⏱️ Lectura: 10 min
Starting in November 2026, Fujitsu will sell a 2-nanometer CPU capable of doubling AI inference performance without relying on a GPU. The Japanese company calls it FUJITSU-MONAKA and pairs it with a server built entirely at its Kasashima plant, designed for sovereign AI infrastructure in Japan and Europe.
📑 En este artículo
The announcement comes amid a global shortage of AI compute and growing demand for platforms that companies can operate within their own borders, without depending on foreign infrastructure.
TL;DR
- Fujitsu announces FUJITSU-MONAKA, a 2nm CPU designed and manufactured in Japan, going on sale globally from November 2026.
- Maximum frequency of 3.8 GHz and memory transfer rate of 8800 MT/s.
- 3D-stacked architecture: compute core on a 2nm process, cache and I/O on 5nm, with Fujitsu’s own ultra-low voltage technology.
- Fujitsu says it doubles AI inference performance and cuts the number of servers needed in half.
- Fujitsu MONAKA Server ships in air-cooled 1U and 2U formats, available in Japan and Europe from November 2026.
- Manufacturing happens entirely at Fujitsu’s Kasashima plant, reinforcing supply chain traceability.
- Includes confidential computing with hardware-based encryption, built for multi-tenant cloud environments.
- The 1U server uses CXL/CDI technology to pool resources across nodes.
Introduction
On September 14, 2026, Fujitsu officially announced the global launch of FUJITSU-MONAKA, its next-generation CPU designed entirely in Japan. This is no minor announcement: Fujitsu is betting that a CPU, without an additional GPU, can sustain real AI inference workloads in data centers constrained by power and space.
What happened
Fujitsu will sell FUJITSU-MONAKA as a standalone product starting in November 2026, aimed at cloud operators, data centers, and server manufacturers looking to integrate the chip into their own equipment. Alongside it, the company is launching the Fujitsu MONAKA Server, a complete server with the CPU already integrated, aimed at enterprises, academia, high-performance computing (HPC), and also the defense sector.
Initial availability covers Japan and Europe. Fujitsu explicitly describes the goal as sovereign AI infrastructure: hardware, manufacturing, and supply chain under national or regional control, rather than depending on external suppliers.
Background and history
Fujitsu isn’t new to this space. The company had already developed the A64FX, the ARM CPU that powered Fugaku, one of the fastest supercomputers in the world for several years. FUJITSU-MONAKA inherits that line of high-performance ARM processor design, but shifts the target: instead of optimizing for pure scientific computing, it now prioritizes AI model inference.
The backdrop behind this launch is broader than a single product. The growth of generative AI has driven demand for compute and electricity at a pace that already causes management problems in several countries. At the same time, many companies need to process confidential information within their own infrastructure, without uploading it to a third-party cloud, and without relying on specialized facilities like liquid cooling. Add to that the geopolitical risk: governments and companies increasingly seek platforms whose supply chain they can verify and control.
Technical details and performance
FUJITSU-MONAKA uses a 3D-stacked architecture that combines two different manufacturing processes in the same package: 2 nanometers for the compute core and 5 nanometers for the cache and input/output (I/O) sections. Fujitsu justifies this mix as a way to balance cost and energy efficiency, rather than manufacturing the entire chip on the most expensive available node.
The chip adds Fujitsu’s own ultra-low voltage operation technology, which, according to the company, is what allows it to sustain high performance without spiking power consumption. The figures the company publishes are a maximum frequency of 3.8 GHz and a memory transfer speed of 8800 MT/s, both described as category leaders.
For AI inference specifically, FUJITSU-MONAKA includes hardware acceleration for matrix operations through dedicated instructions, combined with SVE2 (Scalable Vector Extension 2, ARM’s vector extension) vector operations and software optimization. That combination, according to Fujitsu, achieves double the inference performance compared to other CPUs, to the point that the company says the chip alone, without a GPU, already reaches practical inference performance.
The chip also integrates confidential computing (confidential computing), which hardware-encrypts data and applications while they are in active memory, not just at rest or in transit. This is the piece designed so multiple customers can share the same cloud without exposing their data to each other.
💭 Key point: confidential computing protects data even while it’s being processed in memory, not just when it’s stored or traveling over the network. This differs from traditional encryption at rest or in transit.
At the server level, Fujitsu offers two formats starting in November 2026:
| Format | CPUs per node | Key technology | Best suited for |
|---|---|---|---|
| 1U | 1 or 2 | CXL/CDI for resource pooling | High density and future scalability by pooling memory/compute across nodes |
| 2U | 2 | Standard air cooling | Direct deployment in data centers without water or liquid cooling infrastructure |
The 1U server is the one that adopts CXL/CDI, the technology that allows pooling memory and compute across different nodes as if they were a single shared resource, designed to scale without buying entirely new servers every time more capacity is needed.
Architecture diagram
flowchart TD
A["Compute core (2nm)"] --> B["Cache and I/O (5nm)"]
B --> C["Memory at 8800 MT/s"]
C --> D["Fujitsu MONAKA Server 1U/2U"]
D --> E[("Resource pooling via CXL/CDI")]
subgraph Server
D
E
endHow to get started or try it
The hardware won’t arrive until November 2026, but an infrastructure team can start preparing its stack today, since FUJITSU-MONAKA follows the ARM64 architecture and uses SVE2, an extension that already runs on Linux on other ARM CPUs available in the cloud. Checking whether your current CPU supports SVE2 is the first step:
$ lscpu | grep -i sve
Flags: ... fp asimd ... sve sve2 svebf16 svei8mm ...
