At the "2025 Technology Wind List" annual ceremony held in Beijing on January 15, 2026, Song Jiqiang, Vice President of Intel and President of Intel China Research Institute, delivered a keynote speech titled "Promoting the Implementation of Trustworthy Embodied AI Applications with Heterogeneous Computing."

Song Jiqiang stated that the development of AI capabilities is evolving from foundational large models towards agent-based AI, with a greater emphasis on providing specific functionalities to construct workflows. As an important form of physical AI, embodied intelligence embeds the intelligent capabilities of the digital world into physical devices to interact with the real world, and these types of applications are predominantly inference-based. He pointed out that industry analysts also predict that the focus of AI computing power demand is shifting from training to inference, which will consume a corresponding proportion of computing resources.

Song Jiqiang also discussed how multi-agent systems build complete workflows and achieve parallel operation of multiple streams, thereby creating demand for heterogeneous infrastructure. He explained that the functional support for an AI Agent includes various models, schedulers, and pre-processing modules, which require different types of hardware to provide optimal energy efficiency and cost-effectiveness. While all tasks can technically run on a CPU, it is difficult to balance timeliness and functional effectiveness, thus necessitating a combination of hardware such as high-end GPUs and mid-range GPUs to achieve precise adaptation for models and task scenarios of different scales.

Song Jiqiang further proposed that heterogeneous systems need to possess flexible heterogeneous support capabilities at three levels: the upper layer needs to build an open AI software stack to shield applications from system-level changes and protect investment validity; the middle system infrastructure must be adapted to the needs of small and medium-sized enterprises, providing user-friendly server configuration settings and Ethernet interconnection solutions; the bottom layer needs to integrate continuously evolving diverse hardware, including CPUs, GPUs, NPUs, AI accelerators, and neuromorphic computing devices with different architectures, building a flexible heterogeneous system through layered infrastructure.

Regarding the field of embodied intelligent robotics, Song Jiqiang analyzed various methods for implementing intelligent tasks, ranging from traditional layered custom models to fully end-to-end VLA models. He noted that the industry has not yet determined an optimal solution and is currently in a phase of diverse experimentation. He indicated that traditional industrial automation control solutions focus on reliability, real-time performance, and computational precision, whereas solutions based on large language models lean towards a neural network approach, requiring differentiated computing architectures for support. This can be achieved by utilizing CPUs for high-speed response, NPUs for low-power output, and GPUs for visual and language model recognition, relying on the scheduling of different workloads across heterogeneous chips combining CPU+GPU+NPU.

Song Jiqiang emphasized that the era of embodied intelligent robots is inevitable and will bring challenges in computing power and energy consumption, with heterogeneous computing gradually becoming the core architecture of AI infrastructure. In the future, when the scale of robots reaches millions, they will break through the limitations of industrial scenarios and widely support commercialized and personalized applications, urgently requiring the support of multi-agent systems. He pointed out that the technology stack for running multi-agent systems on physical AI devices still faces many challenges, and heterogeneous computing is a key path to solving the problem of system trustworthiness.