Will SK Hynix's Q3 earnings call drive a reversal in the memory sector's market performance?
- An 8% beat bought today's 10.9% gain.
- Year-on-year revenue rose 328% and operating profit rose 632%.
- The preview shows operating profit of 107 trillion Korean won.
- Guidance is upward, yet the share price is 20% lower than in June.
Will SK Hynix's Q3 earnings call drive a reversal in the memory sector's market performance?
海力士Q3业绩会推动存储板块行情逆转吗
The desk answers已核实Walnut, from public filings
It can't be pushed. What this earnings call can support is a one-day rebound, not a reversal—today the memory stocks in Korea, Japan, and the US have already risen as much as they can (SK Hynix +10.9%, Samsung +9.0%, Kioxia +13.8%, Micron intraday +4.5%), and of the two most important new pieces of information it gave, one is bullish and one is bearish, and they exactly cancel each other out.
First, distinguish which information is new
SK Hynix's third-quarter revenue was 104.6 trillion Korean won, operating profit 83.3 trillion, operating margin 80%. Revenue rose 32% quarter-on-quarter and operating profit rose 38% quarter-on-quarter; year-on-year revenue rose 328% and operating profit rose 632%.
The 80% figure itself is not new. The operating margin in the second quarter was already 76%, and the third quarter merely took another step along the same line. The real increment is "how much it exceeded expectations": the day before the earnings call, Korean brokerages expected operating profit of 77 trillion, and the actual figure was 83.3 trillion, an 8% beat.
An 8% beat bought today's 10.9% gain. That account was settled in one day.
What is easy to read backwards is net profit. Third-quarter net profit was 39.8 trillion, 58% less than the second quarter's 93.9 trillion.
This is not a deterioration in operations—second-quarter net profit was far higher than that quarter's operating profit (60.5 trillion), indicating that part of those three months' profit came from non-operating items. The company said in the same materials that the gains from its Kioxia investment will be returned to shareholders this year. When looking at this company, do not judge by quarter-on-quarter net profit.
One positive item and one negative item, written in the same document
The positive item is new, and for the first time it gives specific numbers: the company has signed long-term contracts with 14 major customers, locking in volume equivalent to more than half of total capacity for 2027 and 2028 combined, with a typical term of five years. This indeed pushes it one step from "a cyclical stock selling spot" toward "a company with contracted revenue."
But the sentence immediately after is more critical—the long-term contracts include a mechanism to "reflect market conditions within a price range agreed with customers." In plain terms: volume is locked, price is not. Output is guaranteed, but the risk of price declines remains on its own financial statements. If this is bought as "long-term contracts guaranteeing prices," that is a mistake.
The negative item is also new, and officially confirmed: next year's capital expenditure will "increase substantially," Yongin Phase 1 will fully enter production, M15X will be ramped to full capacity by year-end dedicated to HBM4, and 321-layer NAND will account for more than half of domestic capacity in the fourth quarter.
What the market has been selling since late June is precisely this. A July 20 report in Korea's Maeil Business Newspaper had a headline to the effect that all manufacturers are expanding production, and concerns about oversupply have reappeared.
Over these four months, SK Hynix fell from 2.919 million Korean won to 1.341 million on October 27, a decline of 54%, and every earnings report during the period was a record.
So the nature of this earnings call is not "the bad news is fully out," but rather placing the latest evidence from both the bullish and bearish sides on the table at the same time: the demand side is locked in for two years, and the supply side is also being opened up in parallel. The disagreement has not been eliminated; it has merely become easier to reconcile.
All three original manufacturers are setting records, and their share prices are still far below their highs
| Company | Latest quarter revenue | Operating margin | From June high | October 28 |
|---|---|---|---|---|
| SK Hynix | 104.6 trillion won | 80.0 | (49.1) | 10.9 |
| Micron (FQ4) | 542.3 yi USD | 80.7 | (20.0) | 4.5 |
| Samsung Electronics (preliminary) | Not disclosed | — | (33.2) | 9.0 |
| Kioxia | Not disclosed | — | (60.4) | 13.8 |
The unit is percent; Micron is intraday price; Samsung currently only has an October 8 earnings preview, with formal details to be released on October 29; the preview shows operating profit of 107 trillion Korean won.
