Summary:
1.Server industry trends and outlook analysis
In 2023, the domestic server market will be squeezed by ordinary servers, while the demand for AI servers will grow, but it will be subject to instability caused by the external environment.
The growth of global AI servers shows an irregular rhythm, and the market share of domestic AI server cards has increased at the end of 23, and the general server market is still rising in 2024.
In 2024, the sales volume of general-purpose servers is estimated to be 2002.2 million units, accounting for 40% of the Internet industry; The AI server market is dominated by domestic cards, and the number of domestic training cards** is about 250,000.
2.Analyze the AI server market and NVIDIA dynamics
China's AI server market is expected to have 500,000 H800 units in demand in 2024, but Nvidia's market share has halved due to export restrictions, and the remaining 250,000 units are shared by HW and Haiguang.
Nvidia will actually ship about 2.03 million AI and H series training cards in 2023, and it is expected to be nearly 4 million in 2024, of which H100 is the mainstay, and it is expected to be 3 million; The A100 was drastically reduced to 200,000 sheets.
Overall AI server sales are expected to grow by about 30% in 2024, while training server purchases will slow in the fourth quarter as major customers turn to inference servers and cards.
3.Nvidia market control and strategic analysis
Nvidia is facing intensified competition, but the training card in the North American market is basically in a monopoly position, AMD and Intel products are difficult to pose a threat, and the self-developed GPUs of Internet manufacturers are mainly used internally.
In the domestic market, due to the competition of HW and other companies, the price of Nvidia H20 has been reduced, and the profit is not high; The profit in overseas markets can reach 50-70%.
Nvidia has adjusted the pricing of the H100 to $22,000 and reserved space for the B and HR versions**, so the B100 and H200** will not fall below customer expectations. There will be no further price reduction of the H100 this year, because mass production has already begun, and the market acceptance is good.
4.ODM market share and profitability analysis
Stable North American market: Taiwan's three major ODM OEMs have a relatively stable market share in North America, and do not cut prices to grab market share to maintain high profits, and maintain order market share within the production capacity.
Changes in profitability: Last year, there was a subsidy for the initial mass production of F7 platform servers, and this year, with the increase in production and the decrease in costs, the gross profit margin has decreased.
New platform products will continue to deliver high gross margins: The H200 and B100 platforms are expected to start mass production in 2025, and although the unit price has decreased, there are still high gross margins due to factors such as initial subsidies.
5.Board and module market development outlook
Server card and module market status: In 2024, H100 boards will still be 100% OEM by FII, 30% of its modules will be OEM by FII, and the remaining 70% will be OEM by Taiwan Flextron. In 2025, this situation is expected to continue, and the B100 board module will continue to be shared by these factories.
H20 board performance and ** changes: The performance of the first version of H20 in November 2021 is comparable to that of H800. Due to the U.S. requirement to reduce computing power, H20 performance was reduced to 70% of H800's, and *s were also lowered accordingly. The amount of smuggling from Nvidia's server illegal channels is very small, and regular customers do not purchase products without warranty.
AMD estimates in the server market: AMD expects to ship 350,000 training cards in 2024, but it is still in the beta phase. AMD's success depends on the migration from the NVIDIA ecosystem to the AMD platform.
2024** and market estimates for the HW Ascend series: about 85,000 RMB for the 910B**, and 150 1.6 million RMB for the whole training server; The 910c** is expected to reach 17,180,000 RMB. HW's market share in the domestic training card market is expected to reach 30 to 40%.
Minutes**: Wen Bagu Research] Mini Program
q&a
Q: How will the market performance of China's server industry perform in 2023? Including sales, shipments, market structure and downstream customer characteristics? And the market outlook for early 2024, the trend change between general-purpose servers and AI servers?
A: In 2023 as a whole, the server industry has experienced rapid growth in the field of AI servers, while general-purpose servers have suffered a squeeze on market share. This phenomenon is mainly caused by several factors: first, end customers spend more money on the purchase of AI servers under a fixed budget; Secondly, with the switching of platforms, the enthusiasm of end customers for the purchase of general-purpose servers is weakened; In addition, with the release of next-generation platforms by companies such as Intel, the cost of related components remains high, making it difficult to increase the purchase of general-purpose servers. In 2023, domestic AI servers will grow rapidly, but the growth will not have a fixed rhythm, partly because GPU card providers have been affected by US sanctions, especially Nvidia GPU cards, which are extremely unstable. However, with the imposition of sanctions on its products by the United States, domestic alternatives such as HW's Kunpeng and Haiguang have begun to gain market share, providing a new growth point for the domestic AI server field. In comparison, general-purpose servers will show a decline of about 15% in 2023, and this decline is likely to continue into 2024.
Overall, the general-purpose server market is still in a negative state, with an estimated shipment of 2002.2 million units. The main impacts include the decline of Internet companies and the control of domestic policies on the Internet industry. At the beginning of 2024, the AI server market is expected to continue to grow, but due to the decline in the performance of Nvidia's performance cards, the number of training servers in the domestic market is expected to be about 30,000.
