Marvell fell 8.81% in one day. In its fiscal 2027 third-quarter report in early December, was data center revenue higher or lower than the previous quarter? In Zhongji Innolight's 2026 annual report (March-April), was module revenue higher or lower than the previous quarter? Did the two companies peak in the same quarter first, or is there a one-quarter difference?
Marvell fell 8.81% in one day. In its fiscal 2027 third-quarter report in early December, was data center revenue higher or lower than the previous quarter? In Zhongji Innolight's 2026 annual report (March-April), was module revenue higher or lower than the previous quarter? Did the two companies peak in the same quarter first, or is there a one-quarter difference?
Marvell一天跌8.81%，它12月初的2027财年三季报里，数据中心收入比上一季度多还是少？中际旭创2026年报（3-4月）的模块收入比上一季度多还是少？两家是同一季度先见顶，还是差一个季度？
The desk answersverifiedWalnut, from public filings
Three questions, one answer: none of them has peaked yet. In Marvell's early-December quarterly report, data center revenue was still rising quarter-on-quarter (estimated 2.49 billion to 2.55 billion USD, more than 15% above the prior quarter's 2.172 billion); the fourth-quarter module revenue carved out of Zhongji Innolight's 2026 annual report was also higher than the third quarter (estimated 31.9 billion to 34.7 billion yuan); the two neither peaked in the same quarter nor are they one quarter apart—their quarter-ends are always one month apart, so comparing them on a "same quarter" basis is itself misaligned.
First, let's be clear about this 8.81%: it has nothing to do with data center revenue. On November 2, Marvell announced an all-stock acquisition of Lumentum, with an enterprise value of 115 billion USD, a fixed exchange ratio of 4.5 Marvell shares for 1 Lumentum share, equivalent to 1,132 USD per share, a 23.9% premium to October 28 (the last trading day before the market began speculating about this). What fell was the price and structure of this deal, plus same-day turnover across the entire AI hardware sector, not orders.
Lay out the arithmetic and it becomes clear. On October 28, Marvell closed at 251.65 USD; on November 3, it closed at 171.97 USD, down 31.7% over four trading days; over the same period, Lumentum fell from 913.95 to 680.16. Under the fixed exchange ratio, one Lumentum share is worth 4.5 times Marvell's share price, which on November 3 was 773.87 USD, while it itself closed at only 680.16, a 12.1% discount. This 12% is the discount the market applies for "closing only in the first half of 2028"—it must pass both sides' shareholder meetings and regulators, with more than a year in between. In an all-stock deal, however much the buyer falls, the seller falls with it; this is an arithmetic problem, not a demand problem.
The data center revenue line itself is accelerating: fiscal 2026 third quarter 1.518 billion USD, fourth quarter 1.651 billion (+9% quarter-on-quarter), fiscal 2027 first quarter 1.833 billion (+11%), second quarter 2.172 billion (+18%), with its share of total company revenue rising from 73% to 79%. Third-quarter total revenue guidance was 3.150 billion USD, plus or minus 5%; in the November 2 announcement, the preview revenue for this quarter was changed to "consistent with guidance, midpoint unchanged, range narrowed to plus or minus 2%," that is, 3.087 billion to 3.213 billion USD.
Calculated at a 79% share, data center revenue this quarter is 2.489 billion USD, +14.6% quarter-on-quarter; calculated at a share two or three points higher (the prior four quarters each rose 2 to 3 points), it is 2.55 billion, +17.5% quarter-on-quarter. Both are estimates; the company does not give segment guidance. Against last year's same period of 1.518 billion, that is +64% to +68% year-on-year, a notch above the second quarter's +46%. Management said in August that the custom business would accelerate significantly starting in the second half of fiscal 2027; the second half is exactly August to January of the following year, covering this quarter and the next.
The numbers on Zhongji Innolight's side are harder. 2026 first quarter 19.496 billion, second quarter 22.282 billion (+14.3% quarter-on-quarter), third quarter 27.769 billion (+24.6% quarter-on-quarter, +171.8% year-on-year), rising all the way. The module business accounts for 99% of revenue (first-half modules 41.355 billion, total revenue 41.778 billion), so the direction of total revenue is the direction of modules.
