AI Storage: HBM Grasping NVIDIA's Achilles' Heel

After experiencing a downturn for nearly 2 years, the storage industry has seen a price increase once again. Storage itself is cyclical, with supply and demand changes directly affecting the price trend of storage chips. What's different this time is that, driven by new applications like AIGC, the storage industry is expected to resonate with a "cycle + growth" pattern.

Referring to the price changes of DDR3 and DDR4 in the past, a clear cyclical pattern can be observed, with approximately a 3-4 year cycle. With the decline in storage prices this time, major manufacturers have started to reduce capital expenditures and adjust production capacity structures in the second half of 2022. Changes in the supply-demand landscape have led to a bottoming out of storage prices in 2023. Coupled with the emergence of new applications such as AI, the storage industry has once again entered an upward cycle.

The current upward cycle of the storage industry: resonance of "cycle + growth"

①Traditional cycle: Smartphones and PCs contribute nearly half of the revenue in the storage industry. As these two major downstream sectors bottom out and recover, they bring about a cyclical rebound to the storage industry;

②New growth in AI: AI applications like ChatGPT have stimulated the demand growth for AI servers, with storage upgrades being an important part of it. The terminal innovation of AI PCs and AI phones has also brought about new storage demands.

The traditional cycle only brings expectations of a bottoming out to the market, while AI can bring higher prospects. Currently, the demand for HBM storage in AI servers has already generated significant revenue and performance. Based on Dolphin Jun's analysis, it is expected that AI servers will bring a demand space of over $10 billion for the storage industry, injecting growth prospects into the storage market. In addition, AI PCs and AI phones are also expected to accelerate the iterative upgrade of storage products.

Currently, the storage industry is mainly dominated by Samsung, SK Hynix, and Micron Technology. Especially in the DRAM market, the three manufacturers collectively hold 95% of the market share. The increase in demand for HBM has expanded the market space for these three manufacturers, but there are differences in market share. SK Hynix, with its breakthrough in HBM3, currently holds a leading share advantage. Although NVIDIA is leading in this wave of AI, it is also constrained by HBM. Currently, due to insufficient capacity, the market is in short supply, and prices are expected to continue to rise. With storage product prices frequently increasing, the performance of the three storage manufacturers is expected to continue to improveAnd the manufacturers who are the first to successfully expand production will enjoy stronger alpha.**

In the next part, Dolphin will conduct valuation calculations on individual stocks and analyze how the additional $10 billion will bring performance elasticity to specific companies.

For Dolphin's specific analysis of the storage industry, please see the following:

I. Storage Chips: Resonance of "Cycle + Growth"

Storage chips, also known as semiconductor storage, include DRAM, Flash, SRAM, PROM, etc. Among them, DRAM and NAND Flash are the main types of storage, accounting for about 99% of the overall market share.

① DRAM: Used for temporary data and instruction storage, it is the mainstream solution for memory in current mobile phones, computers, servers, etc.;

② Flash: Applied in various electronic products' hard drives (SSD, U disk, SD card, etc.), responsible for permanent data storage.

Although both DRAM and NAND Flash are the mainstream products in the current storage market, DRAM has a larger market size and concentration than NAND. With the recent growth of new AI applications, the importance of DRAM has been further enhanced.

1.1 DRAM Products

According to product classification, DRAM can be divided into DDR, LPDDR, GDDR, HBM. The first three types of products are mainly used in the traditional cycle field, while HBM is mainly driven by the AI market. Among them, DDR is mainly used in consumer electronics, servers, and PC fields; LPDDR (low power consumption) is mainly used in mobile devices, smartphones, and automotive fields; GDDR is mainly used in GPU for image processing; HBM is for high-performance computing fields such as AI servers.

It can be seen that the application areas of DRAM are quite extensive, spanning mobile phones, personal computers, servers, automobiles, and other fields. Among many fields, smartphones and PCs still account for nearly half of the market share of DRAM. The recovery of downstream application cycles will drive the price increase of storage chips.

1.2 Recovery of Traditional Cycles

From the perspective of storage chip price changes, the situation of prices hitting bottom in 2023 is similar to the performance of the smartphone and PC markets. Starting from the second half of 2023, the shipments of the two major downstream products have rebounded, driving the increase in demand for upstream chips, thereby starting to push up storage chip prices from the bottom. **Looking at the classification of DRAM products, the impact of traditional cycles here mainly refers to products like DDR, LPDDR, which are expected to drive a new round of storage upcycle**

2. AI Empowerment in the Storage Industry

AI has brought new growth opportunities to the storage industry, covering aspects such as AI servers, AI PCs, and AI phones. Due to the AI wave initially driven by applications like ChatGPT, AI servers have injected a large volume of orders and revenue into the storage industry. From applications to end devices, new products of AI PCs and AI phones have also begun to be launched successively, promising to bring new upgrades.

