After His $7 Billion Valuation Crumbled, He Chose to Start Over | Linear Voice

From cofounder of TuSimple — once billed as the "world's first autonomous driving IPO" — to starting from scratch again after weathering storms, "prodigy" Xiaodi Hou has spent the past decade reflecting on how to balance technology with the realities of the business world.

In October last year, Hou's second venture, Bot Auto, announced the completion of its Pre-A funding round, co-led by Linear Capital. What follows is a rare public retrospective from Hou over this period.

In his view, because a founder must fight until the very end, what he needs is a long-term opportunity to see this through to completion**. That's why he chose to continue building autonomous trucks. But in moving from CTO to CEO,** he shifted from "technology above all" to "operations first," believing that operational efficiency is the lifeline of autonomous driving.** This piece should offer profound insights for technically minded founders.**

"A self-respecting entrepreneur is someone who sees the world as it truly is, and still loves it."

Hou's entrepreneurial story began with TuSimple ten years ago. In 2015, Hou cofounded TuSimple with partners and served as CTO, focusing on L4 autonomous truck R&D. In 2021, TuSimple listed on Nasdaq with a market cap that once exceeded $7 billion. That December, his team completed the world's first public-road driverless heavy-truck test in Arizona — 128 kilometers with no driver or safety operator intervention, sending shockwaves through the industry.

After multiple upheavals, Hou chose to embark on the grueling path of entrepreneurship once more — in July 2023, he founded Bot Auto, again in the autonomous truck space. In October 2024, Bot Auto announced the completion of a $20 million Pre-A round led by Linear Capital, M31 Capital, and others.**

Recently, Hou sat down for an exclusive interview with China Entrepreneur. Reflecting on his TuSimple experience, he said he often reexamines that period. In terms of approach, this venture has brought a shift in mindset — from "technology above all" to "operations first," becoming more pragmatic: restraining expansion, validating the model at minimum cost, and exploring a sustainable business model through grassroots tactics.

As a serial entrepreneur, how does he evaluate his successes and failures over the past decade? Hou believes, "It's all process. Especially since it hasn't ended yet — how can you call it success or failure?" What follows is Hou in his own words:

I left TuSimple on March 8, 2023. After some consideration, I decided to continue building autonomous trucks.

Nobody had money at the time. I hadn't sold any TuSimple stock after the IPO, so I personally had nothing. Some former investors, classmates, and friends helped me scrape together a bit — about $9 million-plus in personal investments. A few investors who deeply believed in my vision put in roughly $6 million. With that $15 million, I started the company. It was enough for our team to run for a year.

One year and three months after founding, we completed our first hub-to-hub demo (a truck leaving one facility, navigating city streets, getting on the highway, exiting, driving more city streets, and arriving at another facility). We raised a small amount more during this period, and the demo cost $16 million total. Compared horizontally with several leading autonomous truck companies: achieving the same demo took Aurora five years, Kodiak six years, with costs reaching hundreds of millions of dollars.

The reasons we could accomplish this at a fraction of our peers' cost, time, and team size: first, the era has reached a stage of major technological development, with abundant new technologies to leverage; second, our team has conviction — many have tasted both failure and success in autonomous driving and know what we're fighting for; third, we see our endgame with extreme clarity: to become an operations-first autonomous driving company. With these things clearly defined, our team focused like a laser on this single point.

Our pilot fleet now has only five or six vehicles — an extremely small scale. Because we have so few trucks, we must prioritize operations, fully mobilizing and utilizing each vehicle.

Operational efficiency is the lifeline of autonomous driving. We sell freight services, not technology — there's a major difference between the two. For a technology-first company, it's quite possible to halt operations at night; as an operations-first company, our vehicles need to run just as efficiently through the night. With the development of large models, we've been pleasantly surprised to find that migrating our daytime perception system to nighttime perception required surprisingly little effort. There are many technical details here, but simply put, scene transfer capability has become much stronger than before. We've now achieved 24×5 operations.

This is my deepest realization from this venture: the battle for autonomous driving will be won or lost on operations.

We can now manage with perhaps 5 people what previously required 50, and we're managing machines, not people. If you center everything around managing people, nothing can scale; if you center it around managing machines, the mission of managing machines is to reduce people, which means our scaling won't hit major pain points.

This is an opportunity created by AI — we can accomplish enormous projects with very few people. AI makes human work more efficient, enables tighter collaboration, and ultimately allows a very small team to complete a large, complex system.

