership to expand to >10 global markets (Wayve press, The Robot Report). London robotaxi trials with Uber begin spring 2026 following accelerated UK AV legislation.

C.5 Waabi commercial trucking

Waabi unveiled the Volvo VNL Autonomous truck at TechCrunch Disrupt 2025, with Volvo Autonomous Solutions and Uber Freight as partners (TechCrunch). The originally planned end-of-2025 driverless launch slipped "the next few quarters" pending validation. Uber Freight CEO Lior Ron joined Waabi in August 2025. NVIDIA DRIVE Thor is being integrated into the Waabi Driver. The strategic claim: the first credible foundation-model-stack production launch on highways, even if slightly delayed.

C.6 Mobileye losing customers and pivoting

Mobileye's 2024 was painful: full-year revenue fell ~20% to $1.65B (from $2.08B in 2023) on customer inventory destocking; ~200 employees laid off (mostly Israel); the LiDAR development unit shut down (Calcalist, Times of Israel); Intel's stake overhang continued. Bright spot: a CES 2024 17-model design-win package with VW Group across Audi/Bentley/Lamborghini/Porsche on SuperVision/Chauffeur/Drive (rolling out from 2026) (Mobileye press). The company's response was twofold — emphasize the "Mobileye Drive" eyes-off product line (broader OEMs at lower margin) and lean harder on REM crowd-mapping as a structural moat. The Mobileye saga is a useful case study of why being a Tier-2 ADAS vendor is harder than it looks: OEMs are gaining confidence in either (a) building their own stacks or (b) buying from Nvidia/Qualcomm with the silicon.

C.7 Trucking AV consolidation

By early 2026 the trucking field has shrunk to roughly five serious independents (Aurora, Kodiak, Waabi, Gatik, Plus) plus OEM-internal efforts (Torc/Daimler, Volvo, Daimler/Hyundai). Embark, TuSimple, Locomation, and Ike are gone or transformed.

C.8 China AV catching up

Huawei ADS 3.0+ is now widely considered the highest-quality consumer ADAS in China [unverified — independent benchmark comparisons vs FSD not located]; ships in HarmonyOS-Smart-Alliance brands (Aito, Luxeed, Stelato, Maextro). End-to-end neural planner since 2024.
BYD's "God's Eye" (天神之眼) democratized advanced ADAS across BYD's lineup starting February 10, 2025, with the entry-level "C" tier reaching ~100,000-yuan models like the Seagull. ~2.3M God's Eye-equipped BYD vehicles in China by November 2025 (CarNewsChina).
Xiaomi entered the EV market with the SU7 in 2024 and is investing heavily in in-house ADAS.
Apollo Go (Baidu) reached 8M+ cumulative rides by October 2024, 9M+ by January 2025, 11M+ by May 2025, and 14M+ by August 2025 — at materially lower per-ride economics than Waymo, partly because of cheaper purpose-built vehicles (RT6 at ~$28k) and lower labor overhead.
Pony.ai (Nasdaq Nov 27, 2024) and WeRide (Nasdaq Oct 25, 2024) both went public, opening Chinese AV companies to US public-market scrutiny.
Horizon Robotics (HK Oct 24, 2024) raised ~$696M as the largest HK IPO of 2024.

The net effect: by 2026 the "China gap" in passenger ADAS/AV is no longer a generation behind — it's plausibly leading on consumer ADAS deployment, even if commercial robotaxi rides are still concentrated in tier-1 cities.

D. Where labeled data is the bottleneck today

This is the section that matters most for Applied Intuition's framing. The honest answer is: labels are not uniformly the bottleneck — they are the bottleneck in specific places, and those places are increasingly the long tail and the supervision-of-supervision layer.

D.1 Perception

3D object detection / tracking. Largely auto-labeled today on the head of the distribution (cars, trucks, pedestrians on cleanly-scanned frames). Human labels concentrate on occluded, partially-truncated, ambiguous-class, and rare-class objects — strollers being pushed in rain, construction barrels in unusual configurations, articulated buses, gritters, road debris.
Semantic / instance segmentation. Pixel labels are expensive. Auto-segmentation models (SAM-style) have collapsed cost dramatically since 2023, but fine-grained driving classes (lane vs lane-marking vs cycle-lane vs bus-lane vs HOV-lane) still need careful curation.
Occupancy networks. A relatively new supervision target (Tesla popularized it in 2022). Ground-truth is generated by accumulated lidar over time on a sensor rig — but for vision-only stacks (Tesla, Wayve, comma) the supervision pipeline is itself a research artifact, and the quality of occupancy auto-labels gates everything downstream.
Lane / road geometry. HD-map-style labels (centerlines, boundaries, connectivity, signage attachment) are some of the most expensive labels per mile because geometry must be globally consistent, not just per-frame.
Traffic light / sign recognition. Long-tail of regional sign types (Texas left turns, European zone-end signs, school bus stop arms) and unusual configurations (cluster signals at multi-lane intersections). Labels here directly affect safety performance.
Occlusion and attention reasoning. Labeling "what is hidden behind the truck and might emerge" is closer to behavior labeling than to perception — a class that is genuinely hard to auto-label.

