The "Water Seller" in the Robotics World is on Fire: How a Company That Doesn't Build Robots Reached a Valuation of Over Ten Billion?
Hello everyone, I'm your financial journalist. Today, we're talking about a story that goes against intuition: while everyone is focused on the physical robots and the "large models" (the brains of robots), a company that specializes in providing data for robots has seen its valuation soar to $1.2 billion in just three months (which is more than 8.5 billion RMB; the article suggests it might exceed ten billion, possibly including future expectations or exchange rate fluctuations, so let's consider this as a high valuation for the sake of the discussion).
This company is called XDOF. It doesn't build robots or develop core algorithms; its work sounds rather mundane: it collects, organizes, and processes video and motion data used by robots in their operations.
Why is capital so eager to invest in it? Behind this is a significant shift in the logic of the Embodied AI industry. Let me break down this phenomenon into five key points to help you understand the reasons.
---
1. The Core Story: Two Rounds of Financing in Three Months, from an Obscurity to a Unicorn
What happened?
XDOF was founded in 2024 by a team from Berkeley, a world-leading institution in robotics research. In June this year, they raised $70 million in their Series A financing. Just three months later, in September, they announced another round of financing, raising their valuation to $1.2 billion.
Why such a rapid growth?
Typically, startups raise funds on an annual basis, but XDOF is doing so on a quarterly or even monthly basis. What does this indicate? It shows that the market is extremely eager for their services.
- Star Investors: Major investors like a16z (Anderson Horowitz Fund, one of the top venture capitals in Silicon Valley) and Thrive Capital have invested in XDOF. These firms usually only invest in the most promising areas.
- Amazing Revenue: Despite being a young company, XDOF already generates nearly $50 million in annual revenue. For a data company that's been around for less than two years, this is incredibly impressive, proving that they are not just selling a vision through a PowerPoint presentation but actually providing valuable services that customers are willing to pay for.
In one sentence: In the field of Embodied AI, data has become the most profitable and scarce resource.
---
2. The Team Behind XDOF: Berkeley Experts with Meta Experience
Who are the Founders?
The three founders of XDOF—Philipp Wu, Fred Shentu, and Nemo Jin—are not just ordinary programmers; they are top researchers with a formal education in this field.
- Philipp Wu (the Core Driver): He comes from the Berkeley Robotics Learning Lab and studied under Pieter Abbeel, a pioneer in robotics learning, who researches how robots can learn tasks like folding clothes and tying knots through observation and trial and error.
- Wu's Background: He has worked at Meta (the parent company of Facebook) as a visiting researcher, focusing on multi-modal robot models, and also at Covariant, a leading company in logistics robotics.
- Technical Foundation: One of their core technologies is the GELLO remote operating system developed at Berkeley. This system allows humans to control robots remotely using a controller while recording every detail of the robot's movements.
Why is their background so important?
In the AI industry, whoever understands data also understands the models. XDOF's team is composed of experts who know exactly what kind of data is useful for robot training and what is not. This technical advantage is unmatched by ordinary data outsourcing companies.
---
3. The Essence of Their Business: Not Just Recording Videos, but Providing “Rehabilitation Training” for Robots
Many people misunderstand: They think XDOF simply hires people to record robots working and sells the videos to AI companies.**
The truth is: This is far more complex than just recording videos.
Why can't it be as simple as recording videos?
- Differences between Internet Data and Physical World Data:
- Training large text models like ChatGPT uses readily available data from the internet (articles, code).
- Training robots for Embodied AI requires data from physical interactions. Did the robot grasp a cup securely? How much force was used? What was the angle of the fingers? If it slipped, was it due to insufficient friction or the wrong angle?
- These “mistakes” and adjustments are the most valuable training materials.
What does XDOF do?
They provide a high-quality data pipeline:
- Simultaneous Recording: They record not only videos but also the robot's joint angles, motor forces, and tactile feedback.
- Data Cleaning: For example, their WARP-RM technology identifies which actions are effective in training and which are unnecessary pauses, allowing the robot to learn faster and more accurately.
- Results: Using XDOF-processed data, the time it takes for a robot to learn to fold clothes has been reduced from 114 seconds to 64 seconds, with a higher success rate.
In one sentence: XDOF doesn't sell videos; they sell data that can directly enhance the intelligence of robots.
---
4. Industry Comparison: The “Scale AI” of Robotics, but More Difficult and Valuable
Why is it called the “Scale AI” of robotics?
- Scale AI is the data provider behind models like ChatGPT and Sora. It doesn't build the models themselves but provides the necessary input. Without Scale AI, models like OpenAI's wouldn't be possible.
- XDOF plays a similar role in robotics. Whether developing the robots (like Tesla's Optimus or Figure) or the large models (like NVIDIA and Google), all need massive amounts of high-quality real-world interaction data for training.
What's the difference?
- Scale AI deals with data from the digital world (images, text), which is relatively easy to process automatically.
- XDOF deals with data from the physical world, which is chaotic and unpredictable. A cup might be tilted, the lighting might change, and the fabric might be wrinkled.
- Therefore, XDOF faces higher barriers: They need advanced hardware (like the GELLO system), specialized algorithms for data cleaning, and a deep understanding of physical laws.
Business Logic:
- Initial Phase: They sell datasets (e.g., ABC-130K, with 130,000 tracks of data).
- Mid-Phase: They sell data collection tools and annotation platforms.
- Long-Term: They aim to become an infrastructure platform, similar to how cloud computing is essential for the internet. All robotics companies will rely on their data pipelines to improve their models.
In one sentence: If robots are the future's iPhones, XDOF is the key supplier of data that ensures the high efficiency of their manufacturing.
---
5. Future Prospects: The Limit of the Data Business
Competition is Rising Globally
The article mentions similar companies in China, such as Guanglun Intelligence (valued over $2 billion) and Mifeng Technology (a spin-off of Zhiyuan Robotics). This shows that the world realizes that whoever controls high-quality data holds the key to Embodied AI.
The Potential of the Data Business
Some question whether the data business is just labor-intensive outsourcing. Is the profit margin low?
XDOF’s model breaks this misconception:
1. Continuous Data Requirement: Models need continuous training, and data needs to be updated regularly. Robots need to learn new tasks constantly.
2. Moving from Selling Data to Selling Solutions: XDOF not only sells data but also the tools and systems for efficiently producing it. It’s like a company that sells oil, selling both the oil and the technology and refineries.
3. Quality Matters: In AI, there’s a saying: “Garbage in, garbage out.” Poor-quality data will result in poor models. XDOF ensures that only high-quality data is used in training.
Risks
- Uncertain Technology Path: The robotics industry is still in its exploratory phase. If simulated data becomes cheaper and more effective than real data in the future, XDOF’s model might be challenged.
- Fierce Competition: Giants like NVIDIA and Tesla might establish their own data teams, potentially squeezing third-party data companies.
In Conclusion
XDOF’s rise marks a shift in the Embodied AI industry from a focus on hardware and concepts to a focus on data and practical applications. For the general public, this means that the future progress of robots will depend not only on engineers writing code but also on data workers and algorithms that remotely control robots and record every detail of their movements in laboratories and factories.
Investment Logic: Businesses that provide the infrastructure for data (like data infrastructure) tend to be more stable and have greater potential for early growth than those that focus on the actual robots.
XDOF’s story illustrates that in the AI era, data is no longer a byproduct but a core asset. Those who can effectively process and utilize this data are becoming new creators of wealth.