By Mengyao, reporting from凹非寺
QbitAI | Official account QbitAI
RSI this, AGI that—no matter how insanely Sam Altman has pushed his own models today, I still have to say one thing—
When it comes to Physical AI,OpenAIhas truly suffered no less than the leading embodied AI vendors!!! (doge)
Remember how in 2020-2021, the unfulfilled OpenAI robotics team went through a round of dissolution, and one of the repeatedly cited real-world difficulties was the lack of training data.
But just a few years later, GPT-6 burst onto the scene, so powerful that the entire internet was stunned into silence! Yet something even more surreal is this:
How on earth is it that you, a general-purpose large language model, can so effortlessly control a robotic arm to complete grasping, placing, and a whole series of embodied tasks!?
△Image source: American independent review agency RoboCurve
So here's the question: back when OpenAI was once stuck on data, how exactly did it get this whole thing going again?
Out of intense curiosity about the technology's development, I actually dug along the trail of public reports and industry leads—and then something shocking, surreal, and absurd emerged—
The Egocentric (first-person perspective) datasets that are red-hot in Physical AI circles right now—OpenAI actually bought them in bulk long ago!
Although OpenAI has not publicly disclosed who this data supplier is, as the Physical AI track continues to heat up, one company has suddenly surfaced amid intense industry discussions:
not some tech giant, but a company calledMaxinsights, which, at most, has only been around fortwo or three years。
Digging further down its customer list, I found that this company has even made it into the supply chains of the world's leading Physical AI players:
Google DeepMind, Figure, 1X……A whole crowd of super-hot top players in the Physical AI circle all appear on its list of partners.
What's more, this company's ARR (annual recurring revenue) has stood at or approached the US$100 million mark for multiple consecutive months~
A Silicon Valley startup only a few years old has actually turned the real operational experience of thousands of people worldwide into a scalable business spanning China and the US market.
So, how exactly did this business get built? Today let's take a good look~
Delivering 2 million hours, breaking into the global Physical AI core supply chain
Honestly, it's really no wonder that Physical AI's demand for large-scale Egocentric Data has surged so fiercely over the past two years.
The reason behind it is very practical: having robots collect manipulation data themselves over and over is costly and slow, and the scenarios they can cover are extremely limited. (Ouch.jpg)
And the power of Egocentric Data happens to fill exactly this big bug in data training—
Simply have a person wear a head-mounted camera, and all the hand movements and human-object interactions in daily life can be recorded.
Lower collection cost, stronger scene diversity, and no dependence on expensive robot hardware —this makes it possible to support large-scale datasets at the level of tens of thousands or even a million hours, nice~
But having said that, many companies work on Egocentric Data, yet only a handful of startups have managed to squeeze into the core supply chain of the world's top model vendors.
After digging deeper, I found that the data business Maxinsights is doing truly goes beyond my usual understanding of startups. (doge)
Just looking at its core data inventory, this company has already fiercely pushed itself to the scale of a "million hours" —
the data delivered to clients alone exceeds2 millionhours, and in addition it holds1.5 millionhours or more of data with hand-tracking and language annotations.
The monthly collection output even reaches450,000hours, and its global collection network extends into2000real environments, covering as many as50000different scenes.
A company just two or three years old has, piece by piece, accumulated humanity's scattered operation experience from around the world into a mine of robot training data~
Of course, no matter how large your data delivery scale is, in the end it still depends on whether customers are willing to pay for this data, and how much value it has delivered in terms of model capability.
So I kept digging and found that, with a data inventory spanning tens of thousands of scenarios, Maxinsights has already done business acrossall six continents of the globe,。
even more intriguing is that they have already squeezed into the supply chains of the world's top Physical AI players—
Google DeepMind, Figure AI, Dyna Robotics, Genesis AI...Pick any one of them, and it's a resounding name in the global robotics world.
In March of this year, the company also brought its business into the Chinese market—the global Physical AI data business really has been made to work by them.
Even more, the data they've sold has already begun to show the "value" of scaled-up pretraining in the training results of cutting-edge robot models.
Take the one released in August (8) this year,Dyna-2for example.
At the time, Dyna Robotics used 1 million hours of human first-person video for pretraining, letting robots learn the laws of physical interaction in the real world.
