Why can’t Nvidia let go of China?

Recently, Nvidia, which has attracted much attention, released its third quarter financial report.

Judging from the performance data, the growth rate of revenue and net profit is quite explosive, and the performance far exceeds industry expectations. However, it has failed to receive positive feedback from the secondary market, and its stock price has performed mediocrely in the past two days. Some people in the industry believe that the main reason is that the stock price has overdrawn the future, and its business in the Chinese market is not optimistic.

Under the relevant U.S. bans, it has been very difficult for Nvidia to do business in China in recent years. Whenever it finds new ways to "make holes through the wall," the U.S. will block them one by one after a period of time. If you ask Nvidia about its mental journey in this "cat-and-mouse game", the answer is probably " the baby feels wronged, but the baby doesn't say anything ."

Although there are many difficulties and obstacles, Nvidia still chooses to inherit and carry forward the tenacity of its founder who suffered setbacks and never gave up easily. On November 21, Nvidia Chief Financial Officer Colette Kress said that due to the impact of policies, the company’s sales to China and other affected regions will drop significantly in the fourth quarter of this year. At the same time, she also confirmed reports that the chip giant is developing new compliance chips for China.

So the question is, why can’t Nvidia let go of the Chinese market?

Source: NVIDIA China Weibo

In recent years, the concepts of virtual currency, metaverse, and AI have emerged, bringing GPU chips, which are mainly used in the gaming field, to the forefront.

In order to better run large-scale cloud AI model training tasks, NVIDIA will launch the A100 chip and H100 chip in 2020 and 2022.

According to public data, the A100 uses a 7nm process, has 54.2 billion transistors and 6912 cores, and its overall performance is 20 times higher than the previous generation.

When it was launched, the H100 was known as the world's most advanced chip , using TSMC's 4nm process, integrating up to 80 billion transistors and having 18,000 cores. It is reported that the performance is at least three times that of the A100.

NVIDIA HGX™ H200: Image source NVIDIA NVIDIA country Weibo

After the launch of the two chips, the feedback from the domestic and foreign markets was not bad.

Unfortunately, in late August 2022, the U.S. government suddenly implemented new licensing requirements for the export of some high-performance GPUs, which took effect immediately. This is like a giant net woven with iron wire, hanging over chip giants such as Nvidia and AMD. It is precisely for this reason that A100 and H100 are on the restricted export list and have no access to the country.

However, among the many giants, Nvidia is a special case. Throughout the past thirty years, this company has never lacked a radical adventurous spirit, and its founder has become increasingly frustrated and courageous in personality. How could he be trapped by the iron network full of loopholes.

September 20, 2022, less than a month after Nvidia received the ban. Huang Renxun publicly stated through the media, " Nvidia will provide Chinese customers with an alternative version based on the Hopper architecture that is not subject to US government restrictions, which is sufficient to meet the vast majority of the needs of the Chinese market ."

In November 2022, Nvidia confirmed that it will launch the A800 GPU, specially for the Chinese market. Based on the A100, this chip reduces the bandwidth of the NVLink high-speed interconnect bus from 600GB/s to 400GB/s, and other parameters have not changed much. In the first half of 2023, NVIDIA will launch a new special product-H800 based on the adjustment of H100, which also imposes restrictions on bandwidth and other aspects.

There is no doubt that the two special products have received positive feedback from many domestic technology companies after their launch - "buy, buy, buy." It’s hard to imagine how Huang Renxun they knew that Nvidia’s special products were selling like crazy in China. Really, it's hard to judge and hard to guess .

The good times did not last long. In October this year, the United States tightened relevant restrictions again. According to reports, the restrictions came into effect on October 23. NVIDIA's filings with the U.S. SEC show that the banned products that take effect immediately include the most powerful AI chips, A800, H800 and L40S.

This time, the mesh was even smaller, and Huang Renxun was even more worried .

However, the ban on the sale of A800 and H800 has been rumored for several months . It is estimated that Nvidia has been mentally prepared for this. No, according to Titanium Media’s report on November 10, Nvidia will launch three AI chips named HGX H20, L20 PCle, and L2 PCle based on H100 specifically for China to cope with the latest chip export controls in the United States. Among them, HGX H20, used for AI model training, theoretically has about 20% of the comprehensive computing power performance of H100, and adds HBM memory and NVLink interconnect modules to increase computing power costs.

Shanxi Securities believes that from a training perspective, H20 has a large computing power gap compared to NVIDIA H100 and A100, but the interconnection speed and CUDA ecosystem under NVIDIA architecture after multi-chip clustering have obvious advantages. For domestic large-scale models, the number of stacked chips can meet the computing power at a scale of hundreds of billions, giving H20 room for imagination in terms of quantity. On the other hand, stacking 2-3 H20 cards can achieve the performance of an A100 card, which will lead to a decrease in cost performance .

At this point, I have to sigh that Huang Renxun's sword skills are precise.

However, there are still uncertainties in reality. No one can guarantee whether the restrictive policies in the United States will be tightened. The so-called H20 that will be launched soon is the final version, and Nvidia’s special supply road can be iterated several times.

Jen-Hsun Huang is the helmsman of Nvidia, a trillion-dollar company, and a shrewd businessman. The reason why he goes to great lengths to launch special products for the Chinese market is obviously not because he is too busy, but because he cannot afford to do it early because there is no profit.