2.1 HBM

The emergence of large models in AIGC has increased the demand for computing chips, thereby driving the requirements for storage chips. Storage chips can affect in two ways: ① Larger bandwidth, which can improve computational efficiency; ② Larger capacity, which can store a large amount of data information under large models. From the product list, HBM products significantly exceed DDR, LPDDR, and GDDR products in terms of capacity and bandwidth.

Although it is possible to achieve sufficient bandwidth by stacking a large number of storage chips, this also occupies most of the chip area. By implementing vertical stacking packaging through 3D packaging technology, HBM can significantly save the occupied area of storage chips.

Under the influence of the AI wave, major technology companies have increased capital expenditures, especially in the procurement of AI servers, where HBM directly benefits. According to Trendforce data, the global shipment volume of AI servers reached 1.2 million units in 2023, an increase of over 35% year-on-year. At the same time, the expected shipment volume in 2024 will continue to increase to over 1.65 million units, with a year-on-year growth rate of 37%. Therefore, Dolphin expects HBM to continue to maintain a high growth performance.

2.2 Market Space Brought by AI Servers

HBM has become the mainstream solution for AI servers, directly driving the rapid growth of HBM. Current AI servers are mainly used for training and inference, both of which have a clear growth trendWhen calculating the demand for HBM space, the performance requirements of AI servers and AI accelerator cards need to be calculated first.

1) Demand for AI Accelerator Cards: Based on industry and company data, Dolphin expects AI servers to continue to grow at a double-digit rate. Benefiting from the increase in training demand, the proportion of AI servers used for training is expected to continue to rise to around 37%. By assuming the average number of accelerator cards used for training and inference respectively, the potential demand for AI accelerator cards in future AI servers is expected to grow to over 9 million units.

2) Demand for HBM: Combining the performance of AI accelerator cards, the demand for HBM can be calculated. Taking NVIDIA A100 as an example, each card requires approximately 80GB of memory. Assuming further increases in memory per card in the future, the overall demand for HBM is expected to continue to increase. Considering the unit price of HBM, the overall market size of HBM is expected to grow to over $12.2 billion, with a compound growth rate of over 50%.

2.3 New Space for AI PC & AI Phone

In addition to AI servers, the emergence of AI PCs and AI phones will also drive the upgrade and growth of the storage industry. From the product classification of DRAM mentioned earlier, AI PCs and AI phones mainly impact the upgrade of DDR and LPDDR, accelerating the iteration of traditional cyclical products.

1) AI PC

According to IDC's expectations, the shipment volume of AI PCs is expected to increase to 167 million units by 2027, with a compound growth rate of over 40%. As PC products upgrade, the share of DDR4 will continue to decrease, and storage products will gradually shift to DDR5 and DDR6 to meet data storage and transmission needs.

Looking at the current pricing of DDR products, the price of DDR4 is approximately 2-3 times that of DDR3, while DDR5 is nearly 10 times the price of DDR4. Product upgrades can directly increase the scale of the storage industry.

2) AI PhoneSimilar to AI PCs, AI phones will also benefit from the AI wave as intelligent terminals. According to Counterpoint Research, it is estimated that by 2024, AI phones will account for 11% of all smartphone shipments; and by 2027, AI phone shipments will exceed 550 million units, accounting for 43%.

Micron expects that the DRAM capacity carried by AI phones is expected to increase by 50-100%. With the upgrade of memory products, the memory market for AI phones is also expected to see a simultaneous increase in volume and price.

Currently, AI servers have formed a scalable income, while AI PCs and AI phones are still in the relatively early stages. With the increase in penetration rate, it is also expected to further accelerate the growth attributes of the storage industry.

III. The "Three Kingdoms" of Storage Manufacturers

Although the DRAM market has significant growth opportunities, the players in the market are very concentrated. Samsung, Micron, and SK Hynix occupy 95% of the market share, creating an almost monopolistic "Three Kingdoms" situation.

1) Technological Capabilities: While Samsung leads the DRAM market, SK Hynix has surpassed in the current popular HBM field. Although Samsung is still ahead in the HBM2E stage, SK Hynix has overtaken in the conquest of technology nodes starting from HBM3. The latest HBM3E product is also the first to become NVIDIA's main supplier by SK Hynix.

2) HBM Market Share: Since SK Hynix was the first to break through HBM3, it occupies most of the HBM market share. However, looking at the production capacity plans of the three manufacturers, Samsung is expected to catch up. By the end of 2024, Samsung's HBM production capacity is expected to reach 130k per month; while SK Hynix is at 120-125k per month. By then, Samsung's market share in the HBM market is also expected to increase.

3) Product Pricing and Performance Expectations: Considering the current industry and company situation, HBM is expected to further raise prices. This is mainly based on the current: ① Market's continuous outlook on AI demand; ② HBM3E's TSV yield still remains low, around 50%, with buyers having demand to lock in the source; ③ Stable suppliers and products providing assurance to customers. With the continuous demand for HBM and price increases, the performance of storage manufacturers is expected to continue to improve.