Fujitsu will sell its 2nm 144 core MONAKA CPU and MONAKA Server from November 2026, aimed at CPU only AI inference in air cooled data centres rather than liquid cooled GPU halls.
What was announced
- Fujitsu will begin global sales of the FUJITSU-MONAKA CPU in November 2026, both standalone to cloud and data centre operators and server vendors, and inside its own MONAKA Server.
- MONAKA Server will be available from November 2026 to data centre operators, enterprises and the academic and HPC sectors in Japan and Europe.
- Fujitsu quotes a maximum operating frequency of 3.8GHz and memory transfer speed of 8800MT/s, and claims twice the AI inference throughput of other CPUs, without naming the comparison parts or publishing benchmarks.
- The launch line up is a 2U model with two CPUs and 1U models with one or two CPUs, both 144 cores, with 24 RDIMM slots on the 2U and 12 or 24 on the 1U, E3.S and M.2 SSD bays and PCIe Gen6 expansion.
- The 1U air cooled configuration is rated for ambient temperatures up to 40 degrees Celsius, and the liquid cooled version for water temperatures up to 45 degrees Celsius.
- The processor is 3D stacked, with 2nm cores plus 5nm cache and I/O, and implements confidential computing in hardware through Arm CCA.
- A 2U multi node model with four nodes and two CPUs per node is planned for academic and HPC use.
The ETON view
The interesting part of this launch is not the core count, it is the cooling envelope. Most UK colocation halls still hand you 6 to 12 kW a rack, which is why so many AI plans stall at the facilities conversation rather than the hardware one. A dense CPU inference node that runs on air at 40 degrees ambient sits inside the estate you already have, with no CDU, no manifolds and no landlord negotiation.
The caveat is that Fujitsu's efficiency and throughput claims have no published benchmark or named comparison behind them, and November 2026 sales into Japan and Europe means real UK availability, pricing and lead times are unknown for now. Arm also moves the software question to the front: your inference stack, hypervisor and licensing all need to be clean on aarch64 before core count matters.
Our read for anyone sourcing capacity this year: treat MONAKA as a 2027 evaluation, not a plan. If the workload is inference and the constraint is power and space, there is a cheaper test available today with current generation x86 or Arm nodes at known price and lead time, and we will quote both new and refurbished so the comparison is on total cost rather than vendor roadmap.
Related infrastructure
Category: Servers · Vendor: Fujitsu · Technology: FUJITSU-MONAKA, Arm, PCIe Gen6, CXL · Last verified: 16 Sep 2026 · ~2 min read
Sourcing this kind of infrastructure? Talk to ETON about availability, lead time and pricing.
