A $14 million seed round is modest by 2026 semiconductor standards. Ubitium's claim is not. The Düsseldorf startup says one reconfigurable RISC-V array can replace the three-chip stack — digital signal processor, neural processing unit, CPU — that almost every autonomous machine carries today.
The company has working first silicon. Its UB108 chip, built on Samsung's 8nm process, packs 256 processing elements and has run signal processing, AI inference and real-time control on the same fabric. The new round, announced October 13, brings total seed funding to $18 million since Ubitium was founded in 2024.
Three jobs, three chips
Every machine that acts in the physical world does three things at once. It senses, it infers and it controls. Today that work is split across a DSP, an NPU and a CPU. Data moves between the three. Latency accumulates. Power and weight rise. On a drone, a warehouse robot or an inspection vehicle, all three are rationed.
Physical AI machines bring a whole new set of computing challenges, and the inference isn't the hardest one. Sensing the real world and controlling the machine in it are just as hard, and today those still run on separate chips.— Hyun Shin Cho, CEO and co-founder, Ubitium
Ubitium's argument is that the split is an artifact of chip history, not a physical requirement. A radar front end and a motor controller both need deterministic timing. An inference workload needs throughput per watt. One array of programmable elements can serve both if the allocation changes at runtime.
A fabric that rewrites itself
The design is a coarse-grained reconfigurable array with unified memory and a standard RISC-V toolchain. Resources shift between sensing, inference and control as the machine's workload changes. Developers target the chip the way they target any RISC-V processor, so the software stack does not need a proprietary compiler.
Ubitium holds 17 granted patents on the core technology. CTO and co-founder Martin Vorbach has worked on reconfigurable computing for more than three decades; the company says its team holds over 250 career patents and silicon experience from Intel, Nvidia, Apple and Texas Instruments.
The trade-off is real. Fixed-function silicon wins on peak efficiency, and a reconfigurable array has to prove that flexibility does not cost too much power or determinism. The 256-element UB108 shows that the concept runs. It does not show that it scales.
$14M seed led by Ibex Investors
256 processing elements in the UB108 test chip
4,096 processing elements planned for the UB410
2027 target for tape-out and evaluation units
The 2027 test
The new capital funds the UB410: a 4,096-element commercial part, a tape-out and evaluation units in 2027. First customers sit in radar and audio, where signal processing and AI already share a board. Drones and robots come after the full perceive-think-act loop.
As we wrote in late September, when SiMa.ai raised its Series C, the edge-inference fight is moving from data-center GPUs to silicon that has to work under tight power budgets. Ubitium is a smaller, earlier bet on the same shift.
Florian Seibel, co-CEO of aerial-robotics company Quantum Systems, invested personally. His stated rationale is weight and integration time: every extra chip costs both, and an autonomous machine has neither to spare.
The $14 million does not validate the architecture. It buys the year in which the UB410 has to.