On September 8, 2026, Arm unveiled three products during the Arm Everywhere China event, marking a clear expansion of its portfolio from CPU designs to graphics, AI infrastructure, and robotics development platforms. The announcements include Arm CSS for Mobile 2 for phones and mobile devices, Arm Neoverse CSS N4 for cloud infrastructure, and Arm Total Design for Physical AI for developing physical AI applications.
Enhancements to AI Processing in Mobile Devices
Arm CSS for Mobile 2 follows the Arm Lumex platform announced in September 2025 and is based on the new C2 CPU processing cluster in its two variants, C2-Ultra and C2-Pro. Arm says that C2-Ultra delivers up to 1.7 times the performance of C1-Ultra, with up to a 15% increase in single-threaded performance and an improvement of up to 38% in power efficiency.
The cluster includes two SME2 units for processing scalable matrix extensions, compared with one unit in the previous generation. According to Arm, this enables small language models to be processed at speeds up to 70% faster. These capabilities are intended to run more AI functions locally on the device instead of relying entirely on cloud processing, but the announcement does not specify particular commercial devices or when they will reach final products.
Arm also announced the Mali G2-Ultra NX GPU, the company's first graphics processor to include a neural rendering accelerator. The unit combines a neural-rendering path with a conventional graphics path, allowing neural rendering to be performed at higher speed or with lower power consumption, depending on the workload. Arm says neural graphics performance per watt is up to four times that of the Mali G1-Ultra, while conventional graphics performance increases by up to 14%.
Neoverse Platform for Infrastructure and Intelligent Agents
Arm Neoverse CSS N4 targets AI infrastructure workloads, including running intelligent agents. The design enables processors with between 8 and 128 cores to be built, with frequencies of up to 3.8 GHz using a 3-nanometer manufacturing process.
The product supports LPDDR6 memory at speeds ranging from 8000 to 12000 MT/s, as well as a PCIe Gen 7 interface. It also uses the CMN S4 interconnect system with a network of up to 16×16, which Arm says doubles bandwidth compared with CMN S3. The design supports assembling chiplets in multi-chip packages, an architecture intended to increase computing density within systems.
Compared with Neoverse CSS N3, Arm attributes to N4 twice the performance, 1.25 times higher performance per watt, and 1.75 times higher memory bandwidth. These remain figures announced by Arm, while their actual impact will depend on licensees' implementations, system design, and the software used.
Framework for Developing Physical AI
Arm Total Design for Physical AI is a program that brings together hardware, software, computing, sensors, and actuators to reduce the complexity of developing robotic systems. The program includes more than 80 partners, including AWS, NXP, Hugging Face, QNX, Siemens, and Unitree Robotics, in addition to other companies such as ECARX, Liquid AI, and Qwen.
As part of the program, Arm announced the Robotics Capability Framework to classify robotic-system capabilities across six levels, beginning with RL0 and ending at RL5. The framework aims to provide a common way to describe system capabilities, without the source providing complete technical details about the tests or criteria that distinguish each level.
What Changes in Practice?
The three announcements combine efforts to bring AI onto devices, expand infrastructure resources, and organize the robotics development chain. The key point is not a single product, but rather Arm's attempt to offer relatively ready-made design packages extending from the phone to the data center and then to the physical system. Nevertheless, the material does not disclose prices, final products, or availability schedules, so measuring the actual commercial impact will remain tied to decisions by device companies, chip designers, and software partners.