The robotics industry is not romantic: the chain is too long, you have to stick your head into the dirt
Xinghaitu founder Gao Jiyang says robotics startup competition is a hexagon: the full machine, supply chain, data, models, channels, terminals. The algorithm propagation cycle is only two to three months — the smallest moat of all.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Waymo's problem is not technology, it is having no founder
Gao Jiyang joined Waymo in early 2019; by the time he left, the perception team had grown from about ten people to seventy or eighty, and the company from 1,000 to nearly 2,000. He judges Waymo to have two problems: first, it is too afflicted by big-company disease — every system is aligned with Google, and it entered big-company status before creating value; second, big-company disease is only the surface, the essence is that Waymo has no Founder — its Founder is Google's Founder, but Google's Founder has no time to directly run this, so the top-down force is missing. He says that starting a company in this kind of industry, being wrong is not what you fear; what you fear is force that is not concentrated and not unified.
— Gao JiyangFour autonomous driving business models, and Huawei is the fourth
Gao Jiyang breaks autonomous driving business models into four types: the first is Waymo's robotaxi, running its own fleet and charging per ride; the second is carmakers selling cars and turning autonomous driving into a software subscription; the third is suppliers like Momenta, providing solutions to carmakers through NRE plus License; the fourth is Huawei, between carmaker and supplier, essentially earning profit at the vehicle level — redefining the car with top-tier autonomous driving experience and cockpit experience, plus its own brand and channels. He believes Waymo's business model itself has no problem, it is just that the cycle is very long, and only today is the dawn of it working visible.
— Gao JiyangLeaving Momenta to start a company, giving up over ten million dollars
Gao Jiyang left Momenta in May 2023. Before that at M he worked on perception, localization and parking systems, infra, planning and control, and his last job was mass-production delivery of the high-speed, high-price NV system to SAIC. He says leaving M to start a company gave up all the completeness; calculated at that point in time, ten million dollars was probably there. Asked whether it hurt, he says not at all, because compared with the thing he wants to do, other things have little value. At his thirtieth birthday at the end of 2022 he figured out he wanted to start a company; the trigger was GPT-3 and InstructGPT making the world believe in AI once again, and mass-production autonomous driving bringing up the sensor and compute supply chain for on-device intelligence.
— Gao JiyangThe first BP was unbearable to look at, angel round 30 million RMB
Gao Jiyang started putting the BP together in August 2023; looking back at that BP he says it was ‘simply unbearable to look at’. What he first wanted to do was last-mile delivery robots, quickly rejected by himself. The first round was 30 million RMB, led by IDG with Baidu Ventures and GSR Ventures following, at a pre- or post-money valuation of 200 million RMB; immediately after, Teacher Wang's CFound fund did a bridge round of ten or twenty million, at a post-money valuation of three or four hundred million. He says the investors at the time said explicitly, ‘the thing you want to do probably won't work, go think it over again’ — that is the angel round: accepting your mistakes and imperfections.
— Gao JiyangDoing the full machine first, because the data closed loop must have a carrier
Gao Jiyang explains why the first thing in starting the company was to tackle the most unfamiliar part, the full machine: the long-term moat is built on a data closed loop in the physical world, and the data closed loop must have a carrier, and that carrier is the full machine hardware; the product to be delivered in the medium and short term is also not an algorithm or a brain, but a physical entity formed by the full machine plus intelligence, with the ability to execute in the physical world. So reasoning backward from both the long term and the medium-short term, the full machine and supply chain must be done well first. The company's theme for the whole of 2024 was the full machine and supply chain, which he calls making up lost lessons.
— Gao JiyangThe cost of an hour of real data is 200 to 250 RMB
Gao Jiyang does the math: to collect an hour of data in the real world, the actual human input is three to four hours, plus robot depreciation (calculated at 100,000 RMB, conservatively scrapped at a 1,000-hour lifespan), the cost per hour is about 200 to 250 RMB. He further says the relationship between data acquisition cost and training cost is roughly 1 to 5 up to 1 to 10 — data obtained for one yuan requires five to ten yuan to train it to clarity, so low data quality means wasting money on training. He also says 100,000 hours of data is the total order of magnitude of one person's interaction with the physical world from birth to eighteen.
