Fujitsu Monaka is finally available to order. Fujitsu said Monday that global sales of its 144-core Arm server CPU start in November, both as a bare chip for cloud operators and server vendors and inside a Fujitsu Monaka Server built at the company’s Kasashima plant in Ishikawa. The servers will be sold in Japan and Europe to data center operators, enterprises, universities and HPC centers, and, notably, to defense customers. Kyodo’s framing was blunter than the press release: Fujitsu says European buyers want AI hardware that’s reliable from a security standpoint, which is a polite way of saying not American and not Chinese.

A quick reality check on “November,” because the headline version flattens it. November is when Fujitsu starts taking orders. According to the Japanese release, the chips themselves ship between January and March 2027, and the servers ship from April 2027, with broader deliveries in Japan and Europe rolling out from there. About 30 companies in finance, telecom, and manufacturing across Japan and Europe already have systems for trials. So this is a sales launch with a real ship date attached, not hardware you’ll see in racks before Christmas. That still beats what Fujitsu implied at Hot Chips in August, where volume production was pointed loosely at 2027.

The silicon is the part I find clever, and it’s worth slowing down on. Monaka is a CPU, not an accelerator: 144 Armv9.3-A cores per socket, up to 288 in a two-socket node, 12 channels of DDR5 at up to 8,800 MT/s, PCIe 6.0 with CXL 3.0, and a top frequency of 3.8 GHz. What makes it unusual is how it’s built. The 144 cores sit on four 36-core compute chiplets made on TSMC’s N2P 2nm process. Each of those is hybrid-bonded face-to-face onto its own SRAM die made on TSMC’s 5nm N5, and that SRAM die holds the entire last-level cache. A separate 5nm I/O die handles memory and PCIe and talks to the stacks through a silicon interposer. Fujitsu uses Broadcom’s 3.5D XDSiP packaging to pull it off.

That split reflects the logic every chip team is fighting with right now. SRAM barely shrinks at new nodes anymore, so spending 2nm wafer area on cache is paying premium prices for almost no density gain. Fujitsu keeps 2nm silicon under 30% of total die area and pushes everything that scales poorly onto cheaper, mature 5nm. The compute die sits on top, closest to the cooler, because it runs hottest. It’s a cousin of AMD’s cache-stacked X3D approach, but closer in spirit to Intel’s Clearwater Forest, where the cache lives in a base tile underneath the compute. I’ve watched Qualcomm, and others, make similar area-versus-cost trades on mobile for twenty years; seeing a server CPU restructure itself this aggressively around the SRAM problem is the most interesting design decision in this launch.

For AI specifically, Monaka leans on the CPU tricks rather than pretending to be a GPU. Each core carries two 256-bit SVE2 vector units, narrower than the 512-bit SVE in the A64FX that powered Fugaku, which is a deliberate step away from pure HPC toward general data center work. Fujitsu adds dedicated matrix instructions with FP8 and INT8 support plus tuned libraries, and claims roughly twice the throughput of competing CPUs. There will be a 350 W air-cooled part and a 500 W liquid-cooled one. Those are vendor numbers, and nobody has run independent production benchmarks yet, so I’m treating them as targets, not results. I also couldn’t find a source for the “less data than an Nvidia stack” line that’s been circulating with this story, so I’m leaving it out.

CPU inference has a hard ceiling, though. Large-model decoding lives and dies on memory bandwidth, and 12 channels of DDR5-8800 works out, on my back-of-the-envelope math, to roughly 845 GB/s of theoretical bandwidth per socket. That’s respectable for a server CPU and still well short of what HBM-equipped accelerators move. A Monaka box isn’t going to serve a frontier model to thousands of users. It can run small and mid-sized models, retrieval pipelines, agent orchestration, and all the pre- and post-processing around a model on hardware that fits in an ordinary rack. That’s where the server design earns its keep: the 1U system is rated to run air-cooled at ambient temperatures up to 40°C, or on 45°C water, and Fujitsu claims it can cut cooling power by up to 80%. The 1U model also supports CDI and CXL, so memory and accelerators can be pooled beyond the box, which is Fujitsu’s answer to the memory squeeze that makes inference painful on any CPU.

