Zhongji Innolight Co., Ltd. (300308.SZ) - Deep Dive Research Report

Prepared 2026-08-04. Most recent verified reporting period: Q1 2026 (released ~late April 2026) plus the 2025 Annual Report. The H1 2026 report is due by August 31, 2026 and had not yet been published as of this date; the company issued a mid-July investor-briefing note and completed a Hong Kong H-share listing on July 30, 2026, both of which are used below.

1. What the Company Does

Zhongji Innolight makes the small optical devices that let computers talk to each other with light instead of electricity. Its core product is the high-speed optical transceiver (also called an optical module): a plug-in component, roughly the size of a thick USB stick, that sits in the front panel of a network switch or a server. It takes an electrical signal, converts it into pulses of laser light, fires that light down a fibre-optic cable to another module somewhere else in the data centre, and converts it back into electricity at the far end. In an AI data centre, tens of thousands of Nvidia GPUs need to be wired together into one giant computer. The wiring between racks is fibre, and at each end of every fibre sits one of these modules. Zhongji Innolight is the largest maker of them in the world.

The company as it exists today is really two histories stapled together. The listed shell began in 1987 as Shandong Zhongji Electrical Equipment, a machinery factory in Longkou, Shandong, that made motor stator-winding equipment and listed on the Shenzhen Stock Exchange in 2012. The business that matters was born separately: InnoLight Technology, founded in Suzhou in April 2008 by returnees from the US optics industry, which rode the cloud-computing build-out of the 2010s to become a serious optical-module supplier. In 2017 the sleepy machinery company bought InnoLight in a reverse-merger-style acquisition, renamed itself Zhongji Innolight, and later spun the legacy machinery unit off. So the "Zhongji" in the name is the old listing vehicle; the "Innolight" is the actual business. Today essentially all the value is the optical-module operation run through the InnoLight subsidiary in Suzhou, supplemented by acquisitions of Chengdu Storm (Tsuhan) in 2020 and Chongqing Junge in 2023 that brought in optical-component and packaging capability.

The value proposition is speed at scale, delivered reliably and on time. Optical modules are graded by data rate: 100G, 400G, 800G, 1.6T (terabit) and now 2.4T on the roadmap. Each generation is materially harder to build than the last because it packs more lasers, tighter optics, and more heat into the same tiny form factor. The hard part is not designing one module; it is manufacturing millions of them at high yield, qualifying them with a customer like Nvidia, and then ramping the newest generation to volume months before anyone else can. Zhongji Innolight's edge is that it repeatedly gets to volume production of the newest, highest-value module first, and can build it in the quantities a hyperscale AI build-out demands.

A concrete example: when Nvidia designs a new GPU cluster, it specifies the optical interconnect it needs (say, an 800G or 1.6T module of a particular type - EML-laser, silicon-photonics, multimode, DR, FR). Innolight's engineers work with Nvidia through a long qualification cycle, sending sample modules that get tested for signal integrity, power draw, heat, and failure rate over thousands of hours. Once qualified, Innolight becomes a named supplier and receives volume orders tied to Nvidia's GPU shipment schedule. Because a training network uses roughly two to three optical modules per GPU, a single large cluster can require hundreds of thousands of modules, and Innolight ships them month after month against contractually committed delivery schedules.

Management's framing in mid-2026 was that 2027 orders from major overseas customers are "locked contracts with monthly delivery commitments, not vague guidance," with some customers already sharing 2028 product roadmaps (July 2026 investor conference). That is the essence of the business: it is now an order-book-visible supplier of a scarce, generationally-improving component to the AI build-out.

2. Business Segments

Zhongji Innolight is, for practical purposes, a single-business company: the design, manufacture and sale of high-speed optical transceivers. It does not report multiple large operating divisions the way a conglomerate would. Rather than invent segments, it is more useful to understand the business along the two real axes management itself uses - product generation and end-application - because that is where the economics and competitive dynamics actually differ.

By product generation (the value ladder)

1.6T modules are the current crown jewel and the fastest-growing line. Innolight was the first supplier to complete Nvidia's 1.6T qualification testing, and by management's own account holds roughly half to sixty percent of the emerging 1.6T market. This is where pricing and margin are richest because supply is scarce: management has repeatedly said only a handful of manufacturers can supply 1.6T stably in volume through 2026-2027, which prevents price wars. The core capability here is being able to manufacture a device with eight or more lasers and silicon-photonics optics at commercial yield, which took years of process learning to reach.

