Access Official Technical PDF

Share Summary:LinkedInEmail

Industrial AI depends on more than a promising algorithm or a successful pilot. It also relies on the software frameworks, computing hardware, data, and specialist expertise needed to develop, train, and maintain AI systems. China’s AI development debates, as documented in a 2019 policy report, offer a useful perspective on these AI infrastructure dependencies—and on the trade-offs involved in building a technology base that can support long-term innovation.

For automation engineers, system integrators, and operations leaders, the central lesson is practical: evaluate the foundations behind an AI application, not only its immediate use case. A model may perform well in a demonstration, yet its development and deployment can remain dependent on resources that are difficult to control or replace.

AI infrastructure dependencies begin below the application layer

AI applications are built on layers of enabling technology. The report describes deep learning frameworks such as TensorFlow and PyTorch as programming tools that help researchers and engineers design, train, and experiment with models. By handling common mathematical operations, frameworks let teams work at a higher level than implementing every model component from scratch.

Chinese alternatives existed, including Baidu’s PaddlePaddle, but the report found that they had not gained comparable traction at the time. This matters because a widely used framework can benefit from a larger community of users and contributors. A smaller ecosystem may find it harder to attract collaborators, improve the software, and compete for developer attention.

In industrial settings, this layer affects how teams develop and maintain machine learning solutions. A framework choice can shape the tools engineers use and the expertise they need. The report does not establish that one framework is suitable for every industrial application; rather, it illustrates why the software ecosystem supporting a model is part of the technology decision.

AI hardware: training and inference are different workloads

The report also distinguishes between training a neural network and inference, or running a model that has already been trained. These workloads do not necessarily rely on the same hardware. At the time, some Chinese AI chips were mainly used for inference, while training large neural networks required powerful hardware capable of carrying out many calculations in parallel.

Graphics processing units (GPUs), particularly those produced by Nvidia, were widely used for this training work. Other platforms, including Google’s Tensor Processing Units and field-programmable gate arrays from companies such as Intel and Xilinx, were also discussed as alternatives. The report noted that these alternatives were not Chinese platforms.

It also described how some Chinese chip companies built products on top of technologies developed elsewhere. DeePhi Tech’s processors relied on Xilinx FPGA frameworks, while Horizon Robotics’ architecture was built on Intel FPGA technology. These examples show why the label “domestic” does not automatically mean a technology stack is independent of external components or platforms.

For industrial AI planning, the distinction between training and inference helps clarify what a proposed system actually needs. Teams should identify where models will be developed, where they will run, and which hardware and software layers those activities depend on. The report does not prescribe an architecture, but it shows why examining these layers can reveal dependencies hidden by an application-level description.

Model performance depends on data and operating conditions

Foundational infrastructure is only part of the challenge. Chinese industry commentary cited in the report described difficulties obtaining useful data, aligning training data with real operating environments, and retaining experienced AI specialists. One account described factory fault-detection work in which records had to be collected manually from tires stored outdoors. The example illustrated how data collection and data quality can undermine a project before model performance is even assessed.

That account also raised a broader issue: training data may not reflect the conditions a model encounters in operation. For automation teams, this makes data readiness and validation essential parts of project planning. Data volume alone does not establish that records are accurate, complete, or relevant to a specific task. Nor does a successful test on prepared data, by itself, demonstrate that a model will behave reliably in its intended operating environment.

The report further noted that business-to-business AI projects can require specialized, on-site work and that senior technical staff may be difficult to retain for lengthy deployments. These constraints are relevant to integration planning: technical expertise must be available not only to develop a model, but also to adapt it to a customer’s data and application.

Foundational research versus rapid implementation

A key debate in the report concerned whether AI leadership depends mainly on applying existing breakthroughs or on contributing to further foundational research. One view, associated with investor Kai-Fu Lee, held that AI had entered an “age of implementation,” in which the priority was applying deep learning to specific uses. Other researchers argued that continued advances in algorithms and methods remained important.

The report described concerns that commercial investment tended to favor applications with more immediate routes to market, while foundational components attracted less attention. It also identified research challenges, including talent moving overseas and the difficulty of building labs that can attract and retain leading researchers. These observations underline a tension familiar to technology planners: near-term deployment and long-term capability-building can compete for resources.

What automation leaders can take from the debate

China’s experience also featured disagreement over the role of government in AI development. Some commentators favored a stronger market-led approach to research funding and resource allocation; others discussed state-backed programs and different routes to innovation. These arguments show that national strategies do not translate automatically into a single, settled development path.

For industrial organizations, the practical implications are less about adopting one national model and more about asking disciplined questions before committing to an AI initiative:

  • Map the technology stack: Identify the frameworks, processors, and other enabling technologies on which development and deployment depend.
  • Separate training from inference needs: Confirm what computing resources each stage requires rather than treating “AI hardware” as a single category.
  • Assess data readiness: Examine data accuracy, completeness, collection methods, and relevance to real operating conditions.
  • Plan for specialist effort: Account for the expertise needed to adapt and support a solution in its intended environment.
  • Balance deployment and foundations: Consider immediate application value alongside the capabilities needed to maintain and advance systems over time.

Conclusion: assess the foundations, not just the promise

Industrial AI projects are shaped by dependencies that sit beneath the visible application: software ecosystems, training hardware, data quality, and specialist talent. China’s AI debate, as captured in the 2019 report, shows that strong commercial activity and ambitious development goals do not necessarily eliminate weaknesses in these foundations. For automation decision-makers, the takeaway is to make those dependencies explicit. A clear view of the full technology and data stack supports more grounded decisions about feasibility, resilience, and the resources required to move from a promising model to a maintainable industrial application.

Access Official Technical PDF

About The Author