Selling the wealthy's chauffeur-and-housekeeper lifestyle to ordinary people
Li Xiang judges that in ten years, the same people will be buying L4 autonomous vehicles and home robots—AI's value isn't replacing jobs, but turning chauffeurs and housekeepers only the wealthy can afford into services ordinary families can consume.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Top AI users aren't your company's former star performers
Li Xiang observed the top 20 employees by token consumption within the company and discovered they weren't conventionally top-tier performers—instead, they were people with weak verbal skills, little access to resources, but exceptional thinking ability. Previously, people like this suffered because they weren't fluent with words; now, as long as they have tokens and a business environment, they can remake everything. Based on this, he recommends all companies avoid layoffs, because the talent standard in the AI era is completely different from the past, and hasty cuts risk eliminating precisely the people who are 'best' by yesterday's metrics.
— Li XiangStrategy teams can build verified demos in a week without engineers
Li Xiang gave his strategy team three problems the company faced—workflow coordination, data access, value measurement—to research how world-leading companies handle them. While studying Anthropic, the strategy team discovered it doesn't write traditional PRD documents but instead builds verified demos directly, so they replicated an entire Cowork system without involving any engineers, then separately verified workflow integration and data connectivity with HR and data warehouse teams. This refutes the assumption that 'strategic roles no longer need business intelligence'—expert teams using AI well reach new heights rather than being replaced.
— Li XiangHumanoid robotics' real market is parts loading, not replacing factory jobs
Li Xiang argues most humanoid robotics startups can't find a business model because they're trying to automate jobs—sorting, bolt-tightening—that assembly lines and cameras have already automated away. The labor that still exists is parts loading: Ideal's factories employ over 10,000 people, with more than 3,000 dedicated to placing components on AGVs; the only operational staff left at automated coffee shops is also loading. He judges this is the critical first battlefield for humanoid robots at scale, a challenge many founders can't meet—possibly due to lack of real factory experience, or perhaps because parts loading 'doesn't look cool enough' for fundraising.
— Li XiangAutonomous driving is the first half of embodied AI
Li Xiang divides autonomous driving into three stages: 2017–2022 was assisted driving (2D vision + rule-based algorithms, tens of Tops processing power); now through ~2027 comes imitation learning for L2/L3 (Transformer vision + end-to-end control, 4B–7B models with 2,000 Tops of compute); after 2027–2028 comes true L4, requiring a stable physics-world foundation model with inference compute approaching ten thousand Tops—several times more than Ideal's current single-chip 1,280 Tops. He believes autonomous driving and humanoid robots will each become $5 trillion markets once mature, which is why robotics companies recruit talent from autonomous-driving teams.
— Li XiangDynamic data-flow architecture is the superior path for automotive inference
Ideal's self-designed Mach chip uses a dynamic data-flow architecture, different from the industry-standard ASIC approach, delivering 1,280 Tops per chip and 2,560 Tops combined on dual-chip Levis—Li Xiang calls this the strongest compute for any automaker's in-house design globally. Before committing, the team spent a year preparing and wrote a 1.4-million-character feasibility report, even consulting Jim Keller and other top chip architects; internally there were two competing approaches, and after settling on the data-flow architecture, the team backing the conventional path left the company. His view: batteries and chips were the only two moats in smart EVs over the past decade, but the real barrier in embodied intelligence is inference silicon, not models.
— Li XiangCutting brake response to 13 milliseconds saves an entire vehicle length
Levis uses the world's first production-grade fully electronic mechanical braking, not traditional hydraulic—all four wheels control independently with response time compressed from traditional mechanical braking's 60–70 milliseconds to 13 milliseconds, nearly twice as fast—equivalent to a 15–20% improvement in human reaction speed, translating into a braking distance difference equal to a full-size SUV's length. Li Xiang stresses this isn't redundancy removal but added safety: even if one wheel's brake fails, the car can still brake normally and execute maneuvers using the other wheels and steering system.
— Li XiangSeparating layers prevents components from reimplementing each other's function
Early this year, Ideal reorganized following top AI companies like Anthropic, restructuring into a body model: the Infra team as the heart, foundation models as the brain, software stack as limbs, hardware stack as the body, plus an independent evaluation team. The root cause was that software and agent teams kept wanting to build models, while model teams wanted to build agents—'like low-order creatures trying to grow a small brain on their limbs.' The announcement came Monday; most staff didn't understand immediately, only the foundation-model and software-stack teams grasped it at once. Li Xiang says AI is a mirror that shows true capability—people without the right skills or those dissatisfied with the company leave on their own, no need for layoffs.
