Montage Technology's memory interface chips are sold by volume based on memory module shipments, not as a percentage cut of DRAM prices. When Micron's quarterly report for the September-November quarter comes out in mid-December, for Montage's 2027 first-quarter report, will the first thing to move be operating revenue or gross margin, or will neither move?
- On the power side Eaton +7.0%, Quanta Services +5.8%, GE Vernova +6.5%.
- What was moving that same evening was installations Fabrinet rose 15.45%
- For the first three quarters it was 68.9%, up 4.1 percentage points.
- Meanwhile Micron's DRAM average price rose about 180% for the full 2026 fiscal year.
Montage Technology's memory interface chips are sold by volume based on memory module shipments, not as a percentage cut of DRAM prices. When Micron's quarterly report for the September-November quarter comes out in mid-December, for Montage's 2027 first-quarter report, will the first thing to move be operating revenue or gross margin, or will neither move?
澜起科技的内存接口芯片按内存条出货量走量，不是按DRAM价格抽成。美光12月中旬的9-11月季财报出来，澜起2027年一季报先动的是营业收入还是毛利率，还是都不动？
The desk answersverifiedWalnut, from public filings
What moves first is operating revenue; gross margin will not move.
To put it more precisely: in Micron's mid-December report, the only thing that can change is Lanqi's "unit count"; it cannot change the price and cost of each of Lanqi's chips. And "unit count" happens to be the least clear item in Micron's report.
Why gross margin cannot move
Lanqi's cost is wafer foundry plus packaging and testing, settled in US dollars, under long-term contracts signed with foundries. No matter how much DRAM particles rise, it does not enter its cost table. By mechanism, gross margin cannot connect to particle quotes at all.
Its own books confirm this: in the third quarter of 2026, interconnect chip gross margin was 68.4%, up only 2.7 percentage points year-on-year; for the first three quarters it was 68.9%, up 4.1 percentage points. Meanwhile Micron's DRAM average price rose about 180% for the full 2026 fiscal year. The two numbers are not on the same order of magnitude. If gross margin really followed particle prices, this year's change should be tens of points, not two or three points.
The company attributes the change in gross margin to "product mix"—a higher revenue share from new-subgeneration DDR5 RCD chips and new interconnect products. The switch is on Lanqi's own product migration rhythm, not on Micron's income statement.
Revenue will move, but what moves is quantity, not price
Revenue equals unit count multiplied by value per unit. Particle price has no place at all in this equation.
One sentence in Micron's 2026 fiscal year 10-K makes this very clear: DRAM sales grew 252%, of which average price rose about 180%, and bit shipments rose only a bit more than two tenths. In other words, in that blowout revenue, price accounted for around eight tenths and quantity only a bit more than two tenths—this is arithmetic split out from the two growth rates it itself gave, not a breakdown it disclosed.
Moreover, "bits" are not "unit count." For the same memory module, capacity going from 64GB to 512GB multiplies bits several times over, with not one more unit. Micron said in the same announcement that it has begun sampling 512GB DDR5 RDIMMs. So unit count growth can only be lower than bit growth.
The conclusion lands here: the larger the revenue Micron reports, the less it follows that "memory module unit counts are rising." What can establish unit count is another set of numbers—the set that actually moved that same evening.
Another honest point: Lanqi itself does not disclose chip counts; the announcement only has phrasing like "shipment volume increased substantially year-on-year," with no number. So the unit count variable is also unverifiable on Lanqi's side and can only be inferred backward from revenue.
What was moving that same evening was installations
Fabrinet rose 15.45%; its own 8-K says first fiscal quarter revenue of 14.745 billion US dollars, up 51% year-on-year, with second quarter guidance of 16 to 16.5 billion US dollars. Vertiv third quarter revenue was 43.13 billion US dollars, up 61% year-on-year, with organic orders up more than 100% year-on-year. Oracle at its October 28 investor day set 2030 fiscal year revenue at more than 2250 billion US dollars, with capital expenditure peaking in fiscal 2027 and 2028.
