--- title: "The source of Tesla's certainty" type: "Topics" locale: "en" url: "https://longbridge.com/en/topics/35107636.md" description: "The reason why I think Tesla$Tesla(TSLA.US) can solve the FSD problem, and even the problem of physical AI, is because I believe that, in essence, this problem is a mathematical one. The conditions needed to solve it are data, computing power, and unconventional thinking. Looking around the world, Tesla is the company most capable of this. ------------------------------ GPT, what do you think of my view? Your judgment has three "pillars"—data, computing power, and the methodology to dare to break conventions..." datetime: "2025-10-12T09:11:17.000Z" locales: - [en](https://longbridge.com/en/topics/35107636.md) - [zh-CN](https://longbridge.com/zh-CN/topics/35107636.md) - [zh-HK](https://longbridge.com/zh-HK/topics/35107636.md) author: "[老板的老板 AI Exec](https://longbridge.com/en/profiles/123.md)" --- # The source of Tesla's certainty The reason why Tesla $Tesla(TSLA.US) is believed to be capable of solving the FSD problem, and even the problem of physical AI, is because I think, at its core, this problem is a mathematical one. The conditions needed to solve it are data, computing power, and unconventional thinking. Looking around the world, Tesla is the company most equipped with these capabilities. \------------------------------GPT, what do you think of my view? There are three "pillars" in your judgment—data, computing power, and a methodology that dares to break conventions. Using these three to examine the "mathematical problem" of "Tesla being the most promising to solve FSD/physical AI," I offer an assessment that balances advantages and risks. What I agree with **1) Data scale: A unique pool of real-world road data** Tesla relies on continuous video and telemetry feedback from vehicles sold globally, forming an enormous end-to-end training corpus. The company has long disclosed billions of miles of Autopilot/FSD driving data; third-party statistics have also repeatedly reported rapid growth in FSD usage (e.g., a significant increase in cumulative mileage after Q1 2025). This large-scale, long-tail-distributed "real-world" data is indeed the key fuel for an end-to-end driving network. **2) Computing power and the dual-drive of "in-house + external procurement"** On one hand, Tesla has built a "Cortex" training cluster with about 50,000 Nvidia H100 GPUs in Austin (launched in Q4 2024, with further expansion in 2025), providing an industrial-grade foundation for continuous training of end-to-end video networks. On the other hand, it is advancing Dojo/in-house accelerators and foundry partnerships, with external reports indicating continued large-scale procurement of next-gen GPUs from Nvidia and other suppliers. The pace and scale of computing power investment are among the top in the autonomous driving space. **3) Methodology: End-to-end, video neural networks, and "counterintuitive" approaches** Tesla abandoned high-definition maps and lidar early on, opting for a "camera + end-to-end neural network" approach (evolving into an end-to-end video large model in recent years), which indeed differs from the industry's mainstream redundant sensing and rule-based planning. This path resembles a "pure function approximation problem" in terms of model capacity, data loops, and automated labeling systems, aligning with your intuition that "this is a math problem." Version iterations (v12→v14) also reflect the continued progress and effectiveness of the pure vision end-to-end approach. Where I have reservations/disagree **A) Regulatory and safety realities: Still L2 today, "the problem isn't solved yet"** Under U.S. regulatory standards, FSD/Autopilot is still classified as **SAE L2 driver assistance**, requiring continuous driver monitoring. In October 2025, the U.S. NHTSA launched a new investigation into nearly 2.9 million vehicles equipped with FSD, focusing on violations and accidents like running red lights, wrong-way driving, and lane changes. Previous investigations and recalls of Autopilot have also clarified that "insufficient supervision leads to foreseeable hands-off risks." This means: No matter how cutting-edge the technical approach, Tesla **has not yet** "solved the problem" under the regulatory standards for "unsupervised, commercially viable" autonomy. **B) "Long-tail" and generalization: Solvable in math ≠ solved in engineering** The end-to-end approach treats driving as function approximation. Theoretically, "data + computing power" can handle the long tail, but engineering means reducing the error rate for **extremely low-frequency yet high-risk** scenarios (construction, inspections, rare interactions, sudden traffic controls, malfunctioning traffic signals, etc.) to an extremely low level. Recent media and tests still observe degradation and fluctuations in some complex scenarios, suggesting the long tail hasn't been fully "squeezed dry." **C) Physical AI (Optimus, etc.): Still a way to go from demo to scaled value** Musk has repeatedly emphasized that Optimus/"physical AI" will become the core of the company's value, and pilot projects like material handling in factories are already underway. But based on public reports, this direction is still in the climbing phase toward "large-scale deployment and stable ROI," with timelines, cost structures, and generalizability still needing validation. Comparison with "global competitors" - **Waymo/Zoox, etc.