The Level 3 Divergence

Honda got there first. Japan certified Honda's Traffic Jam Pilot as the world's first Level 3 system in November 2020.

Mercedes and BMW built the two most advanced Level 3 systems in the world. Both have now stepped back from them within a similar timeframe.

General Motors and Ford are targeting eyes-off Level 3 capability for 2028.

Tesla is skipping the category entirely, betting cameras alone can reach a higher level of autonomy.

Every position here assumes a customer eventually pays for whichever version wins. That assumption turns out to be the least proven part of the rest of this memo

What Actually Changed the Math on Level 3?

BMW and Mercedes each carried lidar options priced at approximately $7,000 as part of their Level 3 packages on the 7 Series and S-Class.

Equivalent lidar-equipped hardware in China now shows up as a $3,100 option on BYD's Seagull hatchback, a car that starts at around $10,300 total.

That pricing gap is connected to Chinese manufacturing scale. Automotive lidar cost as much as $50,000 a unit a decade ago, restricted to research fleets like Waymo's early test cars.

Chinese suppliers Hesai and RoboSense scaled production hard enough to push compact units down to about $200 today. Germany built its first-generation Level 3 systems before that cost curve existed.

Is Germany Actually Retreating?

BMW's exit looks like an actual full retreat. The automaker discontinued Personal Pilot L3 with the 7 Series facelift, reverting the car to Level 2.

Mercedes looks more like a market-specific pause. The facelifted 2026 S-Class ships with Level 2++ Drive Assist Pro in the US specifically. The company’s engineering seems to be directed toward a next-generation Level 3 system developed with Nvidia.

The detail that complicates the retreat framing is that Mercedes just signed a lidar supply agreement with Hesai, the same Chinese company whose manufacturing scale helped cause the German cost problem in the first place. Mercedes didn't decide lidar was wrong. It decided German and American lidar pricing was wrong, and went and bought the hardware that already solved BYD's version of the same problem.

The VW-Rivian joint venture, covered in my prior discussions of the two companies', builds a shared zonal computing architecture that will eventually support automated driving. Rivian's own autonomy stack sits outside that joint venture, however.

It looks like VW is getting the plumbing, not the software running through it - at least not yet.

Why Is Tesla Skipping Level 3 Entirely?

Tesla’s Level 3 economics are unlike anything else in the current landscape. It’s full camera-only set costs a few hundred dollars. Waymo's sixth-generation sensor suite still costs about $20,000 per vehicle, with a fully equipped car approaching $60,000 total.

There is an intriguing counterargument that comes from Waymo, however. The company’s co-CEO Dmitri Dolgov laid out the technical case at a Y Combinator talk in August, without naming Tesla but clearly describing its approach.

Dolgov believes cameras can match human driving and be a very good assist product, since humans only navigate with two eyes. Yet, reaching autonomy well past human reliability requires clearing a bar that weak sensing does not reach because the safety curve flattens out too early.

He explains it as an exponential ladder of reliability. Each additional nine of dependability costs approximately ten times the effort of the one before it. Camera-only systems look most impressive early before the plateau while on the cheap part of that curve.

Interestingly, Tesla's own reported robotaxi numbers show a crash rate approximately three times worse than human drivers, even with a safety monitor riding along.

Tesla has logged approximately 380,000 fully unsupervised miles since removing monitors from its Austin fleet in January. Waymo covers that same distance in about a day, and its 220 million driverless miles show approximately 94% fewer serious-injury crashes than human drivers over equivalent distance.

Elon Musk's timeline on autonomy has a long history of slipping. He told investors in 2019 that a robotaxi fleet was a year out. His current guess for unsupervised FSD reaching customer-owned cars is now the fourth quarter of this year.

Dictate into Slack, email, LinkedIn, or any app and get polished, send-ready text. Wispr Flow strips filler and formats everything. 89% of messages sent with zero edits. Works on Mac, Windows, and iPhone.

Start flowing free

What About the Chip Strategy? Is It Like the Sensor Strategy?

Nvidia has become the default supplier on the in-vehicle side for nearly the entire field: Mercedes, GM's 2028 system, and a long list of Chinese automakers including BYD, Li Auto, NIO, Xiaomi and Zeekr.

GM's move to Nvidia is very interesting. Its original Ultra Cruise compute ran on Qualcomm. GM switched its flagship 2028 autonomy platform to Nvidia in 2025. That’s the same chip Mercedes is adopting.

