Fujitsu MONAKA Puts 144 Arm Cores on Sale in November

A 2nm Arm Server Chip Arrives From an Unexpected Direction

When people list the companies building Arm server CPUs, Fujitsu usually comes up late, if at all. That is about to look like an oversight. Fujitsu has formally launched MONAKA, a 144-core Armv9 server processor built on 2nm compute dies, and it goes on sale in November 2026. For a market that has spent two years assuming AI inference belongs to accelerators, the pitch here is blunt: run it on the CPU.

• Silicon: 144 Armv9 cores across four 2nm compute dies, paired with 5nm cache and I/O chiplets using Broadcom's 3D stacking
• Memory and I/O: 12 DDR5 channels at 8,800 MT/s and PCIe Gen6
• Clocks: 2.1 GHz air-cooled and 2.9 GHz liquid-cooled in shipping servers, with a 3.8 GHz maximum operating frequency
• Timing: standalone processors and 1U and 2U servers go on sale in November 2026, with broader Japan and Europe shipments sequencing from April 2027

Why the Cache Sitting on Its Own Die Matters

The packaging is the most interesting engineering choice. Fujitsu split the design so the cores sit on 2nm dies while cache and I/O live on separate 5nm chiplets, all bonded with 3D stacking.

That is a cost and yield argument as much as a performance one. Cache does not benefit from a leading-edge node the way logic does, and 2nm wafers are expensive. Spending the expensive node only where it pays and hanging everything else off a mature process is the same economic logic that has been quietly reshaping high-end silicon for years. For investors watching the semiconductor supply chain, it is another datapoint that advanced packaging, not just node shrinks, is where differentiation is happening — the same theme running through the GlobalFoundries and Marvell Vermont capacity deal.

What Are Fujitsu's Performance Claims?

Fujitsu says MONAKA delivers twice the AI inference throughput of comparable CPUs while drawing half the power for the same workload, which the company frames as customers running equivalent jobs on roughly 50 percent fewer physical servers. The hardware backing that claim is concrete enough to evaluate: dedicated matrix-multiplication acceleration and dual 256-bit SVE2 vector units per core, aimed squarely at inference rather than training.

Those comparative figures are Fujitsu's own, and no independent benchmarks exist yet. Worth flagging clearly, because "twice the throughput at half the power" is the kind of number that needs a third party behind it before anyone builds a purchase order on it. What is verifiable today is the silicon configuration and the software readiness — compiler support for MONAKA was upstreamed in GCC 15, so the Linux toolchain is already there.

The Sovereign Compute Angle

The go-to-market tells you as much as the spec sheet. Fujitsu is targeting cloud operators, server makers, hyperscalers and — explicitly — institutions buying for sovereign computing, with initial sales focused on Japan and Europe. Select customers in finance, telecommunications and manufacturing are expected to receive systems in the second half of Fujitsu's fiscal 2026.

That is a real and growing market. Air-cooled 1U and 2U options matter here too: a chip that delivers usable inference without a liquid-cooling retrofit is far easier to sell into an existing European or Japanese data centre than one that requires rebuilding the room. Given how much of the current AI build-out is constrained by power and facilities rather than chip supply — a tension visible in the record server market numbers for Q2 2026 — an efficiency-first CPU has a clearer path than the raw specs suggest.

No public pricing has been announced. For anyone tracking Arm's slow encroachment into the data centre, November is the date to watch. More semiconductor and market coverage in our stock trading section.

Sources: The Register — September 17, 2026; Phoronix — September 17, 2026.