从低时延网络到具身智能实训 AI落地进入工程化阶段

This article outlines that China’s AI industry has completed its foundational infrastructure expansion and officially entered the phase of engineering deployment. The core industry shift has moved from focusing on "accessibility" to "performance and usability," with multiple government departments recently releasing a slew of new policies to accelerate real-world AI adoption.

1. China’s AI industry foundation is now fully established. Public data shows that by 2025, the number of AI enterprises in China will exceed 6,000, the core AI industry scale is projected to top 1.2 trillion yuan, and the total intelligent computing capacity will reach 1590 EFLOPS, creating the necessary conditions for large-scale industrial deployment.

2. Three core priority directions for AI deployment have been set out: low-latency computing network construction, embodied intelligence real-environment training, and industry-specific AI adaptation and trusted evaluation. Each direction comes with clear policy targets and implementation timelines: for example, 75% coverage of 1-millisecond latency metropolitan computing circles by 2028, and 10,000-unit embodied intelligence deployment capacity by the end of 2026.

3. Large-scale AI deployment will drive growth across the entire industrial chain, creating new development opportunities for a wide range of sectors beyond AI-native companies, including communication equipment manufacturers, telecom operators, and industrial internet enterprises.

AI has entered the phase of engineering deployment, bringing clear trends and opportunities for brand product R&D and market positioning. Key takeaways for brands:

1. AI-powered products are demanding faster, lower-cost computing access. When developing AI-enabled consumer or enterprise products, brands should track the progress of low-latency network rollout, and prioritize adapting to millisecond-accessible computing resources to reduce product response latency and improve end-user experience.

2. Embodied intelligence is prioritized for B2B and semi-enclosed scenarios. When developing intelligent service robot business lines, brands can prioritize entering industrial, specialized, logistics, inspection and commercial service scenarios. These scenarios easily support clear task definition, controlled operating environments and viable commercial models, with far lower deployment barriers than consumer-facing home robotics.

3. Brands developing industry-specific AI products can participate in policy-backed "data-model synergy" innovation consortia to refine models using high-quality industry data. They should also proactively adapt to AI measurement and evaluation standards to meet enterprise users’ requirements for AI reliability and address adoption concerns.

This article summarizes recent AI industry policies released by multiple Chinese government departments, clarifying growth opportunities and risk directions in the deployment phase. Key takeaways for suppliers:

1. Policies have identified three core high-growth tracks: low-latency computing network construction, real-scenario embodied intelligence deployment, and trusted AI evaluation system development. Suppliers across upstream and downstream segments can align their offerings accordingly. For example, low-latency network construction requires large volumes of 400Gbps/800Gbps backbone transmission equipment and full-optical cross-connect high-speed equipment, so relevant supply chain vendors can prepare to meet growing market demand.

2. The embodied intelligence initiative targets 100 high-value application scenarios and 10,000-unit deployment capacity by the end of 2026. Provincial governments and central state-owned enterprises are each required to roll out at least 10 to 20 priority scenarios, creating a large pipeline of clear project opportunities for both hardware and software suppliers.

3. Policies prioritize industrial, specialized and service scenarios for embodied intelligence deployment. Suppliers can develop scenario-adapted products for these three categories, avoid the still-immature consumer home segment to reduce deployment risk, and actively participate in "data-model synergy" innovation consortia to access collaboration opportunities and policy support.

Engineering-stage AI deployment brings clear opportunities for factories’ digital transformation and business expansion. Key takeaways for manufacturing facilities:

1. AI deployment in factory settings requires extremely low computing latency. Current policies are pushing for the construction of metropolitan millisecond-level low-latency computing access. When building industrial AI and intelligent production systems, factories can leverage the new computing network to resolve long-standing bottlenecks of high computing access costs and slow response, improving the stability of AI applications.

