CAICT’s “Intelligent Terminal” Blue Paper: AI Devices Shift From Features to Core Intelligence

The China Academy of Information and Communications Technology (CAICT) released its “New-Generation Intelligent Terminal Blue Paper (2025)” on March 16, 2026, arguing that AI is moving from a “plus-one feature” to the core intelligence of end devices. The report says AI phones, AI PCs and wearables are entering large‑scale commercialization, and that device roadmaps are being reshaped by heterogeneous computing (CPU+GPU+NPU) and tighter edge‑cloud collaboration. CAICT’s framing places intelligent terminals at the center of China’s consumer electronics upgrade cycle, with new requirements for on‑device capability, software architecture, and ecosystem interoperability.

CAICT is a government‑affiliated research institute under China’s Ministry of Industry and Information Technology (MIIT), and its industry blue papers often signal policy and market direction. In this edition, CAICT describes a shift from “AI + device” to “AI device”, meaning AI is embedded across chips, operating systems, sensing, interaction, and connectivity, rather than running as a separate app layer. That is a clear signal to China’s device makers and supply‑chain partners that “AI terminal” is now the baseline category, not an experimental add‑on.

A key conceptual contribution is the “four‑new” characteristics of AI terminals: cognitive collaboration, scenario prediction, intent‑driven interaction, and service co‑existence. In practice, this means devices should coordinate user context across apps (cognitive collaboration), anticipate multi‑step tasks (scenario prediction), and map intent to actions without explicit commands (intent‑driven interaction). The final feature, service co‑existence, points to a future where device functions are composed from modular services that can be invoked by intelligent agents rather than fixed app screens.

On the hardware side, CAICT highlights the rise of heterogeneous computing as standard design. The report notes that NPU peak performance has crossed 100 TOPS, which enables more on‑device inference for speech, vision, and compact language models. The shift toward CPU+GPU+NPU pipelines is not just about higher benchmarks; it is about latency, privacy, and cost control, allowing more AI workloads to run locally while the cloud handles heavier training or complex reasoning. That on‑device push echoes the appliance market’s race for in‑house AI chips and whole‑home ecosystems.

The software stack is also being rebuilt. CAICT says operating systems are evolving into edge‑cloud collaborative “intelligent hubs,” and that the app ecosystem is transitioning from standalone applications to agent‑callable micro‑services. China is also scaling cloud backbone capacity, such as Shanghai’s 140,000 PFLOPS compute scheduling platform, to support this edge‑cloud split. This implies a re‑architecture of how services are packaged and exposed, with AI agents emerging as a primary interaction layer rather than an add‑on interface. For device vendors, the implication is that AI‑native OS features and developer platforms will become differentiators.

For China’s consumer electronics sector, this blueprint creates a more explicit roadmap. AI terminals are expected to scale across phones, PCs, and wearables, and vendors will need to prove on‑device capabilities rather than rely on cloud features alone. The report’s emphasis on NPU throughput and end‑cloud collaboration suggests that chip selection, thermal design, and power efficiency are now strategic issues, not merely engineering constraints.

The commercial impact is likely to show up in 2026 product cycles. As device makers integrate stronger NPUs and publish agent‑centric features, the competitive bar will rise for both hardware and software. From a supply‑chain perspective, the transition supports demand for advanced mobile chipsets, high‑bandwidth memory, and more efficient power‑management components—all essential to keep AI inference responsive without draining battery life.

Another notable signal is the report’s positioning of intelligent agents as the next interaction entry point, with services exposed as agent‑callable micro‑capabilities rather than monolithic apps. That shift implies a need for standardized APIs, permission models, and safety layers so agents can act on behalf of users without compromising data control. It also changes developer economics: if user journeys move from app stores to agent‑orchestrated workflows, device vendors may prioritize first‑party frameworks to keep service access and data flows inside their ecosystems.

Looking ahead, CAICT’s framing suggests two near‑term outcomes. First, AI‑native terminals will compete on system‑level intelligence (sensing, reasoning, and acting together) rather than on single features. Second, the edge‑cloud split will become more standardized, with more workloads retained on device for privacy and cost reasons while cloud services provide model updates and heavier compute. If this blueprint translates into procurement standards or platform guidelines, it could accelerate a China‑centric ecosystem for AI terminals over the next 12–18 months.

Sources
– CAICT official release: https://www.caict.ac.cn/kxyj/qwfb/bps/202603/t20260316_716534.htm
– IT之家 report: https://www.ithome.com/0/929/608.htm
– Securities Times (STCN): https://www.stcn.com/article/detail/3679116.html
– Shanghai Securities News (Eastmoney): https://finance.eastmoney.com/a/202603163673378166.html