800G modules are the current volume workhorse and the bulk of shipments. Innolight is the global leader in 800G, with an estimated ~35% share of Nvidia's 800G procurement. 800G is more mature, so its margins are lower than 1.6T, but the volumes are enormous. This is the cash-and-scale engine.

400G and below are legacy generations still sold into older cloud and telecom deployments. They are a shrinking share of the mix and largely commoditised.

2.4T / next-generation and in-rack (scale-up) optics are the strategic option. Management has flagged 2.4T and NPO (near-package optics) products as carrying "significantly higher margins than mature 800G and existing 1.6T," with the first small-scale purchases expected in late 2027 and mass adoption in 2028. In-rack "scale-up" optical connectivity - replacing copper inside the GPU rack as copper hits its physical distance limit - is a wholly new market management expects to ramp from 2027.

By technology platform

Within these generations, a critical shift is the move from traditional EML-laser modules to silicon photonics (SiPh), where the optical functions are integrated onto a silicon chip for lower cost and power. Management stated that silicon photonics already accounts for more than half of the company's product by the 2025 annual briefing. This matters competitively because SiPh is capital- and know-how-intensive, and Innolight's in-house SiPh capability (bolstered by its component acquisitions) is a barrier that pure-assembly rivals lack.

Because the company is genuinely single-segment, no segment comparison table is warranted; the value ladder above is the right mental model.

3. Products and Business Detail

The product catalogue runs the full speed range: 10G, 25G, 100G, 200G, 400G, 800G, 1.6T, with 2.4T and 3.2T in development. Within each speed the company offers multiple physical variants tuned to different reach and cost points - EML (electro-absorption modulated laser) modules for longer single-mode reach, silicon-photonics modules for cost and power efficiency, multimode (SR) modules for short in-data-centre links, and DR/FR variants for various single-mode distances. Management described the 1.6T family as spanning "EML, silicon photonics, DR, FR and multimode" types, i.e. a full menu rather than one design. The company also supplies coherent optical modules for longer-distance (metro/DCI) links and is investing in in-rack/scale-up optical connectivity products for 2027.

What makes these hard to make is a stack of process knowledge: laser and photonic-integrated-circuit sourcing and screening, precision optical alignment and packaging (coupling light into a fibre core a few microns wide), high-frequency electrical design, thermal management, and above all high-yield high-volume assembly and test. Each new speed generation requires re-qualification with the customer, a multi-month cycle testing signal integrity, power, heat and long-run reliability. The scarce inputs - EML lasers and DSP chips - are supply bottlenecks the whole industry contends with, which is part of why only a few firms can ship the newest generation at volume.

Manufacturing footprint: production runs across mainland China (the core Suzhou operation plus expanded capacity, including the Tongling Phase 3 facility management said was completed by late 2025), and increasingly Thailand and Taiwan. The overseas manufacturing build-out is strategically important: it lets the company serve US hyperscale customers with product made outside mainland China, a hedge against tariffs and export controls, and proceeds from the July 2026 Hong Kong listing are earmarked partly for "global manufacturing expansion."

Geographies and export markets: the business is overwhelmingly export-driven. In 2025 overseas revenue was 90.58% of the total, concentrated in North America and Europe. The customer base is the US and European cloud/AI ecosystem rather than the domestic Chinese market, which is unusual for a Shenzhen-listed company and central to both its opportunity and its geopolitical risk.

Milestones that reshaped the business: the 2008 founding of InnoLight; the 2017 reverse acquisition that put the optics business into a listed vehicle; being first to pass Nvidia's 1.6T qualification; the 2025 ramp of 1.6T into volume; and the July 30, 2026 Hong Kong H-share listing that raised roughly US$6.8 billion, the largest Hong Kong share sale in nearly seven years, giving the company a war chest and an offshore currency for expansion and M&A.

4. Customers

Who buys. The customers are the builders of AI and cloud infrastructure: Nvidia (both as a direct buyer specifying modules for its reference systems and as the anchor of the ecosystem), and the US hyperscalers - Microsoft, Google, Amazon, Meta - who buy either directly or through the systems Nvidia sells. The customer base is geographically North American and European, matching the 90%+ overseas revenue. Customer concentration is high: the top five customers were 75.98% of 2025 revenue.

