Japan's AI Chip Bet: Can Mokama Compete With US and Chinese Rivals

The Chip No One Is Really Watching — Yet

Fujitsu launched its FUJITSU-MONAKA CPU in November 2026, and almost no one outside Japan is paying attention. That may be its greatest risk. The chip is being sold alongside servers manufactured at Fujitsu’s Kashima factory in Ishikawa Prefecture, targeting cloud providers, data center operators, and server vendors with domestic supply chains in mind. The company is pitching this as the cornerstone of a 3 trillion yen AI business by fiscal 2035.

The timing, however, is brutal. By late 2026, the AI inference chip market was already saturated with NVIDIA H-series and Blackwell systems dominating enterprise deployments, AMD MI300X expanding its foothold, and Apple’s M-series proving that ARM-based inference workloads could run efficiently on custom silicon. Intel was pushing Gaudi for generative AI training and inference, while Google’s TPUs had quietly become the backbone of its own cloud infrastructure. Into this arena walked a Japanese firm selling an ARM CPU optimized for inference workloads — not a general-purpose GPU or a purpose-built ASIC.

What the Specs Actually Mean

The technical architecture is unambiguously serious. The CPU cores are manufactured on a 2-nanometer process, while the SRAM and I/O dies use a 5nm process in a stacked 3D configuration. Fujitsu claims this allows the chip to handle matrix computations at the hardware level, paired with software optimization, to deliver “practical AI inference performance” from the CPU alone. That last qualifier matters enormously — it suggests the company is not claiming to replace GPUs for training, but rather to compete in the inference tier where power efficiency and cost efficiency increasingly matter as model deployment scales.

The server packaging reinforces the efficiency angle. Fujitsu’s MONAKA Server is a 1U unit that operates at room temperatures up to 40°C with air cooling, or 45°C water temperature. The company claims up to 80% reduction in cooling power consumption compared to conventional setups. In markets where electricity and cooling represent 30–40% of total data center operating costs, that is a legitimate value proposition — if the raw inference throughput is competitive enough to justify the hardware switch.

Sovereignty as Strategy

Fujitsu’s CEO, Itaru Saito, framed the launch around the concept of “sovereign orchestration” — building AI infrastructure from Japanese-designed components, Japanese services, and Japanese platforms to ensure data sovereignty and supply chain transparency. The company already offers an enterprise AI model called Takane, a platform called Kozuchi, supercomputing and quantum computing systems, and a consulting service called Uvance Wayfinder. The Kozuchi platform was notably mentioned in discussions around data localization in Southeast Asia, where Japanese firms have been positioning themselves as alternatives to US and Chinese cloud providers.

The domestic manufacturing angle is not just nationalist branding. Fujitsu is selling the traceability argument: knowing where a chip was designed, fabricated, and assembled, with full supply chain visibility. In a world where US export controls have restricted advanced chip sales to China, and where Chinese companies like Huawei have responded with state-backed alternatives, Fujitsu is positioning Japan as a third pole — sovereign but not sanctioned, competitive but not dominant.

The early customers are telling. Financial services, telecommunications, and manufacturing sectors receive先行提供 ahead of general availability in April 2027 for Japan and Europe. Research institutions and defense applications are also on the list. These are precisely the segments where data sovereignty concerns are most acute and where government procurement rules favor domestic suppliers.

The Competitive Reality

Here is where the optimistic narrative meets the hard numbers. An ARM-based CPU doing inference work is fundamentally different from a GPU or TPU. ARM’s strength in server processors — the Graviton chips running AWS infrastructure, the Ampere Altra systems — comes from general-purpose workloads, not dense matrix operations at the scale modern AI inference demands. Fujitsu’s stacked-die approach and hardware matrix acceleration are genuine engineering choices, but they do not erase the architectural disadvantage.

NVIDIA’s inference offerings, including the L40S and the upcoming Rubin generation, are built for exactly this workload. Google’s TPU v5p is optimized for transformer-based inference at scale. The Chinese alternatives — Huawei’s Ascend 910 series and Cambricon’s custom silicon — are aggressively priced and backed by a domestic market of hundreds of millions of inference requests per day. Fujitsu’s target market is a fraction of that scale.

The 2nm process node is also a double-edged sword. Producing at that node typically requires access to TSMC or Samsung Foundry fabs, both of which are under US pressure to restrict capacity allocations to Chinese customers. If Fujitsu is relying on a Japanese fab for 2nm production — which the article does not clarify — it would be operating at a significant yield and capacity disadvantage compared to competitors using mature Taiwanese or Korean fabs.

What Happens Next

Fujitsu’s 1.5 trillion yen target for the “sovereign platform” segment alone is an aggressive number that assumes significant market shift toward localized, traceable AI infrastructure. The company’s strategy hinges on three uncertain variables: whether Japanese and European customers will prioritize sovereignty over raw performance, whether the ARM inference path can close the gap with GPU alternatives on cost-per-inference, and whether Japan’s semiconductor manufacturing ecosystem can support 2nm-scale production domestically or through allied foundries without US or Chinese interference.

The April 2027 shipment date to Japan and Europe is the first real test. If early customers report satisfactory inference latency and cost metrics against their existing GPU deployments, the narrative shifts from curiosity to credible competition. If not, Mokama becomes another well-intentioned national champion project that never escapes the proof-of-concept phase.

Japan’s semiconductor ambitions have a longer history than this launch suggests. The VLIW-based Fugaku supercomputer, developed by Fujitsu and RIKEN, consistently ranked as the world’s fastest system and proved that Japanese engineering could compete at the highest computational tiers. Mokama carries that lineage into the AI era. But supercomputing and AI inference are different battles with different economics. The question is whether the lessons from one translate to the other — and whether the market has room for a third option in a duopoly that is rapidly becoming a monopoly.