D.2 Behavior / prediction

Intent labels. "Is this pedestrian about to step into the road? Is this car about to merge?" These are causal / counterfactual claims and human labelers disagree with each other ~10–25% of the time on hard cases. There is no clean ground truth — only future-frame outcomes, which are noisy.
Trajectory labels. Easier: just look at what actually happened in the next 5 seconds. But high-quality multi-agent scene labels (with social interaction structure) are still hand-curated.
Social interaction / negotiation. Who has right of way, who is yielding, who is bluffing. Fundamentally subjective. The serious teams use ranking / preference labels (this trajectory is better than that one) rather than absolute labels — this is the AV equivalent of RLHF data.

D.3 Safety eval / scenario coverage

This is arguably the largest under-served labeling category. To certify safety, teams need labels for:

Scenario taxonomy (what kind of scenario is this — for slicing eval).
Rare-event tagging (was there a near-miss, a phantom-brake, a hard cut-in, a misclassification?).
Disengagement-cause labels (why did the safety driver take over? — required for regulatory reporting in CA DMV filings).
Counterfactual outcome labels (would the AV have crashed if the human hadn't intervened? — see CA DMV's "OL316" reporting framework, NHTSA Standing General Order data).

The eval-labeling problem is qualitatively different from training labels: it must be reproducible, slice-able, and audit-able. Most general-purpose labeling tools were built for training data and are awkward for safety-eval data. This is a wide-open product space.

D.4 E2E supervision and reward modeling

End-to-end policies (Tesla v12+, Wayve, Waabi, comma, Helm.ai) have a different supervision menu:

Behavioral cloning labels = (sensor-input, expert-action) pairs at scale. The "labeling" is implicit (the human's recorded action), but filtering for expert trajectories is itself a labeling problem (was this driver actually good? was this clip safe? was it legal?).
Reward / preference labels. Same idea as RLHF: humans rank pairs of policy outputs (real or sim) to fine-tune the model. Production AV stacks are starting to use this in 2025 [unverified — public disclosure on production-AV preference fine-tuning is limited; signals from Wayve/Tesla/Waabi research are suggestive but not explicit]. The labeling problem is less "click on the car" and more "watch a 10-second clip and rate the driving."
World-model self-supervision. Generative world models (GAIA-2, Cosmos, Waabi World) are self-supervised on raw video with no labels — but they need meta-labels to slice their generations for evaluation: "show me 1,000 left-turn-against-oncoming-bus scenarios" requires that the source data be labeled by scenario type.

D.5 Multimodal fusion labels

Aligning camera + lidar + radar + map + IMU into a single labeled scene at 10–20 Hz across multi-second clips is mechanically painful. The hard parts:

Temporal consistency of object IDs across frames and across sensors.
Coordinate-frame consistency (sensors are calibrated, but calibration drifts).
Map registration of dynamic actors against a static HD-map context.
Cross-modality conflict resolution (camera says pedestrian, radar says nothing — which is right, and how is the label supervised?).

Auto-labeling addresses temporal consistency reasonably well. Cross-modality conflict and edge-case fusion are still very human-mediated.

E. Implications for Applied Intuition

Applied Intuition occupies an unusual position. It does not run a robotaxi service, does not sell silicon, and does not own a fleet — so it is structurally non-rival to almost every player in sections A.1–A.4. That non-rivalry is the precondition for selling tooling to all of them. (Context: Applied Intuition closed a $600M Series F at a $15B valuation in June 2025, co-led by BlackRock-managed funds and Kleiner Perkins, with 18 of the top 20 global automakers as customers and major US DoD programs (Applied Intuition press, PR Newswire).)

The shape of the industry as of 2026 implies four structural tailwinds for a "data intelligence" play. First, the modular vs end-to-end split is now permanent — Waymo and Wayve will not converge on the same stack — but both camps need the same lower layers (ingest, slicing, auto-label orchestration, scenario libraries, regression eval). A neutral tooling vendor that supports both supervision regimes (per-task labels for modular, behavioral-cloning + preference for E2E) captures the full market rather than betting on one stack winning. Second, the trucking and ADAS markets have OEM customers who are not going to build internal labeling-and-curation platforms — Daimler Truck, Volvo, Stellantis, and the Tier-1s are consolidating around bought tooling, and the bar for "trustworthy enough for safety eval" is high enough to favor a few incumbents rather than DIY. Third, the eval-data and scenario-coverage problem (D.3 above) is genuinely under-served by general-purpose labeling tools and is increasingly the regulatory choke point — owning this layer is the moat, because eval data is by definition longitudinal, audit-able, and stack-coupled in a way that training data is not. Fourth, the same data-intelligence primitives (curation, scenario mining, sim, eval, auto-label orchestration) port directly to non-AV embodied AI — humanoids, warehouse robots, drones, defense — which is exactly where Applied Intuition has been extending. The moat, in plain terms, is: be the system of record for "this stack was safe enough on this slice of the world," across stacks, across modalities, across customers — because that artifact is what regulators, OEMs, and insurers will eventually demand, and no fleet operator will want to build it twice.

Sources

Primary references used during the May 8, 2026 verification pass. Inline citations point to the exact URL that supported each claim; this section consolidates them.

Waymo

Tesla

Cruise / GM

Aurora

Kodiak

Wayve

Waabi

Zoox

Pony.ai / WeRide / Horizon Robotics

Baidu Apollo Go / BYD

Mobileye / NVIDIA / Helm.ai / Nuro / Motional / Bot Auto

Applied Intuition

Regulatory / open data