AI News Feed
Market watch
Companies

Maxinsights Emerges as Egocentric Data Supplier to OpenAI and Physical AI Firms

QbitAI reports Maxinsights has emerged as a supplier of egocentric training data to OpenAI and leading Physical AI developers, with ARR near $100 million.

OpenAI has not disclosed the identity of its embodied-data supplier, QbitAI said. The report, based on public reports and industry leads, said OpenAI had bought large volumes of egocentric data and that Maxinsights emerged as the supplier in industry discussions.

According to QbitAI, Maxinsights has delivered more than 2 million hours of data to customers and holds more than 1.5 million hours of data with hand-tracking and language annotations. Its monthly collection capacity is 450,000 hours, and its global collection network covers more than 2,000 real environments and more than 50,000 scenes.

The report listed Google DeepMind, Figure AI, Dyna Robotics, Genesis AI and 1X among Maxinsights' customers. It said the company entered the Chinese market in March 2026 and has nearly 50 customers, including five top-tier industry clients. Its model is to work closely with leading model teams to set strict data requirements, standardize datasets, and then offer the resulting products to more than 40 other customers. An early collaboration with Google DeepMind began with about 200 hours of bare-hand egocentric data and expanded over time, according to the report.

QbitAI reported that Maxinsights' bookings for the third quarter of 2026 are expected to rise 48% quarter over quarter, and that its ARR has stayed at or near $100 million for several consecutive months. The report said the recurring revenue signals a shift from project-based data sales to a more durable business, as robot-model teams need hundreds of thousands to millions of hours of data per iteration.

Two model launches were cited as examples. Dyna Robotics' Dyna-2, released in August 2026, was pretrained on 1 million hours of human egocentric video; when pretraining data rose from 1,000 hours to 1 million hours, the model's mean squared error for human-motion prediction fell about 12%, and Maxinsights was the largest supplier of that 1 million-hour dataset, QbitAI said. Genesis AI's GENE-26.5, released in May 2026, performed about 20 consecutive cooking steps, including cutting tomatoes and cracking eggs one-handed; some skills required less than one hour of task-specific robot data, and Maxinsights was its exclusive egocentric-data partner, according to the report.

The report said Maxinsights is also working on data formats beyond hand data. In full-body data, it is exploring reconstruction of human motion trajectories from egocentric video while tracking hands and object interaction; in upper-body data, it is combining action, space and semantics; and it has developed MaxSLAM, which it says can reproduce viewpoint changes at millimeter-level precision for tasks such as packing, housework and assembly.