---
title: "Fujitsu MONAKA: 144-Core 2nm Arm CPU for AI Inference Goes on Sale (Nov 2026)"
description: "Fujitsu MONAKA explained: 144-core 2nm Arm CPU, 3.8GHz, 2x AI inference claim, sales from Nov 2026, servers ship Apr 2027. Specs, benchmarks, roadmap, FAQ."
canonical: https://itsupportintokyo.com/en/blog/fujitsu-monaka-2nm-144-core-cpu-ai-inference-2026
language: en-US
published: 2026-11-18T05:40:00.000Z
modified: 2026-11-18T05:40:00.000Z
author: "IT Support in Tokyo Team"
keywords: "Fujitsu MONAKA, FUJITSU-MONAKA, Fujitsu 2nm CPU, 144-core Arm CPU, AI inference CPU, CPU vs GPU AI inference, sovereign AI server Japan, FugakuNEXT, MONAKA-X"
image: https://itsupportintokyo.com/article-assets/fujitsu-monaka-2026/photo-ai-chip.webp
translations: ["https://itsupportintokyo.com/ja/blog/fujitsu-monaka-2nm-144-core-cpu-ai-inference-2026"]
---

# Fujitsu MONAKA: 144-Core 2nm Arm CPU for AI Inference Goes on Sale (Nov 2026)

> Fujitsu MONAKA explained: 144-core 2nm Arm CPU, 3.8GHz, 2x AI inference claim, sales from Nov 2026, servers ship Apr 2027. Specs, benchmarks, roadmap, FAQ.

**Fujitsu's FUJITSU-MONAKA is a 144-core, 2nm Arm server CPU designed in Japan to run AI inference without a GPU.** On Sep 14, 2026 Fujitsu announced it would start selling the chip worldwide in November 2026, for the first time as a standalone processor, together with the made-in-Japan Fujitsu MONAKA Server for Japan and Europe. Fujitsu says MONAKA delivers twice the AI inference throughput of other CPUs and can halve the servers and power needed for the same work. Here is what MONAKA is, what Fujitsu has actually claimed, when you can get one, and whether CPU-only AI inference makes sense for your company.

**Key facts**
- **What it is:** a next-generation Arm server CPU, designed and developed in Japan. Compute cores on a 2nm process, cache and I/O on 5nm, in a 3D-stacked design.
- **Cores:** 144 Armv9 cores per CPU, per Fujitsu's server specifications as reported by [Phoronix](https://www.phoronix.com/news/Fujitsu-MONAKA-Launches).
- **Speeds:** up to 3.8GHz, memory at 8,800 MT/s, per Fujitsu.
- **Claim:** twice the AI inference throughput of other CPUs, halving server count and power for equal work. Fujitsu has not named the CPUs it compared against.
- **Timing:** sales from November 2026; standalone CPUs supplied from January–March 2027 (Fujitsu's fiscal Q4 2026); servers shipping in Japan and Europe from April 2027.
- **Made in Japan:** servers built at Fujitsu's Kasashima Plant in Ishikawa for traceability; sold for enterprise, academia, HPC and defence.

![A large data-centre AI chip package photographed at a low angle on a dark server board, two mirror-like silicon dies in the centre surrounded by memory stacks on a green substrate, lit with blue and green studio light](/article-assets/fujitsu-monaka-2026/photo-ai-chip.webp)

The kind of large multi-die package now used for AI computing. AI-generated image (ChatGPT), not a photo of FUJITSU-MONAKA or any specific product.

## What is FUJITSU-MONAKA?

MONAKA is Fujitsu's first general-purpose data-centre CPU built on the Arm architecture since A64FX, the processor inside the Fugaku supercomputer. Development began in a NEDO-subsidised project on energy-efficient data-centre CPUs ([Fujitsu, 2022](https://info.archives.global.fujitsu/global/about/resources/news/press-releases/2022/0225-01.html)), and Fujitsu first showed the 144-core design publicly in December 2024 ([Tom's Hardware](https://www.tomshardware.com/pc-components/cpus/fujitsu-flaunts-144-core-monaka-cpu-2nm-and-5nm-chiplets-soic-and-cowos-packaging)). The 2026 announcements turn it into a product: Fujitsu's [Sep 14 press release](https://global.fujitsu/en-global/pr/news/2026/09/14-02) ([日本語](https://global.fujitsu/ja-jp/pr/news/2026/09/14-02)) sets out sales, server models and dates.

The chip combines processes by function. The compute cores use a leading-edge 2nm process, while the cache and I/O sit on 5nm silicon, stacked in 3D. Fujitsu says this improves cost efficiency, and that its ultra-low-voltage design gives high performance per watt. For AI, MONAKA adds dedicated instructions that accelerate matrix operations, combined with Arm SVE2 vector units and software tuning. It also supports confidential computing through Arm CCA, which encrypts data in memory so even the operating system or hypervisor cannot read it.

