Micron's chief executive has warned that memory is only going to get scarcer, saying in comments reported by Ars Technica in October 2026 that prices for memory sold for 2027 "are much higher than 2026 prices." The same report frames the expectation among memory industry executives that the shortage will persist into 2028 — a horizon that matters to anyone signing multi-year supply agreements rather than buying at retail today.
The timing could hardly be worse for buyers. Enterprise refreshes, data-centre expansions and cloud contracts are negotiated years before delivery, which means the terms being signed now will be repriced by whatever the market looks like in 2027. The chief executive is telling you, in effect, that you will pay materially more than you do today.
The critical caveat: nobody has attached a number to that warning. Ars Technica's reporting offers no percentage, no contract-versus-spot split, and no separate figures by memory type. Treat "much higher" as a directional signal, not a procurement input. The sensible response is to track spot and contract pricing against your own internal baselines before committing to any 2027 supply agreement.
What is driving the squeeze
This is an allocation problem, not a production failure. HBM — the stacked, high-bandwidth memory that goes into AI accelerators — soaks up wafer capacity that would otherwise feed conventional DDR5. Fabricators are not missing output targets; they are choosing where to spend their most profitable wafers, and the answer is the high-margin tier rather than the commodity one.
That distinction should reshape how buyers think about the shortage. You are not short of memory the way a factory runs short of raw materials. You are short of the specific products that have become the most profitable to build, which is why supply can look plentiful in aggregate while the tier you actually need stays tight.
Readers should note this framing comes from our reading of the reported exchange, not from figures Micron published — but the mechanism it describes is the one the industry has been describing publicly for the past two years.
What it means for the people who buy
Enterprise and data-centre procurement. The question is not what memory costs today, but what your 2027 allocations will cost and whether the current refresh cycle can be re-timed. Where contracts allow, this is the window to test whether the estate you already run can be right-sized before you sign for more capacity at what may be peak pricing.
Cloud customers. Expect memory costs to surface in 2027 renewal conversations, where they are increasingly buried in instance pricing rather than billed separately. Rightsizing — cutting over-provisioned instances, tightening memory-to-vCPU ratios, retiring idle workloads — is likely to be cheaper than absorbing a broad increase across the bill.
Self-hosters. Buy what you need and stop chasing spec sheets. Memory-heavy configurations are the ones most exposed to 2027 pricing risk, so before committing to a bigger build, price out whether a smaller, well-tuned system does the job at materially lower cost — and weigh whether deferring the purchase is cheaper than funding it at next year's prices.
The counterpoint
This is a directional warning, not a prophecy of permanent shortage. Several things could soften the squeeze: new fab capacity expected to come online, any cooling in AI demand, and the ramp of Chinese memory-makers expanding output. None of those is certain, and — if the chief executive's assessment holds — none arrives in time to change 2027 pricing.
That combination is exactly why the signal matters now. Capacity coming online in 2029 does nothing for a contract signed in 2026.
Bottom line
The direction is clear even though the magnitude is not: memory sold for 2027 will cost more than 2026 memory, and outside the fabs, nobody can say by how much. In a shortage, the cheapest capacity is the capacity you never buy. Efficiency, right-sizing and deferral are the levers available today — price those before you price the upgrade.
Source: Ars Technica
美光（Micron）行政總裁警告，記憶體供應只會愈來愈短缺，並在 Ars Technica 於2026年10月報道的評論中指出，2027年供應的記憶體「售價將大幅高於2026年的價格」。同一報道亦指出，記憶體業界管理層普遍預期短缺將延續至2028年——這個時間節點，對任何簽署多年期供應協議、而非今日零售採購的人而言，都至關重要。
對買家而言，時機幾乎壞到不能再壞。企業設備更新、數據中心擴建及雲端合約通常在交付前數年已經開始談判，這意味著現在簽署的條款，將按2027年的市況重新定價。換言之，行政總裁實際上是在告訴你：屆時你將付出遠高於今天的代價。
必須注意的關鍵限制：沒有人為這個警告附上具體數字。Ars Technica 的報道既沒有百分比，也沒有合約價（contract）與現貨價（spot）的拆分，更沒有按記憶體類型劃分的數據。應把「大幅提高」視為方向性信號，而非採購決策輸入。最穩妥的做法，是在簽訂任何2027年供應協議之前，先以自身內部基準追蹤現貨及合約價格走勢。
短缺的成因何在
這是資源分配問題，而非生產故障。HBM——即用於 AI 加速器的堆疊式高頻寬記憶體——大量吞噬了原本會供應傳統 DDR5 的晶圓產能。代工廠並非未達產能目標，而是正在選擇把最賺錢的晶圓投產於何處，答案是高毛利產品線，而非大宗商品級別的記憶體。
這個區別應改變買家對短缺的理解。你所面對的並非如工廠缺原材料般的全面短缺，而是那些製造利潤最高的具體產品出現供應緊張——這正是為何整體產能看似充裕，你實際需要的產品級別卻依然短缺。
讀者須注意，此框架出自我們對上述報道對話的解讀，並非美光公布的數據——但其所描述的機制，正是業界過去兩年公開談及的問題。
對採購方意味著什麼
企業及數據中心採購。 問題不在於記憶體今日的價錢，而在於2027年配額的成本，以及現有更新周期能否重新調配時間。在合約允許的範圍內，此時正是檢視現有基建是否能重新合理化（right-size）的好時機，以免在可能的價格高位簽訂更多容量。
雲端客戶。 記憶體成本預計會在2027年的續約談判中浮現，而且越來越多時已隱藏在實例（instance）定價之內，而非獨立計費。規模優化——削減過度配置的實例、收緊記憶體與 vCPU 的比例、退役閒置工作負載——很可能比全面吸收帳單加價更為划算。
自架服務者。 只買你真正需要的規格，不要盲目追趕規格表。記憶體佔比高的配置，正是2027年價格風險最高的類別，因此在決定擴建之前，應先估算一套規模較小但經過調校的系統能否以顯著更低的成本完成工作——並衡量延後採購是否比以明年價格付款更為便宜。
反面觀點
這只是一個方向性警告，並非預言永久短缺。若干因素可能緩解擠壓：新建晶圓廠產能預計陸續投產、AI 需求可能放緩，以及中國記憶體廠商擴產後的產能釋放。以上皆不確定——而且，若行政總裁的判斷成立，任何一項都無法趕及改變2027年的定價。
正是這種組合，凸顯這個信號為何現階段如此重要。2029年才投產的產能，對2026年簽署的合約毫無幫助。
總結
即使幅度未知，方向已經清晰：2027年供應的記憶體將比2026年更貴，而晶圓廠以外，沒有人能說得出差距有多大。在短缺期間，最便宜的容量，就是你從來不買的容量。效率提升、規模優化及延後採購，就是今天可以使用的槓桿——在估算升級成本之前，先估算這三者。
來源： Ars Technica