Fujitsu is turning its Fugaku supercomputer work into a 2-nanometer CPU business, and the sales pitch is blunt: banks, telecoms and governments want AI hardware they can trust, not only GPUs they can chase.
Fujitsu announced on September 14 that it will start global sales of FUJITSU-MONAKA, a 2-nanometer Arm server CPU developed in Japan, from November 2026. Beating Nvidia at its own game was never the plan. Fujitsu is giving buyers another place to look when the work is AI inference, the day-to-day running of trained models, rather than the expensive job of training them from scratch.
That distinction matters. MONAKA is a 144-core server processor, not a GPU, and Fujitsu is pitching it at sovereign AI infrastructure for customers that care as much about traceability and control as they do about raw benchmark bragging rights. In its own announcement, Fujitsu said the chip reaches a maximum operating frequency of 3.8GHz, supports memory transfer speeds of 8800MT/s, and delivers twice the AI inference throughput of other CPUs by using dedicated matrix-operation instructions alongside Arm SVE2 vector operations.
The chip is not trying to be Nvidia #
The more interesting part sits under the lid. As Tom's Hardware reported from Hot Chips 2026, MONAKA uses a 2-nanometer compute die on TSMC's N2P process, with the last-level cache and input-output work split onto 5-nanometer dies. Fujitsu's own release says the server is manufactured at its Kasashima Plant, which is different from saying every piece of silicon is fabricated in Japan. That split matters. The product is Japan-developed and Japan-assembled into Fujitsu's sovereign server stack, but the most advanced manufacturing still runs through TSMC.
Fujitsu has pulled this design from its high-performance computing playbook. Fugaku, once the world's fastest supercomputer, ran on Fujitsu's Arm-based A64FX processor, and MONAKA is being positioned as the next commercial step from that lineage. It's specific rather than magical: keep the expensive 2-nanometer area focused on compute, move cache and I/O to 5-nanometer silicon, then stack the pieces so the chip can feed AI inference workloads without turning every deployment into a GPU hunt.
TSMC's August Revenue Just Broke NT$500 Billion for the First Time Ever TSMC's August revenue hit a record NT$514.81 billion, or about $16.3 billion, up 53.3% year-over-year and its fourth straight monthly record. The surge is driven by relentless AI chip demand and an unusually fast 2nm ramp tied to Apple's iPhone 18. - TSMC August revenue breaks NT$500 billion record - AI chip demand drives TSMC monthly sales growth
Investors noticed. MoneyDJ reported that Fujitsu shares rose 5.01% to 3,937 yen in morning Tokyo trading on September 14 after Japanese media said the company would export MONAKA for AI use. That reaction sat mainly in Fujitsu's own stock. The market wasn't repricing every chip name in Tokyo, it was treating MONAKA as a company-specific catalyst.
Fujitsu is chasing 250 billion yen in cumulative external CPU revenue through fiscal 2030, based on Japanese media reports around the launch. That isn't giant next to Nvidia's data center business. It is still a real target. If you run a bank in Frankfurt, a telecom network in Singapore, or a ministry system in Tokyo, a smaller number can matter when it comes with a hardware path that doesn't depend on the same GPU queue as everyone else.
The timing helps Fujitsu #
Japan's industrial mood was already turning before the product announcement landed. A Reuters Tankan survey published last week showed Japanese manufacturer sentiment rising to +21 in September from +18 in August, the strongest reading since December 2021. The electronics sub-index jumped to +39 from +24, helped by semiconductor and data center demand. The poll ran from August 26 to September 4. It drew responses from 224 of 510 firms.
Frankly, the chip war has mostly been written as a story about who gets Nvidia GPUs and at what price. MONAKA asks a colder question. Do you want every important AI inference workload sitting on hardware where a foreign export rule, a supply squeeze, or one vendor's roadmap can decide your capacity for you?
Fujitsu is leaning hard into that answer. The company says the MONAKA Server will be offered broadly from November 2026 to data center operators, enterprises, academia and HPC customers in Japan and Europe, as well as the defense sector, and then to select customers in finance, telecommunications and manufacturing in the second half of fiscal 2026 to expand use cases. Shipments in Japan and Europe are set to start sequentially from April 2027, while the standalone CPU is aimed at cloud providers, data center operators and server vendors globally, including the US and Asia-Pacific.
None of this makes Nvidia easy to dislodge. CUDA took years to become the default software layer for AI developers, and CPUs still don't replace GPUs for the heaviest training jobs. But inference is narrower, more repetitive and closer to the customer. If you need a fraud model to run inside a regulated bank or an AI agent to sit near domestic data, the perfect benchmark number may be less important than who controls the box.
That is Fujitsu's opening. MONAKA doesn't need to win the whole AI hardware market to matter. It needs enough cautious, well-funded customers to decide that sovereign compute is no longer a white-paper phrase and start buying servers accordingly.
Arm CEO Rene Haas Says Chip Shortages Are Slowing Down AI Cancer Research Arm CEO Rene Haas told the BBC that AI could eventually crack cancer by modeling how DNA markers drive tumors, but said the breakthrough is being held back by a shortage of chips needed to build AI data centers. He also predicted humanoid robots doing repair and manufacturing work within a decade and downplayed fears of mass AI job losses. - chip shortage slowing down AI cancer research breakthroughs - how chip shortages are impacting AI medical research
Also read: Cohere's Aidan Gomez Calls Anthropic and OpenAI's AI Slowdown a Cartel Play • Roblox stock jumps 11% after RDC 2026 unveils AI creation tools • Apple's iPhone Duo Foldable Costs $1,999 and Ships October 23
This article is posted in Technology News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.