Fujitsu MONAKA: A "Made-in-Japan" AI CPU That TSMC Fabricates

On September 14, 2026, Fujitsu announced FUJITSU-MONAKA, a next-generation CPU it will sell globally from November 2026, alongside the Fujitsu MONAKA Server positioned as "sovereign AI infrastructure." The Hacker News thread hit 240 points, and within an hour the top comments had converged on the same question: where is this chip actually made?

The answer is Taiwan. That does not make MONAKA uninteresting. It makes the marketing more interesting than the engineering, and the engineering is genuinely worth a look.

TL;DR

The sovereignty claim, read carefully

The headline says "Japan-developed 2nm 3D-stacked CPU and server integrated, developed, and manufactured in Japan." Commenters immediately split on how to parse it, and the ambiguity is doing real work. Read one way it means the CPU and the server are both made in Japan. Read the other way, the CPU is developed in Japan and the server is integrated, developed and manufactured there.

The second reading is the correct one. Reporting on MONAKA's manufacturing is consistent: the core die is produced on TSMC's N2P node, with SRAM and I/O dies on TSMC N5. Rapidus, Japan's domestic 2nm effort, is a future consideration rather than the current supplier, and is targeting its own 2nm mass production in 2027.

The chiplet split is a smart engineering decision independent of the politics. The leading-edge N2P node covers less than a third of the total silicon area, with the bulk on mature N5, which cuts cost and accelerates time to market. Fujitsu detailed the architecture at Hot Chips 2026, including the entire cache sitting on a separate 5nm die and a narrowing to 256-bit SVE2, with 350W and 500W SKUs.

Why this matters beyond pedantry. One commenter put the point precisely: the entire purpose of "sovereign" is to avoid being cut off if relations with a supplier country deteriorate. If the silicon is fabbed in Taiwan, the supply-chain exposure that sovereignty is meant to address is unchanged. What Japan gains is design capability, assembly, traceability of the finished server, and no dependence on a US or Chinese vendor. Those are real, and they are not the same as making the chip.

It is worth noting Arm itself is owned by SoftBank, a Japanese company, which gives the ISA dependency a different colour than it would otherwise have.

The engineering, which deserves better than the marketing

Strip the sovereignty framing and there is a coherent product thesis: AI inference that fits where GPUs cannot go.

Matrix acceleration on a CPU. MONAKA provides hardware acceleration for matrix operations through dedicated instructions, combined with SVE2 vector operations and software optimisation, for a claimed 2x AI inference throughput against other CPUs. Fujitsu's pitch is that this makes CPU-only inference practical, so you can deploy at the scale you need in sites with power and cooling constraints.

The cooling numbers are the real product. The 1U server operates in ambient temperatures up to 40C on air, and up to 45C water temperature on water cooling, without specialised cooling infrastructure, reducing server cooling power by up to 80%. For anyone who has tried to put accelerators into an existing colocation facility, this is the constraint that actually binds. Liquid-cooling retrofits are slow, expensive and sometimes impossible in a leased hall.

Memory pooling via CXL. The 1U uses CDI/CXL to allocate memory and accelerators beyond the server's physical boundary, aimed at the memory shortages that throttle inference.

Confidential computing. Arm CCA hardware encryption protects data and applications in memory from privileged software including the OS and hypervisor, which is the credible part of the "sovereign" story: it addresses multi-tenant trust rather than geography.

The specs Fujitsu leads with are 3.8GHz peak and 8800MT/s memory transfer. Note that 3.8GHz is not remarkable on its own; a Pentium 4 reached it in 2004, as one commenter dryly observed. Clock speed is not where this chip's argument lives.

What the announcement does not tell you

The gaps in the release are as informative as the claims, and several drew pointed criticism:

• No tokens per second on any named model. The most useful comment in the thread argued for exactly this: pick a specific open-weight model, publish TPS, and let buyers compare. Without it, "2x AI inference throughput vs other CPUs" is unfalsifiable, because which CPUs, which model, and which quantisation all move the number enormously.
• No memory channel count or SRAM capacity. One commenter estimated roughly 845 GB/s from 8800MT/s across twelve DDR5 channels, against about 1008 GB/s for an RTX 4090 and 8000 GB/s for a B200. Treat that as an unverified back-of-envelope, since the channel count is not in the release, but the order of magnitude is the point: this is memory-bandwidth-constrained relative to GPUs, which is exactly what you would expect and is why the pitch is deployment flexibility, not peak throughput.
• Arm is barely credited. The architecture is inferable only from the SVE2 mention. Fujitsu designs its own microarchitecture on a standard ISA, a pattern it followed with SPARC, and it was a principal architect behind Arm SVE, so this is not a licensing sleight so much as an odd omission.

