Fujitsu to sell 2nm Japan-designed MONAKA CPU and server for sovereign AI from November 2026
- Fujitsu will begin global sales of the FUJITSU-MONAKA CPU in November 2026, offering the chip standalone to cloud and data center operators and server vendors as well as inside its own MONAKA Server.
- The CPU stacks a 2nm core with a 5nm cache and I/O section, runs at up to 3.8GHz with 8800MT/s memory, and adds hardware matrix instructions plus SVE2; Fujitsu says this doubles AI inference throughput versus other CPUs and halves the servers and power needed for the same load.
- The MONAKA Server is designed, developed, and manufactured in Japan at Fujitsu's Kasashima Plant and ships from November 2026 to data centers, enterprises, academia and HPC, and the defense sector in Japan and Europe, with air-cooled deployment and confidential computing in memory.
- The release never names the Arm architecture; commenters point to the SVE2 mention and to a February Fujitsu release saying TSMC would fabricate the 2nm part because Rapidus could not meet the deadline.
- Fujitsu pairs the server with its Kozuchi AI platform and Takane generative AI plus industry-specific models to sell a vertically integrated sovereign AI stack.
Hacker News opinions
This thing is Arm based, they just barely mention it anywhere in the release.
That's what I was hunting for. Mostly I want to know what their toolchain story looks like.
The lack of clarity on the architecture is annoying, and hiding RAM capacity behind slot info doesn't help either. Claiming a 2nm node in Japan smells like marketing, I'd bet the 2nm die is TSMC and only the package is Japanese.
The only reason I could tell it's Arm is the SVE2 mention. That puts it at Armv9 at the earliest.
They don't say it in this article at all, I had to look elsewhere. Weird for an announcement this big, Arm's licensing people would want that front and center.
Fujitsu did the same thing back in the SPARC days, enhanced a CPU on someone else's ABI. They'd rather build on a popular architecture than invent their own.
Calling it sovereign infrastructure while shipping a licensed processor design is a bit much.
Curious how well it actually performs outside Asia.
And whether they'll even sell it there, given the sovereign AI pitch.
Performance should be the same wherever you run the CPU.
Fujitsu used to run their own fabs until they sold them to UMC, and those are stuck around 20nm. So this is probably made at JASM.
It's 2nm, so it has to be TSMC, nobody else has that capacity.
Where are these actually fabbed? The wording is careful, 'designed and developed in Japan' for the CPU but 'domestic manufacturing' only for the server.
A February press release already said TSMC handles the 2nm because Rapidus can't meet the deadline, and Rapidus itself says its 2nm mass production starts in 2027.
Big emphasis on AI inference for a company that doesn't build GPUs.
A GPU is basically matrix math now and this CPU does that in hardware, so the pitch holds up. Memory bandwidth looks low though.
The part I actually like is the lower environmental impact. Liquid cooling is optional and the CPU is the headline instead of a 1000W GPU.
What I want is tokens per second, not GHz and bandwidth. We need one open weight model everyone benchmarks against.
There's a pretty direct link between bandwidth and TPS, you can predict one from the other.
8800 MT/s at 8 bytes across 12 DDR5 channels works out to 845 GB/s. A 4090 is 1008 GB/s and a B200 is 8000, is that the right way to read it?