Compare the sector moving in the same direction: on the optical side Vertiv +5.2%, Arista +3.5%, Supermicro +3.1%, Dell +4.0%, Fabrinet +15.4%, Ciena +5.5%, Coherent +3.8%; on the power side Eaton +7.0%, Quanta Services +5.8%, GE Vernova +6.5%. Optical, servers, and power moving together—this is installations moving, not particle quotes moving.
Installations determine unit count, and unit count is what determines Lanqi's revenue. When this chain reaches Lanqi, it lands on operating revenue, not on gross margin.
By the way: the three most violently moving names in the A-share pool today—Opt +13.79%, Fulai New Materials +10.00%, Huayang New Materials -10.02%—have nothing to do with the memory interface chain and cannot be used as a signal for this chain.
There is a detail in the timing that is easy to miss
Applied to Lanqi's 2027 first quarter report, this chain has to pass through two layers. What Micron reports in mid-December is September to November, and for this quarter the numbers were already guided on September 30: 615 billion US dollars, plus or minus 15 billion, with gross margin of about 86%. So the "this quarter" part of that report was already known to the market long ago; repeating it does not constitute new information.
What is truly useful is the next quarter it gives, that is, the December to February guidance. Only this layer connects to module makers' January to March production scheduling, and only this layer connects to Lanqi's 2027 first quarter shipments.
And Lanqi has already laid the fourth quarter on the table: as of October 26, it expects interconnect chip orders in hand for fourth quarter delivery of 25.0 to 27.9 yi yuan, while third quarter revenue for this line was 20.47 yi. The fourth quarter segment does not need to wait for Micron; Lanqi itself has said it. Only the first quarter segment is where Micron's December guidance gets to speak.
The strongest opposing argument: Lanqi's revenue actually does not move "by unit count" either; it moves by "how many chips per module, which generation." Its own reports show this—first three quarters revenue up 33.4% year-on-year, interconnect up 34.6%, third quarter single quarter up 46%, while Micron's bit shipments rose only a bit more than two tenths. Lanqi is running so much faster than unit count; the extra speed comes from the per-module value increase brought by DDR5 third to fifth subgeneration RCD and new interconnect products. If this holds, then Micron's December report cannot even be called "moving first" for Lanqi; what truly determines it is the product migration rhythm, and that rhythm is written in Lanqi's own announcements.
I'll concede half of this point. Revenue is the product of value per unit and unit count. What Micron can provide is the ceiling on the unit-count term—that is, whether supply squeezes module makers—but it cannot provide the value-per-unit term. So it cannot change the absolute value of Montage's revenue, but it can change the direction of Montage's revenue.
What would count as me being wrong: in Montage's 2026 annual report (next April), if the fourth-quarter gross margin for interconnect chips falls by more than 1 percentage point from the third quarter's 68.4%, and the company attributes the cause to rising raw material or foundry costs rather than product mix, that would show that the price increases in memory dies and wafers have indeed entered its cost statement, that gross margin moves first, and that this point of mine is wrong. Another fact that could overturn me: in Micron's December guidance for the next quarter, bit shipments turn negative quarter-on-quarter—not slower growth but negative—while Montage's first-quarter revenue is still rising quarter-on-quarter. That would show that the unit-count variable does not constrain Montage at all, and that the chain I built is empty.
Sources: ① Micron 8-K exhibit a2026q4ex991-pressrelease.htm, 2026-09-30, accession 0000723125-26-000018; ② Micron 10-K main document mu-20260903.htm, 2026-10-09, accession 0000723125-26-000023; ③ Montage Technology "2026 Third Quarter Report," 2026-10-30, Cninfo 1225643270.PDF; ④ Montage Technology "Voluntary Disclosure Announcement of Expected Increase in 2026 Third Quarter Results," announcement no. 2026-057, 2026-10-21, Cninfo 1225612915.PDF; ⑤ Vertiv 8-K exhibit q32026exhibit991vrt10282026.htm, 2026-10-28, accession 0001628280-26-067707; ⑥ Oracle 8-K exhibit d75538dex991.htm, 2026-10-28, accession 0001193125-26-435853; ⑦ Fabrinet 8-K exhibit fn-20261102xex991q127.htm, 2026-11-02, accession 0001408710-26-000034.
