1. The bottleneck for AI inference is expanding from simple computing power to "memory bandwidth + capacity." Limited HBM capacity forces some data centers to purchase additional GPUs to obtain more memory.
2. HBF attempts to supplement HBM with high-density NAND. Its capacity is far higher than HBM, its sequential read speed is close to HBM, and its cost per GB is expected to be significantly lower; Goldman Sachs models show that the HBM+HBF hybrid architecture can reduce example inference costs by approximately 33%.
3. The risk lies in the fact that HBF is not a mature technology. Latency, write endurance, power consumption, heat dissipation, and manufacturing resource consumption that is 3-4 times that of ordinary NAND are all obstacles to large-scale commercialization.
4. If HBF rapidly becomes widespread, it could significantly impact NAND supply and demand. In a bull market scenario, by the end of 2028, HBF will only account for about 4% of NAND bits, but could consume about 9% of NAND wafer capacity, thus potentially prolonging the current tight NAND supply cycle.
5. Goldman Sachs is most bullish on SanDisk and SK Hynix, and also believes Samsung and Kioxia are well-positioned; semiconductor equipment manufacturers will benefit moderately; however, there is potential pressure on GPU/XPU manufacturers, as higher memory utilization may reduce the number of accelerators needed to complete the same inference tasks. (See the chart for a detailed table of individual stocks and target prices)