2. Humanoid robots and embodied intelligence are prioritized for deployment in industrial scenarios, and policies support real-scenario model training. Factories can leverage their own production environments to participate in national specialized initiatives: they can replace manual labor with robots to cut costs and improve efficiency, and also qualify as training scenario providers to access policy and resource support.

3. Factories can join the policy-backed "data-model synergy" innovation consortia, integrate their own production process, quality standard and operational experience data with AI models to develop better-adapted industrial AI solutions, accelerate in-house digital transformation, and also generate additional revenue by licensing industry data to external stakeholders.

As AI enters engineering deployment, clear industry pain points have created defined market demand for service providers. Key takeaways:

1. The core industry pain point has shifted from insufficient computing supply to low computing access efficiency and high latency. Service providers can develop specialized solutions around low-latency computing networks, helping clients optimize their computing access layout, simplify network hierarchies, and build millisecond-level computing access to meet the requirements of large model, industrial agent and autonomous driving customers.

2. The core bottleneck for embodied intelligence today is the lack of high-quality real-scenario training data, which leaves models unable to adapt to complex, variable real-world environments. Service providers can lead or participate in the construction of real-scenario training spaces, connect scenario owners with model developers, accumulate real-machine data, and offer training data and scenario access services to embodied intelligence enterprises.

3. A core pain point for industry AI deployment today is the absence of standardized trusted evaluation frameworks, which makes clients in sectors like manufacturing hesitant to adopt AI solutions. Service providers can develop evaluation services aligned with national AI measurement systems, help AI enterprises adapt to industry requirements, bridge the gap between laboratory innovation and commercial deployment, and resolve end-clients’ concerns over reliability.

As AI enters the engineering deployment phase, shifting market demand has opened up new development directions for AI platform operators. Key takeaways:

1. Industry demand has shifted from pure intelligent computing center scale expansion to efficient computing network connectivity. AI platforms can no longer compete only on computing scale; they need to expand into computing network development, improve connectivity between computing resources and end scenarios, accelerate construction of high-speed optical transmission systems, optimize computing access layout, and build millisecond-level computing access to reduce access latency and costs to gain a competitive edge in the next phase of industry development.

2. Embodied intelligence deployment requires extensive real-scenario resources and cross-entity collaboration. Platforms can lead the formation of innovation application consortia, build real-scenario training platforms that connect scenario owners, model developers and hardware providers, and attract relevant enterprises to on board, which aligns with policy guidance and helps access additional government resources.

3. Platforms can build industry service platforms centered on "data-model synergy" and trusted evaluation to attract relevant enterprises. They should note that AI deployment is prioritized for B2B scenarios for now, while consumer home scenarios remain immature, so they should avoid early high-risk investments in unproven consumer segments.

This article summarizes the latest developments in China’s AI industry, identifies key industry shifts, and outlines new research directions. Core findings:

1. New industry developments: China’s foundational AI infrastructure is now complete. By 2025, the number of AI enterprises will exceed 6,000, core industry scale will top 1.2 trillion yuan, and total intelligent computing capacity will reach 1590 EFLOPS. The industry as a whole has transitioned from scale expansion to engineering deployment, with focus shifting from "accessibility" to "performance and usability." Three core challenges need to be addressed: efficient computing access, model adaptation to industry scenarios, and robotic operation in unstructured real environments.

2. On the policy front, multiple government departments have released policies across four dimensions: computing networks, embodied intelligence, data-model adaptation, and trusted evaluation, forming a complete policy framework to support AI deployment. The impact of this policy portfolio on industrial adoption is a key area for further research.

3. On the business model front, multiple new trends have emerged: the commercialization path for embodied intelligence differs from that of large language models, relying on real-scenario training, with B2B semi-enclosed expected to achieve viable commercial models first; "data-model synergy" innovation consortia have emerged as a new organizational model for industry AI deployment; and trusted evaluation has become a prerequisite for AI entry into high-requirement industry sectors. All of these are promising new directions for research.