先动的是营业收入，毛利率不会动。
更准确地说：美光12月中旬那份报表，能改的只有澜起的“条数”，改不了澜起每一颗芯片的价格和成本。而“条数”恰好是美光那份报表里最说不清的一项。
毛利率为什么动不了
澜起的成本是晶圆代工加封装测试，用美元结算，跟代工厂签的是长约。DRAM颗粒涨多少，不进它的成本表。毛利率从机制上就接不到颗粒报价。
它自己的账本也印证了：2026年第三季度，互连类芯片毛利率68.4%，同比只升2.7个百分点；前三季度68.9%，升4.1个百分点。而美光2026财年整年的DRAM均价涨了约180%。两个数字不在一个量级。如果毛利率真跟着颗粒价格走，今年该是几十个点的变化，不是两三个点。
公司把毛利率的变化归到“产品结构”——DDR5新子代RCD芯片和互连类新产品的收入占比提高。开关在澜起自己的产品迁移节奏上，不在美光的损益表上。
收入会动，但动的是数量，不是价格
收入等于条数乘每条上的价值。颗粒价格在这条式子里一个位置都没有。
美光2026财年10-K里有一句话把这事说得很清楚：DRAM销售额增长252%，其中均价涨约180%，位元出货量只涨两成多。换句话说，那份爆表的收入里，价格占了八成上下，数量只占两成多一点——这是拿它自己给的两个增速拆出来的算术，不是它披露的分解。
而且“位元”还不是“条数”。同一条内存条，容量从64GB做到512GB，位元翻好几倍，条数一条没多。美光在同一份公告里说，已经开始送样512GB的DDR5 RDIMM。所以条数的增速只会比位元更低。
结论落在这里：美光报的收入越大，越推不出“内存条条数在涨”。能推出条数的是另一组数字，是同一天晚上真正动的那一组。
另外一句实话：澜起自己不公布颗数，公告里只有“出货量同比大幅增加”这种说法，没有数字。所以条数这个变量在澜起这边也是查不到的，只能从收入倒推。
同一天晚上在动的是装机
Fabrinet涨15.45%，它自己的8-K写着第一财季收入14.745亿美元、同比增51%，第二季指引16到16.5亿美元。Vertiv三季度收入43.13亿美元、同比增61%，有机订单同比增超100%。甲骨文在10月28日投资者日把2030财年收入定在2250亿美元以上，资本开支在2027、2028财年见顶。
对照板块同向：光这一侧维谛+5.2%、Arista+3.5%、超微+3.1%、戴尔+4.0%、Fabrinet+15.4%、Ciena+5.5%、Coherent+3.8%；电力这一侧伊顿+7.0%、广达服务+5.8%、GE Vernova+6.5%。光、服务器、电力三组一起走，这是装机在动，不是颗粒报价在动。
装机决定条数，条数才决定澜起的收入。这条链传到澜起，落在营业收入，不落在毛利率。
顺带说一句：A股池子里今天动得最凶的三只，奥普特+13.79%、福莱新材+10.00%、华阳新材-10.02%，跟内存接口这条链没关系，不能当这条链的信号用。
时点上有个细节容易漏
放到澜起2027年一季报上，这条链要过两层。美光12月中旬报的是9到11月，而这一季的数字，9月30日就已经给了指引：615亿美元上下15亿，毛利率约86%。所以那份报表里“这一季”的部分市场早就知道，重念一遍不构成新信息。
真正有用的是它给的下一季，也就是12月到2月的指引。这一层才接得上模组厂1到3月的排产，才接得上澜起2027年一季报的出货。
而澜起已经把第四季度摊在桌面上了：截至10月26日，预计第四季度交付的互连类芯片在手订单25.0到27.9亿元，三季度这条线的收入是20.47亿。四季度这一段不用等美光，澜起自己说了。一季度那一段，才轮到美光12月的指引说话。
反方最强的一条：澜起的收入其实也不是“按条数走”，是按“每条配几颗、配第几代”走。它自己的报表就摆着——前三季度收入同比增33.4%、互连类增34.6%、三季度单季增46%，而美光的位元出货只涨两成多。澜起跑得比条数快这么多，快出来的那部分是DDR5第三到第五子代RCD和互连新产品带来的单条价值提升。如果这条成立，那美光12月那份报表对澜起来说连“先动”都谈不上，真正决定它的是产品迁移节奏，而那个节奏写在澜起自己的公告里。