As for how the fourth quarter would go, the company itself gave the answer at its 9 p.m. conference call on November 3. Asked about the customer pull-in pace from the fourth quarter to the first quarter of next year, the answer was that data center customers' pickup pace is normal and has not slowed, optical module supply overall remains tight, and there has been no situation of goods being produced but not shipped; asked whether core materials in the fourth quarter improved versus the third quarter, the answer was that materials are gradually improving, and shipment volume "is expected to maintain quarter-on-quarter growth for several consecutive quarters." At the same call, it was said that for 2027 "the vast majority of orders already cover the full year," and that for 2028 at least one key customer has already given clear demand guidance.
Based on third-quarter 27.769 billion and fourth-quarter quarter-on-quarter +15% to +25%, fourth-quarter single-quarter revenue is 31.9 billion to 34.7 billion (estimate), and 2026 full-year module revenue is about 100.4 billion to 103.2 billion (estimate). For reference, last year: 2025 fourth quarter 13.235 billion, 29.5% higher than the third quarter's 10.216 billion; the fourth quarter was already its strongest quarter of the year.
So the either-or question of "does it peak in the same quarter first, or is it off by one quarter" — neither answer holds up. Marvell's fiscal quarter is August to October, Zhongji's third quarter is July to September; Marvell's next quarter is November to January of the following year, Zhongji's fourth quarter is October to December. The two companies' quarter-ends are always off by one month, not off by one quarter. Forcing a direct comparison of "that quarterly report in early December" with "the annual report in March to April" compares America's August to October with China's October to December, whereas what truly corresponds to Zhongji's fourth quarter is Marvell's next quarter, and that quarterly report won't come out until March next year.
There's also an indicator you can look at in advance: Lumentum's revenue for the quarter ended September 26 was 12.626 yi USD, +25.5% quarter-on-quarter, of which devices sold to module makers were 7.542 yi (+16.1%) and systems 5.084 yi (+42.4%); its guidance for October to December is 15.75 yi to 16.25 yi USD, +24.7% to +28.7% quarter-on-quarter, and management said this quarter will produce about 3.5 yi USD of quarter-on-quarter incremental volume. The layer that supplies module makers is still accelerating in October to December, so it's unlikely module makers' shipments will turn around this quarter.
Looking at that day's tape again, Marvell wasn't the only one falling. Vertiv fell 13.14%, yet the third-quarter report it delivered on October 28 had organic orders doubling year-on-year, full-year revenue guidance raised by 6.5 yi USD, and fourth-quarter guidance continuing to increase quarter-on-quarter; Quanta Services fell 12.00%, GE Vernova fell 10.03%, Eaton fell 9.69%, while on the same day Microsoft rose 6.6%, Google rose 4.2%, Amazon rose 7.0%. On the other side, A-share semiconductor equipment was rising (Piotech +20.00%, AMEC +16.62%, Sanhuan +14.77%), while US peers were falling (Applied Materials -4.1%, Lam -4.1%). Money flows from the shovel-sellers to the shovel-buyers, and the two markets are pricing the same batch of equipment in opposite directions — this kind of tape is positions changing hands, not orders disappearing.
The strongest point for the bear side: someone in the module layer has already turned negative quarter-on-quarter. Eoptolink's 2026 third quarter single quarter was 109.40 yi, 13.0% less than the second quarter's 125.72 yi — same layer, same quarter, same end demand, and it turned around first. Nor can hard evidence on the order side be found in Zhongji's statements: at the end of the third quarter contract liabilities were only 0.11 yi yuan, meaning essentially no customer used prepayments to lock in volume; and the devices Lumentum sells to module makers were only +16.1% quarter-on-quarter, less than half of its systems business's +42.4%, indicating that module makers' pull-in speed in the September quarter was already slower than the cluster deployment speed. Seen this way, Zhongji's fourth quarter "still able to grow quarter-on-quarter" rests on seasonality and management's verbal guidance, not orders already landed; the final point depends on whether customers take delivery in this one month of December.
