The consensus is that the AI memory shortage can last until 2027—what data could overturn it? In Micron's (MU) quarterly report in mid-December, if inventory days turn back up and next quarter's gross margin guidance is no longer raised, would that mean the shortage story is over?
- Inventory as a share of single-quarter revenue fell from 20.7% to 19.1%.
- At the end of the same quarter last year it was 13.1%.
- Below 86%, meaning the price floor is starting to be tested.
The consensus is that the AI memory shortage can last until 2027—what data could overturn it? In Micron's (MU) quarterly report in mid-December, if inventory days turn back up and next quarter's gross margin guidance is no longer raised, would that mean the shortage story is over?
共识是 AI 存储缺货能撑到 2027 年，什么数据能推翻它？美光(MU) 12 月中旬的财季报表里，库存天数掉头回升、下一季毛利率指引不再上调，是不是就说明缺货故事讲完了？
The desk answersverifiedWalnut, from public filings
No. Neither of these two points overturns "shortage holding up until 2027," and the first one already happened long ago: in Micron's quarter ended September 3, inventory days had already turned upward; there was no need to wait until December. The second point is about prices peaking, not the shortage ending. In this cycle they are two separate variables.
Inventory days have already turned, but what is rising is work in progress
At the end of September 3, inventory was 103.72 yi dollars, single-quarter cost of revenue was 71.82 yi, and inventory days were about 131 days. The previous three quarters were 124.5 days, 123.2 days, and 121.8 days. This is calculated by dividing ending inventory by single-quarter cost of revenue and multiplying by 91 days; it is an estimate, and the company does not disclose this metric.
Broken down, almost all of the increase is work in progress. Work in progress rose from 69.60 yi to 82.52 yi, raw materials rose from 9.86 yi to 11.62 yi, and finished goods rose from 6.21 yi to 9.58 yi. Finished goods account for only 9.2% of total inventory; at the end of the same quarter last year it was 13.1%. Inventory as a share of single-quarter revenue fell from 20.7% to 19.1%.
This is the difference between stocking up and piling up: wafer starts are increasing, but finished goods are not piling up. And the denominator of this days figure is cost, not selling price—single-quarter cost of revenue rose only 12.2%, while revenue rose 30.8%. In a price-increasing cycle, cost inherently lags, so the days figure will passively rise. Using it as the first signal that "the shortage is over" would lead to a direct misjudgment in this cycle.
Gross margin guidance is no longer being raised, and that happened in September
What is really becoming duller is the pace:
| Report date | Next quarter gross margin guidance | Actual result | Beat |
|---|---|---|---|
| 2025-09-23 | 51.5% | 56.8% | 5.3 points |
| 2025-12-17 | 68.0% | 74.9% | 6.9 points |
| 2026-03-18 | 81% | 84.9% | 3.9 points |
| 2026-06-24 | 86% | 87.0% | 1.0 points |
| 2026-09-30 | 86.25% | To be reported | — |
All four times it beat, but the magnitude of the beat narrowed from 5.3 points to 1.0 point; after the guidance itself reached 86%, this step moved only 0.25 point. Revenue is the same: between two consecutive revenue guidance figures, quarter-on-quarter fell from +49.3% to +23.0%, implying only +13.4% for the just-reported 542 yi.
So "next quarter's gross margin guidance will no longer be raised" is not new information to wait for until December; it was already like this in the September 30 report. What it overturns is "another big beat every quarter," not the shortage.
Prices are locked by contracts, volumes are not
Gross margin moving sideways corresponds to another thing this quarter: there is a cap on prices.
Note 16 of the 10-K is very clear. As of September 3, unfilled contractual obligations were about 1340 yi dollars, of which about 130 yi had been recorded as contract liabilities, and about one quarter will be recognized in the next 12 months; last year this item was not yet material. In 2026, cash received from customer contract deposits was 127.47 yi, almost all scheduled to be refunded in 2029 to 2031; there were also letters of credit totaling 70 yi dollars. The contracts are take-or-pay, locking in volume, and most have price floors and ceilings: if the customer meets the minimum purchase volume, the deposit is refunded; if not, Micron can retain all or part of the deposit as revenue.
