Jiwei reported (by Chen Bingxin) that the training of China's AI large models using US cloud services may be restricted. On January 29, the U.S. Department of Commerce officially announced in the Federal Register a proposal to take additional measures to address a national emergency related to significant malicious cyber behavior. The proposal explicitly requires U.S. IaaS vendors to verify the identity of foreign users when providing cloud services, restrict foreign actors' access to U.S. IaaS products, and require detailed reporting of foreign transactions to train AI models to protect U.S. cybersecurity and interests.
U.S. Commerce Secretary Gina Raimondo also said on January 26 that Biden would propose to require U.S. cloud computing companies to determine whether foreign entities are accessing U.S. data centers to train AI models. Cloud service providers such as Amazon, Microsoft, and Google, a subsidiary of Alphabet, are required to actively investigate and disclose foreign customers who develop AI applications on their platforms, and require the disclosure of details such as the names and IP addresses of foreign customers. If this policy goes into effect, the U.S.** could use these requirements to cut off the main avenues for Chinese companies to access data centers and servers.
In this regard, spokesman Wang Wenbin said that the development and governance of artificial intelligence is related to the fate of all mankind, and what is needed is to work together and coordinate response, rather than decoupling and breaking chains, and building fences and walls. We urge the US not to violate the objective laws of scientific and technological development, earnestly respect the principles of market economy and fair competition, and create favorable conditions for strengthening international coordination and cooperation in the field of artificial intelligence.
In recent years, driven by the ChatGPT craze, the number of large models in China has surged. According to statistics, by the end of last year, the number of AI models released in China exceeded 200. At present, the model level of the leading domestic manufacturers is close to GPT35. Various manufacturers are actively deploying to catch up with GPT4. Behind the training of every large model means a massive amount of computing power investment. Moreover, since the fourth quarter of last year, with the steady increase in the number of daily active users of online products, the demand for inference has also increased significantly, including applications such as drawing and generation, and the demand for AI computing power on the inference side has also continued to rise.
According to industry insiders, for head model manufacturers, they generally build their own computing power platforms. Therefore, most of the training of related models adopts a hybrid mode, some of which are completed on their own computing power platform, and some of which are based on cloud computing power. Some domestic large models have even begun to be trained based on domestic computing power such as Ascend, such as iFLYTEK's Xinghuo large model. In this way, the impact of the new regulations issued by the U.S. Department of Commerce on the domestic model training field is still relatively controllable in the short term. However, it is also pointed out that the situation may change when the model is trained towards larger parameters such as GPT5 in the future.
However, for those small and medium-sized model manufacturers who do not have enough funds to build their own AI computing centers, the introduction of the new regulations will have a certain impact. Due to the strict export control of A100, H100 and other high-computing GPU chips required for large model training in the United States, domestic AI large model training is facing a "core shortage" situation, and the computing power gap is relatively large. It is a certain way to obtain the computing resources required for AI model training by subscribing to cloud services. Many small and medium-sized model companies are unable to build their own computing platforms, and even prefer to adopt cloud service models, which can be supported by professionals. Once the new regulations come into effect, it is expected that some companies that use this method to obtain computing power will be cut off by the U.S. Department of Commerce, which will affect the training progress of large models.
But it's also important to see that cloud service leasing at this scale is quite expensive. For example, Nvidia's AI supercomputing service, DGX Cloud, at GTC, has 3 instances per instance$70,000 per month, including 8 A100 or H100 accelerator cards, if you choose to rent 10,000 (1,250 instances) accelerator cards for one year, you need 3.8 billion yuan. As far as the domestic cloud computing power leasing market is concerned, the leasing of cloud computing power is 10% to 15% compared with October last year. At present, the server rent of 8 cards and A100 cards is about 90,000 months, and the annual payment of large customers will be discounted by 6-7 (about 120,000 p per year).
In addition, for some large model training that involves sensitive data or affects ***, the use of US cloud services also has relatively large data security risks. At this point, regardless of whether or not the U.S. restricts it, and to what extent, it may not be sustainable to acquire computing resources through overseas cloud services in the long run.
In the interview, some industry insiders pointed out that the expansion of U.S. export controls, in the long run, will benefit domestic cloud service providers, thereby accelerating the process of China's cloud services going overseas. The advent of the AIGC large model era has made intelligent computing power a common demand, which will accelerate the development of cloud computing services. In 2022, the global cloud computing scale reached 356.6 billion US dollars, and it is expected to exceed 400 billion US dollars in 2023. According to IDC data, Amazon AWS, Microsoft, Google, Alibaba, and IBM account for more than 51% of the global market share. AWS currently has a market share of 86% (after Alibaba Cloud, Huawei Cloud, e Cloud and Tencent Cloud).
Judging from this data, since the market share of overseas cloud service providers such as AWS in China's domestic market is not high, the impact of the implementation of the new regulations on the revenue of US cloud service providers in China may be relatively limited. However, for Chinese cloud vendors, it will be a good thing in the long run, not only to obtain the domestic cloud service market that some overseas cloud service providers "spit out", but also for those Chinese enterprises that go overseas, they originally chose AWS and Azure, and may switch to Chinese cloud service providers in the future, thereby accelerating the process of Chinese cloud services going overseas.
Of course, the development of China's AI cloud computing power is still based on AI chips. At present, some leading model manufacturers in China have begun to deploy AI models based on Ascend. It is reported that the theoretical calculated value of the Ascend 910B card is close to that of the NVIDIA A100, but there are still some uncertainties in terms of ecosystem and yield. However, in the context of the intensification of the US blockade of China, domestic computing power represented by Huawei Ascend and others is expected to develop further.