Goldman Sachs Global Technology: Semiconductors & Storage — High-Performance NAND Could Help Break AI's 'Memory Wall' While Driving Significant Capex Savings
Summaries are public. PDF access requires an active membership; all members have the same access. Sign in
Goldman Sachs examines high-performance NAND (HBF, high-bandwidth flash) as a potential answer to the intensifying memory bottleneck in AI inference. As inference workloads become more memory-intensive and context windows expand exponentially, HBF offers high-density, near-compute storage that can lift GPU utilization and cut the total cost of inference throughput by roughly 30%. Goldman sizes the potential TAM at $100-150 billion by 2030, assuming 15%-50% GPU adoption. In a bull case, HBF bits could account for about 4% of global NAND bit shipments by end-2028 while consuming roughly 9% of NAND wafer capacity given its wafer-size penalty. A single HBF stack provides about 512GB of capacity—around 16x HBM—and HBF can deliver roughly 2.7x energy-efficiency gains at scale, though latency, endurance, thermal management, and manufacturing intensity remain key challenges. Near term, NAND supply should stay tight and HBF may extend the current cycle; Kioxia plans to sample products exceeding 10 million IOPS by end-2026 and a 100-million-IOPS second generation in 2027. Ratings: SanDisk (SNDK) Buy, target $2,200; SK Hynix (000660.KS) Buy, target KRW3,500,000; Samsung Electronics (005930.KS) Buy, target KRW500,000; Kioxia (285A.T) Buy, target ¥38,667; Micron (MU) Neutral, target $1,250; Nvidia (NVDA) Buy, target $300.
Key exhibits
Exhibit 1
Exhibit 2
Exhibit 3
Exhibit 4
Exhibit 5
Exhibit 6
Exhibit 7
Exhibit 8
Exhibit 9