Micron's guidance for the next quarter is clearer in direction: revenue 61.5 billion US dollars (+13% quarter-on-quarter), gross margin 86.25%, earnings per share 38.15 US dollars. Guidance is upward, yet the share price is 20% lower than in June.
The problem is clear by this point: good performance has long since been priced in by the market; what has not been priced in is "how many more quarters this round of price increases can last." The public statements of the three companies are highly consistent—SK Hynix says supply falls short of demand through 2030, and Micron guides for continued increases—so consistent that it does not look like three independent judgments.
This is the bear side's strongest blow: when all three say the same good thing, the information content of that statement itself is very low, and simultaneous capacity expansion by all three is a fact that is happening. What today's bullish candle buys is not performance, but "it has fallen too much."
The real answer lies at the A-share level
SK Hynix's 80% gross margin is not good news for the "memory sector"; it is a profit distribution table: the money in this chain is concentrating toward the leaders and the end that has its own part numbers.
| Company | Q2 gross margin | Q3 gross margin | Change | Remarks |
|---|---|---|---|---|
| 佰维存储 | 55.98 | 42.81 | (13.17) | Inventory 240.5 yi, operating cash flow -100.4 yi |
| 德明利 | 38.84 | 32.89 | (5.95) | Inventory 289.9 yi, 2.4 times net assets |
| 香农芯创 | 13.23 | 7.79 | (5.44) | Distribution, first three quarters revenue 1064.3 yi |
| 北京君正 | 57.12 | 65.93 | 8.81 | Memory chip first three quarters gross margin 56.53 |
| 东芯股份 | 74.49 | 79.31 | 4.82 | First three quarters gross margin 74.01 |
The gross margins of all three module makers and distributors turned down in the third quarter, while the two companies making their own part numbers moved up over the same period. The dividing line is not "whether it is a memory stock," but "whether it has its own part numbers and capacity."
For those making niche DRAM, NOR, and SLC NAND, price increases flow directly into gross margin; for those buying wafers to make modules, price increases first flow into costs.
Longsys's October 15 earnings preview stated this more bluntly than anyone:
Affected by the relatively high price base in the earlier period, the price increases of DRAM and NAND Flash have narrowed significantly compared with the earlier period, and the marginal contribution of memory price increases to the company's profit has slowed markedly.
It itself expects third-quarter revenue of 13.9 billion to 15.9 billion yuan, basically flat with the second quarter's 14.18 billion. The boost to module makers' revenue from price increases had already stopped by the third quarter.
Biwin Storage's quarter-on-quarter figure must be looked at separately: its third-quarter revenue was 61.4 yi, 30% less than the second quarter's 87.6 yi, while over the same period Longsys was flat and Demingli even grew 24%. Peers did not fall together, which shows this is its own delivery and supply rhythm and cannot be treated as industry evidence; but in the same quarterly report, its gross margin was also the one that fell the hardest.
The real risk for module makers is not on the income statement but on the balance sheet. At the end of the third quarter, Demingli's inventory was 289.9 yi, 2.4 times its net assets of 120.3 yi; short-term borrowings were 101.5 yi and long-term borrowings 91.0 yi; operating cash flow for the first three quarters was -136.9 yi.
Biwin Storage's inventory was 240.5 yi (78.7 yi at the end of last year), prepayments rose from 1.5 yi to 73.8 yi, and operating cash flow was -100.4 yi.
Compare with the original manufacturers: SK hynix had net cash of 98.7 trillion Korean won at the end of the third quarter; Micron had cash plus investments of 734.8 billion US dollars, full-year capital expenditure of 273.7 billion US dollars, and free cash flow of 623 billion US dollars.
In the same price upcycle, one side is accumulating cash, the other is borrowing short-term debt to stock up. If prices really turn, the first to run into trouble will not be SK hynix.
What has already risen, and what has not yet been priced
Today's gain, the 8% earnings beat, and the 100 trillion Korean won shareholder return (of which the 40 trillion won buyback was completed last week and the shares are about to be cancelled) — these three have already been bought in the market.
What has not yet been bought is three things: the real meaning of long-term contracts that "lock volume but not price"; the inventory impairment risk of module makers (Biwin only recognized 9,679 wan in inventory write-downs in the third quarter, because prices were still rising and write-down provisions could not be built up); and the reconciliation between new capacity in 2027 and the 50% of demand that has already been locked up.