What would count as me being wrong: in Marvell's early-December quarterly report, if data center revenue is below 21.72 yi USD, i.e. negative quarter-on-quarter, then my judgment that "this quarter is still growing" is directly void; if it lands below 24.9 yi but is still positive, it means I extrapolated the share too optimistically — right direction, wrong magnitude. On Zhongji's side, if the module revenue disclosed in the 2026 annual report is below about 966 yi (first three quarters' modules about 688 yi plus the third quarter single quarter's 277.69 yi — this is an estimated threshold), it means the fourth quarter single quarter is below the third quarter, and I'm wrong. Further out, if 2027 first quarter single-quarter revenue is below 2026 fourth quarter, the company's line about "several consecutive quarters of quarter-on-quarter improvement" is broken, and at that point what's wrong isn't this judgment — demand has truly peaked.
Sources:
Marvell 8-K 2026-11-02 / 0001193125-26-442714 / d81187dex991.htm (merger announcement with Lumentum, fiscal 2027 third-quarter preliminary revenue)
Lumentum 8-K 2026-11-02 / 0001193125-26-442758 / d163169dex991.htm
Lumentum 8-K 2026-11-02 / 0001628280-26-068655 / lite_ex991xq1fy27.htm (fiscal 2027 first-quarter results, second-quarter guidance)
Marvell 8-K 2026-08-27 / 0001835632-26-000022 / q227_8kx812026ex-991.htm (second-quarter results, third-quarter guidance, revenue by end market table)
Marvell 8-K 2026-05-27 / 0001835632-26-000014 / q127_8kx522026ex-991.htm (revenue by end market table)
Marvell 8-K 2026-03-05 / 0001835632-26-000006 / q426_8kx1312026ex-991.htm (revenue by end market table)
Vertiv 8-K 2026-10-28 / 0001628280-26-067707 / q32026exhibit991vrt10282026.htm (third-quarter results, fourth-quarter guidance)
Zhongji Innolight 2026 Third Quarter Report 2026-10-30 / AN202610291831181859
Zhongji Innolight Investor Relations Activity Record Form 2026-010 2026-11-03 / AN202611031831382786
Zhongji Innolight Announcement on Recent Production and Operations 2026-100 2026-10-23 / AN202610231830833072