— Li XiangCars and robots will be sold to the same people in a decade
Li Xiang says Ideal's vision for the next decade is turning what once only top-tier billionaires could own—a dedicated driver, a housekeeper, a cleaner—into services ordinary families can afford through AI and embodied intelligence, which he sees as the greatest benefit of technological progress, not endless job replacement. He also offers a forecast: in ten years, 90% of people buying L4 autonomous vehicles will be the same people buying home robots, because both ultimately serve the same life need.
— Li XiangIn their own words · checked verbatim
Because I think the biggest change in this wave of AI is that it's both productivity and labor.
因为我觉得这一波AI的最大的变化 其实它既是生产力 又是劳动力
Li Xiang6:01
His expression used to be weak, he couldn't access resources, but his thinking is extraordinary—so as long as he has tokens and a business environment, he can remake everything.
他表达能力过去不强 他获取不了什么资源 但他脑子极强 所以只要有Token 然后有这个什么 然后有业务环境 他就可以去改造一切
Li Xiang9:02
An ordinary person's code has usability so poor that large-scale deployment is basically disastrous for them.
就一个普通人 然后做出来的那个代码 那个可用性 差到极致 更不要说往大规模去部署了 那基本上对他们来就是个灾难
Li Xiang10:02
I think it's the business opportunity, the biggest business opportunity—when everyone's chasing humanoid robots, they're actually failing to find a business model.
我觉得是商业机会啊 这是个最大的商业机会啊 今天大家在搞人形机器的时候 其实是找不到商业模式的
Li Xiang20:08
Autonomous driving is the first half of embodied intelligence; humanoid robots are the second half.
就是自动驾驶是 巨身智能的上半场 人形机器人 是巨身智能的下半场
Li Xiang26:16
How to give hundreds of millions and billions of people access to the lifestyle top billionaires have—that's the greatest benefit technology brings.
如何把这些顶级富豪拥有的生活 然后给到更多的人 让几亿人几十亿人 也能消费这些 这是科技进步带来的最大的好处
Li Xiang40:20
We believe models might be competitive, but not necessarily a moat—the final barrier we see is inference silicon.
我们甚至认为模型可能是竞争 但不一定是壁垒 我们能看到最后的壁垒是推理芯片
Li Xiang1:16:12
It's like those low-order creatures that insist on growing a small brain on their limbs—I'd call that primitive.
特别像那些动物 就是这个手脚上非得长个小脑 我说这是个低级动物的做法
Li Xiang1:47:27
Figures
| Levis single-chip compute | 1280 Tops | 54:23 |
| Levis dual-chip compute | 2560 Tops | 54:23 |
| Full drive-by-wire brake response | 13 milliseconds | 1:00:25 |
| Human brake reaction time | 350–400 milliseconds | 53:23 |
| Livis extended-range endurance | over 1600 km | 57:24 |
| Export model starting price | $32,500 | 2:04:36 |
| Ideal factory parts-loading headcount | 3000+ of 10000+ employees | 20:08 |
| Chip feasibility report length | 1.4 million characters | 56:24 |
| Ideal current scale | 30000 employees, over 100 billion yuan revenue | 17:07 |
| Ideal 5–10 year target | 30000 employees, 1 trillion yuan revenue | 17:07 |
Glossary
- Embodied AI
- AI-driven robots and vehicles with perception and cognition operating autonomously in the physical world.
- VLA/World Model
- Foundation models for understanding and predicting physical-world dynamics in autonomous driving, typically trained via imitation learning.
- Full Drive-by-Wire Chassis
- Electronic signals instead of mechanical linkage control steering and braking, allowing independent adjustment of force and response per wheel.
- Dynamic Data Flow Architecture
- A chip compute architecture distinct from mainstream GPU/ASIC designs, where Ideal claims edge-side data movement efficiency is superior.
How to listen
Founders and executives leading teams through AI transformation, and professionals tracking embodied AI commercialization paths and automakers' in-house chip strategies.
The opening chatter about personal AI learning habits can be skipped; substantive content begins with robotics and autonomous driving technology paths.