
14 Sep 2026
Fujitsu launches Japan-made MONAKA CPU and servers aimed at sovereign AI inference
Fujitsu said it will start global sales of FUJITSU-MONAKA, a Japan-designed processor, in November 2026, plus Japan-made servers aimed at sovereign AI inference — running already-trained models on hardware buyers can trace.
While U.S. labs argue about pacing models, a major Japanese vendor is putting a made-at-home CPU and server line on the calendar for AI inference — for buyers who care about power, air-cooled rooms, and supply-chain traceability.
Fujitsu Limited, the Japanese tech company, said via JCN Newswire that it will begin global sales of FUJITSU-MONAKA, a next-generation CPU designed and developed in Japan, starting November 2026. A CPU is the main processor chip that runs a computer. The company will also sell the chip on its own to cloud and data-center operators and other server vendors. That Fujitsu announcement, dated from Kawasaki, is the filing event. This desk is not saying the chips have shipped.
The company will also launch the Fujitsu MONAKA Server, made in Japan at Fujitsu’s Kasashima Plant, from November 2026. Fujitsu says it will be offered to data-center operators, enterprises, and the academic and HPC sectors — high-performance computing, the big scientific machines — in Japan and Europe, and to the defense sector. File those as Fujitsu’s sales plan, not as orders already booked.
Fujitsu claims a top clock of 3.8 GHz — how fast the chip can tick — and memory transfers of 8800 MT/s, meaning 8,800 million data moves a second. It says the chip delivers about twice the AI inference throughput of other CPUs, and better power efficiency. Inference is the everyday step: running an already-trained model to answer a question or make a prediction, not the costly training step. Those speed and power lines are Fujitsu’s. This desk did not run a bake-off.
Fujitsu says the chip uses a 3D-stacked design: a 2-nanometer process for the cores and a 5-nanometer process for cache and input/output. A nanometer is a billionth of a meter; smaller process numbers usually pack more work into less power. 3D-stacked means layers of silicon sit on top of each other. The company says dedicated matrix-operation hardware speeds inference. It also says the chip supports confidential computing — hardware encryption of data in memory — using Arm CCA. How that stack is wired is in Sources.
Fujitsu says the first rack models are 1U and 2U — one or two standard 1.75-inch slices of a data-center rack. It says the 1U box can run on air cooling in rooms as hot as 40°C, and that its cooling tech can cut server cooling power by up to 80%. Those are Fujitsu’s figures. File them as company claims, not as an independent energy audit.
Later on Fujitsu’s calendar: broader CPU availability in the fourth quarter of fiscal year 2026, including the United States and Asia-Pacific. Select customers in finance, telecom, and manufacturing get servers in the second half of FY2026. Shipments then start from April 2027 in Japan and Europe. File the later dates as Fujitsu’s plan. This is not a claim that boxes are already leaving the plant.
Fujitsu also mentions a later vertical stack: the MONAKA server plus its Kozuchi AI platform and Takane enterprise generative model, aimed at “sovereign AI.” Sovereign AI, as Fujitsu uses it here, means keeping the chips, servers, and data on a path a country or customer can trace and run itself. Treat Kozuchi and Takane as a roadmap mention on this release — not a separate product launch today.
Fujitsu’s footnote says the new MONAKA technology is based on work subsidized by NEDO, Japan’s New Energy and Industrial Technology Development Organization. Park that as the company’s subsidy note.
CONFIRMED here: Fujitsu Limited’s 14 Sep 2026 announcement via JCN Newswire. NOT claimed: independent benchmarks, a price list, that shipping has started, or a second newsroom confirmation. Distinct from the already-filed dmatrix-nvlink-fusion-raptor, positron-875-million-series-c, and mistral-x-cloudera — this filing is Fujitsu’s Japan-made CPU and server calendar only.
RELATED
- NVIDIA blog: d-Matrix to connect Raptor XPUs via NVLink Fusion
- Positron raises $875 million to tape out inference silicon
- Mistral partners with Cloudera for sovereign enterprise AI
- Google puts €13 billion into Finnish AI infrastructure
- Amodei: a misaligned swarm could take over the entire internet in 6–12 months