New York startup Midcentury emerged from stealth on September 23, 2026, announcing a $15 million seed round and two core products: a proprietary egocentric dataset spanning more than two million hours of human behavior and a cloud simulation platform called Matrix.
Midcentury’s egocentric dataset
The dataset covers over 50 environments and 20,000 tasks, with every hour annotated with 3D hand pose tracking, depth maps and point tracks. Midcentury says the scale far exceeds existing public research options such as Ego4D v2, which holds roughly 3,600 hours. The company also holds about 50,000 hours of gameplay data with engine-level signals and roughly 69,000 hours of conversational voice data across 25 languages. Because the dataset is strictly proprietary, researchers will not receive free access.
Matrix simulation platform
Matrix is a cloud simulation platform that lets teams construct digital twins of real-world scenarios and builds physics models directly from real data rather than hand-coded rules. Engineering teams can run thousands of parallel tests on GPU clusters and convert test failures into immediate training examples.
Midcentury is positioning first-person human action data as the training foundation for physical AI, arguing that such data can train robots in a manner analogous to how web scrapes trained large language models. The company explicitly leans on Rich Sutton’s “bitter lesson,” favoring brute-force compute and massive datasets over hand-engineered features.
The startup declined to disclose valuation or investor names. The $15 million seed places the round in the top 1% of all-time seed deals within the big data category; by comparison, the median seed round in Q2 2026 was around $4.5 million. A January 2026 SEC filing shows the company sold roughly $8.9 million across five backers, indicating the round closed in stages.
A crowded field is chasing the data layer for embodied AI: competitors including Rerun and Vision Lab have also secured venture funding over the past year. If scaling laws transfer from language models to physical robots, the entities controlling the training data will likely dictate the market.
