Earnings Confirm, Stocks Flee: Marvell and NVIDIA Lock In CPO's First Commercial Year as Taiwan Optical Stocks Hit Limit-Down the Same Week (W23)

2 days ago

8 min read

All three fundamental signals this week pointed the same way - Marvell's earnings, NVIDIA's CPO timeline, and Q1 results from Chinese module makers - yet Taiwan's silicon photonics stocks hit limit-down in the very same week. When the numbers and the share prices start moving in opposite directions, the question is no longer "is this cycle real?" but "what is the market afraid of?"

Introduction

On May 27, Marvell pushed its data center business to 76% of revenue and guided next-quarter growth to re-accelerate to 35% YoY. The same week at Computex, NVIDIA moved Spectrum-X Ethernet Photonics - the product that bonds CPO directly into the switch ASIC - from a show-floor demo onto a 2H26 shipping schedule. Add to that the two Chinese optical module leaders posting Q1 revenue up nearly 3x and 2x respectively. Any one of these three events alone would justify a headline saying "the AI optical story is confirmed."

But Taiwan's market ran the opposite script this week. LandMark Optoelectronics, PCL Technologies and Luxnet hit limit-down intraday, while EZconn, FOCI, TrueLight and Browave at one point fell more than half the daily limit. Fundamentals are so strong that management almost seems embarrassed, yet the stocks started correcting ahead of everyone.

This weekly review tries to answer one question: when earnings, timelines and shipment volumes - three hard indicators - all confirm the optical bull narrative at once, yet Taiwan responds with limit-down moves, what is that time lag telling us? Is it a sign the rally is over, or another shakeout where "the fundamentals are right but the valuation got ahead of itself"?

1. Double Confirmation: Marvell's Earnings and NVIDIA's Timeline Nail Down the Narrative

Starting point: Marvell turns its "twin engines" from story into numbers

For the past year, the bull case on Marvell has been the same pitch - "twin engines of custom ASICs plus optical interconnect." The problem is that this pitch ran for four quarters while staying stuck in future-tense vocabulary like design-ins, sampling and strategic partnerships. This week's Q1 FY2027 earnings call switched it to the past tense for the first time.

Revenue came in at $2.418B, up 28% YoY and above the midpoint of guidance. The real story is not the total but the mix: data center revenue of $1.833B for the quarter, or 76% of the total, with next-quarter guidance at a $2.7B midpoint, up 35% YoY (Source: StockTitan / Marvell 8-K). When a company's revenue mix shifts to three-quarters data center within a year, and next quarter's growth re-accelerates rather than slows, the design-in phase is over and volume shipments have begun. The call also noted that the Celestial AI and XConn acquisitions have started contributing - these two deals are the final pieces of Marvell's optical interconnect puzzle, and they are now flowing into the income statement.

Non-GAAP gross margin was 58.9%, and operating cash flow hit a record $638.8M. Put those numbers back into the supply chain: Marvell's ramp of optical DSPs and Photonic Fabric effectively pulls forward 2H 2026 visibility by a quarter for its downstream Taiwan packaging, test and passive optical component suppliers.

Evidence: NVIDIA moves CPO from slideware onto the shipping calendar

If Marvell confirmed that "demand is really shipping," NVIDIA confirmed at Computex 2026 / GTC Taipei this week that "the next-generation architecture really has a timeline."

The Spectrum-X Ethernet Photonics switch integrates CPO directly into the ASIC, with bandwidth of up to 409.6 Tb/s and availability in 2H26; the companion Quantum-X InfiniBand CPO switch delivers 115.2 Tb/s - 144 ports x 800G built on 24 optical engines based on TSMC COUPE; and the Vera Rubin NVL72 system uses sixth-generation NVLink, ConnectX-9 and Spectrum-X Photonics (Source: NVIDIA Blog, Computex 2026).

The key word here is "date." CPO has been the boy who cried wolf for three years, always held back by vague production timelines. This time NVIDIA wrote 2H26 directly into its supply plan, which opens a clear design-in window for the entire Taiwan CPO supply chain - TSMC COUPE, optical engine packaging, passive optical components, fiber array units (FAU) and fiber. For Taiwanese suppliers, the question is no longer "whether to prepare" but "whether they can be ready in time."

Layer this onto real shipments on the demand side and the picture is complete: Innolight's Q1 2026 revenue was RMB 19.50B, up 192% YoY, with net profit attributable to shareholders up 262%; Eoptolink's revenue was RMB 8.34B, up 106% YoY (Source: BigGo Finance). TrendForce estimates shipments of 800G-and-above optical modules will jump from 24 million units in 2025 to about 63 million in 2026, a 2.6x increase in one year; Yole goes further, estimating that 1.6T modules alone will exceed 10 million units shipped in 2026.

Interpretation: three lines point to one thing - this cycle is no longer expectation, it is delivery

Earnings (Marvell), timelines (NVIDIA) and shipment volumes (the two Chinese leaders plus TrendForce's numbers) used to be discussed separately. This week they showed up in the same frame for the first time, and all of them pointed up.

For STT readers, the significance is this: the 2026 optical rally has evolved from "the market expects 1.6T to ramp" to "earnings prove 1.6T is ramping." That is a big difference - the former is valuation running, the latter is earnings running. Once earnings take the baton from valuation, the supply chain's visibility no longer depends on analyst optimism but on purchase orders. In theory, this is exactly when share prices should feel safest.

But the market sees it differently.

