Arm Server CPUs Enter China, Shipments Expected to Reach One Million Units by 2027
nashnova research
Arm is shifting from licensing IP to selling finished server CPUs for AGI workloads. DIGITIMES forecasts 1 million units shipped by 2027, with China as the key growth market — domestic CPU manufacturing constraints and open commercial-cloud procurement create the window.
Why is Arm suddenly selling finished chips?
Arm's old business was selling blueprints — licensing Neoverse core IP so customers could build their own chips. Now Arm sells ready-made server CPUs targeting artificial general intelligence (AGI) workloads.
This means → Arm moves from collecting royalties to collecting chip revenue, a shift from asset-light to asset-heavy.
Meta co-developed the product and is the lead customer; OpenAI also plans to adopt it. In plain terms = customers skip the time and cost of designing their own silicon; Arm gains a fatter revenue line.
Why does China matter so much for Arm?
DIGITIMES sees China as the largest incremental market outside the US. The logic has two legs: domestic CPUs face manufacturing limits, and commercial cloud providers still have room to buy foreign chips.
China has strong server-CPU design capability. The bottleneck is manufacturing — restricted access to advanced process equipment and technology leaves some domestic chips lagging in performance, efficiency, and volume production.
This reflects a structural gap: design can deliver, but fabrication cannot keep up. That gap is Arm's entry ticket.
What exactly do export controls block?
From 2022, the US tightened semiconductor export controls, adding license requirements for specific equipment needed to produce 16/14 nm and more advanced FinFET chips in China, and restricting US personnel from supporting those processes.
Between 2023 and 2024, controls expanded to etch, deposition, ion implantation, metrology, and inspection equipment, plus certain design and manufacturing software. The Netherlands also introduced export-license controls on some advanced DUV lithography tools.
Huawei and HiSilicon face both the Entity List and the Foreign Direct Product Rule (FDPR). Overseas foundries need US approval to supply controlled chips — the Kunpeng server-CPU line is therefore cut off from advanced foreign fabrication.
How do procurement rules differ between server CPUs and AI accelerators?
In AI accelerators, Nvidia disclosed in August 2026 that H200 shipments to China remained below approved volumes — because Chinese policy requires some government-funded new data centers to use domestic AI accelerators.
Server-CPU rules are more flexible. China's 2023 government-procurement standard requires township-level-and-above Party and government bodies to include "security and reliability" compliance, but does not require all private cloud providers to buy domestic CPUs.
In plain terms = AI accelerators hit a "must be domestic" hard wall; server CPUs face only a "government procurement must comply" soft bar — commercial cloud customers can still freely choose foreign CPUs.
Is the one-million-unit forecast credible?
DIGITIMES estimates Arm's first AGI server CPU will ship 1 million units in 2027, combining all the factors above.
The demand logic: China's AI inference and agent services expand → data processing, tool invocation, and sandbox execution needs grow → server-CPU demand rises.
This means → the forecast hinges on a single variable: how long the domestic-CPU manufacturing constraint window stays open. If domestic process breakthroughs accelerate, Arm's window narrows; if controls persist, the window holds.
市场有风险，内容仅供研究参考，不构成投资建议。