Eight Years of Autonomous Driving Venture Capital: Companies That Started With Too High a Valuation Almost All Died
Mengxing ran the numbers: for non-resource-based frontier tech companies valued above $100 million, there don't seem to be any very successful cases yet. Companies that start too high spend their money maintaining the appearance of a good sale rather than developing quietly and scrappily.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Starting with too high a valuation is a kind of fate
Mengxing did a piece of statistics: for non-resource-based frontier tech companies valued above $100 million, there don't seem to be any very successful cases yet. The mechanism is this — an early valuation that gets pulled up means this is money raised on consensus rather than on performance, so the team is under pressure to keep the appearance of a good sale looking better, and the highest-value-for-money and least falsifiable way to do that is to hire very expensive people at high prices. The company then falls into a state of ‘trying every direction once’, fails in many industries, and three or four years later hits a capital downturn and can only shrink the team and start over. He calls this ‘the fate of starting too high’, and says this cycle has already appeared many times in many frontier tech fields.
— Meng XingThe 2016 wave was three conditions colliding
First, Cruise sold to a big company for $1 billion, and everyone saw that autonomous driving had monetisation potential, while there were at least a dozen carmakers worldwide that wanted to do this. Second, deep learning had developed from face recognition and scene recognition all the way along, and what was blocking autonomous driving at the time was precisely the perception system, and the most convenient way to improve perception was to transplant the image-vision capability over, so most of the people who poured in early were computer vision people. Third, after tech entrepreneurship became the mainstream direction, everyone was looking for the field best suited to tech entrepreneurship — autonomous driving didn't need to convince anyone it was useful, as long as you could build it, and the whole chain was shortened to almost nothing but the technology chain.
— Meng XingVC money can't keep investing past $300-500 million
Mengxing breaks the cycle down very clearly: late 2016 to early 2017 was the peak, and 2019 was the first trough. The mechanism of the trough wasn't a technology problem, but that these companies founded in 2016 had all reached a $300 million to $500 million valuation — a single VC cheque is generally $30 million to $50 million, corresponding to 10% to 15%, which is basically the ceiling for a single lead investor. Beyond that is out of VC range, and you have to enter late-stage growth funds or PE, and PE has a set template: look at revenue, look at user numbers, and autonomous driving companies have none of that. So in 2019 many companies were on the brink of collapse, which was also the period when he joined Didi's autonomous driving unit and observed the industry from the front line.
— Meng XingThe 2021 peak was Tesla spillover
In the second half of 2020 Tesla went from over $40 billion to possibly $600 billion, more than ten times in a year, spilling over two concepts: electric vehicles and autonomous driving. Electric vehicles already had many listed companies, so autonomous driving became another target, bringing in funds aimed at the secondary market, growth funds and even sovereign funds — sovereign funds generally don't invest directly in projects, many act as LPs, but in that wave sovereign funds went directly into autonomous driving companies, even as their first direct investment, with neither data nor financials. Mengxing says he was doing investment banking at the time, and the most aggressive prospectus he saw priced using a PS ratio 60 months out, whereas in all his years he had used at most 12 months forward.
— Meng XingThere has been no second Cruise
Mengxing says this exceeded his expectations: when investing in Momenta and looking at Cruise, he really did think there were many opportunities to sell companies to OEMs, and at no small value, because the judgment was that autonomous driving would become 70-80% of a car's value, and suppliers would gain very strong bargaining power. But this has at least not fully happened to this day, Cruise is a one-of-a-kind case, and no second one has appeared in the world. The acquisitions that did appear in the industry were sales to companies like Amazon, not to carmakers; carmakers are gradually exiting from L4. He adds one line: people have long conflated L4 capability with assisted driving, but L4 is simply suited to providing mobility services.
— Meng XingPeople can't tell sometimes from most times
Mengxing cites Chris Emerson: autonomous driving goes through four stages — never works, sometimes works, most times works, all times works — but people don't find it easy to distinguish these stages. They can distinguish never works from sometimes works, but find it less easy to distinguish sometimes works from most times works, and find it even harder to distinguish most times works from all times works. So once past the first stage, people assume it works all times. And looking at backend data, 30 problems in 100, 10 problems, 3 problems — a simple test can't tell the difference at all, and the gap in between can be a very, very large multiple.