2. Divergence Alert: Fundamentals Running Hot, Yet Stocks Hit Limit-Down First

Starting point: management is still reporting good news while the tape is already fleeing

The most jarring moment in Taiwan this week: LandMark Optoelectronics (3081) saw gross margin reach roughly 48% - an almost luxurious figure for an upstream optical component maker - yet its stock hit limit-down intraday the same week. PCL Technologies (4977) and Luxnet (4979) also hit limit-down, while Elaser, EZconn, FOCI, TrueLight and Browave at one point fell more than half the daily limit (Source: Anue, SinoPac Securities Fengyun Academy).

This is not company-specific bad news. Not a single company released negative news this week. On fundamentals, almost every company's story is still moving in the right direction. This is a classic "strong fundamentals, stock corrects first" divergence - the tape is reacting not to earnings but to something beyond earnings.

Evidence: stretched valuations, plus capex noise starting to appear

To understand this divergence, we first have to admit an uncomfortable fact: these stocks rose too much, too fast. Since the week in early May when the big three - Lumentum, Coherent and Fabrinet - reported "strong results, falling shares," the optical sector has faced structural "good news fully priced in" pressure. When a stock has already priced in two years of future growth, even great earnings only "meet expectations," and in an overheated market, meeting expectations is a sell signal.

More worth watching is the first crack appearing on the demand side. This week the market heard noise that "some cloud giants are turning cautious on 2025-2026 data center procurement budgets" (Source: Sina Finance CPO sector analysis), in contrast to across-the-board optimism on the optical module supply side. In China, the violent single-day market cap losses of the "Yi-Zhong-Tian" group (Eoptolink, Innolight, TFC Communication) also show how crowded positioning has become. When a sector's bull case becomes consensus and everyone is in the same boat, any whisper of "capex may slow" is enough to trigger a stampede.

Interpretation: be clear whether you fear valuation or demand

This is the one thing most worth thinking through this week: is the pullback in optical stocks a "valuation correction" or "weakening demand"?

All current evidence points to the former. On the demand side - Marvell's shipments, NVIDIA's timeline, the Chinese leaders' Q1, TrendForce's 2.6x - there is not a single sign of weakening. The only thing weakening is share prices, and they are weakening because they rose too far, positioning is too crowded, and there is one unconfirmed piece of capex noise.

Short-term divergence between fundamentals and share prices is standard equipment in almost every structural bull cycle. What is truly dangerous is not the divergence itself, but not knowing what you are actually afraid of.

In other words, this week's limit-down moves were a "valuation checkup," not a "death knell for demand." The practical takeaway for readers: if your research thesis is aimed at the 2026-2027 ramp of 1.6T and CPO, this week's market moves should not change your industry view - they only change your entry price. The premise, though, is that you actually watch demand-side indicators rather than infer fundamentals from share prices. Only if that capex noise is confirmed in a hyperscaler's earnings call does the story need rewriting. Until then, divergence is just divergence.

Other Signals Worth Tracking This Week

GlobalFoundries (GFS): on 5/4 it unveiled the SCALE co-packaged optics platform, billed as the industry's first silicon photonics CPO solution compliant with the OCI MSA spec. It has demonstrated native 8λ/16λ bidirectional DWDM and uses detachable fiber for serviceability and known-good-die testing. A foundry moving directly into the CPO module layer sets up a two-track SiPh foundry race with TSMC COUPE (Source: GlobalFoundries press release).
Broadcom (AVGO): Q1 FY26 AI semiconductor revenue was $8.4B, up 106% YoY, with an AI backlog of $73B; the CEO reiterated a view of more than $100B in AI chip revenue in 2027. The Q2 FY26 earnings call is expected in early June and is next week's single biggest catalyst (Source: Tom's Hardware).
TSMC (2330): COUPE uses SoIC-X to stack electronic and photonic dies and enters its volume production year in 2026; the roadmap runs from small pluggables to CoWoS-based CPO and 6.4T optical engines, all the way to 12.8T Optical I/O. TSMC, ASE, MediaTek, Quanta and others have formed the Silicon Photonics Industry Alliance (SiPhIA) (Source: TrendForce).
Accton (2345): continues to supply high-speed Ethernet switches to hyperscalers such as Google and Meta, riding the 800G-to-1.6T upgrade cycle (Source: SinoPac Fengyun Academy).
CPO market size: SNS Insider estimates growth from $91.27M in 2025 to $1.924B in 2035, a 35.7% CAGR (Source: GlobeNewswire).

STT Take and What to Watch Next Week

This week's takeaway is clean: the optical bull narrative was confirmed on fundamentals by three hard facts at once - Marvell's earnings, NVIDIA's CPO timeline and Chinese module makers' Q1 - but Taiwan's market reacted first with limit-down moves reflecting valuation pressure after a big run. This is a valuation correction, not a demand alarm, and the two must never be confused.

What really matters are two checkpoints over the coming weeks. First, what Broadcom says on its Q2 FY26 call in early June - it is one of the few players betting on all three legs of ASIC, switch and CPO. If its CPO progress and backlog stay strong, this week's Taiwan pullback will be defined as a pure shakeout; if instead management's wording hints at any capex caution, the whisper that was only "noise" this week could become the main theme next week. Second, watch the Taiwan upstream names where "fundamentals haven't changed but the stock corrected first" - whether LandMark's 48% gross margin holds, and whether it bottoms first once valuation returns to a reasonable range, will be the best thermometer for judging whether this is a shakeout or a reversal.

Demand is real; valuation simply ran ahead. At times like this, watching purchase orders is far more useful than watching candlestick charts.

This article is for technology and industry trend analysis only and does not constitute investment advice.

STT Market Insights | June 1, 2026 | Compiled by the STT editorial team