Fujitsu's MONAKA CPU: 2nm, 3.8GHz, Claims 2x AI Inference Throughput with Half the Server Count
Fujitsu starts global sales of FUJITSU-MONAKA in November 2026. The chip goes to market as both a standalone processor sold directly to cloud operators and server vendors, and as the core of the Fujitsu MONAKA Server, a complete made-in-Japan AI inference appliance.
The technical claims are specific. The CPU runs on a 2nm 3D-stacked process, operates at a maximum frequency of 3.8GHz, and delivers memory bandwidth of 8800MT/s. Fujitsu says the chip provides twice the AI inference throughput of competing CPUs and halves the server count and power draw needed for equivalent workloads. The server package is designed for air-cooled data centres, bypassing the liquid cooling requirements that are constraining expansion of GPU-heavy AI infrastructure.
Sovereign positioning is central
The chip is designed and developed in Japan; silicon is fabricated by TSMC on its 2nm process, with server assembly at Fujitsu’s Kasashima plant. Fujitsu frames this as a sovereign AI offering, targeting customers for whom supply chain traceability and hardware provenance matter: cloud and data centre operators, enterprises, and academic and HPC sectors in Japan and Europe. The defence sector in both regions is also a listed market.
Fujitsu pairs the hardware with its own software stack: the Kozuchi AI platform and the Takane enterprise generative AI model. The intent is a vertically integrated sovereign AI deployment where chip design, model origin, and operational stack are all under a single Japanese vendor.
Why this matters
Japan’s government-backed AI infrastructure push has so far focused on GPU clusters, most visibly the joint project with NVIDIA announced in July 2026 for 27,500 Rubin GPUs at 140MW. MONAKA runs a parallel track: domestic silicon optimised for inference density rather than training throughput, deployable in standard air-cooled facilities without specialist cooling infrastructure.
CPU-only AI inference servers have struggled to displace GPU-centric architectures for large-scale workloads, but the density and power efficiency argument is real at the inference layer. If Fujitsu’s 2x throughput and 0.5x server count claims hold under independent testing, MONAKA becomes a credible option for organisations running large inference workloads who cannot or will not route through NVIDIA or non-Japanese supply chains.
Pricing has not been announced.