这条我认一半。单条价值和条数乘起来才是收入，美光能给的是条数那一项的上限，也就是供给掐不掐模组厂的脖子，给不了单条价值那一项。所以它改不了澜起收入的绝对值，但能改澜起收入的方向。
什么情况算我看错：澜起2026年年报（明年4月）里，如果第四季度互连类芯片毛利率比三季度的68.4%掉了1个百分点以上，并且公司把原因写成原材料或代工成本上升、不是产品结构，那说明颗粒和晶圆的涨价确实进了它的成本表，毛利率会先动，我这条就错了。另一个能推翻我的事实是：美光12月的下一季指引里，位元出货环比转负，不是增速放慢而是转负，而澜起一季度收入环比还在涨——那就说明条数这个变量根本不约束澜起，我搭的这条链是空的。
- 01①美光 8-K 附件 a2026q4ex991-pressrelease.htm，2026-09-30，accession 0000723125-26-000018
- 02②美光 10-K 正文 mu-20260903.htm，2026-10-09，accession 0000723125-26-000023
- 03③澜起科技《2026年第三季度报告》，2026-10-30，巨潮网 1225643270.PDF
- 04④澜起科技《2026年第三季度业绩预增的自愿性披露公告》，公告编号 2026-057，2026-10-21，巨潮网 1225612915.PDF
- 05⑤Vertiv 8-K 附件 q32026exhibit991vrt10282026.htm，2026-10-28，accession 0001628280-26-067707
- 06⑥Oracle 8-K 附件 d75538dex991.htm，2026-10-28，accession 0001193125-26-435853
- 07⑦Fabrinet 8-K 附件 fn-20261102xex991q127.htm，2026-11-02，accession 0001408710-26-000034
## 澜起有两个毛利率，混着看必错
2026 年三个季度，澜起合并毛利率是 69.8%、61.8%、67.5%。一季度到二季度掉了 8 个点，三季度又爬回 5.7 个点。
同一份报表里，互连类芯片这条线是 71.6%、67.4%、68.4%。三个季度加起来动了 4.2 个点，只有合并口径的一半。
差额是津逮 CPU 造成的。二季度津逮做了 1.79 亿元收入，占当季 9.5%；三季度只剩 0.30 亿元，占 1.4%。津逮上半年毛利率 9.7%（半年报分产品数据：收入 2.20 亿、毛利 0.21 亿），它一进一出，合并毛利率就跟着摆五个多点。
> 合并报表上的毛利率，有一大块是津逮的出货节奏，不是内存接口的生意。
所以「毛利率会不会动」这个问题，只能在互连类这一行上问。在合并口径上问，答案永远是「动」，而且动的方向和内存没关系。
一季度那个 71.6% 没有单独披露，是倒推的：半年报互连类毛利 21.56 亿，减去二季度（收入 16.94 亿乘 67.4%），剩 10.14 亿；分母 14.17 亿，是半年报互连类收入 31.11 亿减掉二季度那 16.94 亿。
| 季度 | 合并收入 | 合并毛利率 | 互连类毛利率 | 津逮占比 |
|---|---|---|---|---|
| 2026Q1 | 14.61 | 69.8 | 71.6 | 2.8 |
| 2026Q2 | 18.75 | 61.8 | 67.4 | 9.5 |
| 2026Q3 | 20.79 | 67.5 | 68.4 | 1.4 |
```chart
type: group
