Fujitsu MONAKA: A 2nm Arm CPU Built for a Global AI Server Push

Fujitsu’s MONAKA program is an attempt to re-establish the company as an international data-center processor and AI-infrastructure supplier. Rather than build and operate a leading-edge fab, Fujitsu is designing an Arm-based server CPU, outsourcing production to TSMC, and planning to sell complete AI servers and supporting development and operations platforms alongside the chip. Sales are reported to begin from 2027 in Japan, the United States and Asia-Pacific. 1718

What MONAKA is

MONAKA is a server processor aimed at AI, high-performance computing and data-center workloads. Fujitsu’s published material describes an Armv9-A design with 144 cores, Arm SVE2 256-bit vector support, and configurations of up to two sockets per node. 15

The design reflects Fujitsu’s supercomputing heritage, but its commercial goal is broader: to turn that CPU expertise into a platform for enterprise and data-center deployments. Fujitsu has said its processor design combines a 2nm-class process, its own microarchitecture and low-voltage technologies to pursue strong power performance. 30

A 2nm compute die paired with 5nm support dies

MONAKA uses a disaggregated, 3D-stacked chiplet approach rather than putting every function on the most advanced manufacturing node. The compute die is made on TSMC’s N2P 2nm process, while SRAM and I/O are placed on separate 5nm dies. The entire last-level cache sits on the dedicated SRAM die beneath the compute die. 713

That arrangement matters because leading-edge logic area is expensive. By reserving 2nm silicon for compute and using 5nm silicon for cache and I/O, Fujitsu is aiming to balance performance, power and manufacturing cost rather than maximize the amount of silicon fabricated on the newest node. 67

Two models for different data-center constraints

Fujitsu has outlined two MONAKA versions, both with 144 cores:

• High-efficiency model: 350W, air-cooled, with a reported 2.1GHz base clock.
• High-performance model: 500W, liquid-cooled, with a reported 2.9GHz base clock. 716

The split is commercially important. Air cooling can make the lower-power model easier to deploy in facilities with limited cooling capacity, while the liquid-cooled model is intended for customers willing to support a higher-power configuration in exchange for more performance. Volume production is planned for 2027. 7

Fujitsu is selling infrastructure, not just a CPU

The go-to-market plan is centered on integrated systems. Reporting says Fujitsu intends to sell MONAKA-based AI servers as well as platforms for AI development and operations, rather than relying only on standalone processor sales. 17

The company’s reported commercial targets are ambitious: 60,000 AI servers and ¥250 billion in cumulative CPU-related revenue by fiscal 2030. Those are company goals, not booked revenue. 1718

Fujitsu has reportedly secured supply to Super Micro Computer and is seeking major U.S. cloud providers as customers. Separately, 24 companies in fields including finance and telecommunications are said to be testing equipment, while Fujitsu is discussing potential adoption with more than 40 additional companies. These figures indicate an active pipeline, but they should not be read as confirmed production deployments or sales commitments. 1819

The longer-term Rapidus option

MONAKA itself relies on TSMC for fabrication. Fujitsu’s longer-range plan, however, points to a possible 1.4nm-class successor, MONAKA-X, around 2029, with Rapidus discussed as the prospective manufacturing partner. 1718

That plan is inherently less certain than MONAKA’s TSMC production path. It depends on Rapidus reaching commercially viable leading-edge manufacturing and on Fujitsu turning the proposed successor into a competitive product. For now, it is best understood as a strategic supply-chain objective rather than a settled production outcome. 27

Why MONAKA matters

MONAKA is not simply a new CPU launch. It is Fujitsu’s effort to operate as a fabless, systems-led provider in a data-center market dominated by established x86 suppliers and increasingly shaped by AI power and cooling constraints.

Its differentiation rests on three connected claims: a 144-core Arm design, a packaging approach that limits the use of 2nm silicon to compute dies, and a server-plus-platform sales model. Whether that translates into meaningful global share will depend on real-world performance per watt, software compatibility, customer validation, TSMC supply, and Fujitsu’s ability to convert its testing pipeline into deployments. The 2027 launch will be the first major measure of whether this supercomputer-derived architecture can become a durable commercial data-center business. 717