Zhongji Innolight 2026 Semi-Annual Report revenue by product data (optical communication transceiver modules 413.55 yi yuan) 2026-08 / tushare main business composition
Eoptolink, Accelink Technologies, T&S Communications 2026 third-quarter report single-quarter revenue (derived by subtracting cumulative figures) / tushare income statement
Nasdaq historical closing prices 2026-10-28 to 2026-11-03 / api.nasdaq.com/api/quote/{ticker}/historical
三个问题一个答案：都还没见顶。Marvell 12 月初那份季报里，数据中心收入环比还是往上（估算 24.9 亿到 25.5 亿美元，比上一季的 21.72 亿多一成半）；中际旭创 2026 年报里拆出来的四季度模块收入，也比三季度多（估算 319 亿到 347 亿元）；两家既不是同一季度见顶，也不是差一个季度——它们的季末永远差一个月，拿「同一季度」去比本身就错位。
先把这个 8.81% 说清楚：它跟数据中心收入没有关系。11 月 2 日 Marvell 宣布全股票收购 Lumentum，企业价值 1,150 亿美元，换股比例固定为 4.5 股 Marvell 换 1 股 Lumentum，折合每股 1,132 美元，比 10 月 28 日（市场开始传这件事之前的最后一个交易日）溢价 23.9%。跌的是这笔交易的价格和结构，加上整个 AI 硬件板块同一天的换手，不是订单。
算术摆出来就明白了。10 月 28 日 Marvell 收 251.65 美元，11 月 3 日收 171.97 美元，四个交易日少了 31.7%；同期 Lumentum 从 913.95 掉到 680.16。固定换股比例下，一股 Lumentum 值 4.5 乘 Marvell 股价，11 月 3 日是 773.87 美元，而它自己只收在 680.16，折价 12.1%。这 12% 是市场给「2028 年上半年才交割」打的折——要过双方股东会、要过监管，中间还有一年多。全股票交易里买方跌多少，卖方跟着跌多少，这是算术题，不是需求题。
数据中心收入这条线本身在加速：2026 财年三季度 15.18 亿美元，四季度 16.51 亿（环比 +9%），2027 财年一季度 18.33 亿（+11%），二季度 21.72 亿（+18%），占公司总收入从 73% 抬到 79%。三季度总收入指引 31.50 亿美元、上下 5%；11 月 2 日公告里把这季的预告收入改成了「与指引一致、中点不动、区间收窄到上下 2%」，也就是 30.87 亿到 32.13 亿美元。
按 79% 的占比算，这季数据中心是 24.89 亿美元、环比 +14.6%；按占比再抬两三个点算（前四季每季抬 2 到 3 点），是 25.5 亿、环比 +17.5%。这两个都是估算，公司不给分部指引。对上去年同期 15.18 亿，同比是 +64% 到 +68%，比二季度的 +46% 还高一截。管理层 8 月说过定制业务从 2027 财年下半年开始显著加速，下半年正是 8 月到次年 1 月，覆盖的就是这季和下一季。
中际旭创这边的数更硬。2026 年一季度 194.96 亿、二季度 222.82 亿（环比 +14.3%）、三季度 277.69 亿（环比 +24.6%、同比 +171.8%），一路往上。模块业务占收入 99%（上半年模块 413.55 亿、总收入 417.78 亿），所以总收入的方向就是模块的方向。
四季度怎么走，公司自己在 11 月 3 日晚上九点的电话会上给了答案。被问到四季度到明年一季度客户拉货节奏，回答是数据中心客户拿货节奏正常、没放缓，光模块供应整体仍偏紧，没出现生产完发不出去的情况；被问到四季度核心物料比三季度是否改善，回答是物料在逐步改善，出货量「有望连续几个季度保持环比提升」。同一场会里说 2027 年「绝大部分订单已覆盖全年」，2028 年已有至少一家重点客户给了明确需求指引。
按三季度 277.69 亿、四季度环比 +15% 到 +25% 推，四季度单季是 319 亿到 347 亿（估算），2026 年全年模块收入约 1,004 亿到 1,032 亿（估算）。参照去年：2025 年四季度 132.35 亿，比三季度的 102.16 亿高 29.5%，四季度本来就是它一年里最强的一季。
