Arm’s Shanghai AI Push Spans Phones, Data Centers and Robotics

Arm’s Arm Everywhere China event in Shanghai on September 8 presented a single thesis: the company wants an Arm-based computing platform to extend from mobile devices to AI data centers and autonomous machines. The event’s three core announcements were CSS for Mobile 2, Neoverse CSS N4, and Arm Total Design for Physical AI.35

The most consequential backdrop was the Arm AGI CPU. Although this production-silicon server processor was announced earlier in 2026, Arm used the event to position it alongside CSS N4 as part of an AI-infrastructure offering that customers can deploy rather than design from the ground up.385

The announcements at a glance

1. CSS for Mobile 2 targets on-device agentic AI and neural graphics

CSS for Mobile 2 is Arm’s second-generation mobile compute subsystem: a pre-integrated platform that combines the C2 CPU cluster with SME2, the Mali G2-Ultra NX GPU, enhanced system IP, physical implementation work, and a developer-oriented software ecosystem.7

Its two headline compute engines are:

• Arm C2-Ultra CPU, which uses the second generation of Arm Scalable Matrix Extension (SME2) for responsive, efficient AI processing on the CPU.5
• Arm Mali G2-Ultra NX GPU, which adds dedicated neural accelerators. Arm says this enables neural graphics to run natively on the GPU.57

The significance is integration rather than a single component: Arm is offering chip partners a more complete starting point for phones that need to run AI experiences and advanced graphics locally.

2. Neoverse CSS N4 gives server-chip partners a configurable foundation

For cloud infrastructure, Arm introduced Neoverse CSS N4, which it calls its most configurable Compute Subsystem to date. The design supports up to 128 cores per die, LPDDR6 memory, and PCIe Gen 7 connectivity.9

Against Neoverse CSS N3, Arm claims CSS N4 can deliver up to:

• 2× performance
• 1.25× performance per watt
• 1.75× memory bandwidth9

Those are Arm performance claims, not independent benchmark results. Their practical implication is that partners can configure a server SoC around a validated Arm platform more quickly, especially for throughput-oriented, scale-out workloads associated with AI services.95

Some reporting describes N3P as an implementation path for CSS N4, but the broader point is configurability: customers may choose process and implementation options rather than treating one manufacturing node as an inherent characteristic of every CSS N4 design.24

3. Total Design expands into physical AI

Arm also extended its Total Design program into physical AI. The new ecosystem brings together more than 80 companies spanning the stack needed for autonomous systems, including compute, AI models, software, sensors, robotics and digital-development tools.18

Its first initiative, the Robotics Capability Framework, is intended to establish a common industry language for describing robotic-system capabilities. The objective is to reduce integration friction as developers move from component selection to deployed autonomous systems.1819

This is an ecosystem initiative rather than a new robot processor. Its value will depend on whether partners use the framework and shared development path in real products.

Why the AGI CPU is the strategic shift

For most of its history, Arm primarily licensed architecture and processor IP to companies that built their own chips. The AGI CPU adds another route: a customer can license Arm IP, adopt a Compute Subsystem, or deploy Arm-designed silicon.3845

The processor is a production-ready, up-to-136-core Arm Neoverse V3 server CPU made on a 3nm process. Arm introduced it for AI data-center infrastructure, where CPUs coordinate accelerators, data movement and surrounding services.3852

That changes Arm’s commercial position in two ways:

1. A faster path to deployment. Organizations that do not want to undertake a custom server-CPU program can buy into a ready platform through Arm’s server ecosystem.38
2. More direct participation in the server market. Arm is no longer limited to supplying the underlying IP for every implementation; it is more directly involved in the finished-silicon option.4538

The move does not mean Arm has abandoned licensing. Instead, it layers silicon onto the existing IP and CSS models. But it can place Arm closer to customers—and potentially in a more complicated relationship with ecosystem partners that have historically designed Arm-based chips themselves.

China: what partners have said about adoption

Arm’s China messaging included early infrastructure adoption plans:

• ByteDance’s Volcano Engine plans to bring the AGI CPU into data centers for next-generation AI services.15
• Lenovo is developing AI infrastructure around the AGI CPU and is among Arm’s named server ecosystem partners.138
• Reporting says Arm’s collaboration with New Unigroup covers both AGI CPU and CSS technology. However, the supplied evidence does not establish a specific public product, configuration or deployment timeline, so those details remain unconfirmed.46

These plans signal interest, but they should not be read as evidence of large-scale deployments or measured performance wins.

What it means for Intel and AMD

Arm’s announcements add competitive pressure to x86 in AI infrastructure, but the evidence supports a nuanced conclusion.

IDC reported $89.7 billion in AI-infrastructure spending in the first quarter of 2026 and said Arm-based rack-scale platforms had overtaken x86 in accelerated servers.31 That result is important because AI systems are increasingly acquired as integrated racks, where the host CPU, accelerator architecture, memory, networking and software stack are evaluated together.

Arm’s advantage is not simply raw CPU performance. Agentic-AI infrastructure can require substantial CPU-side work around accelerators: orchestration, networking, storage, retrieval and data movement. CSS N4 addresses this through a configurable scale-out design, while the AGI CPU supplies a deployable Arm server option for customers that want less custom-silicon work.5952

Still, “Arm overtook x86” in accelerated rack-scale systems is not the same as saying it has displaced x86 across the server market. The result is heavily shaped by AI-focused rack deployments, including platforms using Arm-based host CPUs. Intel and AMD retain extensive enterprise software ecosystems, installed bases and server-platform channels. The likely outcome is workload-specific competition, not an automatic architecture switch.3143

Bottom line

Arm’s Shanghai event was more than a product refresh. CSS for Mobile 2 extends its AI-native platform approach to phones; CSS N4 gives infrastructure partners a configurable route to server silicon; and Total Design for Physical AI tries to make robotics and autonomous-system development less fragmented.7918

The AGI CPU makes the strategy more consequential. By offering IP, subsystem designs and finished silicon, Arm is giving customers three ways to build AI infrastructure—and presenting Intel and AMD with a more complete competitor in the data-center CPU market.3845