Summary of Key Points
On August 4th, SanDisk and SK Hynix jointly released the standard specifications for HBF (High Bandwidth Flash), a new type of storage technology designed for AI computing. HBF fills the gap between HBM (fast but expensive with limited capacity) and solid-state drives (large capacity but slow performance), making it ideal for AI inference applications. By using vertical stacking technology, it offers 8-16 times the capacity of HBM at the same cost and can work in conjunction with HBM. The industry believes that HBF could change the storage competition landscape, particularly since there are many more players in the NAND market. Its future market size may even exceed that of HBM, although the technology is highly challenging, and currently only a few manufacturers are capable of mass-producing it.
1. HBF: The “Intermediate Layer” for AI Storage, Solving the Shortcomings of HBM and SSDs
To put it simply, HBM is like a small snack box on your desk—fast to access but with limited capacity and expensive; solid-state drives are like large storage cabinets in the kitchen—capable of storing a lot but slow to retrieve data. AI inference requires processing large amounts of historical data (such as trained models) while also needing fast performance, which neither HBM nor SSDs meet perfectly. HBF acts as an “intelligent storage solution” that offers 8-16 times the capacity of HBM at the same cost and faster speeds. It uses the same vertical stacking technology as HBM and can be packaged together with GPUs. The division of tasks is clear: HBM handles real-time calculations (referred to as “hot storage”), while HBF manages large amounts of historical data (referred to as “cold storage”). This analogy illustrates how HBF complements existing systems, much like how bookshelves in a library (HBM) complement the entire library (HBF)—you can quickly retrieve books from the shelves, but the books are still stored in the library.
2. Promising Prospects: Market Size Could Exceed HBM’s, with Commercialization Expected by 2027
The industry has high expectations for HBF:
- Kim Jong-ho, the “father of HBM,” predicts that storage will have a greater impact on AI performance than GPUs in the future, and HBM will not be able to meet the growing demand. He expects HBF to become commercially available by the end of 2027 to 2028, initially for AI inference, and later it will coexist with HBM as a standard configuration. By 2038, the demand for HBF is expected to surpass that of HBM.
- Bank of America forecasts a market size of $246 billion for HBM by 2030; according to Kim Jong-ho, HBF’s market size could reach tens of billions of dollars.
- SK Hynix also mentions that complex storage solutions, including HBF, will see accelerated growth around 2030.
3. Changing Competitive Landscape: From a “Triopoly” to a “Multi-Player Arena”
The HBM market was previously dominated by SK Hynix, Samsung, and Micron, but HBF falls within the NAND flash memory segment, which has many more players, such as SanDisk, Western Data, Kioxia, and Yangtze Memory. This means that the HBF industry will no longer be controlled by a few companies. For example, Yangtze Memory has expertise in 3D stacking technology (Xtacking) and could potentially join the market; SanDisk and SK Hynix have already taken the lead in developing standards and setting up production capabilities. The competition will become more intense, but it also opens up opportunities for more manufacturers.
4. Manufacturer Actions
Manufacturers are actively moving forward with their plans:
- SanDisk: Using CBA (Chip Bonding Assembly) technology, they will provide HBF samples to customers by the end of this year and launch AI inference devices equipped with HBF in early 2027.
- SK Hynix: Developing the “H3 architecture” to integrate HBM and HBF alongside GPUs, which is more energy-efficient than pure HBM solutions.
- Samsung: Positioning HBF as a more cost-effective option to reshape the AI inference storage landscape.
- NVIDIA/Google: Collaborating with Samsung and Kioxia on similar products (e.g., XL Flash, Z NAND), expected to be released early next year.
- Yangtze Memory: Has a technical foundation but has not yet announced any specific plans.
5. Technical Challenges: High Barriers to Entry
Entering the HBF market is not easy:
- Technical Difficulty: Only SanDisk and SK Hynix currently have the capability for mass production.
- Production Challenges: The manufacturing cycle takes 9 months, much longer than for conventional storage solutions, and the required production space is three times that of regular production lines. Optimizing transfer efficiency and extending the lifespan of the production lines are crucial to profitability.
These factors mean that it will be dominated by leading manufacturers in the short term, but as technology advances, more players will enter the market.
6. Stock Market Reactions
On August 3rd, SanDisk’s stock price in the US market rose by 6.03% and continued to rise by another 3.2% after the market closed; SK Hynix’s stock price in South Korea fell by more than 3% on the same day. This may reflect concerns about new competition from HBF, given SK Hynix’s dominant position in the HBM market. However, in the long run, these developments are part of a strategic move towards future growth.
---
The translated text maintains the structure and tone of financial journalism, using natural language that is suitable for a global audience. It also adapts technical terms to fit the target culture while ensuring accuracy and consistency in financial terminology.