Who makes the buying decision and on what criteria. Inside these customers, the decision sits with hardware/networking engineering and supply-chain organisations. The criteria are, in rough order: does the module pass qualification (reliability, signal integrity, power); can the supplier ship the required volume on the required schedule; is the price competitive; and is there supply-chain resilience (second-sourcing, geography). The sales cycle is long at the front - qualifying a new generation takes many months - but once qualified, orders recur at high frequency against a delivery schedule tied to GPU shipments.

Why they choose Innolight. Three specific reasons. First, it is usually first to volume on the newest generation, so a customer racing to deploy a new GPU cluster can actually get the modules. Second, it can supply the sheer quantity a hyperscale build demands, which few rivals can match. Third, it has in-house silicon-photonics and component capability that supports cost, power and supply continuity as volumes scale.

Switching costs. These are real but not absolute. Qualification is the moat: a customer that has spent months qualifying Innolight for a given generation does not casually re-qualify a rival mid-ramp. But customers deliberately multi-source - Nvidia uses Innolight alongside Coherent and its Fabrinet/TFC in-house line - so Innolight competes for wallet share every generation rather than holding a locked account. The switching cost protects the current generation's business; it does not guarantee the next.

Concentration and contract structure. The top-five 76% concentration is both a quality signal (it reflects deep integration with the few customers who matter in AI) and a risk (loss of Nvidia wallet share would be severe). On structure, management characterised 2027 orders from major customers as contracts with monthly delivery commitments rather than loose forecasts, which gives unusually good revenue visibility for a hardware component maker - closer to a committed backlog than to spot business.

5. Competitive Landscape

The optical-module industry is an oligopoly at the leading edge and commoditised at the trailing edge. At 800G and 1.6T - where the AI money is - only a handful of firms can ship at volume and yield, and Innolight is the largest of them. The picture below focuses on that leading edge.

Eoptolink (300502.SZ) is the closest peer and the number-three global supplier. It shares Nvidia wallet share with Innolight; together the two Chinese firms are reported to hold roughly 60% of Nvidia's 800G module supply. Eoptolink competes hard on cost/performance and has been strong in LPO (linear-drive pluggable optics). Innolight's edge over Eoptolink is scale, breadth of the 1.6T menu, and depth of silicon-photonics capability; Eoptolink's edge is nimbleness and price.

Coherent (COHR) is the leading US module maker and holds perhaps ~20% of Nvidia's wallet. Independent analysts have flagged Coherent's historical weakness as operational discipline rather than technology; a newer management team is trying to fix that. Innolight tends to win on time-to-volume and cost; Coherent wins where customers want a US-based, non-China supplier.

Fabrinet (FN) is the contract manufacturer behind much of Nvidia's in-house and partner module production (the TFC/Fabrinet line); it is less a branded rival than the manufacturing arm of Nvidia's insourcing effort.

Lumentum (LITE) supplies both modules (via its Cloud Light acquisition) and, critically, the lasers that go inside everyone's modules, so it is part-competitor, part-supplier.

Accelink (002281.SZ), Huawei/HiSilicon, Hisense Broadband, Source Photonics and Huagong Tech round out the Chinese field, generally stronger in telecom and trailing-edge than in leading-edge AI modules.

The two structural threats sit above the module makers. Nvidia's own insourcing (targeting more than half of module production internally over time) is the single biggest competitive overhang - it is the customer trying to become the supplier. And co-packaged optics (CPO), which would integrate the optics directly onto the switch chip and could bypass pluggable modules altogether, is a medium-term disruption that benefits switch-chip makers like Broadcom (AVGO) and Nvidia. Innolight's counter is that CPO adoption is phased and slow (management's own read: NPO/near-package small-scale in late 2027, mass adoption 2028), and that even in a CPO world someone must build the optical engines.

Market-cap figures are rough approximations as of mid-2026, shown only as peer-size reference; they are highly volatile in this sector and should be re-checked.

Barriers to entry are moderate-to-high at the leading edge: customer qualification cycles, silicon-photonics and packaging process knowledge, access to scarce EML lasers and DSPs, and the capital to build volume capacity. They are low at the trailing edge, which is why 400G and below are commoditised. Innolight is strongest at the newest generation and most exposed to Nvidia insourcing and CPO.

