Tesla's robotaxi finally works. Just not on the cars it sold you

Tesla has confirmed a Cybercab launch event, with a contest running through August 23 offering robotaxi riders a chance to attend, and reporting suggests the purpose-built vehicle could enter service before the month is out. After a decade of promises, a Tesla that drives itself with nobody in the front seat is about to carry paying passengers.

It is worth being precise about which Tesla that is. It is not the one in your driveway.

The Cybercab has no steering wheel, was never sold to a private buyer, and according to reporting on its production units runs an upgraded Autopilot board with roughly double the memory of the AI4 computer in customer vehicles. Tesla has not published a detailed specification, and the company should be given room to correct the record. But the direction is consistent with everything else it has said this year.

TSLA closed at $351.12 on August 19, up 4.2 percent on the day, though still roughly 30 percent below its 52-week high of $498.83 and trading under both its 50-day and 200-day averages.

The ladder nobody wants to be standing on

Start with what is now settled. Tesla has confirmed that HW3 vehicles cannot achieve unsupervised Full Self-Driving, and has announced an upgrade path and trade-in discount for the owners affected. Those cars were sold between roughly 2019 and 2023, many with Full Self-Driving purchased outright on the understanding that the capability would arrive over the air.

That is not a small population, and the remedy is not a small undertaking. Tesla has been reported as planning dedicated factory capacity purely to retrofit HW3 cars it had previously said would drive themselves. Building factories to fix cars you already sold is an unusual line item, and it exists because the alternative is worse.

Now look one rung up. HW4, also called AI4, was the hardware that was supposed to be sufficient. On the Q1 2026 call, asked about upgrading HW3 cars, Elon Musk indicated Tesla would rather skip HW4 as the upgrade target and move to the next generation board, saying it would make sense to upgrade essentially any car with less than Hardware 4, and that the destination should be the next-generation AI board rather than AI4.

Read that from the position of an HW4 owner and it is not reassuring. The company is describing AI4 as a way station.

The physics reinforce it. Reporting on AI5 indicates it will draw something like two to three times the energy of AI4, requiring a new electrical and thermal package that AI4 vehicles are not built to accommodate without significant work. A software update can ship overnight. A power and cooling redesign cannot.

The software is what forces the hardware

The mechanism connecting these two things is the model itself.

FSD v15 reportedly carries something like ten times the parameters of current builds. It is being tested on robotaxi vehicles first, with customer availability indicated for late 2026 into 2027, and Musk has previously said robotaxi operations would not expand significantly until v15 arrived. That sequencing is the whole story in miniature: the version that makes unsupervised driving work is being validated on company hardware before it reaches the cars people bought.

Parameter counts are an imperfect proxy for compute demand. Quantisation, sparsity and distillation all change the arithmetic, and a model ten times larger does not automatically need ten times the silicon. But an increase of that magnitude is not absorbed by clever engineering alone. It shows up as memory footprint, as sustained power draw under continuous inference, and as heat that has to go somewhere.

Which is precisely the list of things reported to separate the Cybercab's board from the one in customer cars, and precisely the list that AI4 vehicles are said to struggle with for AI5. The hardware floor is being set by the software, and the software is arriving after the cars were sold.

Why this breaks a specific promise

The robotaxi pitch was never only that Tesla would operate an autonomous fleet. It was that your Tesla would join it.

That was the shape of the argument for years: buy the car, buy FSD, and when the software matures the vehicle earns money while you sleep. It justified the FSD price, it supported residual values, and it turned every sale into a claim on a future network. We looked at the economics of that model when we examined robotaxi fleet density, and the whole structure depends on the installed base being the fleet.

A purpose-built Cybercab running hardware no customer owns is a different business. It is a capital-intensive, vertically integrated ride-hailing operation where Tesla buys the vehicles, Tesla runs the vehicles, and Tesla keeps the revenue. That may well be a better business. It is not the business owners were told they were buying into, and the difference is not a detail.

The uncomfortable version of the question: if the autonomy that works requires hardware Tesla only installed in cars it kept, what exactly did FSD buyers purchase?

What it does to the competitive picture

Waymo reached the purpose-built conclusion years ago. It designs its own vehicles, operates them itself, and sells rides rather than cars. Nobody criticises it for that, because Waymo never told a customer their own vehicle would join the fleet.

Tesla's structural advantage was supposed to be exactly that difference. Millions of vehicles already on the road, already gathering data, and available to be switched on as a fleet the moment the software was ready. No competitor could assemble that footprint at any price. It was the strongest argument in the entire bull case.

If activation requires hardware that exists only in vehicles Tesla owns, the advantage narrows considerably. The installed base remains extraordinary for data collection, and data still matters. But deployment is where the revenue is, and on deployment Tesla is now closer to a conventional competitor: buying vehicles, operating them, and scaling city by city against rivals who have been doing precisely that for years.

Tesla retains genuine strengths here, including manufacturing cost, vertical integration and a charging network none of its rivals can match. Those are real. They are also not the argument that carried the valuation.

The financial edges

Three places this shows up in the numbers.

Deferred revenue. Tesla recognises FSD revenue as capability is delivered. A generation of cars formally excluded from unsupervised operation complicates what "delivered" means for those buyers, and a second generation in doubt complicates it further.

Retrofit cost. Factories built to upgrade already-sold vehicles are capex and cost of goods spent against revenue recognised years ago. That lands on margins already under pressure from the AI build-out, a squeeze we covered through the Q2 capex picture and the analyst target cuts that followed.

Residuals. Used Tesla values have carried an implicit option on future autonomy. If the market concludes that option only attaches to hardware generations still ahead, that premium compresses across millions of vehicles, which affects trade-in economics and therefore new car demand, including for the cheaper model meant to expand the base.

None of these is fatal. All of them are real, and none of them appears in a launch event.

The case that Tesla is doing the right thing

The fair reading deserves equal space, because there is a strong version of it.

Autonomy turned out to be harder than anyone at Tesla believed in 2019, and the compute required grew accordingly. A company that ships unsupervised autonomy on hardware that cannot safely support it is a company that kills people. Conceding that HW3 cannot do the job, and offering an upgrade path rather than pretending otherwise, is the responsible outcome even though it is expensive and embarrassing.

Building a purpose-designed vehicle is also straightforwardly correct engineering. A car with no steering wheel, designed around sensors and compute from the start, will be safer and cheaper to operate than a retrofitted consumer vehicle carrying a decade of design compromises.

Tesla's problem is not that it built a better robotaxi. It is that it funded the effort partly by selling the promise to people whose cars will not receive it.

What to watch

Whether Tesla states an AI4 position. The HW3 answer is now explicit. The AI4 answer is being communicated through hints and hardware roadmaps. Owners of the current fleet deserve a plain sentence, and the absence of one is itself informative.

Cybercab unit economics at launch. Purpose-built means Tesla carries the capital. Watch cost per vehicle and utilisation assumptions, because a company-owned fleet has to clear a much higher bar than one assembled from cars customers already paid for.

FSD v15 on customer hardware. If the version that finally delivers unsupervised operation lands on AI4 cars in a materially reduced form, that is the answer to the AI4 question, delivered by release note rather than by statement.

The Cybercab is going to be a genuinely impressive product, and the coverage this month will reflect that. The story underneath it is that Tesla has quietly stopped promising the fleet will be yours.