In this regard, a senior industry expert Wang Cheng (pseudonym) pointed out to Master Xin that there are three main reasons, namely, the high proportion of business in the Chinese market, the fear of the rise of Chinese AI chip companies taking advantage of the market gap, and the upcoming computing power bubble .

Image source NVIDIA NVIDIA country Weibo

According to Finance Eleven, citing comments from multiple server manufacturers, Nvidia’s actual revenue contribution from the Chinese mainland market is much greater than the financial report data, because companies in Taiwan, such as Asus and Gigabyte, have integrated Nvidia’s chips, and a large number of them eventually Customers are still in mainland China. NVIDIA's third quarter financial report shows that revenue from mainland China (including Hong Kong) was US$4.03 billion, accounting for 22.4%. Revenue from Taiwan, China was US$4.23 billion, accounting for 23.5%. In other words, the total revenue of the two major markets was US$8.26 billion, accounting for 45.9% .

On the other hand, China is developing rapidly and has strong terminal demand. Once it is far away from this growing market, China's local AI chip manufacturers will inevitably quickly fill the gap. Although there is still a large gap in product technology between the two sides in the short term, in terms of long-term impact, this will This trend is dangerous for Nvidia.

For domestic companies, NVIDIA's high-end GPU chips and CUDA ecosystem are not only usable, but also easy to use! But the premise of everything is that it is useful !

When the above premise does not exist, the probability of domestic substitution tends to be infinite. For local chip manufacturers, what they lack most is not products, but the market and opportunities for continuous improvement and iteration based on downstream needs. In this magical land, there are already countless cases where if given a chance, it will definitely work.

Huang Renxun has expressed concern about the rise of China's domestic chip industry chain on multiple occasions. "Restrictions on exports to China will only force Chinese companies to develop and produce their own chips and become competitors of the United States." In June this year, Nvidia CEO Financial Officer Colette Kress also emphasized that although restrictions on the export of artificial intelligence chips to China will not have a significant immediate impact, it " will result in a permanent loss of opportunities for the U.S. industry ."

Source: NVIDIA country Weibo

In fact, without the influence of the external environment, Nvidia may be able to rely on its technological and ecological advantages to greatly delay the growth of its Chinese competitors, just like Microsoft. In 2018, Academician Ni Guangnan once said that starting from Win7, Microsoft has allowed you to install pirated versions. Strictly speaking, this is a "strategy" of Microsoft. Microsoft uses piracy "strategy" to deliberately let Chinese users use it for free, denying the Chinese people the opportunity to develop their own operating systems .

Wang Cheng believes that the current industry may have a computing power bubble , and Nvidia wants to sell more products and make more money while the market is still going strong.

Since October last year, ChatGPT has once again ignited the industry’s enthusiasm for artificial intelligence. Since the beginning of this year, many technology companies, including Microsoft, Meta, Baidu, and ByteDance, have rushed to build large artificial intelligence models, striving to stay on top of industry technologies and hot topics.

But if you want to achieve rapid iteration of large models, NVIDIA's high-end GPU chips are the best choice. As a result, under the tumult, Nvidia's stock price climbed from the bottom to the top of the mountain, becoming another technology company with a market value of trillions of dollars.

Wang Cheng believes that many technology companies are currently promoting large models, but everyone is telling stories but cannot explain how to charge, and large models have not yet found an effective business model . Maybe in about two years, the hot spots caused by ChatGPT will dissipate, and so much computing power will no longer be needed. According to later reports, TSMC Chairman Liu Deyin said at this year’s second quarter results meeting that it was unclear whether the hot demand for AI was a short-term bubble . But NVIDIA's orders are there, and TSMC can only choose to significantly expand production to keep up.

Source: NVIDIA country Weibo

It is worth mentioning that before this AI craze, Nvidia had just experienced a mining machine bubble in the virtual currency circle. At that time, the weak demand for graphics cards from mining machines directly led to Nvidia's sluggish performance. According to public information, Nvidia’s revenue in the second quarter of fiscal year 2023 (the three months ending July 31, 2022) was US$6.7 billion, which was far lower than the US$8.1 billion quarterly revenue guidance it gave in May last year, and the current period Net profit was 30% less than expected, and inventory grew rapidly .

Artificial intelligence is not new. It has been developed for more than 80 years since 1956. During this period, the industry has experienced several ups and downs. As for whether this spring breeze led by ChatGPT is a bubble or a real industry change, I believe experienced Nvidia executives will have their own judgment.

"We don't have to pretend that the company is in danger. Because we can feel that the company has indeed been in danger." A strong sense of worry and anxiety helped Huang Renxun lead the company out of crises several times and successfully stood at the top of the industry .

His words also apply to the integrated circuit industry chain that is currently growing and growing. In the past thirty years, due to a slow start and little accumulation, the domestic industry had to face the fact of falling behind. Now that the external environment is turbulent, local companies should not wait and wait. Instead, they should proactively seize opportunities to seize the market and at the same time unload their impetuous research and development of technology products.

30 years ago, Academician Ni Guangnan called on China’s information industry to master core technologies such as operating systems and CPUs through independent innovation. He believed that core technologies cannot be bought, begged, or begged. Only by achieving core technologies Only by being independent and controllable can we not be controlled by others, and only then can we promote the development of China's information industry .

The content of this article is for communication and learning purposes only. If you have any questions, please contact us at info@gsi24.com.