AI News

Who is supplying embodied data to OpenAI?

QbitAIUpdated Nov 16 18:31

AI summary

The company supplying embodied data to OpenAI and other leading vendors is Maxinsights, a startup founded only two or three years ago. It has delivered over 2 million hours of first-person data to customers and holds more than 1.5 million hours of data with hand tracking and language annotations. Its client list includes Google DeepMind, Figure AI, Dyna Robotics, and Genesis AI, and it supplied the 1 million hours of pre-training data for Dyna-2. In the third quarter of 2026, its bookings are expected to grow 48% quarter-over-quarter, and its ARR has reached or approached $100 million for several consecutive months.

Body · AI translation

This RSI, that AGI—never mind that Sam Altman has already pushed his own models to the absolute limit today, I still have to say one thing—

On the Physical AI front, the hardships OpenAI has endured are truly no fewer than those of the leading embodied AI companies!!! (doge)

Remember back in 2020-2021, OpenAI's robotics team, with its ambitions unfulfilled, went through a round of disbandment, and one of the practical difficulties repeatedly mentioned was that there wasn't enough training data.

But just a few years later, GPT-6 burst onto the scene just like that, so powerful that the whole internet was left stunned and floored! But another thing that feels even more surreal is:

How is it that you, a general-purpose large language model, can so effortlessly control a robotic arm and complete a whole series of embodied tasks like grasping and placing!?

So here's the question: how did OpenAI, once held back by data, manage to get this going again?

Out of strong curiosity about technological development, I actually dug around following public reports and industry leads, and then something absurd, surreal, and outrageous came up—

The Egocentric (first-person perspective) datasets that are white-hot in the Physical AI circle right now—OpenAI actually bought a huge batch of them long ago!

Although OpenAI hasn't publicly disclosed who this data supplier is, as the Physical AI track keeps heating up, one company has suddenly surfaced amid the intense industry discussion:

Not some tech giant, but a company called Maxinsights, which has only been around for two or three years at most.

Then, following the customer list further down, I found that this company has even made it into the supply chain of the world's top Physical AI players:

Google DeepMind, Figure, 1X... a whole host of super-hot top players in the Physical AI circle all appear on its partner list.

What's more, this company's ARR (annualized recurring revenue) has stood at or approached the $100 million mark for several consecutive months~

A Silicon Valley startup founded just a few years ago has managed to turn thousands of people's real-world operating experience worldwide into a scaled business spanning both the Chinese and American markets.

So how did this business get built? Let's take a good look today~

Honestly, you can't blame Physical AI for its explosive, explosive growth in demand for large-scale Egocentric Data over the past two years.

The reason behind it is very real: having robots collect operating data themselves over and over is costly, slow, and can cover only a very limited range of scenarios. (heartbreaking.jpg)

And what Egocentric Data can do happens to fill this big bug in data training—

As long as people wear head-mounted cameras, they can record hand movements and human-object interaction processes in daily life.

Collection costs are lower, scenario diversity is greater, and it doesn't depend on expensive robot hardware—it's capable of supporting large-scale datasets on the order of tens of thousands or even millions of hours, nice~

That said, plenty of companies do Egocentric Data, but few startups can squeeze into the core supply chain of the world's top model makers.

After a deeper look, I found that the data business Maxinsights does is indeed a bit beyond my usual understanding of startups. (doge)

Just its most core data assets alone—the company has already pushed to the "million-hour" scale—

It has delivered more than 2 million hours of data to customers alone, and in addition owns more than 1.5 million hours of data with hand tracking and language annotation.

Monthly collection output even reaches 450,000 hours, and its global collection network has spread into more than 2,000 real environments, covering a full 50,000-plus scenarios.

A company founded just two or three years ago has, bit by bit, gathered the human operating experience scattered around the world into a mine of robot training data~

Of course, no matter how large your data delivery scale is, ultimately it comes down to whether customers are willing to pay for this data and how much value has been realized in model capabilities.

So I kept digging and found that Maxinsights, with its data assets spanning tens of thousands of scenarios, has already done business across six continents worldwide, including North America.

What's even more intriguing is that it has even squeezed into the supply chain of the world's top Physical AI players—

Google DeepMind, Figure AI, Dyna Robotics, Genesis AI... pick any one of them, and they're all big names in the global robotics circle.

In March this year, the company also brought its business into the Chinese market, and it has truly made the global Physical AI data business work.

