Will Abbey had a number for the developers gathered in Shanghai on Tuesday: 350 billion. That, the Arm executive vice president and chief commercial officer said, is how many chips based on Arm technology its global partners have shipped in total. The figure, delivered on a conference stage at the company’s Arm Everywhere China event, framed the announcement that followed.
Arm used the gathering to release the second generation of its mobile client compute subsystem, called Arm CSS for Mobile 2. The platform bundles Arm’s processor designs into a pre-integrated package meant to support agentic AI and AI-native graphics inside a single computing architecture. The company positioned the release as an answer to a specific problem: phone makers and chip designers want to ship devices that run large language models locally, but assembling the silicon to do it has grown harder with every generation.
The new subsystem combines the Mali G2-Ultra NX graphics processor with the C2 CPU cluster. The GPU is the more significant piece. It is the first Mali part to include a dedicated neural network accelerator, allowing graphics processing and neural network computation to run inside the same GPU pipeline rather than across separate blocks of silicon. Arm said the arrangement lifts per-watt performance by as much as four times in neural-graphics workloads, while gaming performance improves up to 14 percent against the previous generation.
The change reflects where mobile workloads are heading. The first wave of on-device AI treated the neural processor as an auxiliary unit, switched in for specific tasks such as photo processing or voice recognition. The next wave, in Arm’s telling, is different: applications increasingly want the neural network to help render scenes, understand context and respond while a game or interface is running. Keeping those computations on one pipeline avoids the inefficiency of shuffling data between separate processors.
Compute subsystems are a relatively young idea for Arm, which built its business licensing individual processor designs to hundreds of customers. The company introduced the format in 2024, offering chipmakers a pre-verified combination of CPU, GPU and system IP tuned for a manufacturing process. The pitch was speed: a partner that adopts a subsystem skips much of the integration work that once consumed months of a design cycle. The approach has spread into data centers as well, where Arm sells similar blueprints for server chips.
For mobile, the second-generation product arrives as the market for premium phones has become a contest over who can run the most ambitious models locally. Qualcomm, MediaTek and Samsung’s system-on-chip business all ship Arm-based designs, and each has been marketing neural performance as a headline feature. A common Arm blueprint does not erase competition between them, but it sets a floor for capability across the ecosystem and compresses the gap between a flagship chip and its cheaper cousins.
The economics help explain the strategy. Arm is paid royalties per chip sold, so its fortunes rise with unit shipments and with the value of what designers build on its architecture. A subsystem that makes it easier for partners to add capable AI hardware could widen the market for Arm-powered devices, from phones to laptops to the growing category of edge devices that run small models on-site. The company’s move into full blueprints also deepens its relationship with partners who might otherwise design more of their own silicon.
The China venue was deliberate. Arm’s licensing partners in the country account for a large share of the world’s smartphones, and the event gave the company a stage to court designers at a moment when trade restrictions have complicated access to advanced American chips. Local chip developers have leaned on Arm’s architecture to build competitive mobile processors, and products such as CSS for Mobile give them a faster path to market at a time when speed matters as much as raw specification.
Tuesday’s announcement did not name the partners that will ship the first devices based on CSS for Mobile 2, and the company said the number for cumulative shipments will keep climbing. The broader question is whether a common platform can stay ahead of the customization that phone makers use to differentiate their flagships. Arm’s answer is to keep the subsystem’s boundaries wide enough that partners can vary the CPU and GPU configurations while trusting the integration underneath.
The trajectory matters beyond phones. If neural-graphics processing becomes the standard way mobile chips are judged, the designs validated in Shanghai will filter into laptops, automotive systems and industrial devices that share the same architectural lineage. Arm is betting that its role in the AI era will be defined less by individual core designs and more by the completeness of the platform it hands to partners. The 350 billion chips already shipped are the base of that argument; what runs on the next generation will test it.