![Simplified diagram of MONAKA's 3D stack: four 2nm compute dies with Armv9 cores on top, a 5nm SRAM cache layer directly beneath them, a 5nm I/O die handling memory at 8,800 MT/s, PCIe and CXL, all on a package substrate](/article-assets/fujitsu-monaka-2026/stack-en.webp)

Our simplified illustration based on Fujitsu's description and press reports of its Hot Chips 2026 talk. Not to scale.

![Three figures: 2nm compute cores with 5nm cache and I/O, 3.8GHz maximum frequency, and 8,800 MT/s memory transfer rate](/article-assets/fujitsu-monaka-2026/specs-en.webp)

From Fujitsu's press release of Sep 14, 2026.

## What Fujitsu claims for AI inference

Fujitsu's headline claim is "twice the AI inference throughput" of other CPUs, which it says halves the servers and power needed for the same load. The press release does not name the comparison processors. In its results materials for the second quarter of fiscal 2026 ([October 2026, PDF](https://resources.global.fujitsu/prir/2026/10/2026Q2-Financial-Summary-en.pdf)), Fujitsu published more specific numbers: 307.0 tokens per second running the open gpt-oss-20B model at batch size 2, 55.2 tokens per second on Gemma 4 31B, 213.8 TOPS of theoretical INT8 performance per server, 1,000 GB/s memory throughput and 86.8% execution efficiency on an HPC application.

![Three figures published by Fujitsu: 307 tokens per second on gpt-oss-20B at batch size 2, 55.2 tokens per second on Gemma 4 31B, and 213.8 TOPS INT8 theoretical server performance](/article-assets/fujitsu-monaka-2026/perf-en.webp)

Vendor figures from Fujitsu's 2Q FY2026 results materials. No independent benchmarks have been published yet.

These are vendor numbers. No independent benchmarks have been published, and Phoronix noted it hopes to test the chip once access is available. For context, they describe mid-sized open models (20 to 31 billion parameters) at small batch sizes, which is typical of internal chatbots, document search and AI agents rather than serving millions of users.

## CPU or GPU for AI inference?

![Gloved fingertips holding a large square server processor with a blank silver heat spreader and chiplet tiles around its edge over an anti-static mat](/article-assets/fujitsu-monaka-2026/photo-chip.webp)

A large server processor package. AI-generated image (ChatGPT), not a photo of FUJITSU-MONAKA.

GPUs remain the standard for training large models and for serving very large ones to many users. MONAKA's pitch is different: run inference where GPUs are hard to deploy. Fujitsu's 1U server can run air-cooled at room temperatures up to 40°C, or water-cooled at water temperatures up to 45°C, without special cooling infrastructure, and Fujitsu says this cuts server cooling power by up to 80%. That matters for companies whose data cannot leave the building, for older data centres without liquid cooling or high power per rack, and for public-sector and defence users who want a traceable, domestically made supply chain.

## Server lineup, dates and roadmap

![A Japanese data-centre engineer in a navy work jacket slides a slim 1U rack server into an open rack in a bright air-cooled server room with rows of racks and blue status lights](/article-assets/fujitsu-monaka-2026/photo-server-room.webp)

An air-cooled server room, the kind of site Fujitsu targets with its 1U server. AI-generated image (ChatGPT), not a photo of a Fujitsu facility or product.

From November 2026 Fujitsu will sell three Fujitsu MONAKA Server models in Japan and Europe: a 1U server with one or two CPUs that uses CDI/CXL to pool memory and accelerators beyond the box, a 2U server with two CPUs for AI agents and inference, and a 2U multi-node server (four nodes, two CPUs each) for academia and HPC work such as surrogate models and fluid analysis. Phoronix reports 144 cores at 2.1GHz on the air-cooled models and 2.9GHz with liquid cooling, twelve memory modules per socket, and PCIe Gen6 support.

![Timeline from 2026 to 2027: Hot Chips presentation and test servers lent to about 45 companies in late August, sales plan announced on September 14, sales of CPU and servers start in November, CPU supplied to clouds, data centres and server makers from January to March 2027, and server shipments in Japan and Europe from April 2027](/article-assets/fujitsu-monaka-2026/timeline-en.webp)

From Fujitsu's Sep 14 release and 2Q FY2026 results. Spacing is not to scale.

Selected customers in finance, telecoms and manufacturing get servers first, in the second half of fiscal 2026. Shipments to Japan and Europe follow in sequence from April 2027, and standalone CPUs reach cloud and data-centre operators and server makers worldwide, including the US and Asia-Pacific, from Fujitsu's fiscal Q4 2026 (January–March 2027). About 45 companies have been testing loaned servers since the end of August. Fujitsu has also partnered with the French cloud provider Scaleway on MONAKA-based AI inference in Europe ([Dec 2025](https://global.fujitsu/en-global/pr/news/2025/12/04-01)), and began building sovereign AI servers in Japan earlier this year ([Feb 2026](https://global.fujitsu/en-global/pr/news/2026/02/12-01)).

