As major AI server manufacturers announce their financial reports one after another, the gross margin has become a key focus for investors. Supermicro's market value has plummeted due to a decline in gross margin, while Quanta and Wistron have performed well.
Supermicro's latest financial report has shocked Wall Street, with a gross margin of 11.2%, which is a significant drop from the previous quarter's 15.5% and also a substantial decrease from 17% in the same period of 2023, leading to a significant reduction in the company's market value.
In contrast, Quanta's gross margin for the second quarter has improved to 8.58%, an increase of 0.1 percentage points from the previous quarter and 0.06 percentage points from the same period last year. Wistron recorded a gross margin of 8.36%, an increase of 1.16 percentage points from the previous quarter and 0.8 percentage points from the same period last year.
Although this comparison is like comparing apples and oranges, as these three companies have different product lines in addition to servers—Supermicro also manufactures supercomputers, Quanta supplies notebook computers and automotive electronics, and Wistron produces notebook computers, desktops, and monitors—the gross margin of AI servers is indeed a concern for companies that adopt the buy-and-sell model, in which the ODM purchases key components, assembles the entire machine, and then sells it to customers.
Advertisement
Different business models among ODMs
In the buy-and-sell model, the higher the unit price of the entire machine, the higher the revenue. However, the assembly fee, due to its relatively high transparency and competitive pressure, cannot be increased accordingly. The gross margin calculation needs to be based on the assembly fee divided by revenue, so the higher the unit price, the lower the profit margin.
In the second quarter of 2024, Quanta's AI server shipments surged, accounting for 50% of the total server shipments, but the company still managed to increase its gross margin. According to Quanta, this was due to exchange rate fluctuations and an increase in shipments of low-cost Chromebooks.
However, the company admits that it will be difficult to repeat this feat in the third quarter, as shipments of high-priced products (including AI servers and high-end notebook computers) are expected to increase. Since Quanta mainly ships AI servers based on Nvidia's H100 chip, some people are wondering what impact the more expensive GB200 will have on the company's gross margin; however, others believe that the focus should be on profit rather than gross margin.
Another major server assembly factory, Wistron, adopts a different approach to ensure gross margin by using the OEM model, where customers purchase key components, and then Wistron assembles them. Although this model may lead to a decrease in revenue, it can also avoid a decline in gross margin.Finally, although Wistron also ships complete server systems like Supermicro and Compal, the GPU modules used for AI servers account for a larger share of the company's shipments. Compared to complete server systems, the unit price of GPU modules is lower, hence the gross margin is higher.
Both Compal and Wistron's notebook computer revenue is declining. According to Compal, notebook computers accounted for 30% of total revenue in the second quarter, and this proportion is expected to decline further in the second half of the year. Since Wistron began producing GPU modules for Nvidia in 2023, notebook computer revenue has also dropped to 30% of total revenue.
Can Supermicro find its own path in the competition?
At the same time, Supermicro's gross margin issue is mainly due to the transformation of the company's customer base. In the past, Supermicro mainly targeted the high-profit enterprise server market, but as the company shifted its strategy to cloud service providers (CSPs) such as Amazon, Microsoft, and Google, although revenue increased, the gross margin has declined.
Although Supermicro is seeking to improve profit margins through liquid cooling and improving the efficiency of its factories in Taiwan and the upcoming one in Malaysia, the company still faces tremendous pressure. Most CSPs have already established long-term partnerships with server ODMs, and the gross margins of these ODMs have long been below 10%. How Supermicro can compete for orders without reducing prices remains to be seen.
The trump cards of AI server assembly market players
Generative artificial intelligence continues to fuel the wildfire spread of the global server business potential. Taiwan manufacturers dominate the global AI server assembly industry, accounting for up to 90%. This is attributed to their excellent skills in server design, research and development, and contract manufacturing. Taiwan is the main beneficiary of this market dominance.
The AI server business opportunity is boundless, and Taiwan's server assembly industry players are naturally eager to capture a significant market share. Although AI has just started, its development speed is rapid, and speed and capital investment are the main variables affecting industry competitiveness.
What are the key advantages of major equipment assemblers such as Compal, Inventec, Wistron, Wistron NeWeb, Foxconn, and Pegatron in the AI server field compared to their competitors, and how can they utilize these advantages?