That’s the market Fujitsu is aiming at, and I think it’s smarter than a head-on fight. Many European hospitals, ministries, regional data centers, and mid-sized banks sit in buildings never designed for 100 kW liquid-cooled GPU racks. They still want to run AI locally, and after two years of arguments about the CLOUD Act, they want to know where the hardware came from. I’ve argued that Europe’s sovereign AI debate has mostly been about cloud contracts. Fujitsu is betting it becomes a procurement question about boxes, with a traceability system that records where each component came from and its manufacturing history. Kyodo reported Fujitsu is in talks with about 40 companies about adopting that system.

The sovereignty pitch deserves some scrutiny, though. “Made in Japan” here means designed, integrated, and assembled in Japan. The 2nm and 5nm dies come from TSMC in Taiwan. The packaging technology is Broadcom’s, an American company. The instruction set is Arm’s, a British company owned by Japan’s SoftBank. Japan’s own 2nm effort, Rapidus, isn’t in this supply chain. None of that makes Monaka a bad product. It does mean “not American, not Chinese” really translates to “a Japanese vendor with a transparent, allied supply chain,” which, for most European buyers, is exactly the point. Sovereignty in 2026 is about control and traceability far more than about owning every fab, and I’d rather vendors said that plainly.

The competition is crowded. Nvidia’s Grace is the Arm server CPU everyone benchmarks against, and it arrives married to Nvidia’s GPUs. AWS runs Graviton at a scale nobody else can match, but you can only rent it. Ampere’s AmpereOne is the merchant Arm alternative, now under SoftBank, which already owns Arm. Intel’s Xeon 6 with AMX and AMD’s EPYC dominate the x86 installed base, and both handle CPU inference well enough that “just use what you have” is Monaka’s biggest real rival. In Europe, SiPearl’s Rhea is the homegrown Arm option and already shows up in EuroHPC AI factory designs, the same program I looked at when EuroHPC’s newest machine went to Bull and AMD instead of Nvidia. Fujitsu’s advantage over SiPearl is scale, a working 2nm part and a finished server line. SiPearl’s advantage is that it’s actually European, which matters to some buyers more than a transparent supply chain does.

Fujitsu’s roadmap has a twist I don’t think got enough attention. Its next part, Monaka-X, is reportedly aimed at around 2029 on a 1.4nm process with NVLink Fusion support, so it can sit next to Nvidia accelerators on Nvidia’s interconnect. That’s the same interconnect play I wrote about when Nvidia put $3.5 billion into MediaTek. Fujitsu is selling Monaka as an alternative to the American stack today while planning to plug into it tomorrow. I don’t read that as hypocrisy. It’s what a realistic CPU vendor does, and it’s a hint that the “non-American” pitch has a shelf life.

For corporations, the calculation is fairly simple. European enterprises with AI workloads that don’t justify a GPU cluster get another option that won’t require a cooling retrofit, from a vendor with a long history of European services. Server makers get a new Arm part to build around if they want one. Telecom operators, where Fujitsu’s partner 1FINITY already works, are a natural early fit. Defense buyers are the wildcard: Fujitsu listing the defense sector explicitly is new language for a Japanese chip launch, and it lines up with how quickly military AI procurement is growing on both sides of the Atlantic.

The financing story is mostly a government story. Monaka is backed by Japan’s NEDO under a green data center program that targets 40% energy savings by 2030, and it inherits R&D from Fugaku, which was itself a state project. That’s a very different funding model from the venture rounds lighting up European AI hardware, or from the private money that just went into Mistral’s €3 billion round. It gives Fujitsu patience a startup can’t afford. It also means the pressure to find paying customers outside Japan is real, because public programs fund development, not a sales channel. Europe’s own answer, between the EU Chips Act, EuroHPC and the startups it funds, is still spread thin across too many bets. A Japanese vendor showing up with a finished server and a defense-friendly story is going to make that fragmentation more visible in procurement meetings, not less.

I’m leaving Monaka’s role in FugakuNEXT, Japan’s next flagship supercomputer, for another day. It’s a big part of why the chip exists, but it’s a different audience.

So does a CPU have a real shot in an AI market built around GPUs? For the workloads Fujitsu is targeting, yes, and I like that it isn’t pretending otherwise. The test comes in the first half of 2027, when shipping systems meet independent benchmarks and European buyers see a price. If Monaka Servers land in a few ministries and banks by the end of next year, Fujitsu will have proven that sovereign AI can be sold as a box. If they don’t, it’ll be another beautifully engineered chip that Europe admired from a distance.

Sources