For the better part of a decade, the autonomous vehicle industry has been two industries pretending to be one. The Waymo-Cruise-Mobileye axis builds vehicles that, by design, do not have a driver. The Tesla axis builds vehicles that, by design, do, and treats removing the driver as a future software update.

These are not the same engineering problem and they are not the same business. The fact that the press, the regulators, and the investor class have spent ten years treating them as variants of the same thing is the single largest source of confusion in this market.

The L2 path and the L4 path

The L2 path, Tesla, Mercedes Drive Pilot, GM Super Cruise, Ford BlueCruise, the legacy OEMs broadly, builds a system whose floor is “the driver is responsible for everything the car does.” Whatever the car ships, it ships with the legal posture that a human is supervising. The cost of being wrong is borne by the human. The benefit of being right is sold as a feature upgrade. The unit of progress is feature richness and intervention frequency.

The L4 path, Waymo, Mobileye’s robotaxi line, the surviving Chinese robotaxi operators, builds a system whose floor is “the car is responsible for everything it does, within an operational design domain it controls.” The cost of being wrong is borne by the operator. The benefit of being right is sold as a ride. The unit of progress is operational design domain expansion and revenue miles.

These produce different vehicles, different organizations, different investor stories, and, critically, different liability postures.

The convergence that wasn’t

The Tesla bet has been, since 2019, that the L2 path converges with the L4 path. The argument is that if you collect enough miles of driving data from a fleet of supervised cars, eventually the supervision stops being necessary, and the same software that supervised drivers were using becomes a driverless robotaxi service. This is the entire premise of the Tesla robotaxi business.

The bet has not, so far, paid off. FSD v14 is better than FSD v13. FSD v13 was better than v12. The trend line is real. The trend line has also been real for six years, and the gap between “really good supervised driving” and “no human required” has not narrowed in a way that you can draw on a chart and extrapolate to a date.

Meanwhile the L4 operators, the ones with no convergence story to sell, who just built driverless cars from day one, are running revenue services in five US cities and approximately every Chinese tier-one city.

What the regulators are forcing

NHTSA’s new framework, the EU’s L4 type approval pathway, China’s Tier-3 robotaxi authorizations: all three are explicitly L4 regimes. They define “vehicle without traditional human controls” and they regulate that category. They do not regulate the L2 path, which falls under existing driver-supervised rules.

The result is that the L2 path and the L4 path now have different regulatory regimes, with different evidentiary requirements, different disclosure obligations, and different liability frameworks. A car cannot, in 2026, be both. The Tesla pivot, “this car you bought as an L2 vehicle is now an L4 vehicle via software”, is a thing the framework does not contemplate. Either the vehicle is certified L4 or it isn’t.

That is the wall the convergence story is now driving into. It is not clear, on the public evidence, whether Tesla has a plan for the wall.

The opinion

There is a version of the next five years in which Tesla’s L2 path produces a robotaxi service and the convergence story turns out to be correct. There is also a version in which it doesn’t, and Tesla ends up running an L4 program structurally identical to Waymo’s, purpose-built vehicles, dedicated operations, geofenced launches, the entire playbook the company spent a decade saying was unnecessary.

The market is currently priced for the first version. The engineering, on the available evidence, is increasingly producing the second.

The driver-seat question is the only question that matters. The industry is, finally, being made to answer it.