The market interprets the decline in this industry chain on November 6 as a de-rating, with orders unchanged. At Nvidia's conference call in mid-to-late November, if the quarter-on-quarter growth rate of next quarter's revenue guidance is lower than the guidance it itself gave last quarter, will this explanation still hold up?

The market interprets the decline in this industry chain on November 6 as a de-rating, with orders unchanged. At Nvidia's conference call in mid-to-late November, if the quarter-on-quarter growth rate of next quarter's revenue guidance is lower than the guidance it itself gave last quarter, will this explanation still hold up?

The quarterly report around November 17 — "next quarter's revenue guidance quarter-on-quarter is lower than the previous quarter's guidance quarter-on-quarter" — this condition almost certainly holds. But its holding doesn't say anything about whether orders are good or bad.

Because what it measures isn't demand, it's the magnitude by which Nvidia itself beat guidance last quarter. For five consecutive quarters, it sold 3.9% to 5.7% more than its own guidance each quarter. The base was raised by its own overselling, so next quarter's guidance quarter-on-quarter naturally sits lower.

This condition already held once three months ago. The FY27 second-quarter guidance given on May 20 implied quarter-on-quarter +11.5%, three points lower than the FY27 first-quarter guidance of +14.5% given on February 25. In the following quarter, Zhongji Innolight's single-quarter revenue quarter-on-quarter accelerated from +14.3% to +24.6%. The inference that "guidance deceleration equals order weakening" was falsified once.

So what really needs watching isn't Nvidia's quarter-on-quarter, it's which layer this chain's gross margin is sinking toward. The third-quarter reports have already given the answer: money has moved from assembly to the chokepoints, and the middle layer is the next to be squeezed. As for "valuation compression, orders unchanged," it's half right and half wrong — orders at Zhongji, Shengyi, and Wus have indeed not changed, while at Eoptolink and Inspur they already have; and the three words "valuation compression" are not what happened on this day at all.

This is an arithmetic problem, not a demand problem

The most important column in this table isn't quarter-on-quarter, it's the last column. In the quarter where guidance was 1080 yi, the actual will most likely land around 1130 yi. For its next quarter's guidance implied quarter-on-quarter to hold at 12.2%, fourth-quarter guidance would have to come in at 1272 yi dollars, 192 yi more than the 1080 yi just given.

The largest single-quarter guidance jump over the past six fiscal quarters was 170 yi dollars. 192 yi is 13% more than that maximum. This number has never appeared, and it won't appear just because "orders are still there."

As long as Nvidia's third quarter lands above the guidance midpoint — by the record of the past five quarters, this is highly likely — then "fourth-quarter guidance quarter-on-quarter lower than third-quarter guidance quarter-on-quarter" is inevitable rather than a signal. Taking something inevitable to verify something that might exist — you can't verify it.

So once this condition holds, what should be watched? Gross margin. Revenue quarter-on-quarter is polluted by its own magnitude of overselling; gross margin is not polluted: it only reflects, for what was sold, how much is price and how much is cost.

November 5 to 6, what fell was not one chain

On November 5 US stocks fell broadly, but the biggest and smallest differed by a factor of two.

The next day A-shares opened and picked up this thread. But where it picked up is very particular.

The ChiNext Index fell 4.0%, this chain fell 7% to 9.6%, equipment barely moved, and wafers were still red.

That equipment didn't fall is the key to this day. At a real capex inflection, the first to be cut are always the equipment sellers: order cancellations land on equipment first, and only afterward on parts and assembly. Yet among the three — NAURA, AMEC, and Piotech — the worst faller was down only 1.2%.

So what the market priced this day wasn't "no more building from now on," it was "once built, who gets the money." The premise of the term "valuation compression" is that similar assets get repriced by the same proportion, but here even the equipment one street over wasn't compressed; what was compressed was only the segment selling to overseas cloud vendors.

The trigger wasn't orders either. On November 5 Reuters reported Nvidia chips piling up in Hong Kong, with China not approving imports. Nvidia itself wrote plainly in its second-quarter earnings release: third-quarter guidance assumes no data center compute revenue from China. If something goes wrong in China, it can't hit the number to be announced on November 17.

