Three things happened in the first quarter of 2026 that belong in the same sentence.

In January, a startup called Tensor stood on the CES floor with a car it says you can own, that drives itself at Level 4, and that folds its steering wheel away when it does. In the same month, Mercedes-Benz paused Drive Pilot, the only eyes-off system American consumers could buy. In March, Sony and Honda cancelled the Afeela 1 before a single customer took delivery.

So the most ambitious version of the personal autonomous car was announced in the same season that the modest version was withdrawn. I think that tells us more than any spec sheet does.

What is actually being sold

Tensor’s Robocar is the clearest example. The company describes it as a Level 4 vehicle designed for private ownership, with a dual mode: drive it, or be driven. The hardware list is extraordinary. Eight NVIDIA DRIVE AGX Thor chips, more than 8,000 TOPS of compute, over a hundred sensors including five lidars, and a foldable steering wheel.

Tensor targeted deliveries for the second half of 2026. The most recent independent write-ups I could find still describe it as an advanced pre-production programme, with the operating zones, weather limits, and regulatory permissions not yet public. Treat every claim in this section as the manufacturer’s until a customer car is on the road.

Tensor is not alone in the ambition. Lucid has said it wants to deliver a consumer-owned Level 4 car using two Thor computers in its midsize line, with the capability arriving a couple of years after launch. Note the difference: two chips against eight.

A word on 8,000 TOPS

I work on navigation for small drones, where the compute budget is measured in watts and every one of them comes out of flight time. From that seat, 8,000 TOPS reads as a statement about redundancy more than intelligence.

TOPS measures how many operations the silicon can perform. It says nothing about whether the model running on it drives well. For comparison, one report puts the Volvo ES90’s computer at 508 TOPS. A sixteen-fold gap in compute is not a sixteen-fold gap in safety.

What a large compute budget does buy is headroom: duplicate stacks that can cross-check each other, and room for bigger models delivered later over the air. That matters for a car someone will own for ten years. It is still a capacity number, and I would not let it stand in for evidence.

The car that earns while you work

This is the pitch that gets attention. You commute, the car drops you off, and then it spends the day on a ride-hailing network earning money. Lyft and Tensor have announced exactly this as a plan, with the first “Lyft-ready” markets targeted for 2027, subject to regulatory approval.

I cannot give you an honest earnings figure, because no personally owned car is doing this yet. What I can give you is the shape of the calculation:

Owner’s net = fares from paid miles − platform commission − energy − cleaning − commercial insurance − extra depreciation − remote support fees

Each term on the right is a job that a fleet operator does today with a depot and staff.

Empty miles. A robotaxi drives unpaid between riders. Those miles still cost energy and wear.
Cleaning and inspection. Somebody has to deal with the spilled coffee and check the car between riders. A fleet does it at a depot. Your car is in a different part of town from you.
Sensor upkeep. A Level 4 system depends on clean, calibrated sensors. Fleets check them on a schedule. Private owners are not known for this.
Remote assistance. Every driverless service I know of keeps humans on call for the moments the car gets stuck. Someone has to staff and pay for that.
Insurance and liability. Personal cover does not extend to carrying strangers for money. When nobody is in the car, the question of who answers for a crash gets sharper.
Operating zone. Level 4 means driverless inside a defined domain. If the profitable part of the city is outside your car’s approved zone, the car cannot work there.

Add these up and the owner is not really a passive investor. The owner is a one-vehicle fleet operator, or else is paying a platform to be the operator, in which case the platform keeps most of the margin.

There is a telling detail in the Lyft announcement. Lyft, through affiliates, reserved hundreds of Robocars for its own fleet operations. The first customer for the personal robocar, in other words, is a fleet.

Tesla has promoted the same owner-earns idea for years. It remains a stated plan rather than something owners can switch on, even as the company runs its own robotaxi service in Texas and Florida.

What happened to Level 3 highway cruising

I originally planned a section here on how eyes-off highway driving in luxury cars is changing the commute. The reporting does not support it.

Afeela was cancelled in March 2026 after Honda’s wider retreat from its EV programme removed the platform it depended on.
Mercedes paused Drive Pilot with the S-Class facelift. Reports cited modest demand and high cost. The system had been limited to about 40 mph on approved roads in Nevada and parts of California.
BMW is reported to have dropped its own Level 3 option with the 7 Series facelift.
MB.Drive Assist Pro, which replaces Drive Pilot, is a Level 2 system. It handles city streets as well as highways, and the driver must supervise at all times.

GM still says an eyes-off Super Cruise will arrive in 2028. For now, what is changing the daily commute is hands-off, eyes-on driving. That is a real improvement in comfort, and it is not autonomy.

Why Level 3 was the awkward rung

Level 3 asks for something humans are bad at: stop paying attention, then resume full control within seconds when asked. It also moves liability to the manufacturer while the system is engaged, which is why the approved conditions were so narrow. A feature that costs thousands, needs extra sensors, and works below 40 mph in good weather is hard to sell.

Level 2++ avoids all of that. It does far more, across far more roads, and the driver stays responsible. From an automaker’s point of view it is the better product and the smaller legal risk.

That leaves a gap in the ladder. The industry now appears to be going from supervised driving straight to driverless, without a stable rung in between. Personal Level 4 is the attempt to make that jump in a consumer car.

The case for the personal robocar

The argument deserves a fair hearing, because parts of it are strong.

Most cars are privately owned, and most places are not dense cities. Robotaxi fleets need ride density to pay for depots, so they concentrate where demand is thick. A personal Level 4 car serves the suburb and the small town that a fleet will reach last, if ever.

It also solves the fleet operator’s biggest problem, which is capital. If owners buy the vehicles, the network grows without anyone financing tens of thousands of cars.

And there is the plain preference argument. People like their own car, with their own things in it, available immediately, with their data kept on board rather than in a fleet operator’s cloud.

The counter is the one I keep returning to. Level 4 is less a feature of a vehicle than a property of a vehicle plus an operation: maintenance, monitoring, remote help, incident response. The UN’s new automated driving regulation reflects this by requiring manufacturers to run a safety management system and monitor vehicles in service. A fleet has that operation by definition. A manufacturer selling to individuals has to build it around cars it does not control.

On the word “agentic”

Tensor calls itself an agentic AI company, and the label fits the pitch: a car that acts on your behalf, parks itself, charges itself, and perhaps earns for you.

My test for an agent is what it can finish without me. By that test a fleet robotaxi is already agentic, inside its zone, because an organisation stands behind it. A personal car becomes agentic only when the same support exists for a fleet of one.

What I am watching

The first customer delivery of a personal Level 4 car, and the published operating domain that comes with it.
Insurance. The first insurer to price an owner-operated driverless car will tell us more about the business model than any keynote.
Whether Lyft’s “ready” markets open in 2027, and whether the cars on them belong to individuals or to fleets.
Eyes-off returning. If GM ships in 2028 at highway speed, Level 3 was early rather than wrong.

My view is that personal Level 4 will arrive, and that it will look less like owning a robotaxi and more like subscribing to a fleet service that happens to park in your driveway.