Speed and resource competition at the manufacturing endHigh-end products consist of core components and peripheral devices, with extremely stringent requirements for the production site. Companies that can quickly establish a complete manufacturing and assembly supply chain at the designated production location within a short period of time are more likely to secure more orders.
However, the artificial intelligence boom seems to be irresistibly affected by geopolitical factors. Insiders in the supply chain have pointed out that unless it is delivered to Chinese customers, the main production will be carried out in areas outside mainland China.
In the AI server market competition, apart from capital, scale, and talent, optimizing resources is also the top priority. Due to the strong demand from customers, Thailand, Vietnam, and Malaysia are the three most favored non-mainland, non-China Taiwan factory establishment regions in Southeast Asia. Currently, both Compal and Quanta have set up factories in Thailand, Wistron in Vietnam, and Wiwynn and Supermicro in Malaysia.
The establishment of manufacturing factories in Southeast Asia is also due to the region's growing demand for AI data center deployment. Local enterprises or governments, as well as large CSP manufacturers, all need to build their own data centers.
In this wave of AI server boom, Foxconn is said to have benefited the most from the first wave of supply chain participants. Due to its overall competitiveness, it can achieve a higher proportion of computing board assembly, and it is unlikely to be overtaken by competitors in the short term. Foxconn currently has more than 200 facilities and offices in 24 countries/regions worldwide. Its adaptability and collaboration with industrial facilities make it difficult for other companies to compete in the short term.
On the other hand, Pegatron is optimizing its existing capacity, integrating server factories that were originally part of different business units, leveraging synergistic effects, and concentrating production in Taoyuan.
In addition, Pegatron is also strategically meeting the diverse market demands by establishing cross-border businesses. Currently, the main focus of the Mexican factory is to serve North American customers. To meet customer requirements, it is expected to significantly increase production volume starting in the third quarter of 2024. In the long run, the production capacity plan of Pegatron's Mexican factory is five times that of its current Taoyuan factory.
Of course, simply increasing the number of factories is not enough. AI technology also plays a key role in the unexpected success of various factories. For example, Foxconn, using its rich experience in the Taoyuan factory, built a new server-specific factory in Mexico with the help of digital twin technology and put it into production in a short period of time. At the same time, the relevant data can be efficiently transmitted to the production lines of other global manufacturers, allowing the same product to be quickly replicated in any factory.
Another place showcasing the application of AI servers is Pegatron's Taoyuan factory. It is known that the Taoyuan factory mainly produces AI servers, but Pegatron has also combined Nvidia's digital cloning technology with 5G to create a smart factory, capable of installing robotic arms and AI flaw detection.
First come, first served, experience wins.Despite iterations, the architecture of AI servers, such as that of Nvidia's Blackwell series of chips, differs significantly from previous generations. Still, the industry recognizes that the lessons learned from previous generations still play a special role in the design and manufacturing of next-generation servers. Therefore, having development and production capabilities is critical to determining the outcome of the AI server competition. The old server manufacturers include Quanta, Wistron, Wiwynn, Inventec, Foxconn, etc. Although the OEM orders are different and each has its main force, the Taiwanese server manufacturers still occupy a dominant position. Quanta Chairman Barry Lam has reiterated that the company worked with Nvidia as early as 2016 to develop GPU servers. In stark contrast to the widespread discussion surrounding artificial intelligence over the past two years, he believes that artificial intelligence has already seen substantial growth and development. As early as 2016, he had already felt the appeal of artificial intelligence. Wistron President Lin Zhijie recently discussed whether the company's relationship with Nvidia has changed with processor iterations. Rather than comment directly, he used metaphors. He added that if Wistron was going to make a product, no one would have noticed it six to seven years ago; however, a manufacturer has endorsed it. How can Wistron give up this cooperation once the product becomes popular? The above remarks also revealed that Wistron and Nvidia have been cooperating for many years to jointly build GPU servers. At a time when generative AI is driving the popularity of AI servers, how can NVIDIA ignore Wistron? In addition to experience, this cooperation also shows that both parties have a deep understanding of cooperation. Ye Sen, chairman of Inventec, said at the shareholders' meeting that Inventec built the first server in Taiwan. Before AI servers became popular, Inventec was the world's largest server motherboard manufacturer. It can be seen that Inventec has rich experience. Development and production experience. Liu Yangwei, chairman of Foxconn, has always said that Foxconn is the world's largest server foundry, with a market share of more than 40%. Foxconn's deep accumulation in the server field is an important reason for its market share. As a latecomer in the server field, Pegatron has always emphasized that "it is never too late to join the field of artificial intelligence." Ahead of COMPUTEX 2024, Pegatron has opened a new server showroom at its Guandu headquarters in Taipei. In addition to seeing the demo, customers will soon be able to test innovative AI server solutions. Tong Shihao said that although Pegatron entered the AI market later than its competitors, it has a deep foundation in electronic manufacturing services (EMS) and has invested heavily in R&D, technology, products, etc. He compared Heshuo to a buried seed that will sprout and grow when there is a suitable growing environment. He is not worried about the lack of customers, because "Currently, there are too many solutions proposed by customers!" ”Heterogeneous Technology Integration Ecosystem
The manufacturing of AI servers requires development and production experience, as well as experience in production across multiple locations. Mastering the ecosystem is an advantage.