Q: What challenges and changes may the Chinese server market face in 2024?
A: In 2024, the main challenge facing China's server market is the negative growth trend of general-purpose servers. Thanks to PC50 platform components, such as DDR 5 memory and CPU** did not fall but rose, and the market improved in general. It is expected that the domestic general server market will maintain a negative growth of about 10%. The Internet industry is the largest market share, followed by carriers, finance and other industries.
The AI server market is in a growth stage, but limited by international policies and chip sanctions, domestic card products such as HW products may obtain more market opportunities. On the other hand, Nvidia announced the launch of a lower-performance H20 replacement card in the Chinese market, with an estimated number of around 250,000 cards.
Overall, the AI server market is still optimistic, but the challenges of general-purpose servers are greater, and the market pattern may change. Special attention needs to be paid to the market competition between NVIDIA and local Chinese card manufacturers, as well as the impact of domestic alternatives on the market structure.
According to research estimates, it was initially expected that the share of H800 cards in the AI server market in Chinese mainland would be about 500,000 in 2024. However, due to restrictions in the United States and the lack of popularity of H20 cards, Nvidia's share of the Chinese market slipped to about 250,000 cards, mainly shared by HW and Haiguang. The total volume of the domestic AI server market in 2023 is expected to be between 120,000 and 140,000 units, an increase of nearly 30% over last year, indicating that the AI server market is showing great growth overall.
Q: How has Nvidia shipped training cards in recent years?
A: In 2023, the actual shipment of NVIDIA's A and H series training cards is about 2.03 million units, including 1 million A cards and 1.03 million H cards. For 2024, it is expected that the shipment volume will be nearly 4 million, of which H cards are the mainstay, and it is expected that about 3 million copies of H100 and about 200,000 pieces of A100 will be shipped.
Q: Can you provide an estimate of Nvidia's shipments on a quarterly basis?
A: Nvidia expects to ship 70,000 units of the A100 and 850,000 units of the H100 in the first quarter of 2024. In the second quarter, 60,000 units were shipped for the A100 and 900,000 units for the H100. By Q3, the A100 will drop to 30,000 sheets, while the B100 and H200 will begin to be in small batches**. Shipments of Series A and H series cards were essentially flat in the fourth quarter, while B100 and H200 continued to grow.
Growth is expected to slow in the second half of the year as customer demand begins to shift to inference servers and cards in the third quarter, and new models B100 and H100 prepare to enter the market. Quarterly shipments are subject to adjustment and are expected to fluctuate around 3%.
Q: What is the status of Nvidia training cards in the global market?
A: Starting from the third quarter of 2023, the production capacity of Taipower Kehuas will be restored, the training card situation in the global market will be improved, and the North American market** will be relatively normal. The Chinese market is recovering slowly.
Q: What is the pricing of the new Nvidia B100 and H200?
A: Expected acceptance** of the B100 and H200 cards in North America is between $37,000 and $39,000 and approximately $29,000 and $31,000, respectively. **Will be determined based on customer acceptance and market feedback.
The two new cards are currently in beta and are scheduled to be gradually introduced to the market in Q3 2024, with the H200 starting in June and the B100 starting in August at the latest**.
Q: What is the competitive state of NVIDIA in the current market? Is there a difference in the performance and profitability of their products in the domestic and foreign markets?
A: Nvidia's training cards in the North American market are basically in a monopoly position, and neither AMD products nor Intel products can pose a threat to NVIDIA. The control of foreign markets is much higher than that of domestic markets, and the profit margin can be maintained overseas by a high percentage or perhaps 70%). However, in China, due to the competition of companies such as HW, Nvidia's local ** has declined, and the profit is not high. In general, there is no reason for Nvidia's training cards to be lowered overseas, because they have strong control over overseas markets.
Q: How is the ** adjustment of the Nvidia H100 card? Is there room for price reductions in the future?
A: The Nvidia H100 card was priced at $24,000 to $25,000 in January this year, and it was adjusted to $22,000 at the end of January. This price adjustment has reserved space for the B version and the HR version. It is expected that the B100 and H200** will not need to be placed much lower than the customer expects. It is mentioned that the products that have not been ordered before will be delivered according to the new **, and those that have already placed an order and have not been delivered will be delivered according to the old **. As for the H100, it is believed that there will be no price reduction this year, because mass production has begun, the market has accepted this product, and there is no need for price reduction.
Q: What is your view on the state of the information innovation market? What is the current replacement situation in the domestic chip market?
