Part 1 — An Economic Policy Based on Data Sunway TaihuLight China has succeeded in producing the first supercomputer to be listed in the ‘Top500’ and the ‘Green500’s’ top 5 using only Chinese components manufactured in China, even though it does not possess the most advanced production technologies at global level. To do so, it relied on preceding technological generations, which it succeeded in implementing in an innovative architecture. Accelerating Setup of European AI Infrastructure There are several types of infrastructure required for development of AI, covering the various phases from research to development and on to marketing of the product itself. In certain AI fields, such as machine learning, life cycles comprise two main phases: the learning phase and the inference phase. The speed and performance of the learning phase are conditioned by the scale of the material resources allocated, in particular as regards dedicated processors (GPUs, for example). Hence, infrastructure size conditions productivity and efficiency of research and development. The second phase, that of inference, has much less need of material resources, and can even be carried out inside embedded peripheries (an AI in a smartphone). Learning and Inference This is basically how AI techniques based on learning work: first of all, they go through a learning phase during which an algorithm seeks out all parameters enabling the model to carry out the required task at the best possible performance level. Once this phase is over and the model’s parameters are set, an inference phase follows in which the task for which the model has been trained during the learning phase is carried out. During the learning phase, one must distinguish between several types of workflows. Cases in which a supercomputer dedicated to AI calculation is fully mobilized (typically with resources numbered in thousands of GPUs) are quite rare and only concern a limited field of research. The great majority of applications require far fewer resources (numbered in dozens of GPUs, for example). This type of need that could well be met by an “AI cloud”. Setup of such an infrastructure requires very considerable investment and is the preserve of a specialized branch of activity: infrastructure, data centers and the cloud itself have to be taken into account. It is therefore a matter of pooling such resources as far as possible, at least for the public authorities overseeing the development of key sectors. 53