In the three weeks before my second venture, I kept thinking about what to do. Many paths were open — robotics, large models. One investor told me: if you do large models, I'll give you money right now.

But I asked myself: what am I actually pursuing? Tesla may bring short-term heat to the autonomous driving industry. When you're cooking over a campfire, do you splash alcohol on the flames — whoosh, immediate blaze — or add two logs?

I wanted a long-term plan, but large models or robotics startups offered only short-term temptation. I wouldn't be like some investors who enter at Series A and exit at Series C, making quick money. As a founder, I will definitely fight to the end, so what I need is a long-term opportunity to see this through to completion.

For me at that time, the most valuable thing was the various lessons from my failures at TuSimple. For example, we bound ourselves too early with certain OEMs, and some strategic partnerships became burdens instead; internal governance and team expansion also need to be handled with extreme care. These experiences combined could provide me with valuable long-term guidance.

But how much of my autonomous driving experience could transfer to large models? I doubted it. How much to robotics? I doubted that too. For instance, I didn't even know who my first customer would be if I pursued robotics. But I knew exactly who my first customer would be in autonomous driving: whoever ships freight, that's who I bill. Autonomous driving is simply cost-per-mile versus revenue-per-mile. When I can push cost below revenue, the company has its foundation.

Regarding my TuSimple experience, I often reflect. As a founder, I had certain limitations in understanding my role. For a long time, I believed that if our technology ranked number one in the world, I had nothing to fear. This technology-will-crush-all mentality, I now see, was rather naive. Business is complex, and human nature even more so.

Technology alone cannot solve everything — a company has many other problems demanding attention. Such as unhealthy overexpansion. We previously assumed scale would reduce costs. That was an illusion. Expanding before achieving unit economics — the more you scale, the more you lose. It's a "death spiral."

As CTO, I believed commercialization wasn't my responsibility, and I didn't want to handle it. But when a company has problems, the ultimate responsibility always falls on the founder.

Looking back at the first-half turbulence at TuSimple, I have two realizations. First, as founder, I was the ultimate person responsible for TuSimple. I should have underwritten the company's strategic development, but my awareness was insufficient then. Second, this turmoil made me realize the importance of risk management and self-protection.

Coming from a technical background, making this transition to management required learning industries I previously didn't understand, even stepping outside my comfort zone to study corporate governance bylaws, what Wall Street thinks, what our logic is. Many of these things I didn't know or consider before becoming CEO — I even mistakenly thought learning them was a waste of time.

Of course, I believe this was all inevitable experience. It also helped me understand what being a CEO means, what being a founder means.

(My role has indeed changed somewhat in this second venture. I pay more attention to investor relations, industry relations, government relations — "unknown territories" for a technical founder — ensuring we can withstand unknown risks. I used to call myself a builder; now I call myself a guardian. I used to focus on technical problems — if you can't solve it, let me at it; now I'm more focused on strategy, securing adequate strategic space and buffer for the technical team. On this foundation, I directly manage less technology and focus more on building the company's resilience.

One final important thing: I align the company's vision. For example, I used to tell the technical team that being technically strong was enough; now everyone needs to consider technology, business, and everything else. In the real world, can our product be seamlessly used by customers? Are our operating costs low enough? I need to make these things clear to every "geek" in the company. This isn't easy, but I think it's where I spend more time now, and it's worth it.

Meanwhile, I've observed an interesting industry phenomenon: most autonomous driving companies that disappeared "died unnatural deaths." In other words, unlike other manufacturing industries where a company might die because its funding chain broke and it simply couldn't continue, in autonomous driving, the "ways to die" are quite varied.

The problem here is that when a company is in danger, with no precedents to follow, all sorts of "weird things" easily happen. In such situations, you must rely heavily on first principles to avoid falling into traps.

Compared to L2, L4 is a clearly defined product. Once an accident occurs, an L4 company cannot excuse it by saying the human took over too late — because there's no one in the vehicle. All responsibility falls on the L4 developer.

In this situation, our definition of the product's applicable scope is extremely strict. Everyone must first acknowledge the boundaries of what their system can do. If the system can deliver functionality and ensure safety, then the second step is profitability.