Management said in the same filing that for contracts with price bands, even if priced at the floor, gross margin would be "meaningfully above our peak quarterly margins in any past cycle" (original text: meaningfully above our peak quarterly margins in any past cycle).
This statement cuts both ways. The ceiling caps gross margin from rising further, and the floor caps gross margin from falling. One supporting evidence: the cloud storage business unit's gross margin this quarter was 83%, exactly the same as last quarter, the only one of the four business units that did not move (core data center 87% to 90%, mobile and client 87% to 90%, automotive and embedded 79% to 84%).
So the way this shortage exits is not a spot price plunge, but contracts expiring without renewal. Using "gross margin peaking" to infer "the shortage is over" omits this layer in between.
The most fragile part of 2027 is in the second half
Fiscal 2026 DRAM revenue was 1006.79 yi, year-on-year +252%, of which average price rose about 180%, and shipments rose only twenty-something percent; NAND revenue was 317.85 yi, year-on-year +274%, average price rose about 200%, and shipments likewise rose twenty-something percent. Revenue growth was almost entirely price, and supply volume barely moved.
Micron's own fiscal 2027 spending plan: first quarter 115 yi, first half total about 250 yi (fiscal 2026 full-year net was only 273 yi), and higher in the second half. The first DRAM wafer from the first Boise fab will be produced in mid-2027, the second fab at the end of 2028, and New York in 2030.
Supply is not absent before 2027; it arrives concentrated in the second half of 2027. The reading of "holding up until 2027" most easily broken is to understand it as holding up for the full year. In addition, Micron is only one of three original manufacturers; the spending pace of the other two is not in this report, and this table can only disprove part of it.
For the December report, look at four things in order
Micron released last year's first-quarter report on December 17; this year, as usual, it should also be in mid-month, and the specific date has not yet been announced. Only three types of data can overturn it: locked volumes, supply that lands, and finished goods that pile up. On the report these correspond to four lines.
- Gross margin guidance for next quarter. Below 86%, meaning the price floor is starting to be tested.
- Customer contract deposit balance. A net decrease appears at 127 yi, indicating customers are cutting volume, not a price issue.
- Whether the 1,340 yi of unsatisfied contract obligations continues to increase. If it stalls, it means the pace of new contract signings has broken.
- Finished goods inventory amount. Only a second consecutive quarter of increase, exceeding 12% of total inventory, counts as a backlog.
If the first two appear at the same time, the shortage story is truly over. Looking at only the first one is not enough.
What the tape says is not the end of the shortage
All four memory names fell together: SanDisk -13.47%, Seagate -8.0%, Western Digital -7.4%, Micron -5.3%. But on the same day it was not only memory falling: network equipment and optical modules that do not buy memory also fell (Arista -8.3%, Ciena -8.2%, Fabrinet -6.1%, Lumentum -8.6%, Applied Optoelectronics -12.80%), and the four equipment names (Applied Materials -7.5%, Lam Research -6.0%, KLA -5.7%, ASML -6.5%) fell together too. Pricing aimed only at "the end of the memory shortage" would not knock down optical modules unrelated to memory.
Intel -28.58% is separate, its magnitude is not on the same scale as the sector, it is the stock's own matter. Qualcomm -11.12%, AMD -10.88%, Dell -10.77%, Coherent -11.34% are the ones moving in the same direction as the sector.
A-shares are more direct: the falling names had turnover from 8.4 yi to 216.1 yi (Eoptolink -14.19%, Victory Giant Technology -13.31% 86.3 yi, Sunway Communication -10.39% 37.8 yi, Robotechnik -10.31% 32.5 yi, Tengjing Technology -10.30% 8.4 yi), while the gainers were upstream materials like copper-clad laminate (Jin'an Guoji +10.00% 40.8 yi) and two small caps under 10 yi (Mankun Technology +20.01% 8.3 yi, Jinyang Precision +10.81% 4.6 yi). This looks more like money moving upward and shrinking-volume gaming, not like someone got new information that the shortage is over.