GigaDevice's and Longsys's third-quarter reports had not yet been released as of October 28; their most complete period is the first half (GigaDevice's second-quarter single-quarter gross margin was 66.57%). These two companies' third-quarter reports are the next test point for this judgment.
What would count as me being wrong
- DRAM and NAND contract prices continue to rise quarter-on-quarter in November and December, and the two original manufacturers' gross margins take another step up in the fourth quarter — then "cannot push through a reversal" would be wrong, and the valuation repair would become a trend.
- The 2027 capital expenditure growth rate SK hynix gives in January next year is lower than 2026's — supply concerns would be removed, and the market would price it as growth again.
- A-share module makers' single-quarter gross margins rebound in the fourth quarter (Biwin back above 45%, Demingli back above 35%) — showing the pressure is not continuing downstream.
- Conversely, if any module maker recognizes more than 5 yi in inventory write-downs in the fourth quarter single quarter, the above judgment that "profits are concentrating upstream" would be verified faster.
One sentence to close: this earnings call proves that memory is still making money; it does not prove that memory can still make money the way it did over the past four quarters. The former is today's bullish candle; only the latter would be called a reversal.
Sources: SK hynix, "SK hynix Announces 3Q26 Financial Results," October 28, 2026; Micron, "Reports Record Fiscal Fourth-Quarter and Full-Year 2026 Results," Form 8-K Exhibit 99.1, September 30, 2026, announcement No. 0000723125-26-000018, and the FQ1-27 guidance table in the same filing; Samsung Electronics, "Earnings Guidance for Third Quarter 2026," October 8, 2026; Biwin Storage, "2026 Third Quarter Report," October 29, 2026, "Announcement on the Provision for Asset Impairment for the First Three Quarters of 2026," announcement No. 2026-101, "Announcement on the Company's Recent Production and Operations," announcement No. 2026-100 (October 