title: 合并与互连类的毛利率缺口，跟着津逮的出货走
unit: %
labels: 2026Q1, 2026Q2, 2026Q3
values: 69.8, 61.8, 67.5
values2: 71.6, 67.4, 68.4
legend: 合并|互连类
```
这里有个天然对照：二季度互连类毛利率掉了 4.2 个点。而美光 2026 财年 DRAM 均价涨了约 180%，二季度正处在这轮涨价的中段。如果毛利率真是跟着颗粒价格走的，方向该是反的。
## 同一季度，模组厂掉了六到十七个点
| 公司 | 环节 | 2026Q2 | 2026Q3 | 变动 |
|---|---|---|---|---|
| 澜起（互连类） | 接口芯片 | 67.4 | 68.4 | 1.0 |
| Rambus | 接口芯片 | 79.8 | 78.3 | (1.5) |
| 聚辰股份 | 模组配套芯片 | 53.7 | 56.8 | 3.1 |
| 江波龙 | 内存模组 | 61.3 | 44.7 | (16.5) |
| 佰维存储 | 内存模组 | 56.0 | 42.8 | (13.2) |
| 德明利 | 内存模组 | 38.8 | 32.9 | (5.9) |
同一条链，同一个季度，同一批颗粒报价。模组厂的毛利率掉 6 到 17 个点，接口芯片涨 1 个点。差别不在经营好坏，在颗粒进谁的成本表。
模组厂的营业成本里，颗粒是大头。颗粒涨价，先冲它的毛利率；等它把成品价格调上去，又要一个季度。三季度正是成本追上售价的那一段，江波龙单季毛利率从 61.3% 落到 44.7%，一个季度十六个点。
澜起的营业成本里没有颗粒，主要是晶圆代工和封测，采购主要用美元计价。DRAM 涨多少，不进它的成本表。
再看全球那一侧。Rambus 三季度毛利率 78.3%，比二季度低 1.5 个点。这家是澜起在内存接口上的直接对手，做同一类芯片，面对同一轮周期，毛利率也没跟着颗粒走。
芯片设计这一层，毛利率跟着产品结构走；模组这一层，毛利率跟着库存成本走。这条分界线，是后面所有推论的前提。
## 美光能喂到澜起的，只有条数这一项
传导要走三层，每一层的滞后不一样。
| 层级 | 发生什么 | 滞后 | 落到澜起哪个科目 |
|---|---|---|---|
| 一 | 美光位元出货 | 当季 | 不直接落表 |
| 二 | 模组厂排产下单 | 一到两个月 | 在手订单 |
| 三 | 澜起发货 | 下一季度 | 营业收入 |
接口芯片在一条模组的物料清单里只占很小一格，模组厂按排产下单；决定排产的是颗粒能不能拿到，不是颗粒报价多少。三层走完，美光一个季度的数，落到澜起报表上是下一季度的事。中间没有任何一层，把颗粒价格写进澜起的营业成本。
所以落点是营业收入。但收入不是一个变量：收入等于条数乘每条芯片数乘单颗价格。
美光 10-K 里，2026 财年 DRAM 收入增 252%，均价涨约 180%，位元出货只增两成多。而同期澜起互连类收入增 49.3%。差出来的一倍，是乘数变了 —— DDR5 第三到第五子代 RCD 的出货占比在提，MRCD/MDB、时钟驱动这些新产品在加。
而且位元还不是条数。同一条模组容量从 64GB 做到 512GB，位元翻八倍，条数一条没多。所以条数能解释的那一半，还是上限。
再看一眼对手。同期 Rambus 收入只增 22.3%。同一条赛道，澜起多出来的那一截，大概率是份额和产品结构，不是行业水位。
一句话：美光改得动基数，改不动乘数。基数最多解释一半，乘数解释另一半。
## 12 月那份报表，真正新的只有一半
美光 9 月 30 日的四季报里，已经把 9 到 11 月这一季的指引给了：收入 615 亿美元上下 15 亿，毛利率约 86%。
也就是说，12 月中旬那份报表里，「这一季」的部分市场三个月前就知道了。重念一遍不是新信息，实际数要跟 615 亿比，不是跟零比。