— Gao JiyangThe data recipe is today's biggest secret
Gao Jiyang says data comes in many kinds: robot-centric real-machine teleoperation data, human-centric 5-meter and collection-glove data, POV head-mounted camera data, third-person internet video, and graphics-based or world-model-generated simulation data. But nobody knows what the proportional relationship among these data should be; this is called the data recipe, and it is where many large model companies' biggest secret lies today. He states clearly that Xinghaitu is mainly based on real data, but whether it is 10,000 hours of real-machine data, 50,000 hours of 5-meter data or 200,000 hours of POV data has to be found by experiment — AI is ultimately still an experimental science.
— Gao JiyangThe algorithm propagation cycle is only two to three months, the smallest moat
Gao Jiyang uses ‘propagation cycle’ to measure the moat of each element: the full machine and supply chain are 12 to 18 months, customer channel building starts from a cycle and large customers take longer, the data system requires another 6 to 12 months on top of the full machine, while the algorithm propagation cycle for first-tier companies is only two to three months. So algorithm investment is large but the moat is small; the moat against being copied is very small. From this he explains why the company's values are called ‘pragmatic innovation’ — innovation must first be pragmatic, idealism cannot become empty fantasy, and the basis on which idealism can be realized is calculating ROI every day.
— Gao JiyangIn their own words · checked verbatim
The essence is that I think Waymo has no Founder. Right, its Founder is actually Google's Founder, but Google's Founder doesn't have the time to directly manage this, so in this there is a top-down force that I think is missing.
本质是我觉得 Waymo是没有Founder的 对 它的Founder其实是Google的Founder 但是Google的Founder又没时间直接去管这事 所以在这个里面就是自伤而下的力量 我觉得是缺失的
Gao Jiyang41:30
I think the thing I care about most is still the thing I want to do, not some money or whatever. And I think compared with that thing I want to do, all these other things have little value.
我觉得我最care的事 还是我想去做的那件事 而不是一些钱什么东西的 然后我觉得和那件我想做的那个事比起来 其他的这些东西都价值不大
Gao Jiyang1:20:49
If my data quality is low, I am actually wasting a lot of money on the training step. So this is what we mean: from a cost perspective, when you calculate this, you also have to find every way to raise the data quality, that is how you save your training cost.
如果我的数据的质量是低的 我其实把很多钱浪费在训练这一步了 所以这是我们就是说 从一个成本的角度 你去算这个事 你也要想尽办法把数据的质量贴上去 你这样才节约你的训练成本
Gao Jiyang1:58:23
The algorithm propagation cycle is two to three months, so its competitive moat is the smallest. Yes, its investment is large, but the moat is small. Right, in innovation its investment is very large, but in preventing being copied its moat is very small.
算法传播周期是两到三个月 所以它竞争壁垒最少 是的 它投入大 但是壁垒小 对 它在创新上面 投入非常大 但是它在 防止被抄袭 这件事上 的壁垒很小
Gao Jiyang2:35:51
I think this industry does not allow such people to exist. If such people exist, they will probably suffer a great deal.
我觉得这个行业 不允许这样的人存在 如果有这样的人存在 可能他会 会有很大的suffer
Gao Jiyang3:03:23
Figures
| Waymo headcount | about 1,000 at joining, close to 2,000 at leaving | 39:28 |
| Xinghaitu bridge round | ten or twenty million RMB, post-money valuation three or four hundred million | 1:27:57 |
| Real data collection cost | 200 to 250 RMB per hour | 1:59:23 |
| Data acquisition to training cost ratio | 1 to 5 up to 1 to 10 | 1:57:22 |
| Algorithm propagation cycle | two to three months for first-tier companies | 2:35:51 |
| Full machine and supply chain propagation cycle | 12 to 18 months | 2:34:51 |
| Xinghaitu team size | from about ten people to over 200 | 2:47:06 |
Glossary
- VLA / vision-language-action model
- A foundation model that takes vision and language as input and outputs actions, driving the body to execute tasks.
- data recipe
- The mixing ratio of different types of data (real-machine, 5-meter, POV, simulation), which determines model performance.
- sim-to-real gap
- The distribution difference between simulation data rendered by graphics and real-world data.
- domain transfer
- Transferring a model trained in one domain to another; Gao Jiyang believes it is better to use data from that domain directly.
- NRE plus License
- The charging model in which a supplier provides autonomous driving solutions to carmakers.
How to listen
Founders and engineers in robotics, autonomous driving and embodied intelligence, plus early-stage investors trying to judge where the moat in this track actually is.
The growth and education section from 1:49 to 1:23 can be fast-forwarded; the core information is concentrated in the full machine, data and organization parts after 1:35.