Midcentury raises $15M seed for physical AI training data
What's the deal? MidcenturyDealroom has a profile for this one. Try Dealroom →, a New York startup, raised $15 million in a seed round and emerged from stealth on September 23, 2026. It launched two products: an egocentric dataset spanning more than 2 million hours of human behavior, and a cloud simulation platform called Matrix. The company declined to name its investors.
Why so much? The round lands in the top 1% of all-time seed deals in the big data category, based on a sample of 7,938 rounds. For comparison, Carta pegs the median second-quarter 2026 seed at $4.5 million. Midcentury didn't disclose its valuation.
What's the endgame? Midcentury is betting that first-person human action data will train robots the way web scrapes trained chatbots. Its dataset spans more than 50 environments and 20,000 tasks, each hour annotated with 3D hand pose tracking, depth maps, and point tracks. That dwarfs public research sets — Ego4D v2 holds roughly 3,600 hours, EPIC-KITCHENS-100 just 100.
The data is proprietary, so researchers won't get free access. The company argues its scale and diversity justify the commercial approach. It also claims 50,000 hours of gameplay data with "engine-level signals," plus about 69,000 hours of conversational voice data across 25 languages.
The product: Matrix lets teams build digital twins of real-world scenarios with physics learned from actual data rather than hand-coded rules. Companies can run thousands of parallel tests on GPU clusters, then turn failures into fresh training examples. It isn't self-serve yet; Midcentury is taking early-access requests and working with unnamed "frontier labs."
What could go wrong? Midcentury leans on the "bitter lesson," the idea from researcher Rich Sutton that brute-force compute and massive datasets beat hand-engineered features. "Human action data has finally unlocked a scaling law for PhysicalDealroom has a profile for this one. Try Dealroom → AI," the company said. Whether that lesson transfers cleanly from language models to robots remains unproven.
Who's behind it? Chief executive officer Chetan Kulhari previously worked at Magic, the AI coding startup that went through Y Combinator in summer 2022. A January 6, 2026 SEC filing listed David Guo as a director and showed roughly $8.9 million sold across five backers, suggesting the round closed in stages. Midcentury Labs was incorporated in Delaware in 2024.
The signal: A crowd of startups is racing to become the data layer for embodied AI. Rerun raised $17 million last March, Vision LabDealroom has a profile for this one. Try Dealroom → pulled in $6 million in June, and Physical and Human ArchiveDealroom has a profile for this one. Try Dealroom → are chasing the same market. If the scaling law holds for robots, the companies that own the data stand to own the field.
Read more: founderland.ai
Image credit: Generated with Gemini
