A couple of days ago I watched Telescope’s review of the Xiaomi N90, which contained the author’s judgment on how this car will affect the auto market over the next year. I thought one comment was well put:

Professionals should dare to make judgment calls.

When I was still very naive, I made some calls about autonomous driving and the auto industry. Many were spectacularly wrong, fully displaying my foolishness. For example:

• In 2020, I predicted that among NIO, XPeng, and Li Auto, XPeng would do best, followed by NIO, with Li Auto last. My judgment was purely from a technical angle: XPeng invested the most in technology, NIO had great design taste but was too expensive, and Li Auto looked the least technical. In fact, Li Auto was the first of the three to turn a profit, and in the end-to-end wave it regained technical leadership thanks to its VLM. NIO found its rhythm again in 2024–2025 after crisis, its product strength has stayed solid, and its recent driving solutions have improved — living up to those four Nvidia Orin chips. From this mistake, I learned that product strength is the core competitiveness: fancy screens, fridges, and sofas are more attractive than cool-but-at-the-time-impractical autonomous driving. Technology changes rapidly, and both persistent long-term investment and correct technical direction choices matter enormously.

• In 2023, I predicted a certain automaker could not succeed at in-house autonomous driving development, since all of its models had supplier technology involved. I made this call because I believed the automaker’s management style couldn’t support high-tech pre-research, and it lacked the professional talent to support algorithm R&D. As of now, its in-house solution has shipped on lower-tier models, and its in-house chips have landed. With hindsight, this automaker is indeed exceptional — it has a persistent desire to self-develop supply-chain technology and has truly reached world-class level on several key components. Although full-stack in-house development isn’t achieved yet, achieving most of it looks like only a matter of time. It may not reach the very top results, but using in-house development to discipline suppliers is a correct strategic choice. That said, I don’t feel embarrassed about getting this one wrong — after all, Musk mocked that automaker too.

Now, after six years in the autonomous driving industry myself, I want to take stock of my views and make predictions for the next two years. The purpose of predicting is to record these thoughts, hoping verifiable judgments will expose my foolishness and update my understanding of the industry. The one thing this industry never lacks is change: FSD V14 is about to enter China, driving suppliers are lining up for IPOs, and LLM technology evolves daily. Correct judgment is hard; I can only predict based on my own experience.

To See the Future, First Look at History

Tesla:

• August 2023: Musk livestreamed an end-to-end version of self-driving on X.
• March 2024: V12 pushed at scale in North America with a one-month free trial — the first large-scale production deployment of end-to-end.
• October 2024: Officially unveiled the Cybercab.
• December 2024: V13 pushed to HW4, 5x the parameters of V12, with a longer context window.
• June 2025: Robotaxi service officially launched in Austin.
• October 2025: V14 released.
• January 2026: Some robotaxis removed safety drivers.
• April 2026: Robotaxi began small-scale mass production, hundreds of units per week.
• To date, there are several statistical views of Tesla FSD safety. The official site shows one minor collision per 2.5 million km and one serious collision per 9 million km in North America. Teslafsdtracker shows one serious disengagement per 1,300 km for FSD V14 in urban areas, mostly lane-navigation related; collision-related disengagements are rare. A community-tracked Tesla robotaxi safety report, robotaxi-safety-tracker, shows an average of 520,000 km per incident as of April 2026.

XPeng:

• March 2023: Began “city-by-city” expansion, supporting 50+ cities by year end.
• May 2024: Announced XNGP would adopt an end-to-end large-model approach, integrating perception, planning, and control into a unified neural network — a typical two-stage architecture.
• Second half of 2024: Began rolling out the end-to-end solution.
• 2025: Promoted VLA and VLA 2.0, with the head of autonomous driving and the CEO making a public bet that by August 30, 2026, XPeng VLA’s overall performance in China would match Tesla FSD V14.2’s performance in Silicon Valley. We shall see.

Li Auto:

• July 2024: Released the end-to-end + VLM dual-system architecture.
• March 2025: Released the MindVLA architecture.
• June 2026: Released its in-house Mach chip and a brand-new VLA architecture, promising that by the end of 2026 Li Auto’s driving performance would match FSD V14. I could not find more detail on how “matching” is defined.

For the Next Two Years, I Believe:

On Tesla

• Tesla’s self-driving remains far ahead, both technically and in revenue. FSD performance will keep improving; urban serious-disengagement MPI will surpass 5,000 km (per Teslafsdtracker).
• Tesla’s installed base is roughly 3 million HW3 cars and 4 million HW4; FSD subscription on new cars is about 50%. Within two years, HW4’s overall subscription rate will exceed 80%, becoming a feature you can’t go back from. Many people will buy a Tesla for FSD — but autonomous driving still won’t be the primary purchase factor for everyone.
• Thanks to FSD’s absolute lead, Tesla Robotaxi will operate in major US cities, far beyond Waymo’s current scale (3,000 units), exceeding 10,000 units.
• Unless Tesla launches new models, its per-car profit still cannot support a $1.2 trillion valuation — i.e., a P/E over 100.

On China’s Domestic Autonomous Driving Technology

• Top domestic systems will be basically on par with Tesla in daily experience and still better at localization, but urban disengagement mileage will remain far behind — at least an order of magnitude (serious-disengagement MPI within 500 km). The main differences come from chip design, data-loop scale, and the iteration time the technology inherently needs.
• Thanks to technical progress and further compute growth, urban NOA will exceed 80% of all assisted driving (highway NOA + urban NOA; currently 50%).
• Vision-only models still won’t be able to operate without safety drivers (typically XPeng’s vision-only models). This is my least certain point, because XPeng has all the elements (data, talent, compute) relative to Tesla. I believe vision-only can reach L4 — it just takes longer than two years.
• Among assisted-driving suppliers there is no clear technical leader — no company’s MPI will be 10x ahead of the others. No hit car model will have autonomous driving as its main selling point.