In the U.S. capital market, the Internet has been tepid recently, and AI computing power has become more and more frenetic. At this time last year, A-shares also set off a wave of AI computing infrastructure hype, but from the beginning of the year to the end of the year, from **back and forth**. Many people are very anxious, worried about whether this year will repeat the mistakes of last year and end up in a mess, so recently, I have frequently received repeated inquiries from enthusiastic fans, hoping that I can systematically explain the domestic computing power market. Satisfy everyone's appetite today.
Description: 1. The focus of this article is to discuss the structure and logic of the domestic computing power market, and it is not recommended that the name of any listed company appear in the article is only to assist in explaining the author's logic, and does not represent a good comment on the company. **There are risks and investment should be cautious. 2. This article does not include the chip link).
Before AGI came up, the server market industry chain was divided into chip manufacturers, machine manufacturers, and manufacturers, and these three levels were distinct, each making its own money, and it would not penetrate each other (cloud service large customer market, no ** business, direct supply). However, after AGI became popular, NVIDIA changed the industrial chain and took advantage of the scarcity of chips to do the work of the whole machine. There are three reasons for this:
1. The GPU computing power chip itself is very complex, and the general downstream foundry does not have high-quality production capacity, so it is more conducive to rapid market introduction by chip manufacturers for integration.
2. The cost structure of AI servers, the value of memory devices is comparable to that of GPU chips, including optical modules, switching chips, PCBs, etc., which are all high-value components, and chip manufacturers can quickly expand their revenue by vertical integration. (NVIDIA Performance +++.)
3. It is conducive to the formation of a closed-loop ecology for chip manufacturers and the construction of insurmountable obstacles for other competitors.
According to IDC statistics, the global AI server market size is expected to be $21.1 billion in 2023 and reach 317 in 2025$900 million, with a CAGR of 22 for 2023-20257%。In terms of procurement volume, according to Trendforce, as of 2022, the total share of Microsoft, Google, Meta, and AWS, the four major cloud service providers in North America, in global AI server procurement, has reached about 66%, and Microsoft and Google have a global share of 19% and 17%, respectively.
So, it's easy to understand that the U.S. market has most of the demand coming from big cloud service providers, to which NVIDIA supplies directly. As NVIDIA's royal foundry (hometown relationship), SMIC produces the whole machine through localization, and then drinks soup and eats meat. FII is also an important player. The only other is optical transceivers, which are also the only industries that Chinese companies can participate in the U.S. computing infrastructure market.
1. The domestic AI computing business model is completely borrowed from NVIDIA's model, and there is basically no such link as an independent machine manufacturer. Therefore, you have seen that in 2023, Inspur Information and Xinhua.
Third, the head server manufacturers such as Super Fusion, ZTE, and Lenovo not only did not grow last year, but shrank significantly. The reason is the change in the business model, which leads to the redivision of the market pie.
2. The domestic AI computing power market demand is highly similar to that of the United States. Internet: Government and Enterprise = 70%: 30%, Public Bidding Market (Internet + Operator): Government and Enterprise = 80%: 20%. In other words, even if you are an OEM, you must participate in the Internet market, otherwise you will only have a very small piece of the cake.
3. Large Internet factories have strong product development and design departments, in addition to AI servers, storage, heat dissipation, communication and other components, adopt a white-box model, and entrust ODM manufacturers to produce**. This piece is mainly the Industrial Fortune Federation, Inspur Information, and Xinhua.
3. The sphere of influence of these old foundries such as Lenovo. The net profit margin is about 4%, because the future growth is very fast, these companies should be able to achieve better growth and ROE through high turnover.
4. For 20% of the government and enterprise market, in view of the uniqueness of the land where we were born and raised, the situation is very complicated, just like small potatoes scattered all over the country. Generally speaking, slightly larger industries (such as banks) are basically directly controlled by chip manufacturers and can be regarded as Internet-like. For the remaining 10% of other industries and ZF, because they lack system design capabilities and technology, they will generally find the general contractor of the company under their own lineage for system integration. If you can't understand it, take a closer look at Sugon, his main business is the general contracting of government and enterprises in the north, and he is a typical representative. The biggest advantage of these companies is that they live like people; The biggest disadvantage is that living in a greenhouse, you can't grow up, and you can't go to the big Internet market to swim. However, it should be noted that the ZF division has limited spending capacity in the coming years, and the growth rate of this segment is in doubt.
5. Compared with computing power servers, the business model of storage servers and communications (switches + optical modules) has not changed, benefiting from the surge in AI demand, storage capacity + transportation capacity, these two pieces are still relatively friendly to domestic enterprises. For example, optical module manufacturers in China, which are the main markets of optical transceiver manufacturers in China, is a boon. Of course, for companies like Zhongji InnoLight that are the first in the American market and the Chinese market, it is not a problem.
6. For the industry of H.
chain, a unique "** mode" has emerged. As mentioned above, H wants to eat the head of the fish to the tail, but H is the banner of domestic AI computing power, the absolute representative of high technology, and the development of new quality productivity is a priority for any place, and H also needs to strive for the support of various places (such as R&D sites, data center infrastructure...).Based on these common interests, there are many ecological chain partners such as "high-tech development". H is responsible for making chips, modules, and boards, and then the local enterprises simply process a chassis, power supply, cables and other auxiliary materials, and orient them to the local government and enterprise market. This model, in essence, is a kind of "* quotient". Someone is about to ask, why do you want to make such trouble, isn't it over if you just do it? The difference is that if the local enterprise only does the best type, it is not a "high-tech enterprise", and it does not represent "new quality productivity".
At present, H has been with 12 places (Sichuan, Henan, Anhui, Hunan, Guangzhou, Shenzhen...).) established a relationship.
Some people compare H's computing industry with smart cars, but it is actually unreasonable. The smart car is H takes the initiative to find others, so he must protect the interests of his partners, otherwise he will not play with you. The computing industry is actively looking for H in various places, and H is relatively passive, because of special circumstances, H can't help himself, more like a piece of Tang monk meat, and every place wants to take a bite.
1. Chip manufacturers are dominating everything, which is the biggest premise.
2. The business model of the AI computing power market has undergone fundamental changes, and chip manufacturers have gone from fish head to fish tail, leaving a sharp decrease in profits for the industrial chain.
3. Help chip manufacturers do more complex work (complete machine production, liquid cooling and heat dissipation, switches, optical modules), you can eat meat and drink soup, and the rest, only gnaw bones.
4. Domestic chip manufacturers, in addition to switches and optical modules, have also done their own work.
5. For some fragmented government and enterprise markets, some enterprises are flowers in the greenhouse, and their lives are more comfortable, but they do not grow much.
6. From the beginning of the year to the end of the year, it is a high probability event to come and go. Most of the targets are basically rotating in place.