The strongest opposing argument: this round of price increases was not earned from a shortage at all, it was customers financing Micron's capacity expansion on its behalf, and the best day on the statements is the endpoint. The buyers' money does not come from end consumers, it comes from the capital budgets of a few mega-fabs; budgets peak in 2027, and take-or-pay turns into preferring to forfeit deposits rather than take goods, the 127 yi of deposits is only a fraction of their annual spending. At the same time Micron has pulled first-half 2027 spending up to 250 yi, and supply lands concentrated in the second half of 2027. Management's line that "even at the floor price it is above the historical peak" describes the price environment at the time of signing, not a 2027 promise. Contracts can lock quantity, they cannot lock that "this electricity actually gets used."
What would count as me being wrong: in that December statement, if next-quarter gross margin guidance is below 86%, and at the same time customer contract deposits show a net decrease from 127 yi, these two appearing together would mean I mistook "customers are pressing for lower prices" for "customers are still asking for goods," and the reality is the demand side is cutting volume. Only gross margin guidance not moving does not count as me being wrong.
Sources:
- Micron Technology, Inc., 8-K Exhibit EX-99.1, a2026q4ex991-pressrelease.htm, 2026-09-30, accession 0000723125-26-000018 (fourth fiscal quarter figures, first fiscal quarter guidance, four business units)
- Micron Technology, Inc., 10-K, mu-20260903.htm, 2026-10-09, accession 0000723125-26-000023 (Note 16 Revenue and customer contract liabilities; capital expenditure arrangements in the Liquidity section; take-or-pay clauses in the Business section)
- Micron Technology, Inc., 8-K Exhibit EX-99.1, a2026q3ex991-pressrelease.htm, 2026-06-24, accession 0000723125-26-000013 (fourth fiscal quarter guidance 86%)
- Micron Technology, Inc., 8-K Exhibit EX-99.1, a2026q2ex991-pressrelease.htm, 2026-03-18, accession 0000723125-26-000004 (third fiscal quarter guidance 81%)
