The Trio of Storage Technologies: How Will HBM4, CBA+HBF, and 3D NAND Reshape the Future of AI Computing Power?

In April 2025, the JEDEC Solid State Technology Association officially released the HBM4 memory specification, JESD270-4, marking a new phase in high-bandwidth memory technology. At the same time, CBA+HBF process innovations are breaking through the “memory wall” bottleneck, with 3D NAND stacking now exceeding 300 layers.

These three technological breakthroughs will not only redefine the boundaries of storage performance but also provide critical support for the advancement of AI computing power.

The release of the HBM4 standard marks a qualitative leap for the AI and high-performance computing sectors. Compared with its predecessor, HBM3, HBM4 features a 2,048-bit interface, a data transfer rate of up to 8 Gb/s, and a total bandwidth of 2 TB/s, while the number of memory channels has doubled from 16 to 32.

This performance boost is critical for handling large-scale AI models. NVIDIA’s Blackwell Ultra chip is equipped with eight HBM3E memory modules, totaling 288 GB of capacity and delivering 8 TB/s of bandwidth. In contrast, HBM4 can support a single-stack capacity of up to 64 GB and accommodates 4-, 8-, 12-, or 16-layer DRAM stacking.

The three major storage giants have already engaged in fierce competition. SK Hynix announced in March 2025 that it had shipped the world’s first 12-layer HBM4 samples and plans to begin mass production in the second half of 2026. Meanwhile, Samsung has developed HBM4 products based on its 1c DRAM process, completed R&D on the sixth-generation HBM4, and obtained certification for mass-production readiness.

Micron is not to be outdone, having already delivered 11 Gbps HBM4 samples and achieving a bandwidth of 2.8 TB/s. According to TrendForce, HBM4 is expected to enter mass production as early as the end of the first quarter of 2026.

As the number of parameters in AI models and the length of their context continue to grow, HBM capacity is no longer sufficient to meet demand. It is precisely against this backdrop that, CBA+HBF technology has emerged as a result. , becoming the core direction for the future development of storage IDM.

CBA (CMOS Direct Bonding to Array) technology involves fabricating logic chips and memory arrays separately and then bonding them together, which significantly increases storage density per unit area and optimizes internal interconnect routing. This technology has become the irreversible technological roadmap for high-layer-count NAND flash memory.

HBF (High-Bandwidth Flash) adopts the packaging design of HBM but replaces part of the DRAM stack with flash memory. Compared with HBM, HBF offers 8 to 16 times the storage capacity and the advantage of non-volatile storage.

SanDisk’s first-generation HBF product is remarkable. : A single stack (16 dies) delivers approximately 1.6 TB/s of read bandwidth, with a total capacity of 512 GB. When reading the 8-bit quantized weights of the Llama 3.1 405B model, HBF’s system-level performance is only about 2.2% short of “HBM with unlimited capacity.”

3D NAND technology is rapidly advancing from over 200 layers to more than 400 layers. By 2025, competition among major manufacturers will have reached a fever pitch:

Kioxia and SanDisk have launched the 10th-generation BiCS FLASH technology, achieving 332-layer stacking , with an interface speed of 4.8 Gb/s and a 59% increase in bit density.

SK Hynix’s ninth-generation 3D NAND technology has reached 321-layer stacking , has commenced mass production of 2TB QLC products.

Samsung’s V9 technology has achieved 290-layer stacking, and the V10 technology is planned to reach Floors 420–430 , mass production is expected in the second half of 2025.

Micron’s ninth-generation technology has achieved 276-layer stacking, and its 340-layer NAND is now in the sample stage.

Technological breakthroughs are not limited to increasing the number of layers. IMEC’s latest research employs Hole-Side Airgap technology to reduce the Z-pitch from 40 nm to below 30 nm, opening up new possibilities for 1,000-layer 3D NAND.

A plasma etching process developed by Kolin R&D and the University of Colorado, among other institutions, has more than doubled the etching speed for deep vias in 3D NAND, laying the foundation for the fabrication of higher-density storage devices.

These three technologies do not evolve in isolation; rather, they work in synergy to build an efficient AI storage architecture. The industry has proposed the “H3 Hybrid Architecture,” which deploys HBM, HBF, and traditional SSDs in a tiered configuration.

HBM is responsible for latency-sensitive tasks. , such as real-time computing and hot data caching; HBF specializes in high-capacity sequential read operations. , such as model parameter retrieval and cold data storage.

Simulation tests show that this architecture delivers up to 2.69 times better performance per watt compared with a pure HBM solution, and increases concurrent query throughput by a factor of 18.8 in a 10-million-token scenario.

SK Hynix and SanDisk have jointly convened the “HBF Specification Standardization Alliance Kickoff Meeting,” establishing a dedicated task force under the Open Compute Project framework to officially launch global HBF standardization efforts.

Breakthroughs in storage technology are reshaping the entire industry value chain. According to SEMI’s forecast, the global NAND equipment market is expected to reach US$13.7 billion and US$15.0 billion in 2025 and 2026, respectively, representing year-on-year growth of 42.5% and 9.7%.

Innovations in new storage architectures, such as 4F2 DRAM and 3D NAND, are creating new growth opportunities for etching, deposition, and bonding equipment.

Domestic companies are also stepping up their efforts to catch up. Yangtze Memory Technologies introduced the Xtacking architecture as early as 2018, pioneering the use of hybrid bonding in mass production of 3D NAND. Hefei Changxin plans to begin deploying CBA technology in 2026, with gradual ramp-up thereafter.

Morgan Stanley forecasts that the global HBM market is on track to grow from US$4 billion in 2023 to US$33 billion by 2027. Meanwhile, HBF, as an emerging technology, is expected to deliver samples in the second half of 2026 and launch a production-ready product for AI inference in 2027.

When SK Hynix unveiled the world’s first 12-layer stacked HBM4 product, its power efficiency improved by more than 40%. Meanwhile, Kioxia and SanDisk’s 332-layer 3D NAND technology has made it possible to transfer a 4K movie in just one second.

The boundaries of storage technology are constantly being pushed. From the bandwidth leap enabled by HBM4 to the architectural innovation of CBA+HBF, and then to the race to increase the层数 of 3D NAND, each breakthrough is delivering stronger support for AI computing power.

The future is here—this trio of storage technologies is now delivering the most powerful symphony of the AI era.