Arm has introduced a new mobile chip platform called CSS for Mobile 2, built specifically to run artificial intelligence agents directly on phones while maintaining console-quality graphics, all within a 1-watt power budget. The platform represents a fundamental shift in how mobile devices will handle AI workloads, moving away from constant cloud connections toward on-device intelligence that can operate continuously in the background.

Why Are Phone Makers Suddenly Focused on Local AI?

The AI landscape has changed dramatically over the past year. Every major AI lab has shipped agent products that can call tools, browse the web, and act across multiple steps without human approval at each stage. Running these agents reliably on a phone today creates a fundamental problem: either devices make constant round-trips to cloud servers, draining batteries and raising privacy concerns, or they run heavily compressed local models that sacrifice accuracy for speed.

Arm's bet is that on-device inference becomes commercially necessary once agents perform continuous background work rather than answering occasional chat queries. No phone battery survives a full day of constant cloud connections, making local processing essential for the next generation of mobile AI experiences.

What Makes CSS for Mobile 2 Different From Previous Generations?

The new platform pairs two key components designed specifically for the AI era. The C2-Ultra CPU cores deliver a 15% single-thread performance improvement over last year's generation, while the platform doubles its Scalable Matrix Extension units for AI math operations. That doubling delivers a 70% speedup when running small language models directly on-device, according to Arm's claims.

The Mali G2-Ultra NX GPU introduces what Arm calls "neural-assisted ray tracing," using AI techniques to cut ray-tracing workload by 70%, enabling 30 frames-per-second ray-traced gaming within that same 1-watt power budget. Early evidence suggests these lab numbers may hold up in real devices: Xiaomi's Xring O3 chip, which pairs the Mali G2-Ultra NX with neural upscaling technology, claims an 85% GPU performance gain over its predecessor in the Xiaomi 18 Fold.

How Should Developers Prepare for This Shift?

• Split Inference Architecture: Design applications to run lightweight agent steps locally on the device while falling back to cloud servers for heavier reasoning tasks, since this is the architecture every major chipmaker is now building silicon to support.
• Optimize for On-Device Models: Begin testing and optimizing small language models that can run efficiently within the power and memory constraints of mobile devices, rather than relying exclusively on cloud-based models.
• Plan for 2027 Availability: Recognize that the earliest devices shipping on CSS for Mobile 2 are expected in 2027, giving developers roughly 18 months to prepare applications before consumers hold devices that run this architecture well.

The ecosystem is already building on these capabilities. Partners including Alipay, OPPO, and Vivo are adopting the new Scalable Matrix Extension technology, while game developers at Infold Games, NetEase, Tencent Games, Unity China, and Unreal Engine are collaborating on neural graphics features.

When Will These Chips Actually Reach Consumers?

It's important to understand the gap between Arm's platform announcement and actual shipping devices. Arm licenses chip designs but does not manufacture or sell chips directly to consumers. Every performance claim in this announcement is Arm's own, unverified by independent testing until partners like Qualcomm and MediaTek actually build and ship devices on the design.

Arm's own timeline suggests the earliest devices will arrive in 2027, meaning there's at least a year between the platform announcement and consumer availability. Phone gaming graphics improvements have a mixed track record of translating headline specifications into real-world battery life gains, and the "70% speedup" figure lacks details about which specific small language models were tested or what baseline performance was used for comparison.

How Does This Compare to Apple's Approach?

Apple has executed this exact playbook for four consecutive generations of its A-series and M-series chips, expanding Neural Engine throughput each year specifically so on-device features like transcription, photo search, and its own agentic Siri work without requiring network round-trips. CSS for Mobile 2 represents the rest of the Android ecosystem catching up to that same architectural strategy.

The key difference is delivery model. Apple owns its entire stack end-to-end, controlling both hardware and software. Arm's approach as a licensable platform gives every non-Apple phone maker access to the same foundation, which is both an advantage and a constraint. Qualcomm and MediaTek will each tune, brand, and ship the design on their own timelines rather than Arm's, meaning performance and availability will vary across manufacturers.

Beyond mobile devices, Arm is also expanding its platform for physical AI applications like robots and autonomous vehicles, bringing together more than 80 industry leaders including AWS, Hugging Face, and Unitree Robotics to establish common standards for describing and comparing the capabilities of increasingly intelligent autonomous machines. This broader push reflects a fundamental industry shift toward distributed intelligence across edge devices, cloud infrastructure, and physical systems, all running on Arm-based computing platforms.