跳到主要內容

記憶體缺貨評測:AI吃光DRAM,2030年才解封 | AI Memory Shortage Review: DRAM Sold Out Until 2030

By Kit 小克 | AI Tool Observer | 2026-10-30

🇹🇼 記憶體缺貨評測:AI吃光DRAM,2030年才解封

先講重點:記憶體缺貨不是你一個人的錯覺。SK海力士(SK hynix)上週公布財報,利潤暴增632%,CEO直接講DRAM跟HBM產能已經賣到2029年,缺貨狀況可能一路撐到2030年甚至更久。這一切的源頭都是AI——資料中心為了塞滿GPU,把全球記憶體產能幾乎整碗端走,留給一般消費者的份額被壓縮到史上新低。

記憶體缺貨多誇張?數字說話

  • SK海力士Q3營收與利潤雙創新高,股價單日暴漲11%
  • 過去一年DRAM現貨價飆漲將近700%
  • HBM(AI晶片用的高頻寬記憶體)需求是供給的2倍
  • SK海力士已跟14家客戶簽長約,鎖住超過一半的2027~2028年產能

為什麼AI會讓記憶體缺貨撐到2030年?

答案很簡單:蓋資料中心的錢比你買一條RAM出的錢香太多。輝達(Nvidia)、OpenAI這些買家要的是HBM——疊層堆疊、專門配給GPU用的高階記憶體,毛利遠高於一般消費級DRAM。SK海力士、美光(Micron)、三星自然把產能優先留給AI訂單,消費級記憶體被排到產線最後面。The Register的報導指出,SK海力士CEO郭魯正甚至說2027年可能是記憶體產業史上供給最吃緊的一年,而且這波記憶體缺貨不會隨AI泡沫話題降溫就消失,因為機器人、自動駕駛等「實體AI」接下來也要搶記憶體。

消費者會被影響到什麼程度?

  • DDR5 RAM今年第一季報價暴漲110%,一條64GB套件從約195美元飆到788美元
  • SSD漲幅更誇張,最高147%
  • 品牌PC整機售價最高上調20%,均價比去年漲17%
  • 美光已宣布縮減消費級記憶體產線,產能優先供應AI與伺服器客戶

現在該不該買電腦、加記憶體?

老實說,如果你近期真的需要升級或換機,現在下手大概比半年後便宜——因為沒人敢保證價格會回頭。但如果只是「怕漲價先囤一波」,風險也不小:記憶體缺貨是供需結構問題,不是短期炒作,價格起伏還會跟著AI資本支出週期波動,現在衝進去買不一定買在低點。Tom's Hardware的深度分析也認為,這波記憶體缺貨本質上是AI基礎建設跟你家電腦在搶同一塊晶圓產能,短期內看不到解方。

好不好用,試了才知道。


🇺🇸 AI Memory Shortage Review: DRAM Sold Out Until 2030

AI memory shortage is not just a meme — SK Hynix posted record profit up 632% last week, and its CEO said DRAM and HBM capacity is sold out through 2029, with the shortage likely persisting through 2030 or beyond. The cause is simple: AI data centers are eating nearly all the world's memory supply to feed GPUs, leaving ordinary consumers with the smallest slice in history.

How Bad Is the AI Memory Shortage, in Numbers

  • SK Hynix stock jumped 11% in a single day after record Q3 earnings
  • DRAM spot prices have surged nearly 700% over the past year
  • HBM (the high-bandwidth memory AI chips need) demand now runs at roughly 2x available supply
  • SK Hynix has locked in long-term deals with 14 customers covering over half of its 2027-2028 capacity

Why Will the Memory Shortage Last Until 2030?

The math is simple: building AI data centers pays better than selling you a stick of RAM. Nvidia, OpenAI, and other buyers want HBM — stacked, high-margin memory built specifically for GPUs — so manufacturers like SK Hynix, Micron, and Samsung route capacity there first, pushing consumer DRAM to the back of the line. According to The Register, SK Hynix CEO Kwak Noh-jung has warned 2027 could be the worst supply year in the memory industry's history, and this AI memory shortage won't ease just because AI-bubble headlines cool off — robotics and self-driving cars are the next wave of demand lining up behind data centers.

How Much Does This Actually Hit Consumers?

  • DDR5 RAM prices jumped 110% in Q1 2026 alone — a 64GB kit went from roughly $195 to $788 at the peak
  • SSD prices spiked even harder, up to 147%
  • Branded PC prices are up as much as 20%, averaging 17% higher than last year
  • Micron has already discontinued parts of its consumer memory business to prioritize AI and server customers

Should You Buy a PC or RAM Right Now?

If you genuinely need an upgrade soon, buying now is probably cheaper than waiting six months — nobody's betting prices will fall back. But panic-buying "before it gets worse" carries its own risk: this is a structural supply-demand shortage tied to AI capex cycles, not a short-term spike, so there's no guarantee you're buying at the bottom either. As Tom's Hardware's deep dive puts it, the AI memory shortage means your next laptop is now competing with a GPU cluster for the same wafer capacity — and that's not solving itself soon.

You'll only know if it's good once you've tried it.

Sources / 資料來源

延伸閱讀 / Related Articles


AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

留言

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code