6. Industry

Demand driver. The industry is now driven almost entirely by one thing: the AI data-centre build-out. Every incremental GPU deployed needs optical interconnect to be lashed into a cluster, and the ratio of modules to GPUs is high (roughly 2-3:1 in training networks) and, management argues, expanding as cluster sizes grow. This makes optical-module demand a geared play on AI capex rather than on the slower telecom cycle that historically defined the industry.

Size and trajectory. The total optical-transceiver market reached about US$23.8 billion in 2025 (LightCounting). Within that, 800G and 1.6T are the fast-growing core: LightCounting has projected the combined 800G+1.6T market at roughly US$14.6 billion in 2026, around 64% of the whole, with those two speeds dominating for the next three years. The specific 1.6T segment is ramping from a few hundred thousand units in late 2024 toward multiple millions.

Position in the supply chain. Module makers like Innolight sit in the middle: below them are component and foundry suppliers (EML/DFB lasers, DSPs from Marvell/Broadcom, TIAs/drivers from Macom, silicon-photonics foundries like TSMC/Tower/GlobalFoundries); above them are the switch/GPU vendors and cloud buyers. Innolight has been integrating downward into components (via its Chengdu and Chongqing acquisitions and in-house silicon photonics) to capture more value and secure supply.

Import/localisation dynamics. This is the mirror image of most Chinese-industrial stories: rather than substituting imports for the domestic market, Innolight is a Chinese company that exports 90%+ of its output to the West. The relevant dynamic is therefore the reverse - Western customers and governments weighing how much AI-critical hardware to source from China, which is why Innolight is building capacity in Thailand and Taiwan.

Regulation. The binding regulatory environment is geopolitical: US export controls, tariffs, and national-security designations. Innolight was added to the US Department of Defense's Section 1260H "Chinese military companies" list (added in the June 8, 2026 update; also referenced in late 2025). Customer qualification standards (reliability, interoperability) are the other, more routine, regulatory layer.

Cyclicality. Historically the optical industry was cyclical, tied to telecom-carrier capex. The current AI-driven demand is a structural growth wave, but it inherits a new cyclicality risk: it is now tethered to hyperscaler AI capex, which is concentrated in a few buyers and could pause sharply if AI-infrastructure spending disappoints. Within the wave, individual product generations follow a price-decline curve, so margins on any one speed erode over time and must be replaced by the next generation.

Tailwinds: AI capex, rising module-per-GPU ratios, the coming scale-up (in-rack) optical market, and 1.6T/2.4T supply scarcity. Headwinds: annual price declines within a generation, customer insourcing, CPO disruption, and the geopolitical clamp on Chinese AI-hardware suppliers.

7. Growth Triggers

All points below are drawn from the company's investor briefings ("concalls"): the Q3 2025 results briefing (Oct 2025), the 2025 annual results briefing (2026-004, Apr 24 2026), IR record 2026-005 (May 15 2026), IR record 2026-006 (Jul 12 2026), and the July 2026 investor conference (Jul 28-29 2026).

• 1.6T volume ramp accelerating each quarter. 1.6T began formal shipment to key customers in Q3 2025 and ramped faster in Q4; management expects sequential quarterly shipment increases through 2026 (Q3 2025 briefing, Oct 2025; reiterated 2026-004, Apr 24 2026). Repeated across multiple briefings.

• 2027 orders locked with monthly delivery commitments; some 2028 roadmaps in hand.

  "2027 orders from major overseas customers are locked contracts with monthly delivery commitments, not vague guidance" (July 2026 investor conference, Jul 28-29 2026).

• 1.6T supply scarcity through 2027 supports pricing. Only a limited number of makers can supply 1.6T in volume during 2026-2027, keeping new entrants confined to lower-margin 800G and preventing price wars (July 2026 investor conference, Jul 28-29 2026).

• Higher-margin next-generation products (2.4T, NPO) coming.

  "Next-gen products like 2.4T and NPO have significantly higher margins than mature 800G and existing 1.6T" - offsetting normal annual price declines (July 2026 investor conference, Jul 28-29 2026).

• NPO / near-package optics deployment phased in from late 2027. Top two customers begin small-scale NPO purchases in the second half of 2027, with mass adoption in 2028 (July 2026 investor conference, Jul 28-29 2026).

• In-rack "scale-up" optical connectivity a new market from 2027. Management said in-rack optical-connection products are expected to ramp from 2027 as copper hits its distance limit (Q3 2025 briefing, Oct 2025).

• Silicon photonics now the majority platform and expanding. Silicon photonics passed 50% of product by the 2025 annual briefing, supporting cost/power competitiveness on future generations (2026-004, Apr 24 2026).