It actually managed to find a Scaling Law: as pretraining data increased from 1000 hours all the way to 1 million hours, Dyna-2's mean squared error in predicting human actions dropped by about 12%.reduced the mean squared error of prediction by about 12%。
More importantly, even when facing robot data never seen during training, the model's prediction performance continued to improve as the scale of human video grew.
And who would have thought—the largest supplier behind Dyna-2's 1,000,000 hours of pretraining data was not some big corporation, but preciselyNow let's talk about the robot foundation model。
Let's now talk about the robot foundation model released in May this year by the Silicon Valley star robotics company "Genesis AI",GENE-26.5。
When the model was released, they put on quite a complex culinary show—
the robot went straight into the kitchen and autonomously completed about20 stepsof continuous cooking operations in one go: slicing tomatoes, cracking an egg with one hand, coordinating both hands... all without any human intervention.
Even many high-difficulty skills required less than 1 hour of task-specific robot data to achieve autonomous execution. (shocked.jpg)
The key here lies in letting the model learn motor 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 thatthe exclusive first-person data partner of GENE-26.5 is also Maxinsights???"
From the procurement budgets of the world's top model vendors, a hundred-million-dollar-level company has grown.
They've gotten into the global core supply chain, delivered data at the million-hour scale, and earned the recognition of top clients.
We've heard plenty of Physical AI stories over the years, but only a business that turns technical demand into real money has truly found the way to commercialization.
And as for that ledger, Maxinsights' latest performance report has already provided the answer.
According to the latest operating data,in the third quarter of 2026, Maxinsights' Bookings(contract signing value)are expected to grow 48% quarter-over-quarter.
The surface meaning of "Bookings" is easy to understand: it measures how much real-money customers are willing to sign contracts for with this company's data services.
But there is more to it than that — after all, with robot training data, it is not simply a matter of collecting more and selling more.
Whether customers are willing to make purchases,behind the scenes it tests whether the data can truly enter the model training pipeline, whether it can steadily improve model performance, and whether the supplier can keep up with the data scale and quality of the next round of training.
Especially for leading Foundation Model teams, model iterations often require hundreds of thousands or even a million hours of data support, placing相当 high demands on data quality, scenario coverage, and delivery stability.
So, in a sense, the growth in Bookings is customers voting with real money for a data company's ability to deliver at scale and its long-term supply value~
Even more worth emphasizing is that this growth in orders comes on top of having already achieved "million-hour"-scale data delivery earlier on!!!
Judging by the momentum of this data factory's continuous expansion —
in the future this company's number of signed orders might just, possibly, perhaps follow in RSI's footsteps toward an "Intelligence Explosion." (laughs)
There is another set of numbers that is well worth savoring —
Maxinsights has now maintained an ARR for "multiple consecutive months"(annual recurring revenue)at or near the 100 million dollar level.
The key point worth emphasizing is that ARR does not measure how much was earned this month; it refers to the revenue that can be stably repeated over a year under the current contract structure.
Consecutive! Multiple months! Maintaining an ARR of nearly a hundred million dollars means that, within this company, this revenue stream is predictable cash flow that can be stably reproduced every month.
And another figure that must be noted is the amount of "100 million dollars".
You should know that $100M ARR is also an industry-recognized maturity threshold, which nearly means that Maxinsights has alreadycompleted its early-stage technical validation and customer validation, and entered the stage of scaling expansion.
But what I want to say more is that the real value of $100M ARR lies, crucially, in the gradually taking shapedata flywheel——
behind it. The Physical AI data business has a characteristic: e! conomies! of! scale! are! significant!
The more customers there are, the richer the accumulated real-world scenarios and training data; the richer the data, the greater the opportunity to improve model training results; once the results are validated, it attracts customers to continue purchasing and place additional 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 be cashed in at scale~
△AI-generated
In addition, what I also want to say is that behind this performance hides an industry account that deserves even more thought.
Physical AI data training has always been a business where it is easy to earn project money but hard to earn long-term money.
The reason is very realistic: data procurement in fields such as embodied intelligence, robotics, and autonomous driving has long had a strong"project-based" character。
Once a customer buys synthetic data, annotated data, or scenario simulation and finishes training one version of a model, the need theoretically ends — settlement is per order, and once the job is done, everyone parts ways.
So for some startups, an awkward situation can easily occur: revenue fluctuates in pulses following orders, and the company can never grow big...