所以「同一季度先见顶，还是差一个季度」这个二选一，两个答案都不成立。Marvell 的财季是 8 月到 10 月，中际的三季度是 7 月到 9 月；Marvell 的下一季是 11 月到次年 1 月，中际的四季度是 10 月到 12 月。两家的季末永远差一个月，不是差一个季度。硬拿「12 月初那份季报」和「3 到 4 月的年报」直接对，比的是美国的 8 到 10 月和中国的 10 到 12 月，而跟中际四季度真正对得上的是 Marvell 的下一季，那份季报要到明年 3 月才出。
还有一个能提前看的指针：Lumentum 截止 9 月 26 日的季度收入 12.626 亿美元、环比 +25.5%，其中卖给模块厂的器件 7.542 亿（+16.1%）、系统 5.084 亿（+42.4%）；它对 10 到 12 月的指引是 15.75 亿到 16.25 亿美元，环比 +24.7% 到 +28.7%，管理层说这季要做出约 3.5 亿美元的环比增量。给模块厂供料的那一层在 10 到 12 月还在加速，模块厂的出货就不太可能在这个季度掉头。
再看这一天的盘面，跌的不止 Marvell。Vertiv 跌 13.14%，可它 10 月 28 日交的三季报是有机订单同比翻倍、全年收入指引上调 6.5 亿美元、四季度指引环比继续增；广达服务跌 12.00%、GE Vernova 跌 10.03%、伊顿跌 9.69%，同一天微软涨 6.6%、谷歌涨 4.2%、亚马逊涨 7.0%。另一头，A 股半导体设备在涨（拓荆 +20.00%、中微 +16.62%、三环 +14.77%），美股同类在跌（应用材料 -4.1%、泛林 -4.1%）。钱从卖铲子的流向买铲子的，两个市场还给同一批设备定相反的价，这种盘面是仓位在换手，不是订单在消失。
反方最强的一条：模块这一层已经有人环比转负了。新易盛 2026 年三季度单季 109.40 亿，比二季度的 125.72 亿少 13.0%，同一层、同一季、同一个终端需求，它先掉头。中际的报表里也找不到订单侧的硬证据：三季度末合同负债只有 0.11 亿元，等于没有客户用预付款锁量；而 Lumentum 卖给模块厂的器件环比只 +16.1%，不到它系统业务 +42.4% 的一半，说明模块厂 9 月那季的拉货速度已经慢于集群部署速度。这么看，中际四季度「还能环比增」靠的是季节性和管理层的口头指引，不是已经落地的订单，最后一分要靠 12 月这一个月客户提不提货。
什么情况算我看错：Marvell 12 月初那份季报里，数据中心收入低于 21.72 亿美元、也就是环比为负，那我「这季还在增」的判断直接作废；如果落在 24.9 亿以下但仍是正的，说明我把占比外推得太乐观，方向对、幅度错。中际这边，2026 年报披露的模块收入低于约 966 亿（前三季模块约 688 亿加上三季度单季 277.69 亿，这是估算出来的阈值），就意味着四季度单季低于三季度，我错。再往后，如果 2027 年一季度单季收入低于 2026 年四季度，公司那句「连续几个季度环比提升」就断了，那时候错的不是这次的判断，是需求真的到顶了。
- 01Marvell 8-K 2026-11-02 / 0001193125-26-442714 / d81187dex991.htm（与 Lumentum 合并公告、2027 财年三季度预告收入）
- 02Lumentum 8-K 2026-11-02 / 0001193125-26-442758 / d163169dex991.htm
- 03Lumentum 8-K 2026-11-02 / 0001628280-26-068655 / lite_ex991xq1fy27.htm（2027 财年一季度业绩、二季度指引）
- 04Marvell 8-K 2026-08-27 / 0001835632-26-000022 / q227_8kx812026ex-991.htm（二季度业绩、三季度指引、按终端市场收入表）
- 05Marvell 8-K 2026-05-27 / 0001835632-26-000014 / q127_8kx522026ex-991.htm（按终端市场收入表）
- 06Marvell 8-K 2026-03-05 / 0001835632-26-000006 / q426_8kx1312026ex-991.htm（按终端市场收入表）
- 07Vertiv 8-K 2026-10-28 / 0001628280-26-067707 / q32026exhibit991vrt10282026.htm（三季度业绩、四季度指引）
- 08中际旭创 2026 年第三季度报告 2026-10-30 / AN202610291831181859
