A New York startup claims it has amassed the largest collection of first-person human action footage ever assembled, and it's betting that data will become as critical to training robots as web scrapes were to chatbots.
Midcentury announced on September 23, 2026, it raised $15 million in a seed round and emerged from stealth with a dual offering: a sprawling egocentric dataset covering more than 2 million hours of human behavior, and a cloud simulation platform called Matrix designed to help robotics teams test and refine their systems at scale. The company declined to name its investors, though an SEC filing from January 6, 2026, showed roughly $8.9 million sold across five backers, suggesting the round closed in stages.
The size of Midcentury's dataset dwarfs what's publicly available in academic circles. Ego4D v2, a widely cited research collection, contains around 3,600 hours of egocentric video. EPIC-KITCHENS-100 has 100 hours. Midcentury's trove is proprietary, which means researchers won't get free access, but the company argues the scale and diversity of its data justifies the commercial approach. The footage spans more than 50 environments and 20,000 tasks, all captured from the first-person perspective that robots actually experience when navigating the world.
Each hour includes 3D hand pose tracking, depth maps, point tracks, and task annotations. Midcentury also claims to have assembled 50,000 hours of gameplay environment data with what it calls "engine-level signals," plus roughly 69,000 hours of conversational voice data across 25 languages. A sample browser on the company's website offers a glimpse of the metadata layers attached to each clip.
Matrix, the simulation half of the business, 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 convert simulation failures back into new training examples. The platform isn't self-serve yet; Midcentury is taking early-access requests and working with what it describes as "frontier labs," though it wouldn't name customers.
CEO Chetan Kulhari previously worked at Magic, the AI coding startup that went through Y Combinator in summer 2022. David Guo is listed as a director on the January SEC filing. The company's LinkedIn page claims between 11 and 50 employees, with LinkedIn listing San Francisco as headquarters while SEC filings and the website list New York addresses. The website footer lists 32 Mercer Street in Manhattan, while the SEC filing shows 169 Madison Avenue as its principal address. Midcentury Labs was incorporated in Delaware in 2024.
In its announcement, the startup invoked what has become something of a rallying cry in machine learning circles. "Human action data has finally unlocked a scaling law for Physical AI," the company said. "It's time to join LLM researchers in taking the bitter lesson pill." That's a reference to the idea, articulated by AI researcher Rich Sutton, that brute-force computation and massive datasets tend to outperform clever hand-engineered features. Whether that lesson transfers cleanly from language models to robots remains an open question.
The $15 million haul is generous by recent standards. Carta data shows the median seed in the second quarter of 2026 was $4.5 million at a post-money valuation around $23.9 million. Midcentury didn't disclose its valuation.
A handful of other startups are making similar wagers. Physical, Human Archive, and Vision Lab (which raised $6 million in June) are all building training-data engines for embodied AI. Rerun pulled in $17 million last March for a multimodal data stack aimed at the same market. Simulation companies like Antioch are approaching the problem from a different angle, replacing real-world data collection with high-fidelity virtual environments.
The underlying bet is that robotics companies will need specialized infrastructure the same way large language models required web crawlers and compute clusters. That thesis is gaining traction as humanoid-robot makers raise enormous sums. Figure AI closed a $675 million Series B in February 2024 at a valuation near $2.6 billion, with backing from Microsoft, NVIDIA, and the OpenAI Startup Fund. Physical Intelligence was reportedly in talks to raise roughly $1 billion at a valuation above $11 billion as of March, according to TechCrunch.
Midcentury said it plans to release three research projects in the coming months: MC-EgoHands for egocentric 3D motion reconstruction, MC-Shade for real-time photorealistic rendering, and MC-PhysBench, described as a long-horizon physics benchmark for world models. All three are listed as "coming soon" on the company's site, though no specific timelines were given.
Whether the company's data moat holds depends on how quickly competitors can assemble similar collections, and whether robotics teams decide to build or buy their training infrastructure. For now, Midcentury is positioning itself as the pick-and-shovel supplier in what it sees as a coming gold rush for physical intelligence.
