Cloud vendor capital expenditure expectations continue to rise, and there is huge room for AI computing power. Looking at the underlying principles, the marginal slowdown of Moore's Law and the scaling law have jointly opened up the upper limit of AI computing power requirements. Stronger models require a larger amount of parameters and data for training, while supporting model training requires more AI computing power. Looking at the actual demand for AI computing power, the capital expenditure of the four largest cloud vendors in North America increased 59% year on year in the first three quarters of 2024; according to TrendForce, AI server shipments are expected to grow 42% year on year in 2024. Driven by demand from cloud operators and sovereign clouds, shipments are expected to increase by another 28%, driving AI servers to nearly 15% of total server market shipments. The AI server market as a whole still has great potential for growth.

The shipping pattern of high-end AI servers is changing, and the industrial chain is expected to be reshaped. According to TrendForce, judging from the types of AI servers and AI chips, it is estimated that in 2024, mainstream AI servers equipped with GPUs will account for 71%, of which Nvidia's market share is nearly 90%. It is expected that in 2025, Nvidia's next-generation Blackwell will replace the Hopper platform as the mainstream of the market. Since GB200 cabinet-level products are more cost-effective than traditional 8-card servers in terms of computing power, connectivity, and power consumption, the main shipping format for Blackwell architecture products has changed accordingly. According to Hon Hai's spokesperson's response at the French conference, it is expected that GB200 servers will be produced in small quantities in the fourth quarter of 2024 and shipped in large quantities in 2025. Changes in product forms and shipping models will also affect the division of labor in the industrial chain.

IFF is deeply involved in Nvidia's AI industry chain and is expected to gain more share in changes in the division of labor in the industry chain. As an important joint R&D partner of Nvidia and a core supplier to cloud vendors, the company occupied a position in the supply of AI servers in the early days of the AI explosion. The business involved various forms of GPU modules, boards, and servers. In the GB200 cabinet-level product line, the company is deeply involved in the manufacture of cabinet-level products from boards to computing trays and exchange trays to complete machines. Stand-alone value and cabinet shipments are expected to increase dramatically along with mass production of GB200 chips, driving a high increase in cloud computing business revenue and overall performance.

Investment advice: We believe that IFF, as Nvidia's core supplier, will benefit from the rapid growth in AI capital expenditure and the continuous increase in computing power investment. In 2024-2026, IFF is expected to achieve revenue of 583.3/748/864.5 billion yuan, an increase of 22%/28%/16% year over year, and net profit to mother of 23.7/29.4/34.5 billion yuan, an increase of 13%/24%/18% year over year. Corresponding PE is 20x/16x/14x respectively. Give it a “buy” rating.

Risk warning: the risk of technological iteration falling short of expectations; the risk of supply chain stability falling short of expectations; the risk of GB200 mass production falling short of expectations; the risk of increased industry competition and declining gross margin.