- 09中际旭创 投资者关系活动记录表 2026-010 2026-11-03 / AN202611031831382786
- 10中际旭创 近期生产经营情况公告 2026-100 2026-10-23 / AN202610231830833072
- 11中际旭创 2026 年半年度报告分产品数据（光通信收发模块 413.55 亿元）2026-08 / tushare 主营构成
- 12新易盛、光迅科技、天孚通信 2026 年三季报单季收入（累计相减）/ tushare 利润表
- 13纳斯达克历史收盘价 2026-10-28 至 2026-11-03 / api.nasdaq.com/api/quote/{ticker}/historical
## 两家不在同一层，账也不是同一天记的
先把位置摆正。Marvell 在这条链上卖两样东西：定制 ASIC 给云厂，DSP 和 TIA 给模块厂。中际买的就是后一样，再加光芯片，做成模块卖给云厂。同一次需求，钱这么流：云厂下单，Marvell 发 DSP，中际做成模块发货，云厂验收。
收入确认点因此错开。器件发出就记收入，模块要等客户验收，中间隔着组装、测试、海运和客户自己的排产。这段时滞大约一个季度。
所以「差一个季度」这句话，在日历上是错的，在链条上是对的，两件事同时成立。财季只差一个月：Marvell 季末在 10 月底，中际在 9 月底。但同一批货的收入，Marvell 比中际早记一个季度。
真正跟中际 2026 年四季度（10 到 12 月）对得上的是 Marvell 11 月到次年 1 月那一季，那份季报要到明年 3 月。12 月初那份覆盖 8 到 10 月，跟中际四季度只重叠了 10 月一个月。
把「财季差一个月」和「链条差一个季度」混在一起，两个差都不大，加起来正好够让人以为是一家领先另一家一个季度。这题的二选一就是这么来的。
## 同行横截面：掉的是一家的量，不是一层的价
四家公司 2026 年三季度的单季数摆在一起，方向很整齐。
| 公司 | 单季收入（亿） | 环比（%） | 单季毛利率（%） | 毛利率环比（pp） |
|---|---|---|---|---|
| 中际旭创 | 277.69 | +24.6 | 46.61 | +0.19 |
| 新易盛 | 109.40 | -13.0 | 46.27 | -1.69 |
| 天孚通信 | 19.23 | +28.4 | 65.05 | +0.39 |
| 光迅科技 | 50.73 | +32.2 | 25.51 | +0.93 |
三家环比增，只有新易盛掉。更关键的是毛利率：中际连续三个季度往上（46.06、46.42、46.61），天孚和光迅也在抬，只有新易盛掉了 1.69 个点。
这个组合说明模块层没有打价格战。如果是需求见顶、产能过剩，先崩的是单价，毛利率要整层往下走。实际是三家毛利率还在抬，只有一家的量掉了。量掉、价没掉，是订单节奏问题，不是这一层的价格问题。
关键在这儿：掉的是一家的量，不是一层的价，这两种情况后续完全不同——量的问题一个季度能补回来，价的问题补不回来。
## 顶部的三个信号，现在亮了几个
第一个，上游器件的环比。Lumentum 截止 9 月 26 日那季，卖给模块厂的器件 7.54 亿美元、环比 +16.1%，它自己的系统业务 +42.4%。器件这头慢于系统那头，说明模块厂拉货的速度已经慢于集群部署的速度。这个信号亮了一半：还在增，但增速掉队。
第二个，二线模块厂。新易盛三季度 -13.0%，这个信号亮了。
第三个，龙头。中际 +24.6%、毛利率还在抬；Marvell 这季的数据中心按占比外推是 24.9 亿到 25.5 亿美元，比上一季的 21.72 亿多一到两成。这个信号没亮。
一个亮、一个半亮、一个没亮，对应的不是「见顶」，是「顶部正在形成、还没走到龙头」。所以问两家是不是同一季度见顶，等于问两支刻度本来就错开的温度计是不是同时到零度。
## 龙头为什么晚到
```chart
type: bar
title: 中际旭创单季收入，到 2026 年三季度还没出现环比负增长
unit: 亿
labels: 2025Q1, 2025Q2, 2025Q3, 2025Q4, 2026Q1, 2026Q2, 2026Q3
values: 66.74, 81.15, 102.16, 132.35, 194.96, 222.81, 277.69
```
三个原因叠在一起。
一是客户结构。中际的订单分散在几家云厂和设备商手里，单一大客户的排产波动传不到整季；新易盛能在两个季度里从 +50.8% 翻到 -13.0%，看起来是它的单一大客户权重更高，一家客户变了，整季跟着变。