![Six tiles: 1U server with one or two CPUs and CDI/CXL air-cooled to 40°C, 2U server with two CPUs for AI agents and inference, 2U multi-node HPC server with four nodes of two CPUs, MONAKA on 2nm in 2027, MONAKA-X on 1.4nm in 2029 planned for FugakuNEXT, and MONAKA-XX on a newer process in 2031](/article-assets/fujitsu-monaka-2026/lineup-en.webp)

From Fujitsu's Sep 14 release and 2Q FY2026 results. Names only; we do not reproduce Fujitsu's logo.

Fujitsu has also set business targets: ¥35 billion of CPU revenue in fiscal 2027 and ¥250 billion in total over fiscal 2027–2030. Its roadmap shows MONAKA-X on 1.4nm in 2029, planned for the FugakuNEXT supercomputer, and MONAKA-XX on a newer process in 2031.

## Should your company consider it?

![Five questions: is your data too sensitive for a public cloud, do you lack liquid cooling or GPU power budget, are your models mid-sized, can you wait for 2027 deliveries, and will your software run on Arm](/article-assets/fujitsu-monaka-2026/checklist-en.webp)

Our checklist, based on what Fujitsu has announced.

- **Is your data too sensitive for a public cloud?** On-premises, confidential AI inference is MONAKA's main pitch.
- **Do you lack liquid cooling or the power budget for GPUs?** The 1U server runs air-cooled up to 40°C.
- **Are your models mid-sized?** Fujitsu's published tests use 20B–31B-parameter models. Very large models still favour GPUs.
- **Can you wait for 2027 deliveries?** Servers ship from April 2027 in Japan and Europe.
- **Will your software run on Arm?** Most Linux, Python and open-model tooling does, and compiler support landed in GCC 15. Still, check your full stack and any drivers before you buy.

## FAQ

**Is FUJITSU-MONAKA a GPU?** No. It is a general-purpose Arm CPU with extra hardware for AI matrix maths. The point is to run inference without a GPU.

**How many cores does MONAKA have?** 144 Armv9 cores per CPU, according to Fujitsu's server specifications as reported by Phoronix.

**Who makes the 2nm chips?** Fujitsu's releases do not name a foundry. Press reports say TSMC makes the 2nm compute dies; we have not seen Fujitsu confirm this.

**When can I buy one?** Sales start in November 2026. Servers ship in Japan and Europe from April 2027; standalone CPUs reach customers from January–March 2027.

**Planning private AI on your own servers?**

We help Tokyo companies choose between cloud, GPU and CPU-based AI, deploy internal chatbots and document search on-premises, and keep them secure, in English and Japanese. Book a free consultation.

[Book a free consultation](/en/contact)

## Sources

Fujitsu, ["Fujitsu launches made-in-Japan next-generation CPU FUJITSU-MONAKA and Fujitsu MONAKA Server for sovereign AI infrastructure"](https://global.fujitsu/en-global/pr/news/2026/09/14-02), Sep 14, 2026 ([Japanese](https://global.fujitsu/ja-jp/pr/news/2026/09/14-02)); Fujitsu, [Consolidated financial results 2Q FY2026](https://resources.global.fujitsu/prir/2026/10/2026Q2-Financial-Summary-en.pdf), Oct 2026; Fujitsu, [sovereign AI server manufacturing in Japan](https://global.fujitsu/en-global/pr/news/2026/02/12-01), Feb 12, 2026; Fujitsu, [partnership with Scaleway](https://global.fujitsu/en-global/pr/news/2025/12/04-01), Dec 4, 2025; Fujitsu, [NEDO energy-efficient CPU project](https://info.archives.global.fujitsu/global/about/resources/news/press-releases/2022/0225-01.html), Feb 25, 2022; Phoronix, Michael Larabel, ["Fujitsu Formally Announces Their MONAKA 144 Core CPU"](https://www.phoronix.com/news/Fujitsu-MONAKA-Launches), Sep 17, 2026; Tom's Hardware, Anton Shilov, [report of Dec 11, 2024](https://www.tomshardware.com/pc-components/cpus/fujitsu-flaunts-144-core-monaka-cpu-2nm-and-5nm-chiplets-soic-and-cowos-packaging). Graphics are our own and use names only. Photo-style images were generated with ChatGPT and do not show FUJITSU-MONAKA, Fujitsu servers or any other specific product. IT Support in Tokyo is not affiliated with Fujitsu. FUJITSU-MONAKA is a trademark of Fujitsu Limited; Arm is a trademark of Arm Limited.

Related: [More tech news](/en/news) · [Nvidia board and Jensen Huang](/en/blog/nvidia-will-not-reinstate-jensen-huang-ceo-nov-17-2026) · [Cloudflare CEO on AI agent traffic](/en/blog/cloudflare-ceo-ai-agent-traffic-website-hosting-costs-1000x)

## Source

Canonical HTML: https://itsupportintokyo.com/en/blog/fujitsu-monaka-2nm-144-core-cpu-ai-inference-2026
Other language: https://itsupportintokyo.com/ja/blog/fujitsu-monaka-2nm-144-core-cpu-ai-inference-2026