EdgeCortix SAKURA-II AI Accelerator and MERA

EdgeCortix is a fabless edge AI company founded in July 2019 in Japan, with its headquarters and lab in Kawasaki, Kanagawa, and offices in Hyderabad, Singapore and Virginia. It designs energy-efficient AI processors and software so that robots, drones, telecom equipment, defence systems and smart infrastructure can run large AI models locally.

The SAKURA-II accelerator, built on the company's run-time reconfigurable Dynamic Neural Accelerator architecture, provides 60 TOPS INT8 and 30 TFLOPS BF16 at about 8 W, with up to 32 GB of LPDDR4x and 20 MB of on-chip SRAM. It ships as an M.2 2280 module or as low-profile PCIe cards of up to 120 TOPS and handles models such as Llama 2, Stable Diffusion, DETR and ViT at batch size 1. The MERA compiler and framework, plus the MERA Studio desktop tool, compile, run and benchmark models across heterogeneous systems; Renesas has added MERA to its RZ/V MPUs. Customers include Kawasaki Heavy Industries and US defence programmes.

• SAKURA-II accelerator: 60 TOPS INT8 and 30 TFLOPS BF16 at about 8 W typical power.
• Large memory: dual LPDDR4x channels with 8, 16 or 32 GB and 68 GB/s bandwidth.
• Form factors: M.2 2280 modules and single or dual low-profile PCIe cards up to 120 TOPS.
• Generative AI support: runs Llama 2, Stable Diffusion, DETR and ViT models at the edge.
• MERA platform: compiler and framework for inference across heterogeneous processors.
• MERA Studio: desktop GUI for compiling, running, benchmarking and demonstrating models.

A Vietnamese telecom operator can add SAKURA-II PCIe cards to edge servers for AI services close to users. A robotics or inspection equipment maker can run vision and language models on an M.2 module inside the machine. A smart-city integrator can deploy roadside units that analyse traffic video locally.

Runs multi-billion parameter generative models at around 8 W, so language and vision models can be deployed in power-limited robots, drones and roadside equipment. Up to 90% compute utilisation, as stated by the company, gives more usable performance from each chip. M.2 and PCIe form factors drop into existing industrial PCs and servers. MERA Studio lets engineers compile and benchmark models without command-line work.

Generative AI is moving into robots, drones, telecom base stations and infrastructure that cannot rely on cloud connectivity. Japan is funding domestic low-power AI chips through programmes such as NEDO, and Asian industrial and defence buyers seek alternatives to GPUs, while Southeast Asian telecom and smart-city projects need compact accelerators for local inference.

Founded in July 2019, headquartered in Kawasaki, Japan, with offices in India, Singapore and the US. Closed an oversubscribed Series B in November 2025, bringing total funding above 110 million USD, with investors including TDK Ventures, CDIB, Jane Street, Yanmar Ventures, NTT Finance and SBI Investment. Also has a NEDO-backed project and a Mizuho credit facility. NASA validated SAKURA-II for orbital and lunar missions.