— Meng XingLarge models have reached autonomous driving's 2018
Mengxing thinks the two started out very similarly: overseas one company first ran out results during its quiet scrappy development phase, a crowd of Chinese entrepreneurs suddenly discovered this could be done and believed they wouldn't be too slow to catch up, so starting valuations were all very high, and it was the best industry geniuses doing it. The only difference is that there are more big companies in large models — in the autonomous driving era Alibaba and Tencent didn't enter directly, it was mainly startups, whereas for many big companies large models are the foundation of their existence. He judges that today's large models are roughly equivalent to autonomous driving's 2018, not yet having gone through a particularly bad phase, but by the cycle it will very likely go through capital contraction. He says the current chill doesn't count as cold yet: everyone can still get money, and no one has yet heard talk of M&A.
— Meng XingDon't count on every industry getting one life extension
Mengxing says the capital cycle, business cycle and talent cycle are three interlocking rings; autonomous driving before 2021 was a very standard development rhythm, and 2021 actually got one life extension, and this life extension won't necessarily happen in every industry. So large models will definitely go through the bad process of 2016-2017 moving to 2019, and the best preparation is not to count on a 2021 appearing — before the industry goes down or bottoms out, you should already be able to get a business closed loop or small closed loop working and enter a positive cycle; if you still can't get in, you need to do capital-level control in advance, including adjustments to the business development path, personnel and finances. He says these lessons are very direct, and don't even need another layer of abstraction.
— Meng XingIn their own words · checked verbatim
I did a piece of statistics — companies valued above $100 million, non-resource-based companies, frontier tech companies, angel investment above $100 million — there don't seem to be any very successful cases yet.
我做过一个统计 估值在一亿美金以上的 非资源型的公司 就是前科技公司 天使投资在一亿美金以上 好像还没有很成功的案例
Meng Xing28:26
So I think this is the fate of starting too high. I hope someone can leap over this fate.
所以我觉得这是一个 起步太高的宿命 我希望有人能夸过这个宿命
Meng Xing29:26
I think when a company's appearance of a good sale becomes its most important value point, it will head into a cycle of building around that appearance of a good sale and constantly improving it.
我觉得就是一个公司 当它的卖向 被成为它最重要的 价值点的时候 我觉得它就会走向 一个围绕着卖向去打造 去不断提升卖向的 一个循环过程中
Meng Xing30:27
I think in all the years I did investment banking before, I used at most 12 months forward, 12 months forward — I never used 60 months forward, I couldn't even imagine this thing being invented.
我觉得我之前 干了这么多年 投行最多用过12个月 Ford 往前看12个月 从来没用过 往前看60个月的 甚至于这东西 我都想象不到能发明出来
Meng Xing41:38
But people actually don't find it easy to distinguish these stages. People can distinguish never works from sometimes works, but find it less easy to distinguish sometimes works from most times works, and find it even harder to distinguish most times works from all times works — the difference between them.
但人其实不太容易分得清 这几个阶段 人能分得清 从来不work 到有的时候work 但不太容易分清 有的时候work 到大多时候都work 更加分不清 大多时候work 到永远都work 这之间的区别
Meng Xing55:48
And I think it's roughly equivalent to today's large models having reached autonomous driving's 2018, roughly that situation, so it hasn't yet gone through a particularly bad phase.
然后我觉得差不多相当于 今天大模型走到了 自动驾驶的18年 大概这个状况吧 所以还没有经历 一个特别不好的阶段
Meng Xing1:00:49
In 2021 it actually got one life extension. This life extension, I think, won't necessarily happen in every industry. I think we were lucky.
21年 其实得到一次续命 这次续命 我觉得不一定是在 每个行业里面 一定会发生的 我觉得我们是 运气是好的
Meng Xing1:01:50
Figures
| Angel-round valuation threshold for frontier tech companies | Above $100 million | 28:26 |
| Trough valuation range for autonomous driving companies | $300 million to $500 million | 40:38 |
| PS ratio forecast period used in investment banking pricing | At most 12 months; in 2021, 60 months appeared | 43:38 |
| Size of Didi's autonomous driving team | About 1,000 people, with 200-300 reporting directly or within the system | 7:11 |
| Number of companies invested during the Shunwei period | About 20 frontier tech companies in three years | 3:07 |
Glossary
- SPAC
- A reverse-merger route to going public that in 2021 let many unprofitable companies list.
- PS ratio
- Price-to-sales ratio, used to price a company off forecast future revenue.
- Gartner curve
- A curve describing a new technology going from inflated expectations to a burst bubble to a steady climb.
How to listen
People watching primary-market investment in autonomous driving and large models, especially investors and founders who want to understand valuation cycles, fundraising rhythm, and how startups die.
The first 0-15 minutes on kindergarten, the CMU family experience and the Shunwei days can be fast-forwarded.