- Micron Technology, Inc., 8-K Exhibit EX-99.1, a2026q1ex991-pressrelease.htm, 2025-12-17, accession 0000723125-25-000044 (second fiscal quarter guidance 68%)
- Micron Technology, Inc., 8-K Exhibit EX-99.1, a2025q4ex991-pressrelease.htm, 2025-09-23, accession 0000723125-25-000024 (first fiscal quarter guidance 51.5%)
- Inventory line items for finished goods, work in process, and raw materials: SEC XBRL companyfacts, CIK 0000723125, data retrieved 2026-11-01
- Market price change and trading value: the most recent trading day's closing data recorded on this site
不是。这两条都推不翻"缺货撑到 2027 年"，而且第一条早就发生了：美光 9 月 3 日那个季度，存货天数就已经掉头向上，不用等到 12 月。第二条说的是价格见顶，不是缺货结束。这轮周期里它们是两个变量。
存货天数已经掉头，但涨的是半成品
9 月 3 日期末存货 103.72 亿美元，单季营业成本 71.82 亿，存货天数约 131 天。前三个季度是 124.5 天、123.2 天、121.8 天。这是我用期末存货除以单季营业成本乘 91 天算的，估算值，公司不披露这个指标。
拆开看，涨的几乎全是半成品。在产品从 69.60 亿加到 82.52 亿，原材料从 9.86 亿加到 11.62 亿，成品从 6.21 亿加到 9.58 亿。成品只占总存货 9.2%，去年同一个季度末是 13.1%。存货占单季收入的比重从 20.7% 降到 19.1%。
备货和积压的差别就在这：投片在加，成品没堆起来。而且这个天数的分母是成本不是售价——单季营业成本只涨 12.2%，收入涨 30.8%，涨价周期里成本天生滞后，天数会被动走高。拿它当"缺货结束"的第一个信号，这一轮会直接误判。
毛利率指引不再上调，9 月就发生了
真正在变钝的是节奏：
| 报告日 | 下一季毛利率指引 | 实际做出来 | 超出 |
|---|---|---|---|
| 2025-09-23 | 51.5% | 56.8% | 5.3 个点 |
| 2025-12-17 | 68.0% | 74.9% | 6.9 个点 |
| 2026-03-18 | 81% | 84.9% | 3.9 个点 |
| 2026-06-24 | 86% | 87.0% | 1.0 个点 |
| 2026-09-30 | 86.25% | 待报 | — |
四次全超，超的幅度从 5.3 个点收到 1.0 个点；指引本身走到 86% 之后，这一步只挪了 0.25 个点。收入同样：连续两次给出的收入指引之间，环比从 +49.3% 掉到 +23.0%，对刚报完的 542 亿只隐含 +13.4%。
所以"下一季毛利率指引不再上调"不是 12 月才要等的新消息，9 月 30 日那份报表里已经是这个样子。它推翻的是"每季再超一大截"，不是缺货。
价格被合同封住了，量没被
毛利率横着走，在这个季度对应的是另一件事：价格有上限。
10-K 第 16 注写得很直白。9 月 3 日尚未履行的合同义务约 1340 亿美元，其中约 130 亿已作为合同负债入账，约四分之一在未来 12 个月确认；去年同期这项还不重大。2026 年收到客户合同定金现金 127.47 亿，几乎全部安排在 2029 至 2031 年退还；另有信用证合计 70 亿美元。合同是 take-or-pay，锁数量，多数带价格上下限：客户完成最低采购量，定金退；没完成，美光可以把定金全部或部分留下当收入。
管理层在同一份文件里说，带价格带的合同即便按下限价定价，毛利率也会"明显高于历史上任何一个周期峰值的单季毛利率"（原文：meaningfully above our peak quarterly margins in any past cycle）。
这句话是双向的。上限封住了毛利率继续往上，下限封住了毛利率往下掉。一个旁证：云端存储事业部这一季毛利率 83%，和上季一模一样，是四个事业部里唯一没动的（核心数据中心 87% 到 90%，移动与客户端 87% 到 90%，汽车与嵌入式 79% 到 84%）。
所以这轮缺货退场的方式不是现货价跳水，是合同到期不续。用"毛利率见顶"推"缺货结束"，中间漏了这一层。