Where the money is: the staircase of single-quarter gross margin

This staircase has only one rule: gross margin is inversely proportional to "whether this layer can be substituted." Wafers, HBM, copper-clad laminate formulations, optical chips — if it can't be expanded, it can't be expanded, and the price increases are all their own. Assembly and complete machines — anyone can assemble, and not a cent of the price increase lands on their own books.

Micron's gross margin this quarter was 86.8%, and SK Hynix's third-quarter operating margin rose from 76.3% to 80.0%. Both are raising prices, and neither one's capacity is in anyone's twelve-month plan.

Foxconn Industrial Internet's third-quarter report states clearly: cloud computing business revenue up 75% year-on-year and up 25% quarter-on-quarter, AI-GPU rack shipments up 15% quarter-on-quarter, AI-ASIC up 80% quarter-on-quarter. The business is very good. But gross margin is 6.99%, exactly the same as last quarter, within 6.5% to 7.4% for five quarters, and net margin is 4.34%.

What it earns is the money of turnover, not the money of price spread. Revenue up 25%, profit up 26%; if revenue rises only 10%, profit rises only 10%, with no buffer in between.

How the position shift happens

Positions don't shift during the upcycle. Price increases pass down from the end, each layer adds a markup, and each layer's gross margin looks good — on this table for third-quarter 2026, almost all are expanding.

Shifting happens only during deceleration. The end no longer accepts markups, and the middle layer closest to the end is the first to lose markup power; but its upstream chokepoints are still rising — copper-clad laminate up 9 points quarter-on-quarter in the third quarter is evidence that upstream is still marking up.

So the price spread of the middle layer is squeezed from both ends: purchase prices move to the upstream's rhythm, selling prices move to the downstream's rhythm, and the middle portion thins out. This is "profit structure changing position"—not that the total shrinks, but that distribution moves from the middle toward both ends.

So to judge where the money ultimately stays, you don't need to look at whose revenue growth is highest; just look at one thing: whether this layer's capacity can be added within twelve months.

If it can't be added, pricing power is in one's own hands, and the ceiling on gross margin is determined by supply; if it can be added, pricing power is in the customers' hands, and the ceiling on gross margin is determined by competition. On this chain, wafer capacity can't be added, HBM can't be added, and high-end copper-clad laminate and optical chips can't be added either.

The middle layer can be added—module makers' expansion cycle is six to nine months, PCB twelve months, assembly three months. That is their position.

Why the middle layer gets hit first

Optical modules and PCB sit right in the middle of this chain. Upstream are chips and copper-clad laminate; downstream are cloud vendors' annual re-quoting.

In Zhongji Innolight's third-quarter report, operating cost accounted for 53.4% of revenue, the bulk of which was purchased chips and components. If selling prices fall 10% and purchase prices don't move, gross margin directly drops 5.9 points. On this chain, no segment is more price-sensitive than this one.

At the conference call on the evening of November 3, the company addressed both ends. On orders: "The vast majority of customers have not revised down demand or orders; a small number of customers have product structural adjustments; and some customers' 1.6T demand has even increased compared with before" "The vast majority of orders for 2027 already cover the full year."

On prices: "The annual price decline for products is limited; for some products, prices have even not seen any decline."

For now, neither end has broken. But not every company on this chain is like this.

Eoptolink's third-quarter revenue fell 13% quarter-on-quarter, inventory rose 49% in a single quarter, gross margin fell for two consecutive quarters, and net profit fell 19%. Same end demand, same quarter, same layer of business, and it was the first to turn.

The ratio of inventory to single-quarter revenue is an earlier indicator than gross margin: Zhongji 0.86 quarters, GigaDevice 0.72, Shengyi 0.66, Foxconn Industrial Internet 0.49, Eoptolink 1.59. Only this one company has more than a quarter's worth of goods sitting in the warehouse.

This is the most dangerous part of the phrase "this chain." Orders unchanged is true for Zhongji; for Eoptolink it is no longer true. Using one sentence from the leader to represent an entire layer is equivalent to reading divergence as uniformity.

Who is currently a false beneficiary

The first is server assembly. Foxconn Industrial Internet and Inspur: one has the fastest quarter-on-quarter revenue growth, the other has already turned negative, but both have gross margins below 8%, the lowest tier on the entire chain. They will not earn one more point of gross margin because "AI servers sell well"; they will only earn a bit more flow-through money because they sell more.