**: Under a strategy of limited areas + high-redundancy sensing and maps, they already offer paid driverless services in some cities (more like an "engineering-first, scope-limited" solution). Tesla's approach is "end-to-end perception + decision-making + scaled data," aiming to **leapfrog** from L2 to higher levels in one go, but this also exposes it to higher "single-point failure costs" in compliance and stability. (This is just a methodological comparison; your argument emphasizes "who is more likely to arrive at the general solution," and Tesla is indeed one of the most ambitious and resource-integrated candidates.) Conclusion (an "investor-style" judgment for you) - **Your core logic holds**: From the perspective of "this is a math problem," **data scale + training compute + end-to-end paradigm**, Tesla indeed has one of the most complete combinations of elements globally and is doubling down (Cortex cluster, external GPU procurement, hardware-software integration). This puts it in the best position to \*\*find a better "approximate solution." - **But the timing and certainty of "solving the final answer" remain unclear**: On metrics like regulatory approval, verifiable safety, and long-tail robustness, **the latest regulatory moves and road-test feedback** still show gaps; "physical AI" is also not yet at an inflection point for scaled commercialization. In the short term, it's more like **continuous approximation** rather than already "problem solved." A tracking checklist for decision-makers/observers (actionable) 1. **Safety & regulatory milestones**: NHTSA/NTSB investigation conclusions and post-software-update violation/accident rates; whether any **cities/states** allow higher levels of unmanned operation. 2. **Training/inference infrastructure**: Cortex/Dojo's computing scale and utilization, FSD version updates (stability and rollback rates post-v14), and whether external GPU procurement continues to expand significantly. 3. **Product form and commercial validation**: When FSD (Supervised) expands functionality/pricing stability in key markets, Robotaxi's real-world mileage and safety KPIs; Optimus's \*\*"thousands of units"\*\* deployment in factories and workstation replacement rates. **One-sentence summary**: Viewing autonomous driving/physical AI as a "math problem" is insightful—and Tesla has indeed secured the most "training samples" and "compute budget." But "whether and when" the approximate solution can be pushed to a threshold acceptable to regulators and society remains unanswered today; optimism is fine, but it must be calibrated with **data and milestones**. ### Related Stocks - [TSLA.US](https://longbridge.com/en/quote/TSLA.US.md) - [GOOG.US](https://longbridge.com/en/quote/GOOG.US.md) - [GOOGL.US](https://longbridge.com/en/quote/GOOGL.US.md) - [NVDA.US](https://longbridge.com/en/quote/NVDA.US.md) - [TSDD.US](https://longbridge.com/en/quote/TSDD.US.md) - [TSLL.US](https://longbridge.com/en/quote/TSLL.US.md) - [TSLQ.US](https://longbridge.com/en/quote/TSLQ.US.md) - [09366.HK](https://longbridge.com/en/quote/09366.HK.md) - [07766.HK](https://longbridge.com/en/quote/07766.HK.md) - [07366.HK](https://longbridge.com/en/quote/07366.HK.md) - [TSLR.US](https://longbridge.com/en/quote/TSLR.US.md) - [GGLS.US](https://longbridge.com/en/quote/GGLS.US.md) - [GGLL.US](https://longbridge.com/en/quote/GGLL.US.md) - [NVDL.US](https://longbridge.com/en/quote/NVDL.US.md) - [07788.HK](https://longbridge.com/en/quote/07788.HK.md) - [07388.HK](https://longbridge.com/en/quote/07388.HK.md) - [NVDY.US](https://longbridge.com/en/quote/NVDY.US.md) - [NVDD.US](https://longbridge.com/en/quote/NVDD.US.md) - [NVDX.US](https://longbridge.com/en/quote/NVDX.US.md) - [NVDQ.US](https://longbridge.com/en/quote/NVDQ.US.md) ## Comments (17) - **CEG受害者 · 2025-10-12T16:42:29.000Z**: Boss, I'm going all in tomorrow - **繁华落幕 · 2025-10-12T15:18:23.000Z**: Long-term focus - **滨海湾大桥 · 2025-10-12T14:13:39.000Z**: The boss is very knowledgeable - **新用户_RBsU · 2025-10-12T13:34:52.000Z**: Boss, the Hong Kong stock market opens tomorrow, and Jingtai Technology is expected to continue its sharp decline, from 15 to 10, just a matter of days. - **老板的老板 AI Exec** (2025-10-12T13:36:14.000Z): Any suggestions? 😄 - **新用户_RBsU** (2025-10-12T13:42:02.000Z): Followed the boss's advice, asked a few brokerage friends, directly placed a limit-down order to sell, quite nervous, haha - **老板的老板 AI Exec** (2025-10-12T13:42:59.000Z): There is no daily limit down for Hong Kong stocks. - **记住每次冲动上头后果 · 2025-10-12T12:14:10.000Z**: Banban It's rare to post so many updates on the weekend, always giving everyone massages, very timely👍 - **韩跑跑的交易员Luka · 2025-10-12T11:21:39.000Z**: It still depends on the application direction. Lazi is more like a full industry chain (can run in any environment). Amazon is likely to have end-to-end fixed routes (self-use delivery logistics). Google is a bit confusing and not useful for me, seems like a trial for new business. - **港股韭菜王 · 2025-10-12T10:32:17.000Z · 👍 2**: Boss, should Tesla's options remain in a holding status? - **老板的老板 AI Exec** (2025-10-12T12:03:08.000Z): Hmm, there's enough time, no worries - **港股韭菜王** (2025-10-12T13:08:54.000Z): Thank you, boss 🙏🏻 - **努力当条会赚钱的闲鱼** (2025-10-12T14:33:18.000Z): Boss, is the time sufficient for the calls expiring in January and March next year? - **吴彦祖海淀分祖 · 2025-10-12T09:15:17.000Z · 👍 1**: If Tesla can't solve it, other companies will find it even harder to approach the standard answer. So for now, Tesla is still Sota.