2027 年最脆弱的地方在下半年
2026 财年 DRAM 收入 1006.79 亿，同比 +252%，其中均价涨约 180%，出货量只涨二十几个百分点；NAND 收入 317.85 亿，同比 +274%，均价涨约 200%，出货量同样二十几个百分点。收入增长几乎全是价格，供给量几乎没动。
美光自己给的 2027 财年开支安排：第一季 115 亿，上半年合计约 250 亿（2026 全年净额才 273 亿），下半年更高。博伊西第一座厂第一片 DRAM 晶圆 2027 年年中产出，第二座 2028 年底，纽约 2030 年。
供给不是 2027 年之前不来，是 2027 年下半年集中来。"撑到 2027 年"最容易被打破的读法，是把它理解成撑满全年。另外美光只是三家原厂之一，另两家的开支节奏不在这份报表里，这张表只能证伪一部分。
12 月那份报表，按顺序看四样
美光上一年度第一季是 12 月 17 日发的，今年按惯例也在中旬，具体日子还没公告。能推翻它的只有三类数据：锁住的量、落地的供给、堆起来的成品。报表上对应四行。
一、下一季毛利率指引。低于 86%，是价格下限开始被试探。
二、客户合同定金余额。127 亿出现净减少，说明客户在退量，不是价格问题。
三、尚未履行的合同义务 1340 亿有没有继续加。停住，说明新签合同节奏断了。
四、成品存货金额。连续第二个季度上升、占总存货过 12%，才叫积压。
前两条同时出现，缺货故事就真讲完了。只看第一条不够。
盘面说的不是缺货结束
存储四家一起跌：闪迪 -13.47%、希捷 -8.0%、西部数据 -7.4%、美光 -5.3%。但同一天跌的不只是存储：不买存储的网络设备和光模块也在跌（Arista -8.3%、Ciena -8.2%、Fabrinet -6.1%、Lumentum -8.6%、应用光电 -12.80%），设备四家（应用材料 -7.5%、泛林 -6.0%、科磊 -5.7%、阿斯麦 -6.5%）也一起。只针对"存储缺货结束"的定价，不会把跟存储无关的光模块一起打下去。
英特尔 -28.58% 另算，幅度跟板块不在一个量级，是个股自己的事。高通 -11.12%、AMD -10.88%、戴尔 -10.77%、Coherent -11.34% 才跟板块一个方向。
A 股更直白：跌的几只成交额从 8.4 亿到 216.1 亿（新易盛 -14.19%、胜宏科技 -13.31% 86.3 亿、信维通信 -10.39% 37.8 亿、罗博特科 -10.31% 32.5 亿、腾景科技 -10.30% 8.4 亿），涨的是覆铜板这种上游材料（金安国纪 +10.00% 40.8 亿）和两只不到 10 亿的小票（满坤科技 +20.01% 8.3 亿、金杨精密 +10.81% 4.6 亿）。这更像钱往上挪、缩量博弈，不像有人拿到了缺货结束的新消息。
反方最强的一条：这轮涨价根本不是缺货赚的，是客户替美光融资扩产，报表最好的一天就是终点。买方的钱不是终端消费者给的，是几家超大厂的资本预算；2027 年预算见顶，take-or-pay 就变成宁可损失定金也不要货，127 亿定金对它们一年的开支只是零头。同时美光把 2027 上半年开支拉到 250 亿，供给在 2027 下半年集中落地。管理层那句"按下限价也高于历史峰值"，讲的是签约当时的价格环境，不是 2027 年的承诺。合同能锁数量，锁不住"这些电真的被用掉"。
什么情况算我看错：12 月那份报表里，如果下一季毛利率指引低于 86%，同时客户合同定金从 127 亿出现净减少，这两条一起出现，就说明我把"客户在压价"错看成了"客户还在要货"，实际是需求端在减量。只看毛利率指引不动，不算我看错。
- 01Micron Technology, Inc.，8-K 附件 EX-99.1，a2026q4ex991-pressrelease.htm，2026-09-30，accession 0000723125-26-000018（第四财季数字、第一财季指引、四个事业部）
- 02Micron Technology, Inc.，10-K，mu-20260903.htm，2026-10-09，accession 0000723125-26-000023（注 16 收入与客户合同负债
- 03流动性章节资本开支安排
- 04业务章节 take-or-pay 条款）
- 05Micron Technology, Inc.，8-K 附件 EX-99.1，a2026q3ex991-pressrelease.htm，2026-06-24，accession 0000723125-26-000013（第四财季指引 86%）
- 06Micron Technology, Inc.，8-K 附件 EX-99.1，a2026q2ex991-pressrelease.htm，2026-03-18，accession 0000723125-26-000004（第三财季指引 81%）
