On May 11, 2026, Config, a startup headquartered in Seoul and San Jose, announced an oversubscribed $27 million seed round at a valuation above $200 million, bringing total raised to $35 million. Samsung Venture Investment led, with strategic checks from ZER01NE Ventures (Hyundai Motor’s venture arm), LG Technology Ventures, and SKT America — the venture arms of South Korea’s largest manufacturers, arriving together.
Config does not build robots. It builds the data layer for robotics foundation models (RFMs).
The TSMC of Robot Data
Config positions itself as the TSMC of robot data: TSMC fabricates chips for Apple, Nvidia, and AMD without competing with them, and Config wants to be the neutral data supplier for everyone else’s robot AI — so any company building proprietary robot intelligence does not have to collect from zero.
Founded in January 2025, the company’s co-founders include Seo, COO Jack Bang, and three others with backgrounds at Waymo, Google, and Naver. Config already generates revenue from large manufacturers, system integrators, and customers in agriculture and defense; named peers include Physical Intelligence, Generalist AI, and Skild AI.
Why Robot Data Costs More Than Internet Text
CEO Minjoon Seo — previously a researcher at Meta and chief scientist at TwelveLabs — puts it plainly: LLM training data is internet text, cheap and abundant. Robotics data has to be physically gathered, which means robots, facilities, and human operators. Robot AI is structurally more expensive to develop than chatbots.
Config’s approach is recording humans performing physical tasks, both in controlled studio environments and in the field. Nearly 300 staff across Seoul and Hanoi handle data production, and the company has accumulated more than 100,000 hours of human motion data — over 30 times AgiBot World, the largest comparable open-source dataset at roughly 3,000 hours.
Every usable hour of demonstration time requires scheduling, supervision, and annotation — costs with no equivalent in scraping web text. That asymmetry is why data, not compute, is emerging as the gating resource in physical AI.
Convert the Data, Not the Model
The technical differentiation is sequencing. Most teams train models on human motion data and then adapt the model to robots. Config inverts it: the data is transformed before training to fit how robots move. Seo’s analogy is language pedagogy — you cannot teach Korean using only English-language materials.
His words: “The data must be converted, not the model. This conversion technology is Config’s core technical differentiator.” Whether that ordering holds up across different robot morphologies is exactly the kind of claim only real deployments can test.
Why Korea’s Manufacturers Stepped In
The cap table is the signal. Samsung, Hyundai, LG, and SK investing simultaneously in a data company reflects how manufacturing-heavy economies — Korea, Japan, China, Taiwan — are converting industrial bases into physical AI advantage. The underlying assumption: large manufacturers will want proprietary robot AI rather than depending entirely on outside model vendors, and Config sells the foundation for that path.
Other backers include Pieter Abbeel, Covariant co-founder and UC Berkeley professor, plus Mirae Asset Ventures, Korea Development Bank, GS Futures, Kakao Ventures, and Z Ventures.
Next: A Million Hours and Robot-as-a-Service
The funding has three jobs: scale data collection in Vietnam and Seoul toward one million hours; grow the enterprise platform business to $10 million ARR by the end of 2027; and launch a cloud-based robot-as-a-service offering that lets companies run Config’s foundation model without onboard hardware.
Robot-as-a-service is the aggressive one: it moves Config up the stack from supplier to platform, betting that manufacturers would rather rent intelligence than staff the pipeline to build it. If the million-hour target and the ARR goal both land, Config becomes the default on-ramp for industrial robot learning in Asia — and a natural partner, or target, for the same conglomerates funding it now.
For developers and product teams, the significance is this: the robotics foundation model race is replaying the LLM playbook, but “data supply” has become a separate, earlier-monetizing layer. When 100,000 hours of physical data becomes something you can purchase, robot application teams no longer collect from zero — differentiation shifts to data quality and conversion technology, not model size.
Sources
- Korea’s biggest manufacturers back Config, the TSMC of robot data — TechCrunch
- Config
- AgiBot World — Hugging Face
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.