Rebellions Rebel100 NPU adopted by ai& for 100+ racks, 40MW

The announcement marks the first large-scale deployment of Rebellions' NPUs in Japan, coming about six years after the Korean startup was founded. Rebellions positions its NPU as specialized for AI inference, claiming it matches GPU performance while operating with lower power consumption and cost. The ai& deal is being presented as part of that push, with ai& itself founded in March of this year to provide full-stack AI data center services focused on Japanese data sovereignty.

ai& already operates three AI data centers in Osaka and Tokyo, with another Tokyo site about to come online, and it offers a mix of NVIDIA GPUs, AMD GPUs, and Tenstorrent CPUs. Adding Rebellions' NPUs expands that menu of processors so customers can choose the best fit for their purpose. The Rebel100 chip itself uses a chiplet design with four chips on a package, manufactured on Samsung's 4nm process, with 144GB of HBM3e memory. Marshall Choi said the water-cooling requirement is not absolute, as air cooling can just barely handle it, which opens the possibility of installing in existing Japanese data centers that lack water-cooling support.

Rebellions and ai& are framing this as a step toward strengthening Japan's domestic AI compute capacity. The outcome will hinge on whether the planned racks are actually delivered and powered on schedule to reach the 40MW across the five sites, and on whether customers adopt the open-source software stack as intended. For now, the deal gives Rebellions a reference customer in Japan and gives ai& another lever to differentiate its full-stack offering.