Everyone Says This Tiny Startup Feeds OpenAI’s Robots. The Fine Print Tells a Quieter Story.

October 22, 2026, San Francisco. Maxinsights’ funding announcement appeared on Business Wire with a restrained headline: a $30 million seed round led by Peak XV. On the same day, a long Chinese article began circulating with exclamation marks in its headline. It described the company as OpenAI’s secret embodied-data supplier, the invisible choke point in the global Physical AI supply chain, and a two-year miracle with ARR near $100 million.

Two things happened at once, but they were not saying the same thing.

In the funding announcement, founder Yuguang Yuan said something else. He said Dyna-2 was still improving after more than 1 million hours, but “broad human-level adaptability is another challenge” and might “require 10 million hours or more of diverse data.” The Chinese narrative mostly skipped that sentence. It was not exciting enough, not engineer-enough to fit a “mine” story.

But it was the most honest sentence in the whole affair.

I. What Stands Up First

Maxinsights did not appear from nowhere. It was founded in 2024 in the San Francisco Bay Area by Yuguang Yuan and Kun Wang. Its core team comes from Waymo, Google, Meta, XPeng, and Bosch. The work is concrete. It asks real people to wear head-mounted, wrist, and chest cameras and wash dishes, make milkshakes, and wash cars at normal speed in real homes and restaurants. Then it turns the first-person video into training data for robot foundation models.

The hardest customer relationship to dispute comes from Genesis AI. Co-founder Johnson Wang said in a statement that “a significant portion of the data” for GENE-26.5 was provided by Maxinsights. Maxinsights’ CEO was more specific: all 200,000 hours of egocentric data used in GENE-26.5 were supplied exclusively by Maxinsights. On the Dyna Robotics side, there is a cross-checkable claim. Maxinsights was the largest data supplier behind Dyna-2, and Dyna-2 was indeed pretrained on more than 1 million hours of human first-person video. Its technical report showed sustained improvement in prediction error as scale increased from 1,000 hours to 1 million hours.

Both customer relationships are anchored by public statements from the customer side and by model technical reports. They hold up.

On data scale, the company’s own account is this: more than 2 million hours collected cumulatively, of which 1.5 million hours are “model-ready,” monthly processing capacity of about 450,000 hours, and more than 5,000 contractors across six countries. These are company figures, not third-party audited, but the definitions are clear.

The funding is also verifiable. A $30 million seed round led by Peak XV came with two product launches: MaxLoop, an agent-in-the-loop system that scores and corrects data quality in real time during collection, and MaxGlove, a sensor glove that captures tactile force and contact signals.

Those are the foundation. The problem sits in what was stacked on top.

II. The Arithmetic of $100 Million

The most widely circulated Chinese long-form piece said Maxinsights’ ARR “stood at or approached the $100 million mark for several consecutive months.” Multiple Chinese secondary sources repeated the sentence almost word for word.

A search through English first-hand reporting, including the funding announcement, Forbes’ CEO interview, and TNW’s coverage, turns up no ARR figure. What English reporting does confirm is “more than 30 customers, 8 of them on recurring contracts.”

The ARR number may not be fake. It may be a figure the company disclosed privately to investors and that Chinese channels obtained and amplified. Put it back on the market’s ruler, though, and the problem becomes sharp.

According to Global Info Research estimates, global commercial deliveries of embodied-intelligence datasets in 2025 totaled about 1.6 million equivalent usable hours. The weighted average accepted delivery price was about $58 per hour, and industry gross margin was about 34%. Multiply those out, and the global commercial embodied dataset market in 2025 was on the order of $90 million to $100 million. Another market report put the 2025 market size at $95.49 million.

Maxinsights’ claimed ARR alone is close to the entire category’s global commercial delivery value for a full year.

For a company founded in 2024, with more than 30 customers of which only 8 are on recurring contracts, revenue close to the entire market total logically requires an almost monopolistic share. Or it requires its “ARR” definition to include things well beyond “commercial delivery of embodied datasets”: upfront recognition of project-based contracts, annualized conversion of multi-year framework agreements, non-recurring data-processing services.

Those definitions are common in SaaS. In a data-services category still dominated by project-based work, they need extreme caution.

The tension between funding size and revenue narrative is sharper. If a company had truly reached close to $100 million ARR, the pricing logic of a $30 million seed round would be highly anomalous. Either the investors were extremely conservative, or the quality of the ARR number is far lower than it appears.

