Arm CSS for Mobile 2 Unveiled as AI-Native Platform for Agentic AI and Cinematic Mobile Graphics

Cambridge-headquartered Arm has announced CSS for Mobile 2, a new compute subsystem it describes as built from the ground up for the demands of agentic AI and high-fidelity mobile graphics. The platform combines the company’s new C2 CPU cluster with the Mali G2-Ultra NX — Arm’s first GPU to carry dedicated neural accelerators — into a single, pre-validated reference design intended to shorten the time it takes chip makers to bring new mobile processors to market.

The announcement came via Arm’s official account on 8 September 2026, which posted that CSS for Mobile 2 is “built for the system-level demands of agentic AI and cinematic mobile graphics,” a description consistent with materials published on Arm’s newsroom and product pages.

What CSS for Mobile 2 Actually Is

CSS stands for Compute Subsystem. Rather than licensing individual CPU and GPU cores separately, Arm bundles them — along with system interconnect IP, physical implementations, and a software stack — into a coherent reference design. The idea is that a chip maker can take CSS for Mobile 2, validate it once, and build a finished smartphone system-on-chip (SoC) around it without having to assemble and verify each component from scratch. That reduces engineering time and, in theory, gets new handsets to consumers faster.

CSS for Mobile 2 is built on the C2 CPU cluster, which includes C2-Ultra and C2-Pro cores running on the Armv9.x architecture. Those cores include support for Scalable Matrix Extension 2, or SME2 — a hardware capability that accelerates the matrix and tensor maths at the heart of modern AI workloads. Arm says the C2 cluster with doubled SME2 capability delivers up to around 1.7 times the AI performance of the previous-generation cluster across tested models, and roughly 70 per cent faster performance on the latest small language models. Both figures come from Arm’s own product communications and have not been independently verified.

The Mali G2-Ultra NX: Neural Accelerators Inside the GPU

The GPU side of the platform is where things get genuinely new. The Mali G2-Ultra NX places dedicated neural accelerators directly inside the GPU’s shader cores, rather than routing AI workloads to a separate chip block. Arm says this reduces the amount of data that needs to travel across the chip, cutting memory traffic and latency for graphics-linked AI tasks.

That architecture enables what Arm calls neural graphics. Neural Super Sampling uses AI to upscale lower-resolution frames in real time; Neural Frame Rate Upscaling generates intermediate frames to raise the apparent frame rate of games and video. Third-party technical reporting based on Arm briefings suggests these techniques can roughly double frame rates and cut external memory traffic by somewhere between 33 and 50 per cent in some test cases — though again, those figures are unverified vendor metrics rather than independent benchmarks. Arm also claims up to four times higher performance per watt for certain neural graphics workloads compared with a previous Mali generation; that figure, reported in media coverage and Arm’s own materials, should be treated with the same caution.

Agentic AI on a Handset

Arm’s framing around “agentic AI” is worth examining. The company isn’t just talking about running a chatbot on your phone. Agentic AI refers to systems where an AI model can coordinate across multiple applications and tasks — booking a restaurant, checking your calendar, and sending a message, all in sequence, without you touching each app individually. Running that kind of workload on-device, rather than in the cloud, requires sustained, efficient compute that can handle context across several operations at once.

CSS for Mobile 2 ships with a developer-ready software ecosystem including KleidiAI, Arm’s AI software library, alongside wider AI tools and resources intended to help silicon partners and OEMs implement the platform in real products. The software layer matters as much as the silicon; without optimised libraries, the hardware gains don’t reach end users.

René Haas, Arm’s chief executive, has spoken publicly about the company’s direction towards AI-native mobile hardware, though no specific quote from him was issued alongside this particular announcement.

Some technology commentators have already pushed back on the marketing language, noting that terms like “agentic AI” and “AI-native graphics” describe software behaviours as much as hardware capabilities, and that the performance uplift figures — 1.7x here, 4x there — are vendor-supplied numbers that independent testing may not replicate in every real-world scenario.

There are also broader questions. More powerful AI hardware can drive shorter handset replacement cycles, raising concerns about electronic waste. Privacy advocates will point out that on-device AI cuts reliance on cloud servers — which is good for data sovereignty — but also enables more sophisticated local processing, which requires clear user controls and strong governance under UK data protection law.

When Will Consumers See It?

CSS for Mobile 2 is an IP platform, not a finished chip. Arm licenses the design to semiconductor companies, who then build their own SoCs around it. Those chips then go into handsets designed by manufacturers. That pipeline typically takes 18 months to two years from IP announcement to consumer devices appearing on shelves. No chip maker has yet publicly confirmed a product based on CSS for Mobile 2.

What This Means for Kent Residents

Arm’s designs already power the overwhelming majority of smartphones used by people across Kent, so any future handsets built on CSS for Mobile 2-based chips will directly affect local consumers — potentially bringing faster on-device AI assistants, smoother gaming, and improved battery life under heavy AI workloads, though those benefits won’t arrive until chip makers and handset manufacturers complete their own development cycles. For Kent residents who rely on mobile apps for travel, health, or local services, the shift towards on-device AI processing could mean more responsive tools that work reliably even with patchy connectivity. Higher device prices remain a realistic concern if AI-native hardware is concentrated in premium tiers, at least in the near term.