24, 2026); Demingli, "2026 Third Quarter Report," announcement No. 2026-093, October 26, 2026; Beijing Junzheng, "2026 Third Quarter Report," October 28, 2026, "Investor Relations Activity Record," No. 20261027 (October 27, 2026); Dongxin Shares, "2026 Third Quarter Report," October 27, 2026; Shannon Chip Creation, "2026 Third Quarter Report," October 27, 2026; Longsys, "2026 First Three Quarters Earnings Forecast," announcement No. 2026-117, October 15, 2026; stock prices and drawdowns from Yahoo Finance daily lines (000660.KS, 005930.KS, MU, 285A.T) and A-share closing quotes on October 28, 2026; broker expectations quoted from 이투데이, October 27, 2026, "SK하이닉스, 3분기 영업익 77조 전망."
推不动。这场业绩会能撑起的是一天的反弹，不是一轮逆转——今天韩国、日本、美国的存储股已经把能涨的都涨了（海力士 +10.9%、三星 +9.0%、铠侠 +13.8%、美光盘中 +4.5%），而它给出的两条最要紧的新信息，一条朝多、一条朝空，正好互相抵掉。
先分清哪些是新信息
海力士三季度收入 104.6 万亿韩元，营业利润 83.3 万亿，营业利润率 80%。收入环比涨 32%、营业利润环比涨 38%；同比收入涨 328%、营业利润涨 632%。
80% 这个数字本身不新。二季度的营业利润率已经是 76%，三季度只是沿着同一条线又走了一步。真正的增量是"超出预期多少"：业绩会前一天韩国券商给的预期是营业利润 77 万亿，实际 83.3 万亿，超出 8%。
8% 的超预期，换来今天 10.9% 的上涨。这笔账一天就付完了。
容易看反的是净利润。三季度净利 39.8 万亿，比二季度的 93.9 万亿少了 58%。
这不是经营变差——二季度净利比当季营业利润（60.5 万亿）还高出一大截，说明那三个月的利润里有一块来自非经营项。公司在同一份材料里说，Kioxia 投资拿到的收益今年要还给股东。看这家公司，别用净利润环比下判断。
一条利多、一条利空，写在同一份文件里
利多这条是新的，而且第一次给了具体数字：公司已经和 14 家大客户签了长约，锁定的量相当于 2027、2028 两年总产能的一半以上，典型期限五年。这确实把"卖现货的周期股"往"有合同收入的公司"推了一步。
但紧接着的那句话更关键——长约里设了一个"在与客户约定的价格区间内反映市场行情"的机制。翻成大白话：锁量，不锁价。产量有保证了，价格下跌的风险还留在自己报表里。这一条如果被当成"长协保价"来买，就是买错了。
利空这条也是新的，而且是官方确认：明年资本开支要"大幅增加"，龙仁一期全面投产，M15X 年底拉满专做 HBM4，321 层 NAND 四季度占国内产能过半。
市场从六月下旬开始卖的，恰恰是这件事。7 月 20 日韩国《每日经济》一篇报道的标题大意是：所有厂商都在扩产，供给过剩的担忧重新出现。
这四个月里海力士从 291.9 万韩元跌到 10 月 27 日的 134.1 万，跌掉 54%，期间每一份财报都是创纪录的。
所以这次业绩会的性质不是"利空出尽"，而是把多空两边的最新证据同时摆上了台面：需求侧锁住了两年，供给侧也在同步放开。分歧没有被消除，只是变得更容易对账了。
三家原厂都在创纪录，股价离高点还差一大截
| 公司 | 最新季度收入 | 营业利润率 | 距六月高点 | 10月28日 |
|---|---|---|---|---|
| SK海力士 | 104.6万亿韩元 | 80.0 | (49.1) | 10.9 |
| 美光（FQ4） | 542.3亿美元 | 80.7 | (20.0) | 4.5 |
| 三星电子（预告） | 未披露 | — | (33.2) | 9.0 |
| 铠侠 | 未披露 | — | (60.4) | 13.8 |
单位是百分比；美光为盘中价；三星目前只有 10 月 8 日的业绩预告，正式明细 10 月 29 日才出，预告营业利润 107 万亿韩元。
美光给的下一季指引方向更明确：收入 615 亿美元（环比 +13%）、毛利率 86.25%、每股收益 38.15 美元。指引朝上，股价却比六月低 20%。