Both explanations point to the same conclusion. ARR is the number in this narrative that most needs to be interrogated, and the one that least survives a gentle push.

III. The OpenAI Evidence Chain

The headline of that long-form piece was itself the biggest hook: “Who is supplying embodied data to OpenAI?”

The logic ran like this. OpenAI disbanded its robotics team in 2021 because it did not have enough training data. A few years later GPT-6 could control a robotic arm. Therefore OpenAI must be buying egocentric data at scale. And Maxinsights is that supplier.

That OpenAI disbanded its robotics team in 2021 for lack of data is documented. Wojciech Zaremba, who led the robotics team at the time, said explicitly that the lack of sufficient training data was the core reason for shutting it down. Between “OpenAI needs egocentric data” and “OpenAI buys from Maxinsights” lies an entire evidence chain’s worth of gap.

Maxinsights’ own supplier page lists customers including OpenAI, Google DeepMind, Meta, Figure, 1X, Skild, Genesis AI, and Dyna Robotics. It is a long list, but “has served” and “core supplier” are concepts of different orders. Of the more than 30 customers, only Genesis and Dyna have public customer testimony or collaboration details. The other names, including OpenAI and DeepMind, currently have no public confirmation from the customer side, no disclosed procurement contracts, no acknowledgments in model technical reports, no jointly signed first-hand materials.

The long-form piece’s method was this. First acknowledge that “OpenAI has not publicly disclosed its suppliers.” Then use “industry leads,” “surfaced,” and “what many people don’t know is” to complete the leap. That method treats “not denied” as “confirmed.”

A data company listing OpenAI on its website customer list is not the same as OpenAI’s robotics team using that company’s data at scale in its training pipeline. The two are separated by procurement scale, usage mode, whether the data entered the core training process, and whether it was repeatedly purchased.

The more defensible formulation is this: Maxinsights claims to serve OpenAI, but there is currently no independent confirmation from OpenAI’s side. The distance between that sentence and “OpenAI is buying heavily” is the distance between PR and reporting.

IV. The One Million Free Hours

Data scale is the most intuitive moat in the Maxinsights narrative: 2 million hours collected, 1.5 million hours ready, monthly processing of 450,000 hours. In a data-scarce industry, that volume sounds like a mine.

Over the past year, a structural change has occurred in the egocentric data market that the long-form piece completely ignored. The cheapest tier of data has already become free.

Build AI open-sourced Egocentric-10K in November 2025, then released Egocentric-100K in December. That release included more than 100,000 hours of first-person video from more than 14,000 factory workers, under Apache 2.0, free for commercial use. By April 2026, that number had been pushed to about 1 million hours. Guanglun Intelligent, together with Hugging Face, released EgoSuite-Open100K. The first 10,000 hours are online, with hand and body pose annotations, and the plan is to reach 100,000 hours. China’s RoboMIND grew from 100,000 trajectories to more than 400,000, with global cumulative downloads surpassing 20 million by September 2026.

“Having egocentric video” is no longer scarce in itself. Free data is rapidly approaching commercial datasets in both volume and coverage.

Paid data must answer a question free data cannot. What do you provide beyond those 1 million hours of free, Apache 2.0-licensed, commercially usable factory first-person video?

Maxinsights’ answer is embedded in its product structure: four layers of annotation (environment, task, subtask, instruction), plus hand pose, body motion, object interaction, depth, tactile force signals, MaxSLAM’s millimeter-level spatial reconstruction, and grasp-force data collected by MaxGlove. These are things the free datasets do not have. Build AI’s 100K version is 256p resolution and contains no inertial data, hand pose, or contact labels.

From “having video” to “data a model can directly use to train a VLA or world-action model,” there is a river of annotation and physical-signal density in between.

This pushes the question from “how many hours do you have” to “what exactly is in each of your hours.” If most of the 2 million hours are raw or semi-raw video with low annotation density, their training value cannot be measured simply by hour count. Industry procurement standards are shifting from “trajectory count, recorded hours” to “task coverage, cross-embodiment transfer, value of failure samples, model success-rate improvement.”

Maxinsights’ moat is not the number 2 million. It is whether it can prove that each of its hours contributes a quantifiably larger amount to model gradients than each free hour.