What's more, this sold data has already begun to show the "value" of large-scale pretraining in the training results of frontier robot models.

Take Dyna-2, released in August this year.

At the time, Dyna Robotics used 1 million hours of human first-person video for pretraining, letting the robot learn the laws of physical interaction in the real world.

The result was that it actually found a Scaling Law: as pretraining data increased from 1,000 hours all the way to 1 million hours, Dyna-2's mean squared error in predicting human actions dropped by about 12%.

More importantly, even when faced with robot data never seen during the training phase, the model's prediction performance continued to improve as the scale of human video grew.

Who would have thought, then? Behind Dyna-2's 1 million hours of pretraining data, the largest supplier is not some big company—it is Maxinsights.

Let's now talk about GENE-26.5, the robot foundation model released in May this year by Genesis AI, a star robotics company in Silicon Valley.

When the model was released, they put on quite a complex cooking show—

The robot went straight into the kitchen and autonomously completed about 20 consecutive cooking operations in one go: chopping tomatoes, cracking eggs with one hand, coordinating both hands... with no human takeover throughout.

Even many highly difficult skills could be executed autonomously with less than 1 hour of task-specific robot data. (Shocked.jpg)

The key here lies in letting the model learn motion and physical interaction experience in advance from large amounts of human first-person video and glove sensor data:

What many people don't know is that GENE-26.5's exclusive first-person data partner is also Maxinsights???

It has squeezed into the global core supply chain, delivered million-hour-scale data, and won recognition from top customers.

We've heard many Physical AI stories over the years, but only those that can turn technical demand into real money have truly grasped the ropes of commercialization.

And Maxinsights' latest performance report has already given the answer to this account.

According to the latest operating data, in the third quarter of 2026, Maxinsights' Bookings (signed order value) are expected to grow 48% quarter over quarter.

The surface meaning of "signed order value" is easy to understand: it measures how many contracts customers are willing to sign, in real money, for this company's data services.

But there is more to it than that. After all, robot training data is not simply a matter of the more you collect and the more you sell, the better.

Whether customers are willing to buy tests whether the data can truly enter the model training process, whether it can stably improve model performance, and whether the supplier can keep up with the data scale and quality required for the next round of training.

Top Foundation Model teams in particular often need hundreds of thousands or even millions of hours of data to support model iterations, and they have quite high requirements for data quality, scenario coverage, and delivery stability.

So the growth in signed order value is, to some extent, customers using real money to vote for a data company's large-scale delivery capability and long-term supply value~

What deserves even more attention is that this order growth was achieved on top of the "million-hour"-scale data delivery already realized earlier!!!

Judging by the momentum of this data factory's continuous expansion—

In the future, the number of orders this company signs will probably, maybe, possibly follow RSI's pace into an "intelligence explosion." (laughs)

There is another set of numbers that is very worth savoring—

Maxinsights has now maintained ARR (annualized recurring revenue) at or near $100 million for "multiple consecutive months."

It should be emphasized that ARR does not measure how much was earned this month; it refers to revenue that can be stably repeated over a year under the current contract structure.

Consecutive! Multiple months! Maintaining ARR close to $100 million means that this company has predictable cash flow that can be stably reproduced every month.

And another number that must be noted is the figure "$100 million."

You should know that $100M ARR is also a mature threshold widely recognized by the industry, almost meaning that Maxinsights has completed early technical validation and customer validation and entered the stage of large-scale expansion.

But what I want to say more is that the value of $100M ARR lies in the gradually forming data flywheel behind it—

Physical AI data business has one characteristic: scale! effects! are! significant!

The more customers there are, the richer the real-world scenarios and training data accumulated; the richer the data, the greater the chance of improving model training results; once the results are validated, it can attract customers to continue purchasing and add orders.

In this way, a positive cycle of business growth is formed, and $100M ARR is precisely a strong signal that this business model has begun to pay off at scale~

In addition, I also want to say that behind this performance, there is actually a more thought-provoking industry account.

Physical AI data training has always been a business where it is easy to earn project money but difficult to earn long-term money.

The reason is realistic: data procurement in fields such as embodied intelligence, robotics, and autonomous driving has long had a strong "project-based" attribute.

Customers buy synthetic data, labeled data, and scenario simulation, train a version of the model, and in theory the demand is over—settled per order, disbanded when done.

