Tesla versprach, dass 4 Millionen Autos selbst fahren könnten. Sie können es nicht.

Key Takeaways

• HW3 is officially dead for unsupervised FSD: Musk confirmed on Tesla’s Q1 2026 earnings call that Hardware 3 has only 1/8th the memory bandwidth of Hardware 4 and “simply does not have the capability to achieve Unsupervised FSD.”
• 4 million cars need surgery: The fix requires replacing both the computer and the cameras. Tesla plans to build dedicated “micro-factories” in major cities because service centers are too slow.
• The legal bill is mounting: Tesla faces $2.7 billion to $14.5 billion in total lawsuit exposure across 21 legal tracks. A California court has already ordered $10,600 refunds per FSD owner.
• The treadmill never stops: In the same earnings call, Tesla announced HW4 Plus, doubling the RAM. The same obsolescence cycle that killed HW3 is starting for HW4.

The Seven-Year Promise Just Died on an Earnings Call

On April 22, 2026, during a quarterly earnings call, Elon Musk said the sentence that vaporized seven years of promises to roughly 4 million Tesla owners:

“Hardware 3 simply does not have the capability to achieve Unsupervised FSD.”

That sentence was not a surprise to anyone who has been paying attention. But hearing it from Musk himself, on a recorded earnings call with Wall Street analysts on the line, made it irreversible. The promise is dead. The legal clock is running.

Since 2019, Tesla sold millions of vehicles with Hardware 3 (HW3) alongside an explicit claim: these cars had “all the hardware needed for full self-driving capability.” Buyers paid between $5,000 and $15,000 for the Full Self-Driving (FSD) package based on that claim. Some waited seven years.

Now Musk is telling them: the hardware was never going to be enough. HW3 has only one-eighth the memory bandwidth of HW4, “and memory bandwidth is one of the key elements needed for Unsupervised FSD.”

One-eighth. Not a marginal shortfall. A chasm.

The Physics of the Broken Promise

To understand why this cannot be fixed with software, you need to understand what memory bandwidth actually does in a self-driving system.

A self-driving car runs neural networks that process camera feeds in real time. Each frame from each camera generates a tensor, a multidimensional array of numbers that the neural network multiplies, transforms, and feeds forward to produce a driving decision. The speed at which the processor can shuttle those tensors between memory and compute units is the memory bandwidth. It is the fundamental throughput bottleneck of inference.

HW3’s Full Self-Driving Computer (FSD Computer) was designed in 2018 and began shipping in April 2019. It was a custom chip built on Samsung’s 14nm process with two redundant System-on-Chips (SoCs), delivering 144 TOPS (Tera Operations Per Second) of neural network performance. At the time, it was genuinely impressive.

But neural networks got bigger. Much bigger. Tesla’s FSD models in 2026 are orders of magnitude more complex than the 2019 architecture anticipated. The compute can be squeezed and optimized, but the memory bandwidth (the speed of the pipe feeding data to the processor) is a physical property of the silicon. You cannot overclock it. You cannot optimize around it. It is baked into the chip.

HW4, which began shipping in late 2023, uses a newer process node and fundamentally wider memory buses. The result: 8× the memory bandwidth. That is not a tuning difference. It is the difference between a garden hose and a fire hydrant.

No software update bridges that gap. The chip has to come out.

The Surgery: Computer, Cameras, and Factories

Replacing the computer alone would be complex enough. But Musk added a detail on the earnings call that makes the retrofit far more invasive:

“You also need to replace the cameras, unfortunately, to go to Hardware 4.”

HW4 uses a different camera system with higher resolution sensors and different mounting positions. The computer and camera suite are designed as an integrated system. Swapping one without the other does not work.

This means partial vehicle disassembly: removing interior panels to access the FSD computer behind the glovebox, disconnecting and replacing the camera harnesses routed through the car’s body, and recalibrating the entire vision system afterward. It is not a 30-minute service appointment. It is a multi-hour surgical procedure per vehicle.

And there are roughly 4 million HW3 vehicles on the road.

Tesla’s service center network cannot absorb this volume. Musk acknowledged as much:

“We’re going to have to set up micro-factories… in major metropolitan areas… because if it’s done at the service center it is extremely slow to do so and inefficient.”

Building factories. Not to make new cars. To fix old ones. Tesla has provided no timeline for when these micro-factories will be operational, no cost estimate for the retrofit program, and no commitment on how many HW3 vehicles can be processed per day.

The Bill: $2.7 Billion to $14.5 Billion

The legal system is not waiting for Tesla to build those factories.

Tesla currently faces 21 distinct legal tracks across seven categories, with estimated exposure ranging from $2.7 billion on the conservative end to $14.5 billion on the high end. The FSD-related claims alone account for a significant chunk of that range:

Category	Low Estimate	High Estimate
Autopilot/FSD crash litigation	$1.0B	$5.0B
FSD false advertising	$100M	$500M
Robotaxi securities fraud	$1.0B	$5.0B

In August 2025, a California federal judge certified a class action covering all FSD purchasers since October 2016 who opted out of arbitration. The ruling accepted a “full-refund theory,” meaning class members could recover the $5,000 to $15,000+ they paid for the software.

A separate California court has already ordered Tesla to refund $10,600 per owner in a related ruling.

In December 2025, the California DMV issued a ruling calling Tesla’s “Full Self-Driving” branding “actually, unambiguously false.” Tesla’s response was to sue the DMV rather than change the name.