问题到这里就清楚了：业绩好，市场早就定过价了；没定过价的是"这轮涨价还能持续几个季度"。三家公司的公开说法高度一致——海力士说供不应求持续到 2030 年，美光指引继续涨——一致到不像三家独立判断。
这是反方最有力的一击：三家都说同一句好话时，这句话本身的信息量就很低，而三家同步扩产是正在发生的实事。今天这根阳线，买的不是业绩，是"跌太多"。
真正的答案在 A 股这一层
海力士 80% 的毛利率不是"存储板块"的利好，它是一张利润分配表：这条链上的钱，正往年头和手里有自己料号的那一端集中。
| 公司 | Q2毛利率 | Q3毛利率 | 变化 | 备注 |
|---|---|---|---|---|
| 佰维存储 | 55.98 | 42.81 | (13.17) | 存货240.5亿，经营现金流-100.4亿 |
| 德明利 | 38.84 | 32.89 | (5.95) | 存货289.9亿，是净资产2.4倍 |
| 香农芯创 | 13.23 | 7.79 | (5.44) | 分销，前三季收入1064.3亿 |
| 北京君正 | 57.12 | 65.93 | 8.81 | 存储芯片前三季毛利率56.53 |
| 东芯股份 | 74.49 | 79.31 | 4.82 | 前三季毛利率74.01 |
三家模组厂和分销商三季度毛利率全部掉头，同期两家做自有料号的往上走。分水岭不是"是不是存储股"，是"有没有自己的料号和产能"。
做利基 DRAM、NOR、SLC NAND 的，涨价直接落到毛利率；买晶圆做模组的，涨价先落到成本。
江波龙 10 月 15 日的业绩预告把这件事说得比谁都直白：
受前期较高价格基数影响，DRAM 和 NAND Flash 价格涨幅较前期明显收窄，存储价格上涨对公司盈利的边际贡献显著趋缓。
它自己预计三季度收入 139 亿到 159 亿元，和二季度的 141.8 亿基本持平。涨价对模组厂收入的推动，到三季度已经停了。
佰维存储的环比要单独看：它三季度收入 61.4 亿元，比二季度的 87.6 亿少了 30%，而同期江波龙持平、德明利还增了 24%。同行没一起掉，说明这是它自己的交付和供货节奏，不能当成行业证据；但同一份季报里，它的毛利率也是掉得最狠的那一家。
模组厂真正的风险不在利润表，在资产负债表。德明利三季度末存货 289.9 亿元，是净资产 120.3 亿元的 2.4 倍，短期借款 101.5 亿、长期借款 91.0 亿，前三季度经营性现金流 -136.9 亿元。
佰维存储存货 240.5 亿元（去年底 78.7 亿），预付款项从 1.5 亿涨到 73.8 亿，经营性现金流 -100.4 亿元。
对照原厂：海力士三季度末净现金 98.7 万亿韩元；美光手上现金加投资 734.8 亿美元，全年资本开支 273.7 亿美元，自由现金流 623 亿美元。
同一场涨价，一头在攒现金，一头在借短债囤货。价格真拐头，先出问题的不会是海力士。
已经涨掉的，和还没定价的
今天的涨幅、8% 的盈利超预期、100 万亿韩元的股东回报（其中 40 万亿的回购上周已完成、股份即将注销），这三件在盘面上已经被买走了。
还没被买走的是三件：长约"锁量不锁价"的真实含义；模组厂的存货减值风险（佰维三季度只提了 9,679 万元存货跌价，因为价格还在涨，跌价准备提不动）；2027 年新增产能和已经被锁掉的那 50% 需求之间的对账。
兆易创新、江波龙的三季报截至 10 月 28 日还没出，它们最完整的一期是上半年（兆易二季度单季毛利率 66.57%）。这两家的三季报是这个判断的下一个检验点。
什么情况算我错了
- 十一月、十二月 DRAM 和 NAND 合约价继续环比上涨，且四季度两家原厂毛利率再上一个台阶——那时"推不动逆转"就错了，估值修复会变成趋势。
- 海力士明年一月给出的 2027 年资本开支增速低于 2026 年——供给担忧解除，市场会重新按成长给价。
- A 股模组厂四季度单季毛利率回升（佰维回到 45% 以上、德明利回到 35% 以上）——说明压力没有继续往下游走。
- 反过来，如果哪家模组厂四季度单季计提存货跌价超过 5 亿元，上面"利润往上游集中"的判断会被更快验证。
一句话收尾：这次业绩会证明的是存储还在赚钱，没有证明存储还能像过去四个季度那样赚钱。前者是今天这根阳线，后者才叫逆转。