Historically, servers have primarily focused on computational performance. To collaborate with customers, the next generation of servers prioritizes not only processing capabilities but also the vertical integration capabilities of all key chips and components.
Wiwynn's recent reinvestment provides insights into its ecosystem strategy. Wiwynn invested in ZutaCore, a company that develops semiconductor liquid cooling solutions. ZutaCore is known for its waterless operation and heat dissipation using a two-phase coolant. The use of two-phase direct cooling can improve heat dissipation and minimize the pump power required for liquid cooling.
Wiwynn Technology also announced plans to invest in Andra Capital, a private equity fund mainly focused on artificial intelligence software, to gain a deeper understanding of consumer profiles and enhance Wiwynn Technology's overall ecosystem.
Foxconn is undoubtedly an industry leader in the field of ecosystem construction. The Foxconn Group consists of Foxconn Industrial Internet (FII) and its subsidiary, Hongbai Technology, which is responsible for designing and producing various servers.
Foxconn Interconnect Technology (FIT) is also a member of the group and has invested heavily in the development and supply of high-speed connectors and lines. In addition, the investment in Amax Engineering has enhanced Foxconn's ability to provide a full range of liquid cooling solutions from single cabinets to data centers.
Foxconn has a diversified subsidiary portfolio with strengths in multiple specialized fields. It possesses extensive technical expertise, resources, and strong capabilities in mold, component, and mechanical part manufacturing. This involves all stages of design, development, and manufacturing, setting a significant entry barrier for other market competitors.
Pursuit of AI Talent Pool
Throughout the entire product development design process, from the initial stage to the final production testing, each step requires a significant investment of human resources for collaborative development. The growing popularity of AI servers also highlights the increasing demand for skilled professionals.Wei Chuan's Chairman Lin Shi-Min stated that the shortage of hydropower in Taiwan, China, is an important issue, but there is still room for development. However, the lack of talent remains a key problem.
Wei Chuan's Chairman Lin Jianjun also pointed out that no matter how many orders are placed, a lack of manpower is futile. Especially now, major companies are expanding their AI-related talents. In addition to manufacturers in Taiwan, China, multinational companies have also set up R&D centers in Taiwan.
Wei Chuan has about 5,000 R&D personnel, and the number is still increasing. Wei Ying Technology is also expanding its talent pool, currently having 1,500 employees in Taipei, and is expected to establish a new headquarters.
Recently, Quanta Cloud Technology (QCT) President Yang Wei-Xiong said that there is a huge gap in AI server talents. Quanta's server team has grown by 10% this year and will continue to expand.
Rapidly catching up, Renbao regards talent recruitment as a key to operations. Chairman Chen Ruilin said that the current server team has about 700 people, and compared with competitors' thousands of server teams, Renbao's scale will continue to expand.
Foxconn's Fushida Group focuses on network communication server business, with about 190,000 employees, including more than 30,000 R&D personnel.
In early 2022, Compal expanded its business department and established a new department BU18 specifically responsible for server manufacturing business. At that time, there were rumors in the industry that Compal was poaching hundreds of qualified server technicians, including engineers from Foxconn, Wei Chuan, and Quanta, which are companies that manufacture servers.
Comments