The most dramatic precedent came in August 2025 with Benavides v. Tesla: a Miami jury awarded $243 million, including $200 million in punitive damages, for a fatal 2019 Autopilot crash, finding Tesla 33% liable. Tesla had rejected a $60 million settlement offer before trial.

And the pressure is going international. A Dutch owner launched hw3claim.nl, a platform to bundle HW3 owners across the European Union into a collective claim against Tesla. Within one week, more than 4,000 verified owners from 29 countries had signed up, seeking €6,800 per claim.

The Treadmill: HW4 Plus and the Pattern That Never Stops

Here is the detail from the same earnings call that should alarm every current Tesla owner, not just those with HW3:

Tesla announced HW4 Plus (also called AI4.5). It doubles the RAM per chip from 16GB to 32GB (64GB total system memory), with a roughly 10% increase in compute power and memory bandwidth. Samsung is “currently preparing the modifications” to begin production.

This means Tesla is already developing the successor to the hardware it currently claims is “sufficient for unsupervised FSD.”

Musk even hinted at the inevitable endpoint: “At some point the AI4 hardware is going to get like so old that it’s like, okay, the only reason to keep the factory open is for AI4.”

The pattern is now visible:

Year	Hardware	Promise	Outcome
2019	HW3	”All hardware needed for FSD”	Cannot do unsupervised FSD (confirmed April 2026)
2023	HW4	Claimed sufficient for unsupervised FSD	HW4 Plus announced April 2026
2027	HW4 Plus	TBD	?

As one analysis put it: “Tesla has burned through all of its credibility on hardware sufficiency claims.”

The structural problem is not dishonesty. It is that neural network scaling consistently outpaces the hardware roadmap. Every generation of FSD software demands more memory bandwidth, more compute, and higher-resolution camera inputs than the previous generation. The promise (“this hardware is enough”) is a bet against the pace of AI research. So far, the pace of AI research has won every time.

The Subscription Pivot: Solving Future Liability by Eliminating Ownership

Tesla ended one-time FSD purchases on February 14, 2026. Every new vehicle delivered after that date can only access FSD through a $99 per month subscription. Existing one-time purchasers were grandfathered in.

This is not just a business model shift. It is a legal architecture shift.

Under the old model, a buyer paid up to $15,000 for a capability Tesla promised the hardware could deliver. When the hardware proved insufficient, Tesla faced claims for a retrofit, a refund, or both. That is the FSD equity problem that has haunted Tesla for years.

Under the subscription model, a subscriber pays $99 per month for FSD as it exists at the time of use. If the car’s hardware cannot run next year’s neural network, Tesla simply stops supporting it. No broken promise. No refund. The liability evaporates.

The subscription model solves Tesla’s future liability exposure. It does nothing for the roughly 770,000 people who already paid between $5,000 and $15,000 under the old model.

The Steel Man: Why Tesla Deserves Some Credit

This article would be incomplete without acknowledging the counterargument.

Tesla is the only automaker on Earth that has shipped a consumer autonomous driving system operating at anything close to this scale. FSD (Supervised) now has 1.28 million active users as of Q1 2026, a 51% year-over-year increase. No competitor (not Waymo, not Cruise, not Mobileye) has anything approaching that installed base outside of geofenced robotaxi zones.

The hardware treadmill is a genuine engineering problem, not a scam. Neural networks in 2026 are fundamentally different from what anyone predicted in 2019. Apple does not upgrade your iPhone 6 to run the latest computational photography pipeline. Nvidia does not patch your GTX 1080 to run current-gen ray tracing. Hardware generations become obsolete. That is the nature of computing.

And Tesla IS committing resources to the problem. Free upgrades (computer plus cameras) for FSD purchasers, micro-factories to handle the volume, and V14 Lite shipping to HW3 owners by late June 2026 as a stopgap. The company beat earnings expectations in Q1 2026 with $22.4 billion in revenue and $0.41 EPS (adjusted) versus $0.37 expected. Automotive margins climbed to 19.2% excluding regulatory credits. Tesla has the financial capacity to execute a multi-billion-dollar retrofit program.

The question is not whether Tesla can build the factories. The question is whether the hardware promise model (“pay now for capability that arrives later”) is fundamentally broken.

The Precedent That Changes Everything

The answer to that question might come from a courtroom, not an earnings call.

If the California class action succeeds under the full-refund theory, it establishes a precedent: automakers cannot sell hardware capability promises that their chips cannot deliver. The refund exposure for roughly 770,000 one-time FSD purchasers at $5,000 to $15,000 each ranges from $3.9 billion to $11.6 billion before legal fees.

That is a number large enough to change how the entire autonomous driving industry prices its products.

For comparison, General Motors paid roughly $1.2 billion for its ignition switch recall covering 30 million vehicles, but the fix was a $0.90 part. Tesla’s fix requires new silicon, new cameras, and purpose-built factories. The per-car cost will be orders of magnitude higher.

The irony is structural: Tesla’s HW3 article from December 2025 asked whether this moment was coming. Now it has arrived. The speculation is over. The hardware is dead. The lawsuits are live. And in the same breath Musk used to kill HW3, he announced the chip that may follow the same path for HW4.

The treadmill does not stop when you get on the next version. It stops when someone admits that selling hardware promises for software that does not yet exist is a bet that has not yet paid off for the buyer.