Japan’s AI chip push shifts to edge and Physical AI

Japan’s AI semiconductor sector is emerging as a set of specialized efforts rather than a single attempt to recreate NVIDIA’s GPU model. EdgeCortix’s RAIDEN leads the current attention with a chiplet platform for Physical AI, combining DNA-X acceleration, RISC-V cores, MERA 3.0 software and scale-out links for robots, aerospace, defense and edge servers.

RAIDEN X4 is listed at up to 3.36 PFLOPS of FP4 compute with 2:4 structured sparsity, up to 256GB LPDDR5X and 548GB/s memory bandwidth, but power, process, foundry and price remain undisclosed. Customer samples are scheduled for Q1 2027 and mass production for the second half of 2027, so claims that it has surpassed NVIDIA should be limited to nominal Sparse FP4 peak comparisons with Jetson AGX Thor, not real-world model performance or efficiency.

Other domestic players are targeting different bottlenecks: LENZO is developing low-power CGLA/TokenProcessor technology from NAIST; Tokyo Artisan Intelligence is moving from the Sting Ray test chip toward Manta Ray; Preferred Networks is pursuing MN-Core L with 3D-stacked DRAM; ArchiTek, DMP, Floadia, Renesas, TIER IV, Socionext, Sony, Fujitsu and others are focused on embedded vision, reconfigurable computing, Compute-in-Memory, custom chiplets and automotive software-defined systems.

The key test for 2027 will be whether these roadmaps move from tape-out, samples and design wins into measured performance, customer evaluation and volume shipment. LeapMind’s dissolution in 2024 underscores that strong IP is not enough without software support, capital, manufacturing access and sustained demand.