Today, many Robotaxi companies haven't truly begun talking about profitability. Behind continuous operations, most still rely heavily on subsidies, which get folded into operating costs; autonomous driving systems are extremely complex and frequently break, requiring returns to manufacturers for repairs at multiples of standard auto repair costs; and many L4 Robotaxis still have a remote teleoperator behind them, judging complex road situations. Replacing drivers with teleoperators doesn't save that much cost.

Although autonomous driving companies now claim one remote operator can supervise 3 vehicles, 5 vehicles, only when that number rises to a certain level — say, one person managing 10 vehicles — can autonomous driving truly show profit potential. Yet today, to my knowledge, no company has achieved this.

Compared to taxis, trucks spend most of their time on highways where traffic rules are very clear, drastically reducing the need for on-the-spot human judgment via remote control.

So in my view, autonomous taxis face relatively large commercialization challenges; autonomous trucks face much less pressure in this regard. In autonomous trucking, one person managing 10 or 20 vehicles could be quite manageable. Of course, there are even simpler scenarios than autonomous trucks, such as mining operations.

Compared to Robotaxi, I believe autonomous trucks and autonomous mining vehicles will reach profitability sooner. Based on our company's development over roughly the past two years, I estimate that by 2027, our per-mile operating costs — including remote operators, data transmission, hardware maintenance, and all other mileage-related costs — can fall below human driver costs. This is our major signal for profitability on highways in that specific scenario by 2027. Importantly, this "profitability" doesn't depend on scale; scaling comes later.

Often, people believe scale is the primary channel for cost reduction. But in my view, using scale to reduce costs is precisely the "root of all evil" behind why autonomous driving has progressed so slowly and burned so much money. I believe that if an autonomous driving system cannot achieve profitability on a single vehicle, it has no qualification to pursue scale. This is the path to defeat: the larger the scale, the greater the cost drain, meaning more subsidies. Fighting others with subsidies is the typical capital feast. If you meticulously rebuild software systems and operational processes from the ground up with operations as the priority, undergoing fundamental transformation, we found that a very small fleet can drive costs extremely low.

Cost reduction requires extreme automation and designing systems with operations as the priority. For example, can data upload and download save a few seconds or minutes? Does sensor calibration need to be redone? Can the detection process be both fast and accurate? When repairs are needed, how can hardware be designed modularly so fixes take minutes rather than leaving the truck in the shop for a week?

Many assume that with lidar being so expensive, autonomous driving cannot profit in specific scenarios. In fact, if we consider a vehicle's operation, hardware costs represent a very small proportion of operating costs. A heavy truck's lifespan might exceed 1 million miles. In the US, spending $10,000 on sensors — $10,000 divided by 1 million miles — adds just one cent per mile to operating costs, essentially negligible. Often people overestimate hardware costs and underestimate the cost savings from smooth operations.

Currently, autonomous driving companies have two different business models: one is like Microsoft's model, selling software (Windows) to people who already have computers — install Windows and you're ready. The other is like someone buying a Lenovo laptop with both software and hardware included, ready out of the box.

But we now believe that whether software or hardware, autonomous driving systems are still in relatively early stages. It's hard to find a customer who can skillfully operate our autonomous driving system, because such systems will constantly face various problems.

Looking at history, Karl Benz invented the first automobile, but with no repair shops, who fixed it? His wife. In the 1960s–70s, IBM sold mainframes to banks, requiring dedicated IBM engineers to handle repairs, operations, and maintenance. These examples offer a major insight: we must acknowledge that in the early commercialization stage, autonomous driving will inevitably face everyday operational "bread and butter" issues. Early business models must include "service" as part of everything.

In my view, in a context of high automation and high unmanned operation, L4 companies will inevitably grow larger and larger, with winner-take-all dynamics — there isn't really a middle ground. Because managing machines is economies of scale, while managing people is the opposite, diseconomies of scale. Globally, we don't currently see any massive truck freight companies, because driver management is extremely difficult. Managing people inevitably means more people, more management headaches.

We may never have seen a logistics giant managing 50,000 truck drivers, but managing 50,000 computers has been validated by hundreds or thousands of companies — it's not an insurmountable challenge. So, autonomous driving represents a shift from diseconomies of scale to economies of scale.

In short, entrepreneurship is not purely a technology competition, but a comprehensive discipline concerning human nature, governance, and long-termism.

Today, the autonomous driving industry is still working to solve the equation between "economies of scale" and "safety boundaries." When this industry truly cracks the business model, it likely won't be the highest-valued player that wins, but the team that first achieves "positive cash flow per vehicle" that reaches the other shore first.