Nearly ten years ago I acquired Nvidia shares, believing then its GPU-based hardware would be required for autonomous driving capability that I thought would soon occur. It seems my direction was sound, but my timing was off. This Nvidia consolidation is beginning to happen now.

Ford and Tesla are the two real holdouts on in-vehicle silicon. Ford says it built its next-generation compute in-house and hasn't named a supplier publicly. Tesla designs its own inference chips entirely.

But separate the car from the data center and Tesla's independence looks little thinner. Tesla shut down its in-house Dojo training supercomputer and has increased its reliance on Nvidia GPUs for the training runs that actually produce the FSD model. The company most committed to vertical independence on sensors is one of NVIDIA's larger customers where reliability is built in the data center rather than the car.

China is hedging everything. NVIDIA chips still flow into Chinese vehicles today, but BYD's own Xuanji chip and NIO's NX9031 exist inside a much larger US-China chip fight. Beijing has committed more than $50 billion to reducing their exposure by 2030.

Does Any of This Actually Sell a Car?

This is the real test all of the automoumus bets in this memo will be judged on. Detroit's lidar strategy, Tesla's camera-only wager, China's courts and speed, all assumes a customer eventually pays for the result, and that assumption has surprisingly little direct evidence behind it for Level 3 specifically.

McKinsey found approximately a quarter of buyers say they would very likely choose an advanced autonomous feature on their next vehicle, and the most interested customers would pay up to $10,000 for genuine highway autonomy.

Real payment behavior on the Level 2 systems today already tells a more complicated story. GM's Super Cruise renewal rate has climbed sharply, from approximately 2% of the earliest trial cohort choosing to pay when their free period ended, to an estimated 30% to 40% of trial customers renewing now.

Tesla shows the same acceleration. More than 55% of new North American Tesla deliveries in the second quarter of 2026 included an FSD subscription, and Musk has said some buyers now come into stores wanting Full Self-Driving specifically.

This is real evidence that customers will pay for hands-free convenience. It is not evidence about Level 3 specifically. Every one of those numbers comes from Level 2 systems, hands-free but eyes-on, and nobody selling a Level 3 product yet has published data showing the jump to eyes-off changes a purchase decision rather than just raising the price of a feature people already buy.

The regional gap may be the most best insight available now. Approximately 66% of Chinese respondents say they would buy a Level 3 vehicle, against approximately 20% in Western markets, according to one 2026 survey. If that gap holds, the cost collapse and the regulatory speed may only matter commercially in the one market already showing up to buy.

This analysis extends TaaSMaster's cost architecture framework, first applied to the GM/VW divergence in “GM Funds Transition. VW Warns of Crisis. What's the Difference?” The Rivian dimension builds on “The Same Admission, Two Different Partners,” TaaSMaster's comparison of Stellantis and Volkswagen's Chinese partnership structures. The chip self-sufficiency thread connects to “Chinese Exports Aren't Growth.”

Lidar and cost data: Carscoops, Autoblog, and IEEE Spectrum reporting on Hesai and RoboSense pricing.

Honda and the first Level 3 system: Honda Global press materials, Electronic Design, and The Drive on the Traffic Jam Pilot type approval and 2021 launch.

German retreat: WardsAuto, Electrive, Reuters, and Mercedes-Benz Group press materials on Drive Pilot, Personal Pilot L3, and STLA AutoDrive.

Tesla and Waymo safety data: Electrek reporting on Waymo co-CEO Dmitri Dolgov's Y Combinator remarks and Tesla's reported robotaxi crash data. NHTSA's Engineering Analysis EA26002 filings.

Compute strategy: Forbes (Sam Abuelsamid), TechCrunch, and NVIDIA's own DRIVE platform materials on GM's shift to NVIDIA Thor. TechCrunch and Engadget on Tesla's Dojo shutdown and AI5/AI6 roadmap.

Regulation: UNECE Regulation 157 materials, NHTSA reporting to Congress, and MIIT's national safety standard GB 44721-2026.

Consumer research and real adoption data: McKinsey's ACES consumer research and Demand Local's regional willingness-to-buy survey. GM Authority, Automotive News, InsideEVs, and Subscription Insider on Super Cruise renewal rates and Tesla's Q2 2026 FSD attach rate and Musk's remarks on purchase intent.

If you have a perspective or disagreement, reply directly. I read every response.