Who this is actually for

The go-to-market tells you the intended buyer more clearly than the press release does. Sales run to data centre operators, enterprises, academia and HPC in Japan and Europe, plus the defence sector, with early access for select customers in financial services, telecommunications and manufacturing.

That is a procurement-driven market, not a performance-driven one. These buyers have constraints that make an air-cooled, traceable, non-US, non-Chinese inference platform worth paying for even at a throughput disadvantage:

• A bank that cannot send inference off-premises for regulatory reasons
• A defence buyer with supply-chain provenance requirements
• A European public institution under data-residency rules
• An HPC centre with a fixed power envelope and no liquid cooling

For these, "good enough inference, in a rack you already have, from a supplier your procurement office will approve" beats "faster inference you cannot deploy." That is a real market, and it is the same logic driving sovereign AI programmes across Europe and Asia.

Fujitsu is also stacking a vertically integrated story on top: its Kozuchi AI platform and Takane enterprise generative AI, plus industry-specific models, all running on domestic infrastructure.

The wider pattern: everyone's sovereign chip runs through Taiwan

MONAKA is not an outlier, and singling Fujitsu out would miss the point. Every national-champion silicon programme currently faces the same wall: there is exactly one company producing leading-edge logic at volume, and it is in Taiwan. Design sovereignty, packaging sovereignty and assembly sovereignty are all achievable today. Fabrication sovereignty at 2nm is not, for anyone outside TSMC's own fabs.

That makes "sovereign" a spectrum rather than a binary, and it is worth being explicit about which rungs a given programme has actually climbed:

Five of seven is a genuinely strong position, and considerably more than most buyers get. The problem is that the one missing rung is the one that a geopolitical disruption would actually sever, which is why the marketing reaches for a word the supply chain has not earned yet.

Fujitsu has a credible path to closing it. It is reportedly working with Rapidus on a future 1.4nm part with fabrication handled domestically, and Rapidus is pricing aggressively against TSMC to win exactly this kind of business. If that lands, the sovereignty claim stops being a stretch. The honest framing for MONAKA today is "sovereign except for the silicon," which is less catchy and more accurate.

For buyers, the practical question is narrower than the rhetoric. Ask what failure you are insuring against. If it is a US export-control action against your jurisdiction, a Japanese-designed Arm CPU assembled in Japan genuinely helps. If it is a Taiwan Strait contingency, it does not help at all, because the dependency is identical to everyone else's. Those are different risks and they deserve different procurement answers.

The honest scorecard

Credible: the cooling and power-efficiency claims, the chiplet design decision, Arm CCA confidential computing, CXL memory pooling, and the procurement-driven market thesis. These are specific and checkable.

Unverified: every performance number, because no benchmark methodology, model or comparison baseline was published. Fujitsu did present in more detail at Hot Chips 2026 than many vendors do, which is to its credit, but the press release is not where that detail lives.

Overstated: the sovereignty framing. Designed in Japan, assembled in Japan, fabbed in Taiwan. If Rapidus reaches volume production and MONAKA moves there, the claim becomes true. Today it is aspirational, and the wording is careful enough to suggest Fujitsu knows it.

None of this is unusual. Nearly every "national champion" chip project is a design-and-assembly story sitting on top of a TSMC dependency, because there is currently no alternative at the leading edge. The useful posture is to buy MONAKA for its cooling envelope and its procurement profile, and to discount the sovereignty language until the fab moves.

Related on explainx.ai

• How to read AI benchmarks — the method for discounting unverifiable vendor performance claims
• AI token costs and enterprise governance — why per-watt and per-rack economics increasingly decide deployments
• GLM-5.3 built its own inference stack — another sovereign-hardware inference effort, on Chinese accelerators
• CoreWeave's first Vera Rubin cluster — the GPU-side deployment MONAKA is implicitly competing against
• Nebius Madrid AI hub — European sovereign-adjacent capacity buildout

All specifications and claims are from Fujitsu's September 14, 2026 press release unless otherwise noted. Fabrication details come from reporting on MONAKA's TSMC N2P and N5 process usage and from Fujitsu's Hot Chips 2026 presentation; Fujitsu's own launch release does not name a foundry. Performance figures are vendor-stated and not independently verified. The memory-bandwidth estimate cited is a commenter's calculation from an unpublished channel count, not a specification.