Therefore, the fact that a data training company can turn its revenue into a monthly recurring ARR shows that it has overcome several major hurdles of this business —
a customer base diverse enough, customers willing to pay for it, new data from model iterations able to keep the flywheel running, and the capability for sustained self-generated growth.
From the ability to "take orders" to the confidence to "keep making money", Maxinsights has clearly now delivered a convincing report card.
Beyond data supply: Maxinsights co-defines next-generation data with foundation models
Of course, on the Physical AI track, there is a real issue that is easily obscured by the currently hot data procurement business —
the data customers are willing to pay for today may not be the data next-generation robots need most.
After all, model capabilities are iterating so fast now that even top model teams sometimes have to explore while training what robots need to learn next and what data they lack.
But what is interesting is that while many data vendors are still figuring out how to sell their data, what Maxinsights does goes beyond merely "supplying data" —
itdirectly participates in building the Physical AI data infrastructureworking together with top Foundation Model teams to figure out what data the industry actually needs.
Currently Maxinsights already has, worldwide, close to50 clients,including 5 of the very top leading enterprises in the industry.
Yet compared with the common on-demand customization and per-order data delivery model, Maxinsights' business model is somewhat special.
Specifically, team members first work deeply with these leading clients, repeatedly polishing the product according to the most demanding data requirements, accumulating high-quality, relatively standardized datasets, and then bringing this mature data product to more than 40 other clients~
Take Maxinsights' early collaboration withGoogle DeepMindas an example.
At the time, the two sides together explored bare-hand first-person data collection, starting from roughly 200 hours and gradually building up the delivery scale bit by bit.
What is quite intriguing, however, is that the cooperation between Maxinsights and its clients did not come to an end with the data delivery—
After the model team raised requirements, the two sides would still explore data solutions together, then Maxinsights would organize the collection and processing, and continuously optimize the data product based on feedback from the client's model training.
In this way, the originally on-demand data purchasing business gradually turned into a long-term partnership in which both sides jointly explore data needs and continuously develop and iterate on data products.
This partnership model has even extended toText, Upper Body, Full Bodyand other new data directions such as UMI~
△AI-generated
From meeting customer needs to turning one customer's high standards into the shared standard of many more customers — this kind of capability is truly rare across the entire industry.
In the data business of Physical AI, being able to take on orders is one thing, but getting your own data standards adopted by more and more customers is what offers the chance to turn the business into infrastructure.
Of course, daring to take on business from global giants, Maxinsights has also accumulated a set of "technical foundations" spanning data collection, processing, and delivery at scale —
For example, in the Full Body direction.
The company has already begun exploring the reconstruction of full-body human motion trajectories from first-person videos. The system can restore the human body's 3D pose while tracking hand movements and object interactions.
This way, the learning targets for robots have the opportunity to extend from local hand operations to more complex whole-body coordinated movements:
In the Upper Body direction.
Maxinsights' exploration goes even deeper into the joint data representation of "motion–space–semantics".
From making a latte to delivering coffee, every step of the operation can be precisely mapped to specific actions and task instructions — effectively equipping robots with a more refined and useful hands-on textbook for learning complex operations and training VLA models, be like:
In the direction of high-precision spatial perception, Maxinsights has also self-developed a large model calledMaxSLAM.
It can reconstruct the viewpoint changes during human operations with millimeter-level precision, letting robots clearly see the spatial movements behind actions like packing boxes, tidying up household chores, and repair and assembly, providing precise data for subsequent model training.
Its real-world test 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, with orientation error as low as 0.57°:
Honestly, after learning about Maxinsights, I did feel a tiny bit emotional.
Because by 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 stepping in to place their bets.
But hype is one thing; a company that can produce data at scale and with quality, while also keeping the world's top model makers continuously paying, is ultimately one of the few.
From data delivery at the million-hour scale, to being a core supplier for global leading players like OpenAI, Google DeepMind, and Figure AI, to exploring next-generation data needs together with frontier model teams —
Taken together, these achievements give us a clear view of the growth path of aleading startup in the Physical AI data and infrastructure track.
How far physical AI is from entering real life at large, no one can say for sure right now.
But in the supply chain that provides them with training data, some players have clearly already used their unique capabilities to occupy a hard-to-replicate ecological niche in the industry ahead of time.
Perceive the trend early; achieve through solid work.
Maxinsights clearly understands this well.