二是成本或价格。毛利率三个季度连升，说明要么卖出去的模块单位价值在往上走，要么单位成本在往下走，两条都指向同一件事：不是靠降价抢量。降价换来的增长，毛利率一定先掉；它没掉，说明这一轮不是价格驱动的。
三是季节性。2025 年四季度 132.35 亿，比三季度的 102.16 亿高 29.5%，四季度本来就是一年里最强的一季。按三季度 277.69 亿、四季度环比 +15% 到 +25% 推，四季度是 319 亿到 347 亿。
## 同一次需求，在报表上是倒着出现的
传导链是：云厂资本开支 → 定制 ASIC 和 GPU → 光模块 → DSP 和光芯片。收入确认的方向跟需求流的方向相反，越上游越早记，越下游越晚记。
落到科目上：Marvell 落在「数据中心」这个终端市场一行，公司不给分部指引，只能按占比外推；中际落在「光通信收发模块」这一产品行，占收入 99%；新易盛的 -13.0% 落在收入，不在毛利率；云厂落在资本开支。
这条链的意义在于：所谓见顶，在每一层报表上是同一个事实的不同时点。上游先减速、二线先掉量、龙头最后转，这是分工决定的顺序，不是巧合。所以「谁先见顶」的答案，跟「谁在链条上游」是同一个答案。
## 反方最强的一条：这条链没有订单簿
最硬的反驳不是「行业要跌」，是「你根本看不到订单」。
新易盛二季度环比 +50.8%，三季度 -13.0%，中间没有任何报表信号——没有合同负债异动，没有存货异常，没有指引下调。它直接从加速翻到负增长。
中际三季度末合同负债 0.11 亿元，等于客户没有付一分钱锁量。它四季度环比增 15% 到 25%，靠的是季节性加上 11 月 3 日电话会上管理层的口径——数据中心客户拿货节奏正常、光模块供应仍偏紧，出货量「有望连续几个季度保持环比提升」——不是已经签下的订单。
所以真正的风险不是四季度掉，是四季度不掉、2027 年一季度掉，而掉之前报表上什么都看不到。这个风险用公开数据排不掉，只能用上游器件的环比去逼近，这也是为什么第一个信号比第三个重要。
第二条反方针对方法本身：拿股价当需求指标不成立。8.81% 那天跌的是固定换股比例下的收购折价和 AI 硬件的仓位换手，跟数据中心订单没关系。它意味着的是，市场那天定价的是「4.5 股 Marvell 换 1 股 Lumentum 值不值」，不是「订单还在不在」。
所以股价既不能证实也不能证伪收入，用它推季报，方向从一开始就错了。
## 什么算我看错：四条证伪条件
四档，越往后越致命。
- 方向：Marvell 12 月初那份季报的数据中心收入低于 21.72 亿美元、环比为负，我「这季还在增」的判断作废。
- 幅度：中际 2026 年报模块收入低于约 966 亿元，我错。判据是全年模块收入要大于前三季的 688 亿加三季度单季的 277.69 亿，全年低于这个数，等于四季度单季低于三季度。
- 结构：中际年报的累计毛利率低于前三季的 46.39%、而四季度收入仍环比增，说明四季度是拿毛利换收入，顶部先在利润表出现、收入表看不出来，这套「收入没见顶就是没见顶」的框架要改。
- 链条：Marvell 数据中心环比为负、而中际四季度还环比增，说明我把上游领先的时间估短了——不是一个月，也不是一个季度，是两个季度，整条链的时点要重排。
前三档都能用一份文件核，第四档要等两份文件都出齐。
来源：
Marvell 8-K 2026-11-02 / 0001193125-26-442714 / d81187dex991.htm（与 Lumentum 合并公告、2027 财年三季度预告收入）
Marvell 8-K 2026-08-27 / 0001835632-26-000022 / q227_8kx812026ex-991.htm（2027 财年二季度业绩、数据中心收入 21.72 亿美元）
Marvell 8-K 2026-05-27 / 0001835632-26-000014 / q127_8kx522026ex-991.htm（按终端市场收入表）
Lumentum 8-K 2026-11-02 / 0001628280-26-068655 / lite_ex991xq1fy27.htm（2027 财年一季度器件 7.542 亿、系统 5.084 亿，二季度指引）
中际旭创 2026 年第三季度报告 2026-10-30 / AN202610291831181859
中际旭创 投资者关系活动记录表 2026-010 2026-11-03 / AN202611031831382786
中际旭创 2025 年年度报告（2025 年四季度单季收入 132.35 亿元）
新易盛、天孚通信、光迅科技 2026 年三季报（单季收入与单季毛利率由累计数相减，单季毛利额=累计收入×累计毛利率后差分）
单季口径：本期累计减上期累计，一季报直接用累计值；数据源 tushare 利润表与 marketdata.fin