- 07Micron Technology, Inc.，8-K 附件 EX-99.1，a2026q1ex991-pressrelease.htm，2025-12-17，accession 0000723125-25-000044（第二财季指引 68%）
- 08Micron Technology, Inc.，8-K 附件 EX-99.1，a2025q4ex991-pressrelease.htm，2025-09-23，accession 0000723125-25-000024（第一财季指引 51.5%）
- 09存货成品、在产品、原材料分项：SEC XBRL companyfacts，CIK 0000723125，取数 2026-11-01
- 10盘面涨跌幅与成交额：本站记录的最近一个交易日收盘数据
## 共识里其实是两个命题
"缺货撑到 2027"拆开是两件事。一是物理题：2027 年之前缺口不消失。二是价格题：2027 年之前价格不跌。
两条强度差很远。缺口还在的时候，价格可以先横住 —— 中间隔着一层合同、一层库存。美光的毛利率指引只碰得到第二条。
## 同行先交了卷
四家财季末差一个月：美光截到 9 月 3 日，闪迪、西数、希捷截到 10 月 2 日。多算这一个月，内存两家还是横着走，硬盘两家还在往上走。
| 公司 | 单季收入(亿美元) | 非GAAP毛利率 | 下季毛利率指引 | 库存天数 |
|---|---|---|---|---|
| 美光 | 542.3 | 87.0 | 86.25 | 131 |
| 闪迪 | 110.3 | 85.1 | 85.0 | 166 |
| 西数 | 43.8 | 58.1 | 60.5 | 88 |
| 希捷 | 42.3 | 57.8 | 未给 | 86 |
读法不是谁高谁低，是谁还在加价。闪迪 85.1% 配 85.0% 的指引，说的是 NAND 价格在 9 月这季就到头了。美光 87.0% 配 86.25%，同一句话，只是水平更高。
硬盘两家说的相反。西数指引 60.5%，比刚做出来的 58.1% 高 2.4 个点；希捷没给毛利率指引，可收入指引 48 亿比 42.3 亿高 13.4%，每股收益 9.20 比 7.68 高两成。
硬盘是这条链上的泄压阀：冷数据先放硬盘，放不进去才往上压到 NAND 和 DRAM。泄压阀自己在加价，说明压力没找到出口。
所以"毛利率指引不再上调"不是结束的信号，是"涨价停在这一层、下一层接着涨"的信号。
## 库存天数：四家里三家在涨
按期末存货除以单季营业成本乘 91 天算（公司不披露，这是估算）：美光从 122 天到 131 天，西数 80 到 88，希捷 83 到 86，闪迪 178 到 166。
分母是成本，不是售价。涨价周期里成本滞后售价两三个季度，天数会被动走高，跟仓库里堆没堆货没关系。
把分母换成收入，同一个季度给出相反答案。去年 8 月底存货 83.55 亿、单季收入 113.15 亿，按收入算是 67 天；今年 9 月存货 103.72 亿、收入 542.29 亿，只有 17 天。毛利率从 45.7% 走到 87% 的过程里，存货周转快了近四倍，按成本算的天数却从 121 天涨到 131 天。两个口径打架，说明这个指标已经被毛利率污染，不能单独用。
逆着走的只有闪迪：存货涨 12%，收入涨 23%，分母跑得比分子快。如果天数是积压的先行指标，最该出问题的就是它 —— 纯 NAND、涨价最猛、天数最高。结果毛利率 85%、下季指引持平、数据中心收入涨 61%。
天数是高的，生意是好的，两件事同时成立。所以"库存天数掉头"当不了缺货结束的第一信号，它只说明成本在追价格。
## 缺货是被谁吃掉的
闪迪这份报表把传导链露出来了。数据中心 29.8 亿到 48.0 亿，涨 61%；边缘 54.3 亿到 58.3 亿，涨 7%；消费 5.6 亿掉到 4.0 亿，环比 -29%、同比 -39%，只剩收入的 3.6%。
```chart
type: group
title: 闪迪的收入在往数据中心挪，消费端被价格挤出
unit: 亿
labels: 2026Q4财季, 2027Q1财季
values: 29.8, 48.0
values2: 5.6, 4.0
legend: 数据中心|消费
```
不是需求变弱，是需求被价格排了队。原厂把晶圆从消费端挪到服务器，消费端买不到也买不起，"需求消失"的是排在队尾的那部分。
边缘只涨 7%，信息量不比数据中心小。手机和 PC 用同一批颗粒、付同一个价，价格一上来采购就被压平。数据中心还能涨 61%，因为它是唯一不把存储成本当敏感项的买家。
美光分部同方向：核心数据中心 115.24 亿到 180.02 亿，涨 56%；移动与客户端 115.21 亿到 131.14 亿，涨 14%；汽车与嵌入式 46.34 亿到 68.24 亿，涨 47%。四个分部毛利率 83% 到 90%，没有一个往下掉。