来源：SK hynix《SK hynix Announces 3Q26 Financial Results》2026年10月28日；Micron《Reports Record Fiscal Fourth-Quarter and Full-Year 2026 Results》Form 8-K 附件99.1，2026年9月30日，公告编号 0000723125-26-000018，以及同文件 FQ1-27 指引表；Samsung Electronics《Earnings Guidance for Third Quarter 2026》2026年10月8日；佰维存储《2026年第三季度报告》2026年10月29日、《关于公司2026年前三季度计提资产减值准备的公告》公告编号2026-101、《关于公司近期生产经营情况的公告》公告编号2026-100（2026年10月24日）；德明利《2026年第三季度报告》公告编号2026-093，2026年10月26日；北京君正《2026年第三季度报告》2026年10月28日、《投资者关系活动记录表》编号20261027（2026年10月27日）；东芯股份《2026年第三季度报告》2026年10月27日；香农芯创《2026年三季度报告》2026年10月27日；江波龙《2026年前三季度业绩预告》公告编号2026-117，2026年10月15日；股价与回撤来自 Yahoo Finance 日线（000660.KS、005930.KS、MU、285A.T）与 A 股 2026年10月28日收盘行情；券商预期引自 이투데이 2026年10月27日《SK하이닉스, 3분기 영업익 77조 전망》。
## 同一天的两张报表
香农芯创的三季报是 10 月 27 日挂出来的，比海力士业绩会早一天。单季收入 462.2 亿元，二季度是 364.4 亿；单季毛利额 36.0 亿元，二季度是 48.2 亿；单季净利 17.0 亿元，二季度是 23.2 亿。收入涨了两成七，毛利额掉了两成五。
收入涨、毛利额跌这两件事同时发生，只有一个解释：加价率被压掉了。分销商赚的是买进卖出之间的差价比例，上游晶圆涨价，它的成本先动，售价要滞后才跟得上，中间那段就是缺口。这个缺口在报表上不落在"营业收入"，落在"营业成本"——收入被涨价抬高了，成本被涨价抬得更高。
而这家公司六周前的说法是另一套。9 月 18 日的投资者关系活动记录表里，有人直接问"目前 SK 海力士的产能仍然紧张，公司拿货节奏是否正常"，回答是：
> 目前公司经营活动正常开展，与核心供应商 SK 海力士合作关系良好、稳定。产品价格受供求关系等多种因素影响，公司对存储产品价格持谨慎乐观态度。
"谨慎乐观"是 9 月 18 日的措辞，10 月 27 日的报表是另一回事。它没有说错——绝对额还在历史高位，前三季净利 53.4 亿元是去年同期的十几倍。但边际方向已经反了。看周期，方向比绝对数重要。
## 一年存货，两个方向
判断上游涨价要多久打到下游报表上，有个笨办法很好用：期末存货除以单季营业成本，得到这批库存"能撑几个季度"。这个数越久，说明账面记的还是越早的便宜货，涨价要更晚才反映出来；同时也说明压在库里的钱越多。
算法写清楚：单季营业成本 = 本期累计营业成本 − 上期累计营业成本，一季报直接用累计值。存货取期末余额。
| 公司 | 期末存货（亿） | 单季营业成本（亿） | 可撑（季） | 折天 |
|---|---|---|---|---|
| 佰维存储 | 240.53 | 35.13 | 6.85 | 623 |
| 德明利 | 289.86 | 78.14 | 3.71 | 337 |
| 东芯股份 | 14.55 | 3.72 | 3.91 | 356 |
| 北京君正 | 35.99 | 10.72 | 3.36 | 306 |
| 香农芯创 | 69.05 | 426.22 | 0.16 | 15 |
| 深科技 | 35.22 | 36.27 | 0.97 | 88 |
除深科技取二季度数（三季报未出）外，其余都是 2026 年 9 月 30 日数。
这张表最容易看错的地方是：同样是一年的存货，含义完全相反。
模组厂的存货是买来的晶圆。晶圆涨价，它的存货成本跟着涨，售价要等下游客户接受。所以它的存货是敞口——库里的货越贵，下一季毛利率越难看。做自有产品的那几家，存货是自己流片、自己封测出来的，晶圆涨价时生产成本没变（产能是自己的、折旧是固定的），卖价却跟着市场走。所以它们的存货是蓄水池。
同样是"三十几亿存货"，在两种生意里一个是风险、一个是利润。这就是为什么君正、东芯三季度毛利率往上走，佰维、德明利、香农往下走。分水岭不是"是不是存储股"，是有没有自己的产能。
顺带说清议价权从哪来：晶圆是最紧的一环，谁手里有产能，谁定规则。原厂可以要求预付、可以锁五年长协、可以把营业利润率做到 80%；模组厂只能排队拿货、垫资囤货；分销商连差价比例都保不住。海力士那 80%，不是需求的奖赏，是卡点的租金。
这里有个结构性的东西必须点破：A 股没有一家上市公司拥有 DRAM 晶圆产能——君正、兆易都是设计公司，产能靠代工；长鑫、长存都还没上市。所以这个板块里所有公司都站在下游——原厂的超额利润，落到 A 股报表上只能以成本的形式出现，不会以收入的形式出现。这也解释了为什么每轮存储涨价，先炒的总是模组和分销：它们收入弹性最大，涨得最像；但利润弹性属于上游，它们拿不到。板块指数会涨，个股利润不一定跟。
## 佰维把 2.4 个季度变成了 6.9 个
```chart
type: line
title: 佰维一单位存货能撑的季度数
