Ubitium Raises $14 Million to Develop Universal Processor for Physical AI

Ubitium announced on October 13, 2026, that it raised $14 million in new seed funding to develop a “universal processor” for Physical AI. Ibex Investors led the round, joined by OpenOcean, existing investors KBC Focus Fund, Runa Capital and Inflection, as well as Florian Seibel, co-founder and co-CEO of Quantum Systems, as an individual investor. The funding brings the company’s total seed funding since its founding in 2024 to $18 million.

Ubitium’s announcement addresses a key design challenge for Physical AI semiconductors: how far the separate roles of DSPs, NPUs and CPUs can be consolidated. According to the company, its first chip will be manufactured on Samsung’s 8nm process, and the UB410 is planned to have 4,096 processing elements. Manufacturing conditions and simpler integration will be important considerations alongside performance. The company’s choice of high-resolution radar and audio systems as initial customers suggests that deployment will begin in the perception layer rather than with general-purpose applications. In robots and drones, power, weight, latency and component-count constraints are stringent in addition to computing demands. The announcement raises the question of how much of those constraints one chip can absorb.

The novelty in this announcement is not the $14 million in funding itself, but that the money will go toward commercializing a design that shifts processing to a single reconfigurable processor, rather than building on the existing approach of combining DSPs, NPUs and CPUs. Ubitium says it has already demonstrated with the UB108 — which has 256 processing elements and is built on Samsung’s 8nm process — the simultaneous operation of signal processing, AI inference and real-time control. The focus now appears to have shifted from research-stage claims to the commercial UB410, with 4,096 processing elements, and the planned delivery of evaluation units in 2027. A related article from this newspaper, Amazon Unveils Physical AI Toolchain on AWS, Adopts NVIDIA’s Physical AI Stack, showed how cloud-based tools are being developed to support work on real-world applications. Ubitium differs in that it is seeking to consolidate the computing platform inside machines onto one chip, rather than improving the software development environment. In other words, the Physical AI value chain is seeing separate efforts to build out upstream development environments and hardware platforms at the edge. In terms of architecture, Ubitium is promoting reconfiguration on a common architecture, in contrast with conventional designs that separate perception, inference and control. This could matter for applications with substantial signal-processing workloads, such as high-resolution radar and audio. For drones and autonomous machines with tight power, weight and latency constraints, the key question is how much the approach can reduce data transfers between chips and integration burdens. Other unresolved questions include whether the UB410’s 4,096 processing elements will be sufficient for real-world requirements, how many developers will be able to test it when evaluation units are delivered, and how the company will manage costs and supply based on manufacturing on Samsung’s 8nm process. Another question is whether the company’s “single processor” can truly replace all computing in Physical AI. The announcement identifies initial customers in applications serving the perception layer, pointing to an initial focus on sensor-related markets. The next points to watch are how much of perception, inference and control can be integrated onto one piece of silicon, and which functions will be implemented when evaluation units are delivered.

Ubitium aims to combine processing that is often divided among DSPs, NPUs and CPUs into a single processor for machines that perceive, think and act. The company says its first silicon was taped out one year after its founding and demonstrated signal processing, AI inference and real-time control running on the same architecture. According to primary information, the first chip, the UB108, has 256 processing elements and was manufactured on Samsung’s 8nm process. The UB410, planned as the company’s first commercial product, is designed to have 4,096 processing elements, with evaluation units scheduled for delivery in 2027. Initial customers are expected to include applications in the perception layer, such as high-resolution radar and audio systems. The company says the intended end applications include autonomous systems such as drones.

Primary information says the UB108 was manufactured on Samsung’s 8nm process. For Japanese companies, the issue is therefore more likely to be the combination of sensors, audio and edge components designed around a reconfigurable computing architecture than process design itself. However, the announcement does not directly mention Japanese companies or the Japanese market, and its specific impact on Japan remains unclear.