Summary: Config announced a $27 million oversubscribed seed round on May 11, 2026 at a valuation above $200 million, led by Samsung Venture Investment with strategic backing from LG, SKT America, and Hyundai’s ZER01NE. The startup supplies a neutral data layer for robotics foundation models and says it has accumulated more than 100,000 hours of human motion data. It plans to scale toward one million hours, reach $10 million ARR by the end of 2027, and offer cloud-based robot-as-a-service.
Config, which keeps its headquarters in Seoul and San Jose, said on May 11, 2026 that its oversubscribed, $27 million seed round had priced the company above $200 million, with $35 million raised to date. Samsung Venture Investment led. Signing on as strategic investors were LG Technology Ventures, SKT America, and ZER01NE Ventures (Hyundai Motor’s venture arm) — the investment arms of South Korea’s biggest manufacturers, all on one cap table at once.
Config builds no robots of its own. What it supplies is the data layer beneath robotics foundation models (RFMs).
The TSMC Playbook, Applied to Robot Data
In Config’s telling, its position mirrors TSMC’s in semiconductors: Apple, Nvidia, and AMD have their chips made by TSMC, which never competes with them, and Config likewise plans to be a neutral supplier of robot data to anyone — so a company wanting robot AI of its own never gathers data from scratch.
The company began in January 2025. Its co-founder group holds Seo and COO Jack Bang alongside three others, with histories at Naver, Waymo, and Google. Money is already arriving — paid by customers in agriculture and defense, by system integrators, and by large manufacturers — and the companies chasing the same work include Skild AI, Generalist AI, and Physical Intelligence.
The Price Gap Between Robot Data and Internet Text
CEO Minjoon Seo — a former Meta researcher who later served as chief scientist at TwelveLabs — describes the economics without decoration. LLMs feed on internet text: cheap to get and available in enormous supply. Robotics data cannot be downloaded; it has to be captured in the physical world, with human operators, facilities, and robots all required. That is why developing robot AI costs more than developing chatbots right from the start.
Config films humans carrying out physical tasks in controlled studio settings and out in the field; just under 300 people across Seoul and Hanoi run that production, and the library of human motion data has passed 100,000 hours. AgiBot World, the biggest open-source dataset in this field, holds roughly 3,000 hours — Config’s stock is more than 30 times that size.
Each demonstration hour worth using brings overhead: somebody has to schedule it, supervise it, and annotate it — line items that scraping web text never carries. That imbalance is the reason data, rather than compute, is turning into the resource that gates progress in physical AI.
Fix the Data First; the Model Stays Untouched
The genuine technical split is a matter of sequence. The usual route trains a model on human motion data and only afterward adapts the model to robots. Config turns that order around: before training begins, the data itself is transformed to match the way robots move. Seo borrows his analogy from language learning — English-only materials will not get anyone to Korean.
In his own words: “The data must be converted, not the model. This conversion technology is Config’s core technical differentiator.” Whether the sequence keeps working across robots with different body designs is something only deployments in the real world can decide.
What Drew Korea’s Manufacturers Into the Round
The investor list explains itself. Samsung, Hyundai, LG, and SK all putting money into one data company at once reflects how economies built on manufacturing — Korea, Japan, China, Taiwan — aim to turn their industrial bases into an edge in physical AI. The bet underneath: the big manufacturers eventually want robot AI they control, not total reliance on outside model vendors — and the groundwork for that route is what Config sells.
Also in the round: Mirae Asset Ventures, GS Futures, Z Ventures, Korea Development Bank, and Kakao Ventures, plus Pieter Abbeel, who co-founded Covariant and teaches at UC Berkeley.
Ahead: A Million Hours of Data, and Robots as a Service
The money is pointed at three targets. Collection in Vietnam and Seoul scales toward a million hours. The enterprise platform business is to reach $10 million in ARR before 2027 ends. And a cloud-based robot-as-a-service will let companies execute Config’s foundation model even when the robot itself carries no added hardware.
Robot-as-a-service is the boldest of the three: it steps Config from supplier up to platform, wagering that manufacturers would sooner lease intelligence than build — and staff — a pipeline of their own. Land the million hours and the ARR goal, and Config stands to become the standard first stop for industrial robot learning across Asia.
The meaning for developers and product teams: the race to build robotics foundation models is running the LLM script again, except that data supply has split off into a layer of its own — one that starts earning money earlier. Once 100,000 hours of physical-world data is a product sitting on the market, teams building robot applications no longer have to assemble data themselves, and what separates winners becomes data quality and the technology used to convert it, not model size.