Original Source: Brain Extreme

Image Source: Generated by Wujie AI

On October 17, the United States strengthened the ban on AI chips for the Chinese market. It specifically includes performance and density as export control standards, with chips exceeding 300 teraflops of single-chip computing power and performance density exceeding 370 gigaflops per square millimeter being included in the prohibited export list.

Although this ban also affects high-end AI chips provided by companies such as AMD and Intel, due to NVIDIA's dominant position in the global AI chip field, this chip ban is also referred to as the "NVIDIA ban."

The news immediately sparked heated discussions in the AI industry. However, most of the discussions focused on when the ban would be implemented, whether there would be a buffer zone, which specific GPU models would be affected, and how long the ban would last. This has led to an obvious macro background being overlooked: since the start of the trade friction, although there have been fluctuations in Sino-US economic and trade relations, chip control for the Chinese market has continued to increase rather than decrease. In particular, the ban on high-end AI chips has been firmly implemented amidst various controversies.

At this point, it seems that the AI industry must form a basic consensus: abandon illusions and prepare for struggle.

Instead of focusing on which GPUs are banned and whether there is a possibility of lifting the ban, it is better to re-examine where Chinese AI computing is headed in the era of chip control.

This article hopes to start from the current industrial situation and think about the future development of AI computing together.

Current Situation

First, we must understand a question: why did the public opinion and the reaction of the AI industry to the NVIDIA ban this time not seem as severe as the chip blockade in 2018 and 2019 when it just started? It seems that the only issue that sparked some debate among gamers and related merchants was whether the consumer-grade graphics card RTX 4090 would be banned.

The macro logic behind this is that the industry is very unwilling to see high-end AI chips being banned, but in fact, it has long anticipated this situation. On the one hand, the US has been pushing for chip blockades against China for many years, and some high-end GPUs from NVIDIA had already been banned before. The industry's reaction had long shifted from astonishment to calm acceptance. In addition, after the explosion of ChatGPT, the global high-end GPU market suddenly surged this year, followed by repeated statements from the US side to push for a comprehensive ban on high-end AI chips to China. Several months later, the final ban was no surprise.

In order to cope with this foreseeable ban and the objective promotion of large model development, many Chinese technology, finance, and automotive companies concentrated on hoarding high-end NVIDIA GPUs from the end of last year to the first half of this year, which has already caused a shortage of GPUs in the market. In other words, for many small and medium-sized Chinese technology companies and AI startups, high-end GPUs were already difficult to obtain, and the ban did not change much after the ban.

Another reality is that high-end AI chips are not irreplaceable by domestic production. Since the beginning of the trade friction in 2018 and 2019, the domestic AI chip industry has accelerated its development. This has led to the fact that although it is difficult to replace high-end NVIDIA GPUs in AI training demand, they are not irreplaceable.

In addition, AI chips are not as consumer-oriented as mobile chips, and Huawei has already made breakthroughs in mobile chips. All these signs together have led to a calm and even somewhat unsurprised attitude towards the ban, both from the public and the industry.

However, it must be objectively recognized that this ban is not without harm to the Chinese AI industry. On the one hand, it is very difficult to replace NVIDIA GPUs in the short term, and there are significant challenges in terms of chip production capacity and ecosystem compatibility. Moreover, the ban will directly harm manufacturers who use NVIDIA products on a large scale in AI servers and other fields.

The more critical issue is the future. If the ban persists in the long term, Chinese AI computing will gradually decouple from the global high-end chips, which may bring about very complex long-term negative impacts, such as:

1. After the update and iteration of NVIDIA's high-end GPUs, will the development of Chinese AI computing power be disconnected?

2. After the development divergence of underlying computing power, will the Chinese AI industry lag behind in the development of large models and other software technologies?

3. This AI chip ban has already shown the characteristics of a large-scale blockade. Will this kind of technological blockade only stay in the field of AI chips? Will digital basic capabilities such as general computing power, storage, and basic software become the next targets of attack?

In short, this AI chip ban is a prepared game for China, rather than a surprise attack. To successfully pass this test, every card in hand must be played well, with coordinated efforts to reduce the intensity of damage and increase the chances of long-term development.

Currently, there are three "breakthrough plans" that China's AI computing must simultaneously focus on.

Plan 1: Make Good Use of the "Buyer" Identity

There is a simple truth that the logic of the commercial market is determined by supply and demand. However, in the Sino-US technology trade represented by chips, we often fall into a mindset that the vast majority of the game rules are set by the US government and companies. We think that if they want to sell, we will buy, and if they don't want to sell, we have no choice.

As the largest "buyer" in the global chip market, Chinese companies have no say, which is very absurd.

In fact, the most direct harm of the ban on AI chips for the Chinese market is to the US tech giants represented by NVIDIA. Currently, the largest market demand for NVIDIA's AI chips comes from China. Previously, NVIDIA CEO Jensen Huang explicitly stated, "If we are deprived of the Chinese market, we have no contingency plan. There is no other China in the world."

In this situation, we can see that there is a clear contradiction between US tech companies pursuing commercial interests and the US government pursuing political interests. US tech companies always try to find ways to oppose the ban and circumvent it. For example, NVIDIA has been introducing special versions of GPUs for the Chinese market to cope with the embargo policy since last year.

