Tesla's Launch Felt Like an Annual Meeting: The Robotaxi Story Is Told to Shareholders
This launch event didn't discuss any technology; it felt more like an annual meeting for longtime shareholders. What's truly valuable is its two paths to monetize idle vehicle time: as a Robotaxi, or as distributed compute.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Supervised is a cleverly evasive word
Meng Xing believes Musk replaced the L2/L3/L4 framework with supervised and unsupervised, which is a clever evasion. In the past, L4 meant two things: the car has the capability, and it bears responsibility in an accident. But FSD is 'I have this function, but I don't bear this responsibility'—under Chinese regulations it might still count as an L2 device. So unsupervised is more like: after the takeover rate is low enough, you can randomly pick a period to sleep and with high probability won't encounter a scenario requiring takeover. It's essentially an improvement in L2 assisted driving capability, not a true Robotaxi.
— Meng XingA car used 10 hours out of 168 is not a reason for driverless
Musk gave two reasons: an average passenger car is used only about 10 hours out of 168 hours a week, and after autonomy usage frequency would increase 5 to 10 times; and autonomous driving is 10 times safer than human driving. Meng Xing says both are theoretically reasonable, but the first is not a reason for driverless—it's more a reason for the sharing economy. Cars are underutilized, which is exactly what Didi, Uber, and car-sharing have been solving; it's an old story told for over a decade in the ride-hailing era. The second, surpassing human safety, is a potential possibility, but today all autonomous driving companies are still chasing humans, and Tesla itself is far below actual human driving conditions.
— Meng XingCost dropping from $1 to $0.2 means humans are 80%
Musk said Cybercab operating cost drops to $0.2 per mile, with total cost including taxes and fees at $0.3 to $0.4. Meng Xing did the math: from $1 with a driver to $0.2, a drop of 80%, implies human cost accounts for 80%, which is an overestimate. In the ride-hailing system, human cost is the highest proportion, but generally estimated around 50%. And after removing the human, you still have to add back compute chips, sensors, insurance, operations and maintenance, etc. His more rational estimate is a net reduction of 30% to 40%, from $1 possibly to 60-70 cents.
— Meng XingThe shepherd model ignores that individual capital is most expensive
Musk talks about how in the past a ride-hailing driver could only drive one car, but now you can manage a flock of Cybercabs like a shepherd, being a car partner holding assets to earn money. Meng Xing sees this as financial leverage, so Tesla doesn't have to buy cars itself. But he tends to think it's wishful thinking: the ride-hailing system already has CPs (car partners), asset-holding companies; in China there may be thousands, large ones with thousands of cars, small ones with dozens. The reason the industry hasn't become individuals holding a batch of cars is that the cost of individual capital is not cheaper than a company's; many CPs have special resources or low capital costs, e.g., automakers setting up their own CPs with existing financing lease businesses. Individuals borrowing is definitely the most expensive.
— Meng XingDriverless turns hidden costs into explicit costs
Meng Xing says managing a car takes a lot of thought: where to wait for jobs, parking, fines, car washing, maintenance, repairs, charging and refueling. When a driver brings their own car, these costs aren't counted because they'd maintain, wash, and fuel the car anyway. But without a driver, these become very explicit costs, and they're not easy to solve. For example, charging: the system needs to know it can't take more orders when low on battery, know where charging stations are, drive there—many are in underground garages, which may not have maps, and you have to back in, but most Robotaxis today don't have automatic reversing capability, generally only under remote safety driver monitoring. Then you have to take the charging gun, open the cover, return it after charging, and pay. Waymo's choice is to build large depots with centralized staff, but that brings a series of complex issues like deadhead miles, real estate site selection, depot density, and automation level.
— Meng XingStock fell, Uber rose instead
On the night of the launch, Tesla fell 4% to 5% in after-hours trading, while Uber rose 8%, and about 10% after the open. Meng Xing says this is the result of secondary market investors voting with their money, believing Tesla didn't deliver what it said—at least not expressed at the launch. Conversely, Uber, as the most threatened competitor, may now be safer.