所以这条链上受损的不是原厂，是消费电子整机、模组厂、手机 OEM —— 买同一个颗粒、付同一个价，卖的是价格敏感产品，涨价传不到终端，只能自己吞。
末端是云厂的资本预算，不是消费者的钱包。"缺货什么时候结束"等价于问"云厂预算什么时候见顶"，而这件事不在美光的报表里。
## 反方最强的一条：尺子是坏的
一个聪明的空头不跟你争毛利率。他会说：86% 是合同下限价托着的，不是现货托着的。合同锁了量、锁了价带，这张利润表是 2026 年签的字，跟 2027 年的供需没关系。拿一张滞后两年的表证明"现在还很缺"，是循环论证。
这个反驳我接受一半，它比"毛利率见顶"强得多。
接受的那一半：毛利率指引确实失去区分能力了。下一季指引 86.25%，闪迪 85.0%，内存两家都横着走。一个"预期之中就会横着走"的数字，出来不算证据，不出来才是证据。
不接受的那一半：合同锁的是价格，不是需求。合同写的是最低采购量，客户完不成，定金留下当收入。
但得给空头再补一刀：130 亿合同负债，只等于美光一个季度收入的 24%、单季营业成本的 1.8 倍，还是几家客户分摊。摊到任何一家超大规模厂商头上，抵不上它一个季度的存储账单。定金是会计约束，不是商业护城河，客户真要走，赔得起。
## 12 月该看的三行
第一行，客户合同定金。9 月 3 日合同负债 130 亿，2026 财年收进 127.5 亿，几乎全部安排在 2029 到 2031 年退。净减少，就是客户在退量，不是客户在压价。
第二行，未履行合同义务。1340 亿，约四分之一在 12 个月内确认。停止增长，说明新签合同断了，比毛利率早一两个季度。
第三行，成品占总存货的比重。现在 9.2%，去年同期 13.1%。连续两季上升、并且过 12%，才叫积压。
前两行同时出现，缺货故事才算讲完。只看毛利率指引，等于看一个已经不会动的指针。我错在哪也很具体：12 月季报里定金净减、RPO 同时不增，就是我把"客户在压价"错看成了"客户还要货"。只有毛利率指引不动，不算我错。
时间上也是美光先报，闪迪和西数要等到 1 月底。美光定调、硬盘两家验证：如果美光说量还在、价停了，而西数把 60.5% 又往上修一格，压力还在往下传就被确认了一次。
## 2027 的危险在下半年
美光自己给的 2027 财年开支：第一季 115 亿，上半年 250 亿，下半年更高；2026 全年净开支 274 亿。上半年的钱已经等于去年一整年。
钱变成 wafer 要时间。博伊西第一座厂首片 DRAM 晶圆 2027 年年中，第二座 2028 年底；纽约第一座今年 1 月才破土。供给不是 2027 年之前不来，是下半年开始集中来。
边界要说清：这张表只有美光一家。三星和 SK 海力士的节奏不在这份文件里，而它们在 HBM 上份额更大。所以用美光报表只能证伪"全年都紧"，证伪不了"上半年还紧"。
还有一层：被 10-K 点名为竞争对手的长鑫、长江存储，如果真把价格打下来，先松的多半是消费端和小容量，不是 HBM 和高端服务器条。消费端跌价和 AI 端缺货，可以在同一个季度里同时成立。
## 最后
回到最初：库存天数掉头回升、毛利率指引不再上调，是不是缺货故事讲完了。不是。前一条是成本在追价格，后一条是价格停在合同上限，两条都发生在 9 月，都不指向需求。
真正会先说话的不是美光，是西数下一季的毛利率。泄压阀还堵着，压力就还在。
来源：
1. Micron Technology, Inc.，8-K 附件 EX-99.1，a2026q4ex991-pressrelease.htm，2026-09-30，0000723125-26-000018（第四财季收入、毛利率、四个事业部、第一财季指引）
2. Micron Technology, Inc.，10-K，mu-20260903.htm，2026-10-09，0000723125-26-000023（注 16 履约义务与合同负债；2027 财年资本开支；博伊西投产时点）
3. Sandisk Corporation，8-K 附件 EX-99.1，sndkq1-27ex991xpressrelease.htm，2026-10-29，0001628280-26-068243（第一季收入、毛利率、分部、第二季指引）
4. Western Digital Corp.，8-K 附件 EX-99.1，a4ex991-pressreleaseq127.htm，2026-10-22，0001628280-26-066882（第一季收入、毛利率、第二季指引）
5. Seagate Technology plc，8-K 附件 EX-99.1，2026-10-27，0001137789-26-000300（第一季收入、毛利率、第二季指引）
6. SEC XBRL companyfacts，CIK 0000723125 / 0002023554 / 0000106040 / 0001137789，2026-11-01（存货金额；库存天数为估算）