The Chinese market absorbs about one-third of the production capacity of US tech companies, and the link between the two is inseparable in the long term. Faced with the increasingly obvious comprehensive technological blockade by the US, the Chinese market should also actively make good use of its "buyer" identity, making its own behavior more distinct and predictable.

Avoid creating an impression of "welcome when sold, helpless when not sold."

The "buyer" identity should be a position with strength and anger.

Plan 2: Use Cloud Instead of Cards, Concentrate Computing Power

For the foreseeable future, the US ban on AI chips to China is likely to only strengthen, and this timing coincides with the critical stage of the development of large models in AI. Many industry insiders believe that although the development of large models is fast, it has not shown the rapid momentum of previous technological trends, and the main reasons are lack of investment and lack of computing power.

So how to solve the computing power gap for the Chinese AI industry under the ban? The first emergency plan is for companies to increase the configuration and investment of AI computing power in the cloud and promote the use of cloud instead of cards.

In fact, under the trend of the possible ban on high-end AI chips, several major public cloud providers in China have taken actions to strengthen the hoarding of high-end NVIDIA GPUs. On the one hand, this is because the cloud providers themselves need to increase investment in large models and open up the MaaS market, so they have a direct demand for AI computing power. On the other hand, after GPUs are transformed into cloud resource pools, they can be reused in the long term, which is an offensive and defensive situation for cloud providers. Therefore, in the first half of this year, there was a situation where all high-end AI chips on the market flowed to cloud providers, and small and medium-sized enterprises found it difficult to obtain them.

Objectively speaking, this concentration of high-end AI chips to the cloud is beneficial for the overall Chinese market to cope with the ban on AI chips, and it also conforms to the strategic thinking of "Eastern Calculation and Western Algorithm."

Another trend favoring cloud-based AI computing power is that as the parameters of large models and the amount of data used continue to increase, local card pool training has become increasingly tight. It is necessary for training with thousands or tens of thousands of cards in the cloud to become the main development direction in the future, so enterprise users will naturally be more proactive in moving towards the cloud.

At the same time, cloud-based AI computing power will not only stay at the level of hoarding high-end NVIDIA GPUs. Next, domestic independent AI computing power entering the cloud is the general trend. With the promotion of relevant policies, cloud providers are increasing their procurement of independent AI chips. According to IDC data, in the first half of 2023, China's AI servers had already used 500,000 self-developed AI accelerator chips. In terms of service-oriented independent AI computing power, Huawei Cloud has already launched the Ascend AI Cloud Service. The combination of cloud-based and independent AI computing power will see significant development under the background of the ban on AI chips.

In addition, in recent years, under the background of "Eastern Calculation and Western Algorithm," a large number of AI computing centers using independent AI computing power have been established in various regions, and overall, China's cloud-based AI computing power is stable in supply and reliable in guarantee.

However, many enterprises still prefer to purchase local AI computing power. On the one hand, this is because the market for NVIDIA GPUs is tight, and they hold their value particularly well, and can even be used as core assets for enterprises. On the other hand, cloud-based AI computing power often faces issues such as queuing, downtime, and lack of software services.

How to further improve the user experience of developers in using cloud-based AI computing power is the direction that public cloud providers need to focus on next.

Plan 3: Ignite Explosive Growth of Domestic AI Computing Power

Facing a new round of AI chip bans, what is the biggest confidence of the Chinese AI industry? Is it the accustomedness after many years, or the abundance after hoarding a large number of cards? Neither. The most crucial point is that after many years of development, the Chinese AI chip industry has made tremendous progress. NVIDIA's high-end GPUs are indeed still important, but there are already alternative options.

According to the data previously released by IDC, the shipment volume of AI accelerators in China in 2022 was approximately 1.09 million units, with NVIDIA holding a market share of 85%, Huawei's Ascend occupying 10%, Baidu's Kunlun at 2%, and Cambricon and Horizon Robotics each at 1%. From this, it can be seen that domestic AI computing power has achieved a certain degree of market share, rather than just being a concept or theory. At the same time, it should also be noted that domestic AI chips still have a long way to go in terms of core performance, software ecosystem, and shipment capacity. Under the objective conditions of the NVIDIA ban, domestic AI computing power must overcome these difficulties in the short term and accelerate its own growth and maturity cycle. To achieve this goal, several things are very important: 1. Forming industry consensus to avoid conceptual confusion. 2. Moving towards large-scale commercial use to avoid chip-making based on PowerPoint presentations. 3. Strengthening the software ecosystem and enhancing migration capabilities. 4. Increasing support for "main brands" to form a large-scale effect. The NVIDIA ban is a situation that the Chinese AI industry generally does not want to see, avoid as much as possible, and even has some deeply hidden issues today. However, the situation is stronger than individuals, and in the accelerating process of deglobalization and AI competition, similar bans are likely to increase rather than decrease in the future. Avoiding, circumventing, and remaining silent will not solve the problem. Only by facing it calmly and striving for self-improvement can we fundamentally solve the problem. Under the ban, where will Chinese AI computing go? The answer is that we have no choice but to give the world a second choice.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。