So for some startups, an awkward scene easily occurs: revenue fluctuates in pulses with orders, and then the company can never grow big...

Therefore, the fact that a data training company can turn revenue into monthly sustained ARR also shows that it has overcome several major hurdles in this business—

The customer base is sufficiently diverse, customers are willing to pay it, new data brought by model iterations can continue to operate, and it has the ability to continuously generate cash.

From the ability to "take orders" to the confidence to "keep making money," Maxinsights has clearly now delivered a convincing report card.

Of course, in the Physical AI track, there is a reality that is easily obscured by the currently booming data procurement business—

The data customers are willing to pay for today is not necessarily the data the next generation of robots needs most.

After all, model capabilities are iterating too fast now; what robots need to learn next and what data they lack—sometimes even top model teams have to figure it out while training.

But interestingly, while many data vendors are still figuring out how to sell data, Maxinsights does more than just "supply data"—

It directly participates in building the data infrastructure for Physical AI, and works with top Foundation Model teams to figure out what data the industry actually needs.

At present, Maxinsights already has nearly 50 partner customers worldwide, including 5 of the most top-tier companies in the industry.

Compared with the common on-demand customization and order-based delivery data supply, Maxinsights' business model is somewhat special.

Specifically, team members first work deeply with these leading customers, repeatedly polishing products according to the strictest data requirements, accumulating high-quality, relatively standardized datasets, and then promoting this mature data product to more than 40 other customers~

Take Maxinsights' early cooperation with Google DeepMind as an example.

At that time, the two sides worked together on bare-hand first-person data collection, starting from about 200 hours at first, and then bit by bit scaled up delivery.

But what is quite thought-provoking is that the cooperation between Maxinsights and its customers did not end with data delivery—

After the model team raises requirements, the two sides also explore data solutions together, then Maxinsights organizes collection and processing, and continuously optimizes data products based on customer model training feedback.

In this way, what was originally an on-demand data procurement business gradually became a long-term cooperation in which both sides jointly explore data needs and continuously develop and iterate data products.

From meeting customer needs to making one customer's high standards the common standard for more customers—this capability is indeed rare across the entire industry.

In the Physical AI data business, being able to take orders is a skill; only by getting one's data standards adopted by more and more customers can one have the chance to turn the business into infrastructure.

Of course, to dare to take business from global giants, Maxinsights has also accumulated a set of "technical foundations" from data collection and processing to large-scale delivery—

For example, in the Full Body direction.

The company has begun exploring the reconstruction of full-body human motion trajectories from first-person video, and the system can restore the human body's three-dimensional pose while tracking hand movements and object interactions.

In this way, the robot's learning target also has the opportunity to extend from local hand manipulation to more complex full-body coordinated movements:

In the Upper Body direction.

Maxinsights' exploration even goes deep into the joint data expression of "action—space—semantics."

From making lattes to delivering coffee, every step can precisely correspond to specific actions and task instructions, equivalent to preparing a more refined and useful practical teaching material for robots to learn complex operations and train VLA models, be like:

In the high-precision spatial perception direction, Maxinsights has also developed a large model called MaxSLAM.

It can restore the perspective changes during human operations with millimeter-level precision, allowing robots to clearly see the spatial motion behind actions such as packing, tidying up household chores, and repair and assembly, providing precise data for subsequent model training.

The measured results are also quite impressive. In an internal evaluation covering four types of real tasks and 20 operation sequences, MaxSLAM's position trajectory error was only 4.24 millimeters, and its pose error was as low as 0.57°:

Honestly, after learning about Maxinsights, I do have a tiny bit of reflection.

Because today, Physical AI data collection is no longer a niche business at all, and in the past two years, various tech giants have also begun to place bets.

But despite the boom, companies that can make data large-scale and high-quality, and can also get top global model makers to keep paying, are ultimately few.

From million-hour-level data delivery, to being a core supplier to global leading players such as OpenAI, Google DeepMind, and Figure AI, to jointly exploring next-generation data needs with frontier model teams—

These achievements tied together let us see clearly the growth path of a leading startup in the Physical AI data and infrastructure track.

But in the industrial chain that provides them with training data, some players have clearly already relied on unique capabilities to occupy a hard-to-replicate ecological niche in the industry ahead of time.

Recognize the trend first, achieve results in practice.

Maxinsights clearly understands this well.

Source

QbitAI