存储周期分歧 ⟪LINK23⟫
GS 喊出 higher for longer，但买方质疑 CSP 资本开支占比不可持续 多头看好 LTA 锁定和 HBM 挤占产能，空头认为超大厂算不过账
- GS 极度乐观: GS 认为这轮周期比 17-18 年更强，2027/2028 供应依然紧张。传统 DRAM 价格今年涨 300%[1]H-Cap · 06-06"…• Conventional DRAM price increase was 100% in Q1, 50% Q2 • 300% YoY price increase this year. 30% rise in 2027 • NAND: 280%…"，NAND 涨 280%[1]H-Cap · 06-06"…300% YoY price increase this year. 30% rise in 2027 • NAND: 280% YoY price increase this year. $2.1 /GB • HBM: $1.5 /GB. 50%…"，目标 P/E 从 5-6x 提至 9x[1]H-Cap · 06-06"…P/E because market doesnt believe in longevity. 5-6x PE, vs 9x expected by GS • This cycle is higher for longer. 2017-18 w…"。
- 多头逻辑：产能挤占 + LTA 延长 HBM = 3-4x 常规 DRAM 产能[1]H-Cap · 06-06"…24x growth in token consumption by 2030 • Supply: HBM takes 3-4x capacity of regular DRAM, limited supply growth • DRAM supp…"。客户主动签 3-5 年 LTA + 预付定金[1]H-Cap · 06-06"…ow not supplier but driven by customers, who are asking for 3-5 years and making big prepayments • 2027 even tighter than this ye…" → 降低周期性。
- 空头反驳：算不过账 JPM 预计 2027 年存储占 CSP 硬件 capex 将达 73%[2]Terminal Junkie · 06-06"…ex moves from the mid-teens historically to 52% in 2026 and 73% in 2027. That’s a pretty insane shift and basically tells y…"。群内买方直言这不可持续，CSP 不可能把超一半的 capex 花在存储上，"Something has to break"[3]John · 06-06"Good piece. Point 3 should be negative to memory. My own estimate using MU’s last ER numbers is over 33% of hyperscalers…"。
- 算法优化冲击: TurboVec 等算法创新压缩内存使用，引发需求担忧，但也有人认为这只是老调重弹，对实际需求影响有限[4]techbro · 06-07"Lol, how often do we have to have algorithmic innovations that compress memory usage but have no effect on demand. Seems…"。