The Data Foundry: How Config is Becoming the TSMC of the Robotics Revolution

The global race to achieve artificial general intelligence (AGI) has largely been defined by digital prowess—large language models (LLMs) that can compose poetry, code software, and summarize legal briefs. However, a new, more tangible frontier is emerging: physical AI. As Asia’s industrial giants pivot toward automation to maintain their competitive edge, the bottleneck for the next generation of robotics is no longer just the hardware, but the "data layer" required to make those machines think, react, and operate in the real world.

Enter Config, a Seoul- and San Jose-based startup that is quietly positioning itself as the indispensable foundation for robotics. By securing $27 million in an oversubscribed seed round led by Samsung Venture Investment, the company has signaled a structural shift in how robotics AI is being built, financed, and deployed.

The Structural Shift: Manufacturing Meets Intelligence

Asia has long been the world’s manufacturing powerhouse, with South Korea, Japan, Taiwan, and China serving as the bedrock of global supply chains. Unlike the West, which often prioritizes service-oriented or software-first economies, these nations are deeply rooted in large-scale production. This structural foundation is now influencing the trajectory of AI investment.

For industrial titans like Samsung, Hyundai, and LG, the future is not merely about optimizing assembly lines with traditional automation; it is about deploying intelligent, foundation-model-driven robots capable of handling complex, unstructured tasks. However, these companies are wary of relying on black-box software from third-party vendors. They want proprietary AI that understands their specific manufacturing environments.

Config has tapped into this demand. By positioning itself as the "TSMC of robotics," the startup is not trying to build the robots themselves. Instead, it is building the high-quality, specialized data infrastructure that allows manufacturers to train their own bespoke AI models. Just as TSMC manufactures chips for Apple, Nvidia, and AMD without competing with them, Config provides the underlying data intelligence that powers a diverse array of robotic systems.

Chronology of a Data-First Strategy

Config was established in January 2025 by a team of industry veterans led by CEO Minjoon Seo, a former researcher at Meta and chief scientist at TwelveLabs. Alongside three co-founders with high-level experience at Waymo, Google, and Naver, the team identified a fundamental flaw in the prevailing approach to robotics: the "garbage in, garbage out" problem.

• January 2025: Config is founded with a mission to bridge the gap between human intent and robotic execution.
• Early 2025 – Present: The startup establishes a significant operational footprint in Seoul and Hanoi, hiring a workforce of nearly 300 individuals dedicated to the granular task of data production.
• May 2026: Config closes an oversubscribed $27 million seed round. The round, which values the company at over $200 million, brings its total funding to $35 million.
• Strategic Expansion: The company begins signing major industrial players, including system integrators and firms in the agriculture and defense sectors, as its inaugural customer base.

The Economics of Physical AI: Why Data is the Barrier

To understand why Config is attracting such massive institutional backing, one must look at the economics of robotics. Training an LLM is expensive, but the "raw material"—text from the internet—is virtually infinite and easily accessible.

In contrast, robotics AI is defined by extreme scarcity. "Every piece of training data has to be physically collected," explains CEO Minjoon Seo. "You need the robot, the facility to run it, and people to operate it."

This physical requirement creates a massive cost barrier. As companies push for more capable, autonomous robots, the expenses associated with gathering, labeling, and cleaning motion data balloon. Config’s business model centers on lowering this barrier. By centralizing the data collection process in low-cost, high-efficiency environments in Vietnam and South Korea, the startup creates economies of scale that individual robot manufacturers cannot replicate on their own.

The Technical Differentiator: Translation Over Training

Most robotics teams currently train AI models on raw human motion data and hope for a seamless transfer to a robotic frame. Config contends that this is fundamentally flawed.

COO and co-founder Jack Bang suggests that the industry is treating robotics data like a generic input, whereas it should be treated like a specialized language. If you try to teach a student Korean using only English textbooks, the result will be a poor understanding of the language. Similarly, applying raw human motion data to a non-humanoid or differently articulated robot often leads to "model drift" and inefficiency.

Config’s core differentiator is its "conversion technology." Rather than trying to force a model to learn from raw data, Config transforms the data before the training begins. By normalizing and translating human physical motion into a format that is mathematically optimized for robotic actuators and sensory inputs, they ensure that the foundation models are far more efficient and capable from the moment training begins.

Supporting Data: By the Numbers

The scale of Config’s data operations is arguably its most significant competitive moat. To date, the startup has accumulated over 100,000 hours of high-quality human motion data. To put this in perspective, it is more than 30 times the size of AgiBot World, the largest comparable open-source dataset, which sits at approximately 3,000 hours.

The company’s roadmap is equally aggressive:

• 1 Million Hours: The current funding round is specifically earmarked to scale operations in Vietnam and Seoul to reach a milestone of 1 million hours of collected data.
• $10 Million ARR: The leadership team is targeting $10 million in Annual Recurring Revenue (ARR) by the end of 2027.
• Robot-as-a-Service (RaaS): Config is developing a cloud-based platform that allows companies to utilize its foundation models without requiring expensive, heavy-duty onboard hardware, further lowering the barrier to entry for smaller manufacturers.

Strategic Backing and Official Responses

The participation of South Korea’s biggest industrial players—Samsung, Hyundai (via ZER01NE Ventures), LG, and SKT—is a loud endorsement of Config’s thesis. These companies are not merely financial investors; they are strategic partners who represent the exact market Config is targeting.

Other investors, such as UC Berkeley professor and Covariant co-founder Pieter Abbeel, provide the academic and technical validation required to compete in the high-stakes world of robotics research. Financial heavyweights like Mirae Asset, Korea Development Bank, and Kakao Ventures round out a cap table that is as stable as it is ambitious.

In exclusive comments, leadership emphasized that Config’s goal is to remain a neutral party. By refusing to enter the hardware manufacturing market, they maintain the trust of their customers, positioning themselves as a vital, horizontal layer in a vertical-heavy industry.

Implications for the Future of Automation

The rise of Config signifies a broader maturation of the robotics sector. We are moving away from the era of "bespoke, fragile automation" and into the era of "general-purpose physical intelligence."

If Config succeeds in its mission, the implications for global manufacturing are profound. First, it could lead to a massive democratization of robotics. By providing the data "fuel," Config allows mid-sized enterprises to build AI capabilities that were previously reserved for the world’s most well-funded tech giants.

Second, it shifts the focus of the AI arms race. While much of the world remains fixated on the next iteration of ChatGPT or Claude, the real-world value of AI will increasingly be measured by its ability to navigate a warehouse floor, harvest a crop, or weld a chassis with human-like dexterity.

However, challenges remain. The company must prove that its "conversion technology" can scale across vastly different hardware architectures. Furthermore, they will face stiff competition from other emerging players in the physical AI space, such as Physical Intelligence, Generalist AI, and Skild AI.

Ultimately, Config is betting that in a world of increasingly complex machines, the most valuable commodity is not the robot, but the intelligence that guides it. As they race toward their goal of 1 million hours of motion data, the industry is watching closely to see if the "TSMC of robotics" can indeed build the foundation for the next industrial revolution.