· 供应链分析

内存价格预计涨至2028年,AI结构性需求支撑利润,即使利润率下降行业仍被看好

涉及标的:

中文翻译

内存周期。 大概会像这张图一样吗? 以SK Hynix、$SNDK、三星、$MU等为例: -> 价格持续上涨至2028年 -> 需求增长具有持久性 -> 2028年后价格回落 -> 产能增加 * 需求增加 * 利润率下降 = 利润依然可观。 例如: 2026年Q1:DRAM、NAND价格上涨 - NAND价格环比上涨100%+,DRAM上涨70%+。 2026年Q2:DRAM NAND价格上涨 - 三星Q2 NAND价格再次上涨100%+,DRAM上涨。 Q3至2028年价格持续上涨。 1. 反方观点:「在2027年下半年没有任何情景会让内存价格回调,考虑到超大规模云厂商的采购意愿丝毫未减」 2. 英特尔CEO:「2028年前内存短缺不会缓解」。 然而人们误解了的是: -> 内存需求是结构性的,由AI驱动。 -> 但价格不是(结构性的)。 2026年极端内存短缺情况下,我们很可能持续看到价格上涨。 但价格会在2028年开始下跌。人们混淆的是: -> AI带来的极端需求不会导致价格跌到零。 -> 更多产能不会导致需求突然归零。 供给增加 * 价格 * 需求增加 * 更低运营利润率 = 利润依然可观。 运营利润不会像现在这样环比增长10000%+。 但如果SK Hynix通过产能增加、利润率下降但保持稳定运营利润,年度运营收入约1000亿美元以上,市值4000亿美元: 对比1000亿 -> 2200亿 -> 900亿 -> 1200亿的模式。 那就看起来被低估了。 我不认为会出现末日派所预测的1000亿->1800亿(2027年)->100亿的情况,即因需求下滑(如智能手机)和利润率下滑而亏损运营。 主要需要关注两件事:软件/内存使用效率是否变得极高,或者超大规模云厂商资本支出突然消失(AI不再是回事)。 GPU训练/推理也是同样的逻辑。 但我最主要关注的是超大规模云厂商资本支出预测,作为第一指标。不是断章取义地从三星高管那里摘取几句话来预测两年后的运营收入。 AI从根本上改变了「商品」内存的本质,类似于2023年的GPU。

英文原文

The Memory Cycle. Is probably going to look like this chart? With SK Hynix, $SNDK, Samsung, $MU and others: -> Price Hikes until 2028 -> Demand Increase Permanent -> Price Decreases After 2028 -> Increased Capacity * Increased Demand * lower margin = High Profit Anyway. For example: 2026 Q1: DRAM, NAND Price Hikes - NAND prices 100%+ Q/Q, DRAM up 70%+. 2026: Q2: DRAM NAND Price Hikes - Samsung hikes Q2 NAND prices 100%+ Again, DRAM up. Q3 -> 2028 Price Hikes. 1. Counterpoint: "There is no scenario where memory prices correct in the second half [of 2027], given that hyperscaler purchasing intent remains unbroken" 2. Intel CEO: "No Relief on Memory Shortage Until 2028". However what people misunderstand: -> Memory Demand Is Structural with AI. -> Prices are not. We'll likely keep seeing price hikes with the extreme memory shortage in 2026. But prices start to fall in 2028. What people conflate is: -> Extreme demand for AI will not cause prices to go to 0. -> More capacity will not cause demand to suddenly go to 0. More Supply * Price * More Demand * lower operating margin = more profit anyway. Operating income will not be 10000%+ Q/Q like now. But if SK Hynix is producing a more steady ~$100B+ operating income Y/Y at a $400B MC from increased capacity but lower margin: Compared to $100B -> $220B -> $90B -> $120B. Then that itself looks undervalued. I don't see a world where it ends up being $100B-> $180B (2027) -> $10B what doomposters are projecting, where they operate at a loss from both demand downturn (eg. smartphones) and margin downturn. The main two things is to look out for is if software/memory usage gets extremely, extremely efficient or hyperscaler capex suddenly disappears (AI is no longer a thing). Same could be said around GPUs for training/inference. But, I would mainly be looking out for hyperscaler capex projections as the #1 indicator. Not random out of context quotes taken from Samsung executives to signal operating income two years out. AI has fundamentally changed what the "commodity" memory is, similar to GPUs back in 2023.

在 X 上查看原推 ↗