Shanghai Outlines New AI and Computing Power Roadmap

A recently released development plan for Shanghai's software and information technology services sector has charted a clear course for the region's artificial intelligence and computing infrastructure. The emphasis on integrating chip technology with mainstream AI models stands out as a particularly significant strategic move.

Exploring the Next Generation of AI Architectures

The plan advocates for a multi-path exploration beyond the dominant Transformer architecture. It explicitly encourages research into alternative frameworks such as State Space Models, variants of Recurrent Neural Networks, and Liquid Neural Networks. This signals a deliberate shift towards fostering architectural diversity in future AI systems.

Concurrently, the document calls for early-stage investment in frontier foundational technologies like Physical AI, World Models, Quantum Intelligence, and Brain-inspired Computing. These areas represent ambitious attempts to equip machines with a more nuanced understanding or simulation of real-world dynamics.

Key Hardware Focus Areas for Computational Power

To support these advanced models, the plan identifies several hardware components where supply capacity must be strengthened:

  • Computing Chips: Including high-performance GPUs and NPUs.
  • Quantum Chips (QPUs): Tapped as a potential path for future computational breakthroughs.
  • High-Speed Interconnects: Technologies like CPO to alleviate communication bottlenecks in large-scale clusters.
  • High-Bandwidth Memory (HBM): Crucial for handling the massive data loads of large models.
  • Heterogeneous Servers: Integrating diverse computing units to boost overall efficiency.

A central objective is to achieve deeper integration between domestically developed chips and prevailing large language models. This goal touches on both technological sovereignty and the future cost-efficiency structure of the AI computing ecosystem.

Critical Software and System-Level Challenges

Advancements in hardware must be matched by progress in software stacks and system technologies. The plan highlights several key technical challenges:

For data processing, it points to novel storage-retrieval methods and data-model collaboration techniques. The efficiency and precision of information extraction from vast datasets are fundamental to enhancing model capabilities.

At the model level, breakthroughs are needed in high-precision heterogeneous processing, native multimodal fusion, and dynamic value alignment. These technologies are key to developing more reliable and controllable AI systems. The overarching aim is to establish an automated, complex reasoning system that spans the entire lifecycle of corpus data—from collection and processing to training and inference.

The release of this plan indicates Shanghai's strategic push to extend its AI ambitions from the application layer down to the foundational levels of chips, computing clusters, and core software. Its implementation will significantly shape the domestic AI industry landscape in the coming years.