A: In the information innovation market, operators have basically completed the replacement, and it is expected that by 2025, the financial industry will also complete nearly 100% replacement, although some core technologies such as databases may not be completely replaced. At present, most of the framework bidding of large Internet companies such as China Mobile use domestic chips. In terms of operators, HW Shengteng accounts for about 60% of the share, and about 30% of Haiguang belongs to other manufacturers such as Cambrian. In the financial sector, Haiguang accounts for about 50% of the share, and HW about 40%. Overall, the promotion of the start-up industry has provided support for some major domestic chip manufacturers, including HW, Haiguang, Feiteng, Cambrian, etc.
Q: Will there be any changes in the competition in the server COEM market this year? For example, the situation in the North American market, as well as changes in the share and profitability of OEMs.
A: From the perspective of the North American market, the share of the three major OEMs in Taiwan, Foxconn, Quanta and Inventec, is very stable, because these OEMs usually do not compete for market share through price reduction strategies. Instead, it is more focused on maintaining stable high profits. As for profitability, last year there were indeed some subsidies for early platform R&D, resulting in additional R&D expenses converted into profits for OEMs, for example, the AI training server of the H100 platform was sold at about $27 to $290,000 last year, with a gross profit margin of about 7% to 8%, while this year's unit price dropped to $23 to $250,000, with a gross profit margin of about 5% to 6%. Despite this, North American customers still have better margins in small batches, while margins for high-volume production, although downgraded, are still higher than those for general-purpose servers.
Q: Regarding the new platforms H200 and B100, how is the production situation this year? Is there a situation with or without subsidies and industrial reunions on these new platforms?
A: This year, with the growth of customer demand, the increase of existing customers and the addition of new customers, the situation of OEMs will improve. Foxconn, for example, has stolen about 5 percent of the market share, and it has also taken market share of all of Apple's Effie Chey servers. As for the new platform, the subsidy situation for H200 and B100 will continue, because there are usually subsidies provided by cloud vendors during the R&D and low-volume production stages of new models. In the case of the B100, the new platform uses a liquid-cooled system and could cost close to $500,000. Trial production is expected to begin in the third quarter and mass production will begin in January 2025 with new series production**.
Q: What is the current OEM situation and future changes of H100 boards and modules?
A: The foundry of H100 board is currently 100% responsible for the FII**, as for the module, the foundry share of H100 module is borne by the Barka module industrial consortium 30%, and the remaining 70% is responsible for Taiwan Flextron. In the future, i.e., in 2024, the situation is expected to remain like this, but by 2025, a new board, the B100 board, may be launched, and the module foundry is still expected to be shared by the industrial consortium and Flextron. As for the H800 board, it will no longer be produced in the future.
Q: What is the change in the performance of the H20 board?
A: The h20's ** is about $12,000 to $14,000, while the channel provider's ** is about $1,000 higher, about $15,000. In terms of performance, because when Nvidia launched the H20 in November, its performance and ** remained at a high level, but then due to the requirements of the United States, Nvidia reduced the computing power of the H20 at the end of December, resulting in a decrease in its performance, and ** also decreased. The current H20 performance is about 20 to 30 percent lower than that of the H800.
Q: What information is there about the production and market demand of the AM DH 100 board, and AMD's shipment forecast for 2024?
A: AMD is still in the testing stage, AMD gives the performance of the M 300 on par with the NVIDIA H 100, and expects the next generation of M350 to surpass the H100. However, customer testing is still ongoing as the migration of the software platform is not yet fully complete. As for shipments, AMD internally expects to ship around 350,000 M 300 units in 2024, based on customer testing.
Q: What is the information on the ** and estimated shipments of the Ascend series products 910 and 920?
A: In the Ascend series of products**, the 910B is roughly 85,000 RMB, and the whole training server is about 150 to 1.6 million RMB. For the 910c in 2024, the monomer price is expected to reach 17 to 180,000 RMB. In terms of market capacity, assuming that the annual training card market is about 500,000 units, HW may occupy 15 to 200,000 units, becoming the main share of the market.
Q: Is it possible that FII's advantage in chip foundry will be replaced by new chips from Google or AMD?
A: FII has already taken into account the market share of Nvidia chips in its market forecast for 2025, because the growth of training cards has slowed down since the fourth quarter of this year, and the growth is expected to be smaller by 2025. As a result, FII is considering how to utilize the capacity that may be available and has approached AMD to discuss the future foundry situation to make up for the capacity gap. In addition, although FII is a registered enterprise in the mainland, its chip and server foundries are not located in the mainland, in Taiwan and Mexico, respectively, so they will not be affected by US sanctions.
Q: What are the advantages of FII and Taina's two OEM companies?
A: In terms of NVIDIA chip foundry, FII has a great advantage in training cards, but its share in inference card foundry has not yet reached the vast level. At present, the majority is the strongest in the field of server processing, followed by the Industrial Fortune Union. Nvidia's demand for inference cards will grow in the future, while the demand for training cards will decrease, unless there is a serious problem with the vast number and industrial Fortune Union, Nvidia will not easily change the ** business.
Minutes**: Wen Bagu Research] Mini Program