The long-form piece did not ask that question. It treated hour count as the finish line.

V. A Crowded Table

The long-form piece portrayed Maxinsights as an invisible champion that “suddenly surfaced,” as if it were almost the only scaled player in the category.

The competitive landscape is far more crowded.

Scale AI’s revenue in 2025 was close to $1 billion and is expected to exceed $1 billion in 2026. Its Physical AI Data Engine is positioned as an end-to-end data engine for robotics and autonomous driving, and its partnership with Universal Robots is already collecting visual and force-feedback data on industrial robots. Guanglun Intelligent completed a strategic financing of RMB 1 billion in 2026, with a reported valuation above $2 billion. It pursues a hybrid route of “synthetic data + simulation + real collection,” while using free open-source datasets as the top of its funnel. Build AI is using a free open-source strategy to seize the floor price of data supply. MovasAI claims to have collected more than 100,000 hours of egocentric data in real work scenarios including kitchens, warehouses, and retail. Encord, Labelbox, Appen, TaskUs, and other annotation and data-infrastructure companies are also extending into Physical AI.

Maxinsights does differentiate itself. It bets on a narrow, deep route of “real people, real scenes, natural speed, deep physical-signal annotation.” It does not build mock environments. It does not ask operators to slow their hands. It collects in real homes and commercial spaces at normal working pace, then uses MaxLoop for real-time quality control at the collection end and MaxGlove to add the tactile dimension. This route has a structural advantage in covering real-world edge cases. Robots are deployed into real kitchens, not set kitchens, where accidents happen every few minutes.

Substantive differentiation does not equal an unreplicable niche. Guanglun Intelligent’s hybrid route, Scale AI’s end-to-end engine, and the floor-level squeeze from free open-source datasets are all compressing the pricing space and customer bargaining power of a pure egocentric data supplier from different directions.

Maxinsights has more than 30 customers and only 8 on recurring contracts. In a category where free data runs into the millions of hours and giants have revenue in the billions, its “irreplaceability” must be re-proven with every funding round and every customer renewal. A PR piece cannot establish it once and for all.

VI. Where It Actually Stands

With those pieces in view, Maxinsights’ position becomes clear.

It is a company that has done solid engineering in the right direction. Real-scene collection, a four-layer annotation system, MaxLoop’s collection-end quality control, MaxGlove’s tactile dimension, and MaxSLAM’s millimeter-level spatial reconstruction are not packaging. They are engineering investment. Public confirmation from Genesis and Dyna also shows that its data did enter the training pipelines of frontier models and produced effects that can be cited in technical reports.

But it is not what that long-form piece said it was.

It is not the secret supplier “OpenAI is buying heavily from,” at least not on evidence from OpenAI’s side. It is not a scaled machine with “ARR approaching $100 million,” at least not without a more transparent definition. The entire category’s annual commercial delivery is under $100 million, so that number needs far more scrutiny. It is not an “unreplicable ecological niche.” Free data is spreading across the bottom by the million-hour. Guanglun and Scale are squeezing from both sides. Its own recurring-contract customers number only 8.

It is an early-stage company. A seed round, more than 30 customers, 2 million hours collected, two verifiable top-tier customer relationships, and a data collection and annotation process that works in real scenes. These are verifiable. They are enough to support it in continuing to move forward.

The signpost for “continuing to move forward” is not the “$100 million ARR” in that long-form piece. It is what the CEO himself said in the funding announcement. One million hours only proves that scaling is still working. Ten million hours is the threshold.

On the road to 10 million hours, the 2 million hours that today is written as a “mine” is only a number at the starting point. Who can hold quality, annotation density, privacy compliance, and unit economics at that scale is the dividing line for this category.

October 22, 2026, the day the funding announcement went out. Somewhere in an office in the San Francisco Bay Area, one of Maxinsights’ 5,000 contractors was in a real family’s kitchen, wearing a head-mounted camera, a wrist camera, and a chest camera, washing a bowl at normal speed.

That bowl would be recorded as a first-person video, pass through four layers of annotation, hand-pose extraction, and MaxSLAM spatial reconstruction, and eventually become one sample in some robot foundation model’s training pipeline. How many hours, how many dollars, how much model gradient it is worth does not depend on the exclamation marks in that long-form piece. It depends on what it tells the robot that the free bowl does not.