$ cat /proc/cpuinfo | grep -m1 "Features"If the sve2 flag shows up, a binary compiled with those extensions will be able to take advantage of the same kind of vector acceleration that FUJITSU-MONAKA uses for inference. The next practical step is compiling an inference engine with those flags enabled, for example llama.cpp:
$ git clone https://github.com/ggml-org/llama.cpp
$ cmake -B build -DGGML_CPU_ARM_ARCH=armv9-a+sve2
$ cmake --build build --config Release -j$(nproc)
$ ./build/bin/llama-bench -m model.gguf -t $(nproc)That command compiles the runtime with explicit SVE2 support and runs a local tokens-per-second benchmark. Once Fujitsu’s hardware becomes available, the same build pipeline can be used to measure the chip’s real performance against current infrastructure, without relying on marketing figures.
💡 Tip: if your inference workload already runs on ARM CPUs with SVE2 (for example, Graviton or Ampere instances in the cloud), porting and benchmarking it on FUJITSU-MONAKA will take minimal effort once the chip hits the market.
Impact and analysis
Fujitsu’s bet connects to a real tension in the industry: expanding AI infrastructure costs energy and space, two resources that don’t grow at the same pace as demand. A chip that cuts the number of servers needed for the same workload in half, according to Fujitsu’s published figures, directly attacks that bottleneck, especially in locations without access to industrial liquid cooling.
The other axis is technological sovereignty. Fujitsu isn’t just selling a chip: it manufactures the entire server at its Kasashima plant, which lets it document the origin of every component and its manufacturing history. For governments, banks, and the defense sector, that traceability matters as much as raw performance, because it reduces dependence on supply chains they can’t audit end to end.
Fujitsu also plans to integrate the MONAKA Server as the foundation of a vertically integrated sovereign AI model, combining its Fujitsu Kozuchi platform and the Takane enterprise model with industry-specific models. The idea is for the customer to control not just the hardware, but also the traceability and security of the AI layer running on top of it.
The limitation Fujitsu doesn’t solve with this announcement is the software ecosystem: a CPU with dedicated AI instructions only performs well if inference frameworks (llama.cpp, PyTorch, ONNX Runtime, etc.) are optimized for those specific instructions. Migrating workloads that currently run on GPUs to a CPU-based architecture, no matter how fast, involves engineering rework that doesn’t show up in any press release.
⚠️ Heads up: the figures claiming Fujitsu doubles inference performance and cuts servers in half are the ones the company itself published in its announcement. Until independent benchmarks exist on real hardware in third-party hands, it’s best to treat them as the manufacturer’s starting point, not a verified result.
What’s next
Fujitsu confirmed global sales of FUJITSU-MONAKA as a standalone chip starting in November 2026, along with the availability of the MONAKA Server in Japan and Europe for data centers, enterprises, academia, HPC, and defense. The company has not yet published pricing or an expansion timeline to regions beyond those two initial markets. The next real reference point will come when the first independent benchmarks appear, typically 2 to 6 months after the hardware reaches customers.
Try it yourself: if you work with AI inference on ARM64, run lscpu | grep sve2 on your current infrastructure today to find out if you’re already one step away from taking advantage of the same optimization path FUJITSU-MONAKA uses.
📖 Summary on Telegram: View summary
Frequently Asked Questions
What is FUJITSU-MONAKA?
It’s Fujitsu’s new CPU, designed and manufactured in Japan, built for AI inference. It uses a 3D-stacked architecture with a 2nm core and 5nm cache/I/O, reaching a maximum frequency of 3.8 GHz with 8800 MT/s memory transfer.
When will it be available?
Fujitsu confirmed global sales of the chip as a standalone product, and the launch of the Fujitsu MONAKA Server, starting in November 2026.
How does it differ from using a GPU for AI?
Fujitsu states that FUJITSU-MONAKA delivers practical inference performance using only the CPU, thanks to dedicated instructions for matrix operations and SVE2. That makes it possible to deploy AI in locations with power or space constraints, without needing a dedicated GPU.
What is the Fujitsu MONAKA Server?
It’s the complete server that integrates the FUJITSU-MONAKA CPU, manufactured entirely at Fujitsu’s Kasashima plant. It comes in 1U (one or two CPUs, with CXL/CDI) and 2U (two CPUs) formats, both air-cooled.
Which markets does it cover first?
Initial availability, starting in November 2026, is for Japan and Europe, focused on data centers, enterprises, academia, HPC, and the defense sector.
What is the confidential computing it includes?
It’s hardware-based encryption that protects data and applications while they’re actively being processed in memory, not just when stored or in transit over the network. It’s designed for cloud environments where multiple customers share the same infrastructure.
References
- Fujitsu Global: official launch announcement for FUJITSU-MONAKA and the Fujitsu MONAKA Server.
- ARM Developer: documentation on the ARM architecture and the SVE2 vector extension used in the chip.
- Compute Express Link Consortium: CXL specification, the pooling technology used by the 1U server.
- Wikipedia: definition and general context on confidential computing.
📱 Enjoying this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day. @programacion
Imagen destacada: Foto de Brian Kostiuk en Unsplash
0 Comments