— Meng XingWhen the car isn't driving, there are two ways to monetize
Musk said every Tesla has 1000 watts of inference compute capability, used 50 to 60 hours a week, and the rest of the time it can become distributed inference compute resource, adding up to over 100 watts of compute. Meng Xing says this could mean the world's largest compute capacity, many times larger than the largest supercomputing center; Li Bin has said similar things. He calls it option value—a potential, undeveloped but readily monetizable upside. So after the car is sold, it shows value when driving, and when not driving there are two additional ways to monetize: as a Robotaxi, or as a compute center. Both possibilities are huge opportunities.
— Meng XingThe robot is too smooth, 80% likely teleoperated
The only thing exceeding expectations was Optimus, but Meng Xing says it's a bit too good to be true. He observed: when the robot played rock-paper-scissors, its right hand played, and the left hand had an unnecessary small movement at the end—very human-like. Unless it's truly end-to-end imitation learning that learned these small movements, these actions shouldn't be there. He talked to many robotics people yesterday, and about 80% confidence says it was teleoperated. Teleoperation is generally used in training robots to imitate and collect data, which also explains why the robot's legs didn't move—walking while moving would make teleoperation very difficult.
— Meng XingIn their own words · checked verbatim
What he's doing today with this supervised is a very interesting approach, because it breaks away from our previous context and framework of describing L2, L3, L4.
他今天用这个supervised 是一个很有意思的方式 因为他脱离了以前我们对L2L3L4这种描述的语境和框架
Meng Xing9:08
The first reason is actually not a reason for driverless driving. I think it's more like a reason for the sharing economy.
第一个理由其实不是无人驾驶的理由 我觉得更像是共享经济的理由
Meng Xing15:14
If you count all of these, and it can still drop 80%, that means the human proportion might be 90% or even higher in this cost.
把这些都算上 以后 还能降80% 那就说明人的这个占比可能是90% 会更高在这成本
Meng Xing26:18
For people who haven't done ride-hailing operations, or heavy-asset operations, I think this is a relatively wishful-thinking model.
对于没有做过网约车运营的人 而言或者重资产运营的人而言 我觉得这个是一个相对两样想当然的模式
Meng Xing31:20
So this is a result of secondary market investors voting with their money on their feelings about the launch event.
所以这是一个二级市场投资人 用钱来去投票 对于发布会的感受的一个结果
Meng Xing39:23
So I think this is what we should call an option value—a potential, undeveloped but readily monetizable upside value.
所以我觉得这是一个 我们应该叫一个option value 就是潜在的一个 并没有被开发出来 但随时可以去潜在可以变现的上升价值
Meng Xing42:31
I think with high probability, about 80% confidence, this should be the result of teleoperation.
我觉得大概率认为吧 就80%的信心指数 认为这应该是摇操作做出来的结果
Meng Xing51:39
Figures
| Cybercab estimated price | Under $30,000 | 17:15 |
| Model 2 / Robotaxi BOM cost | Under 100,000 RMB | 20:50 |
| Cybercab operating cost | $0.2 per mile, total cost including taxes and fees $0.3-$0.4 | 25:20 |
| US ride-hailing charge per mile | $2.8-$3.8 | 27:18 |
| Tesla after-hours drop | 4%-5% | 39:23 |
| Uber rise | 8%, about 10% after open | 39:23 |
| Inference compute per Tesla | 1000 watts | 41:30 |
| Robovan seats | Musk said 20 seats, on-site count about 16 | 43:05 |
| Confidence in teleoperation judgment | 80% | 51:39 |
Glossary
- BOM
- Bill of Materials, the cost of a car's component materials.
- CP / car partner
- In the ride-hailing system, an asset-holding company that buys cars and rents them to drivers.
- Remote Assist
- In the L4 framework, a cloud-based safety operator provides auxiliary decision-making to the car in complex scenarios.
- option value
- An undeveloped but readily monetizable potential upside value.
- teleoperation
- A person uses a robotic arm or device to control a robot's movements in real time, often used to collect imitation learning data.
How to listen
Founders and investors focused on autonomous driving and the Robotaxi business model, especially those who want to understand the gap between Tesla's narrative and the operational reality of ride-hailing.
The beginning 0:00-5:00 covers the launch atmosphere and delays; can be fast-forwarded.