Foxconn Industrial Internet's own quarterly report has another layer: AI-ASIC server shipments increased 80% quarter-on-quarter, AI-GPU racks increased 15% quarter-on-quarter. The mix is shifting toward ASIC. The unit price and assembly fee of ASIC racks are both lower than those of GPU racks; the more the revenue mix shifts that way, the harder it is to defend the 7% blended gross margin.

The second is Inspur Information. On November 6 it fell only 2.3%, six points less than the chain, but its third-quarter report is the ugliest of this batch of reports: revenue fell 8.1% quarter-on-quarter, net profit fell from 23.5 yi to 9.8 yi, inventory surged from 527 yi to 1,187 yi, and contract liabilities surged from 284 yi to 771 yi.

Inventory and advance receipts both doubled at the same time, indicating that goods came in and orders were taken, but revenue was not recognized and profit was not realized. This shape needs fourth-quarter delivery of goods to match up; if they cannot be delivered, the 1,187 yi of inventory is the floor for impairment.

The third is memory modules. Module makers like Demingli earn not processing fees, but the time spread between "goods stockpiled when cheap" and "selling prices after they rise." In the third quarter, Demingli's revenue still grew 24% quarter-on-quarter, gross margin fell from 38.9% to 32.9%, and the goods in the warehouse were held for 337 days relative to current-quarter operating cost. The time spread will run out; the position will not.

Conversely, the three positions that truly get the money are very clear: Shengyi in copper-clad laminate (41.6%, up 9 points in one quarter), Yuanjie in optical chips (83.4%), and GigaDevice in memory design (70.3%). The common point is—capacity was already full before the price increase, and it cannot be expanded after the price increase either.

Nvidia itself has already cut its gross margin guidance by one point

There is another number no one wants to look at. Nvidia's third-quarter gross margin guidance is 74.0%, a full point lower than the second-quarter actual 75.0%. The guidance sequence is 72.0%, 73.5%, 75.0%, 75.0%, 75.0%, 74.0%—the first downward turn.

Whether this one point is the yield cost of the Vera Rubin ramp or a price concession to customers, the fourth-quarter guidance in the November 17 quarterly report will give the answer. The two things fall on the same line and mean completely opposite things: the former is a short-term cost, the latter is pricing power beginning to loosen.

The answer from upstream leans toward the first. TSMC's fourth-quarter guidance is 523 yi to 543 yi US dollars, with the midpoint up 14.1% quarter-on-quarter, and gross margin 65% to 67%, less than one point lower than the third quarter's 66.4%. What it sees is Nvidia's fourth-quarter wafer starts; it has not slowed down.

One layer further upstream, Lianya Optoelectronics' September revenue was 5.41 yi Taiwan dollars, up 190.1% year-on-year and up 3.8% quarter-on-quarter, the ninth consecutive month of positive quarter-on-quarter growth. This layer is the shortest board in the entire chain; MOCVD tools take more than a year from order to mass production, so capacity cannot be added in the short term. October revenue will be posted on November 10, the earliest health check for this chain.

Wafers have not decelerated, epi has not decelerated, and Nvidia itself gave up only one point of gross margin. Put these three things together, and what they point to is not "orders are gone" but "the distribution of price is changing position."

The strongest counterargument

Some will say: the middle layer is not being squeezed, it is expanding share. The unit price of 1.6T is higher than 800G, and once silicon photonics' share rises, module makers' costs fall further; Zhongji itself said at a meeting that "gross margin on silicon photonics products is better than traditional solutions." Copper connections turning into optical connections means more optical components per machine. Demand is expanding, so why should the middle layer be squeezed?

This rebuttal carries weight. In Q3, Zhongji's gross margin was indeed still rising, and the reason was the rising share of high-end and silicon photonics.

But it cannot explain two things. First, another company at the same layer has already turned negative; if demand is expanding and share is expanding, not just one company would reverse. Second, once CPO arrives, the value of pluggable modules will shift toward FAU and external light sources; Zhongji acknowledged in the Q&A that in CPO solutions, "the main links it can participate in are FAU and ELS external light sources." From a whole-machine supplier to a two-component supplier, both unit price and gross margin have to be recalculated.