unit: 倍
labels: 2025Q3, 2025Q4, 2026Q1, 2026Q2, 2026Q3
values: 2.71, 2.44, 3.79, 4.71, 6.85
```
2025 年三季度末，佰维的存货够它用 2.7 个季度。一年后，够用 6.9 个季度，623 天。
同期它的单季营业成本几乎没动，一直在三四十亿，三季度 35.13 亿；存货却从 78.7 亿涨到 240.5 亿，涨了两倍多。存货增速是成本增速的十几倍——这不叫为下一季订单备货，备货不会备到一年半以后。
德明利是个干净的对照。它的存货也从 70.6 亿涨到 289.9 亿，但单季营业成本同步从 32.1 亿涨到 78.1 亿，所以"能撑几个季度"一直稳在 3.7：一季度 3.80、二季度 3.74、三季度 3.71。这是跟着生意规模一起长。佰维是从 2.44 冲到 6.85，这是押。
钱怎么出去的，三季报写得很清楚。佰维三季度末预付款项 73.75 亿元，二季度末 20.94 亿，一个季度多付出去 52.81 亿；同期存货增加 58.85 亿。新增库存里有九成是"先付钱、后到货"。它把现金交给了上游，换回一堆要一年半才卖完的货。
这批货还带着一个隐性成本。240.5 亿存货，前三季度一共计提存货跌价损失 9,679 万元，占存货 0.40%。提不动是因为准则要求按成本和可变现净值孰低计提——价格还在涨，可变现净值高于成本，账面上找不到计提的理由。
所以这个风险不会慢慢渗出来，只会在价格拐头的那一个季度一次性出现。粗算：存货跌价 5%，是 12.0 亿，相当于佰维三季度单季净利 16.02 亿的四分之三。
这是第一件还没被定价的事。
## 最强的反方，和它成立的条件
反方最有力的一击不是"你太悲观"，而是：涨价周期里，库存就是护城河。
它的证据很硬。佰维三季度毛利率 42.81%，比 2025 年四季度的 31.91% 还高 11 个百分点——高价库存不但卖出去了，而且卖得比涨价前更赚。73.75 亿预付款是锁未来的货，不是压手里的陈货。再往上，14 家大客户愿意锁五年、锁掉 2027 和 2028 两年一半以上的产能；客户比厂商更清楚自己要多少货，若他们判断 2027 年过剩，不会签这种字。
这个反方我认一半。
它成立需要一个条件：售价涨幅不低于晶圆成本涨幅。三季度这个条件已经不成立，而且是三家同时不成立——佰维毛利率掉 13.17 个点、德明利掉 5.95 个点、香农掉 5.44 个点。连"固定加价"的分销环节都转不出去，说明下游终端不接受这个价。
我不认的那一半是：长约解决的是量，不是价。锁量之后模组厂的收入确定性确实提高，但毛利率的弹性被原厂拿走了。收入能见度变好，不等于利润变好——这两件事在报表上是两个科目，一个在利润表第一行，一个在第三行。
## 现在该看什么，什么情况算我错了
四季度有三个数会自己说话。
- 佰维四季度单季毛利率回到 45% 以上、且存货环比不再增加——反方赢，我错。
- 任何一家模组厂四季度单季计提存货跌价超过 5 亿元——这条链"利润往上走"的判断会被更快验证。
- 海力士明年一月给的 2027 年资本开支增速低于 2026 年——供给担忧解除，市场会重新按成长给价，那时"推不动逆转"也错了。
还有一个检验点没到。兆易创新和江波龙的三季报，截至 10 月 29 日凌晨都还没挂出来，最新的仍是半年报，江波龙另有一份 10 月 15 日的业绩预告。这两份不出来，A 股这一层的证据链就是缺的——上面六家里，深科技用的还是二季度数，它三季报同样没出。
一句话收尾：这次业绩会证明的是原厂还在赚钱，它没有证明这条链上其余环节也还在赚钱。恰恰相反，同一天的两张报表给出了相反的答案。前者是业绩会那天的那根阳线，后者才叫逆转。
来源：SK hynix《SK hynix Announces 3Q26 Financial Results》2026年10月28日；Micron《Reports Record Fiscal Fourth-Quarter and Full-Year 2026 Results》Form 8-K 附件99.1，2026年9月30日，公告编号 0000723125-26-000018；香农芯创《2026年三季度报告》2026年10月27日、《香农芯创科技股份有限公司投资者关系活动记录表》2026年9月18日；佰维存储《2026年第三季度报告》2026年10月29日、《关于公司2026年前三季度计提资产减值准备的公告》公告编号 2026-101，2026年10月29日；德明利《2026年三季度报告》2026年10月26日；北京君正《2026年第三季度报告》2026年10月28日；东芯股份《2026年第三季度报告》2026年10月27日；深科技《2026年半年度报告》（三季报截至2026年10月29日未披露）。存货、营业成本、预付款项、净利均取自上述定期报告合并报表；单季数由本期累计减上期累计计算，一季报直接用累计值。