Ubitium Raises $14M to Build One Chip for Physical AI

Düsseldorf-based semiconductor startup Ubitium has raised $14 million in new seed funding to build what it calls a universal processor for Physical AI, bringing its total seed financing to $18 million since the company was founded in 2024. According to Tech.eu, the round was led by Ibex Investors, with participation from OpenOcean, existing investors KBC Focus Fund, Runa Capital and Inflection, and Florian Seibel, co-founder and co-CEO of European defense technology company Quantum Systems, who invested personally.

The announcement, reported on October 13, 2026, arrives as investor attention shifts from the GPUs that train foundation models toward the Physical AI silicon that will let robots, drones, and vehicles perceive and act in the real world. Ubitium's thesis is that autonomous machines face a computing problem data-center chips were never designed to solve: sensing, inference, and control have to happen together, continuously, and in real time.

One architecture for perceiving, thinking, and acting

Physical AI systems need three kinds of computing to work together so they can perceive, think, and act in the real world. Today those functions are typically divided among several types of processors: digital signal processors handle sensor data, neural processing units run AI models, and CPUs manage control. Shuttling data between those chips increases system complexity, latency, and power consumption, according to The AI Insider.

Ubitium's answer is a reconfigurable processor architecture whose programmable processing elements can dynamically allocate computing resources between sensing, AI inference, and real-time control, depending on the machine's workload. The design pairs unified memory with standard RISC-V development tools, so manufacturers can program the chip with existing toolchains rather than learning a proprietary stack.

The company already has hardware evidence. Its first chip, the UB108 with 256 processing elements manufactured on Samsung's 8-nanometer process, demonstrated signal processing, AI inference, and real-time control running on the same reconfigurable array. According to FinSMEs, the UB108 was taped out just a year after the company's founding, and the new funding will advance the first commercial processor, the UB410, which will pack 4,096 processing elements, with engineering samples and evaluation units targeted for 2027.

The bottleneck is not inference

Ubitium chief executive and co-founder Hyun Shin Cho argued in the funding announcement that the industry has been looking at the wrong bottleneck: for machines that move through the physical world, sensing the environment and controlling the machine are just as demanding as running the model, and all three still run on separate chips. The company's pitch is that a single Physical AI processor can do all three, a claim the UB108 was built to prove in silicon within a year.

The founding team brings deep chip-industry credentials. According to FinSMEs, the company holds 17 granted patents covering its core technology with more pending, while the team carries more than 250 career patents in computing, over 350 peer-reviewed publications, and silicon experience from Intel, Nvidia, Apple, and Texas Instruments. Seibel's personal investment from Quantum Systems, a European defense technology company building drones, also signals where the earliest customers may come from: autonomous systems that need the full perceive-think-act loop.

Why Physical AI silicon is attracting capital

The raise is part of a broader capital rotation. While most AI infrastructure money still chases training compute, companies like Ubitium represent an opportunity beyond GPUs, as Physical AI demand grows for processors optimized around latency, sensor integration, energy efficiency, and predictable control rather than maximum model-training performance. Agent infrastructure stacks are maturing in the cloud at the same time the physical layer is getting its own dedicated silicon.

The risk is the one that shadows every chip startup: the gap between a demonstration chip and a commercial product. The UB108 proves the architecture works; the UB410 has to prove it can be manufactured at cost, at volume, and with software support that device makers will trust. Initial applications are targeted at high-resolution radar and audio systems, with longer-term goals in the autonomous systems sector, including drones.

For now, Ubitium has secured the capital to find out whether one reconfigurable chip can replace a stack of specialized ones in Physical AI hardware. As AI funding rounds keep scaling up, the $14 million seed is modest in size but precise in its bet: the next scarce resource in AI is not training compute but the Physical AI silicon that lets machines act on what their models perceive.