This can be verified at the earliest in the second half of 2027, when NPO ramps up. Before that, it is the most substantial opponent to my judgment.

When it can be seen, and what to watch

Under what circumstances I am wrong

One, the FY2027 Q3 report around November 17 (8-K Exhibit 99.1 "Q3 FY2027 PR"): fourth-quarter revenue guidance below 1100 yi US dollars, meaning less than 2% added to the 1080 yi guidance. That would mean demand really has fallen, and the claim that "orders have not changed" is wrong.

Two, the actual gross margin in the same filing: GAAP below 73.5%, breaking below its own guidance floor. That would mean this one point is price concession, not ramp cost, and "derating" must be rewritten as "earnings hit," and my entire logic must be reversed.

Three, TSMC's Q4 2026 earnings call in mid-January 2027 (6-K earnings release): first-quarter revenue guidance turning negative quarter-on-quarter, or gross margin guidance below 63%. The wafer end confirms a reversal, and my statement that "upstream is still accelerating" does not hold.

Four, Lianya Optoelectronics' monthly revenue on November 10 and December 10, 2026 (Market Observation Post System MOPS): two consecutive months of negative quarter-on-quarter growth. The shortest board starts to loosen, and the shortage narrative across the entire chain ends here.

Five, Zhongji Innolight's 2026 annual report in March to April 2027 (CNINFO): fourth-quarter single-quarter gross margin below 45%, or inventory exceeding 300 yi with inventory turnover days longer than in Q3. Then the statement that the module layer "is still expanding gross margin" is wrong.

Six, Coherent and Fabrinet's December-quarter earnings in early February 2027 (8-K Exhibit 99.1): single-quarter revenue turning negative quarter-on-quarter by more than 10%, or revenue falling while inventory days continue to rise. Optical chain orders have already decoupled from Nvidia's financials, and the order of "look at financials first, then at order cuts" is reversed.

There will be a first answer in a week and a half. No need to wait for it to tell us "whether it will happen"; it will only tell us whose account the money is sitting in now.

Source: NVIDIA "Q2 FY2026 Earnings Release", 2025-05-28, 0001045810-25-000115; NVIDIA "Q3 FY2026 Earnings Release", 2025-08-27, 0001045810-25-000207; NVIDIA "Q4 FY2026 Earnings Release", 2025-11-19, 0001045810-25-000228; NVIDIA "Q1 FY2027 Earnings Release", 2026-02-25, 0001045810-26-000019; NVIDIA "Q2 FY2027 Earnings Release", 2026-05-20, 0001045810-26-000051; NVIDIA "Q3 FY2027 Outlook" (Exhibit 99.1 to the Q2 FY2027 earnings release), 2026-08-26, 0001045810-26-000073; TSMC "3Q26 Earnings Release with Guidance", 2026-10-15, 0001046179-26-000726; Micron "FQ4 FY26 Earnings Release", 2026-09-30, 0000723125-26-000018; SK hynix "3Q26 Financial Results", 2026-10-28; Zhongji Innolight "2026 Third Quarter Report", 2026-10-30, Cninfo 1225643953; Zhongji Innolight "Investor Relations Activity Record", 2026-11-03, No. 2026-010, Cninfo 1225656565; Eoptolink "2026 Third Quarter Report", 2026-10-30, Cninfo 1225642947; Shengyi Technology "2026 Third Quarter Report", 2026-10-30, Cninfo 1225644504; Wus Printed Circuit "2026 Third Quarter Report", 2026-10-28, Cninfo 1225627804; Foxconn Industrial Internet "2026 Third Quarter Report", 2026-10-30, Cninfo 1225639532; GigaDevice "2026 Third Quarter Report", 2026-10-30, Cninfo 1225641959; Montage Technology "2026 Third Quarter Report", 2026-10-30, Cninfo 1225643270; Yuanjie Semiconductor "2026 Third Quarter Report", 2026-10-31, Cninfo 1225648199; Inspur Information "2026 Third Quarter Report", 2026-10-29, Cninfo 1225637120; Demingli "2026 Third Quarter Report", 2026-10-26, Announcement No. 2026-093; LandMark Optoelectronics September 2026 revenue, disclosed before 2026-10-15, Market Observation Post System (MOPS).