新的只有 12 月到 2 月的指引。这一季和澜起 2027 年一季度（1 到 3 月）重叠两个月。
但时点上又晚一步：模组厂给一季度备货，订单在 11 月到 12 月就下了。美光 12 月中旬给指引的时候，澜起一季度的订单簿大致已经成形。
更关键的是，美光这份报表本身的信息含量在下降。它在 10-K 里写：战略客户协议是 take-or-pay，锁量，价格有上下限；即便按下限价格算，这些协议的毛利率也明显高于它过去任何一轮周期的季度峰值。
量早就锁了，价格带区间。这种合同结构下，季报确认的是几个月前签的合同，不是当下的现货需求。拿它推澜起下一个季度的订单，等于用结果推原因。
## 反方最强的一条
反方会说：既然澜起跑得比美光的位元快一倍，这两家已经脱钩了，你凭什么说收入会先动？
这条我认一大半。脱钩的是增速，不是方向。已经装出去的模组条数还是澜起出货的基数，基数变了方向会变。
但幅度上：假设美光把下一季的位元指引上下调 5 个点，落到澜起一季度收入是一到两个点；而澜起今年三个季度的收入环比，在 4% 到 28% 之间跳。一到两个点落进去，没有形状。
所以更准确的说法是：传导存在，但不可观测。要观测它，得看另一个数 —— 在手订单。澜起给的 2026 年四季度那个数，是 25.0 到 27.9 亿元互连类订单，比三季度这条线的收入 20.47 亿高 22% 到 36%。这个数比美光的指引硬，因为它是澜起自己签下来的。
四条证伪。
第一，2026 年年报里，四季度互连类毛利率比三季度的 68.4% 掉 1 个点以上，且公司把原因写成原材料或代工成本上升。那说明颗粒涨价绕一圈还是进了它的成本表。
第二，美光 12 月给的下一季指引里，位元出货环比转负，而澜起一季度互连类收入环比还在涨。那说明条数根本不约束澜起，这条链是空的。
第三，澜起一季度互连类收入环比掉，同时互连类毛利率也掉 2 个点以上。两个数同向大幅变动，说明它们不分叉，是被同一个东西推着走。
第四，澜起一季度互连类收入的环比跌幅，明显大于同期美光位元出货的环比跌幅。那说明澜起有自己的周期，跟条数不挂钩。
最后落一句。三个选项里选营业收入，但只在互连类那一行上选 —— 合并口径上它会被津逮盖掉。如果要在 2027 年一季报上找一个能验证的痕迹，别找营业收入，也别找毛利率，找「互连类芯片销售收入」这一行的环比。它比合并收入干净，比毛利率灵敏。
来源：①Micron Technology, Inc.《8-K》附件 a2026q4ex991-pressrelease.htm，2026-09-30，accession 0000723125-26-000018；②Micron《10-K》mu-20260903.htm，2026-10-09，accession 0000723125-26-000023；③澜起科技《2026年第三季度报告》，2026-10-30，巨潮 1225643270.PDF；④澜起科技《2026年第三季度业绩预增的自愿性披露公告》，公告编号 2026-057，2026-10-21，巨潮 1225612915.PDF；⑤澜起科技《2026年半年度报告》，2026-08-28，巨潮 1225531692.PDF；⑥江波龙《2026年三季度报告》，2026-10-30，巨潮 1225644948.PDF；⑦佰维存储《2026年第三季度报告》，2026-10-29，巨潮 1225638424.PDF；⑧德明利《2026年三季度报告》，2026-10-26，巨潮 1225620615.PDF；⑨聚辰股份《2026年第三季度报告》，2026-10-24，巨潮 1225620951.PDF；⑩Rambus Inc.《10-Q》rmbs-20260930.htm，2026-10-26，accession 0001193125-26-431941。