Analysis
Arm introduced its next mobile chip platform this week, CSS for Mobile 2, built around two priorities that used to trade off against each other: running AI agents locally and delivering console-quality graphics, The Register reported Tuesday. The platform is a licensable design, meaning phone makers still have to fabricate their own silicon around it -- the earliest devices are expected in 2027. Pulse has previously covered Arm's push into AI-capable silicon as the company positions its licensable designs against custom hyperscaler chips.
What's actually new in the silicon
The flagship compute cluster pairs two new C2-Ultra CPU cores -- Arm claims a 15% single-thread gain over last year's C1-Ultra -- with six efficiency cores and two Scalable Matrix Extension units for AI math, doubled from the prior generation. Arm says that doubling delivers a 70% speedup running small language models directly on-device. The new Mali G2-Ultra NX GPU adds neural-assisted ray tracing and upscaling, claiming a 70% cut in ray-tracing workload through AI techniques, enough for 30 frames-per-second ray-traced gaming inside a 1-watt power budget -- a fraction of what a discrete GPU draws for comparable output.
“Arm says that doubling delivers a 70% speedup running small language models directly on-device.”
Why the target is agents, not chatbots
Every major AI lab has spent 2026 shipping agent products that call tools, browse the web and act across multiple steps without a human approving each one. Running that reliably on a phone today means either a fast round-trip to a cloud model -- with the latency, battery and privacy costs that implies -- or a heavily quantized local model that trades accuracy for speed. Arm's bet is that on-device inference becomes commercially necessary once agents are doing continuous background work rather than answering occasional chat queries, because no phone battery survives a day of constant cloud round-trips. Qualcomm and MediaTek, the two chipmakers who will actually ship phones on Arm's designs, both face the same pressure from Apple, whose own silicon roadmap has pushed on-device AI performance every generation since the first Neural Engine.
The gap between a platform announcement and a shipped phone
Arm licenses designs; it does not fabricate or sell chips to consumers. Every performance claim in this announcement is Arm's own, unverified by independent silicon until a partner actually builds and ships a device on it, and that gap runs at least a year given Arm's own 2027 timeline. Phone gaming graphics improvements have a mixed track record of translating headline specs into real day-to-day battery life, and "70% speedup" on unspecified small language models is a number without a baseline model size or task attached.
For AI application developers building for mobile, the practical read is to start designing for a split inference architecture now -- lightweight agent steps running locally, heavier reasoning falling back to the cloud -- because that is the architecture every major chipmaker is now building silicon to support, roughly eighteen months before consumers will hold a device that runs it well.
The comparison point nobody in the announcement mentions
Apple has run this exact playbook for four straight generations of its own A-series and M-series silicon, expanding Neural Engine throughput each year specifically so on-device features -- transcription, photo search, and now its own agentic Siri work -- do not depend on a network round-trip. Arm's CSS for Mobile 2 is the rest of the Android ecosystem catching up to that same architectural bet, delivered as a licensable platform rather than owned end-to-end silicon, which is both Arm's advantage -- every non-Apple phone maker can build on it -- and its constraint, since Qualcomm and MediaTek will each tune, brand and ship it on their own timelines rather than Arm's.