• Capacity expansion. Tongling Phase 3 completed by late 2025; management flagged increased investment in 3.2T, silicon photonics and coherent technology; Hong Kong listing proceeds (July 2026) earmarked for global manufacturing expansion, R&D, supply-chain upgrades and acquisitions (Q3 2025 briefing, Oct 2025; H-share prospectus, Jul 2026).

• GPU-to-module ratio expanding. The traditional ~1:3 ratio is widening as clusters grow, so module shipment growth outpaces GPU growth (July 2026 investor conference, Jul 28-29 2026).

8. Key Risks

• US national-security designation and escalation risk (high probability, potentially severe). Innolight is on the DoD Section 1260H "Chinese military companies" list (June 2026 update). By itself 1260H is a Pentagon-procurement restriction, not an import ban, so the direct commercial hit is limited today. The mechanism that would hurt is escalation: if the designation graduated into export controls, an Entity List placement, or pressure on US hyperscalers to de-risk from Chinese optical suppliers, Innolight could lose access to the very North American customers that are 90%+ of its revenue. This is the defining risk of the whole thesis - the company's best market and its home government are on opposite sides of a widening technology conflict.

• Customer concentration and Nvidia insourcing (moderate-to-high probability, severe if it turns). Top-five customers are 76% of revenue and Nvidia is the anchor. Nvidia has stated an intent to insource more than half of module production over time. If Nvidia's in-house/Fabrinet line takes share faster than expected, or a hyperscaler dual-sources aggressively toward Coherent or Eoptolink, Innolight loses wallet share generation by generation. The qualification moat protects the current generation but must be re-won each time.

• AI-capex cyclicality (uncertain probability, severe). Demand is now geared to a handful of hyperscalers' AI spending. Management's visibility ("2027 locked orders") is genuinely good, but the entire industry is exposed to a scenario where AI-infrastructure spend pauses or digests. A single-buyer-driven demand curve can reverse quickly.

• Technology disruption from CPO (low near-term, rising medium-term probability). Co-packaged optics would integrate optics onto the switch and could eventually shrink the pluggable-module market that is Innolight's entire business. Management's own timeline (NPO small-scale late 2027, mass 2028) buys time, and Innolight is investing in the adjacent technologies, but a faster-than-expected CPO transition led by Broadcom/Nvidia would compress the pluggable TAM.

• Per-generation price erosion (high probability, moderate drag). Every module generation follows a declining price curve. Margins on 800G are already lower than on 1.6T. The business model requires continuously climbing the value ladder (1.6T → 2.4T → NPO) to hold blended margins; a stumble in ramping the next generation on time would let price erosion win.

• Component supply bottlenecks (moderate probability, moderate drag). EML lasers and DSPs are chronic industry constraints. Innolight's vertical integration into silicon photonics and components mitigates this, but a shortage of a key input could cap the volume upside precisely when demand is strongest.

• Founder/large-shareholder selling and dilution (see Sections 10-11). The controlling shareholder group has been trimming, and the July 2026 H-share issuance meaningfully increased the share count. Ongoing supply of stock is a valuation and sentiment risk even if operations are strong.

9. Walk the Talk

The six briefings used, oldest to newest: H1 2025 briefing (Aug 2025); Q3 2025 results briefing (Oct 2025); 2025 annual results briefing 2026-004 (Apr 24 2026); IR record 2026-005 (May 15 2026); IR record 2026-006 (Jul 12 2026); July 2026 investor conference (Jul 28-29 2026). The most recent is within a week of this report.

The through-line across these briefings is a management team that has been consistently accurate and, if anything, conservative relative to the outcome. In the second half of 2025 the central promise was that 1.6T would move from qualification into real volume. At the Q3 2025 briefing management said 1.6T had begun formal shipment to key customers in Q3 and would ramp faster in Q4. That is exactly what showed up in the numbers: Q3 2025 saw profit growth accelerate sharply, and the trajectory carried into a very strong Q1 2026, where revenue roughly tripled year-on-year and profit rose even faster. The guide that "each quarter's 1.6T shipments will step up sequentially" has, so far, been delivered rather than missed.

At the 2025 annual briefing (Apr 24 2026) management said silicon photonics had already passed 50% of product and that "1.6T and 800G are the current revenue mainstays," with 1.6T "in smooth production, with quarterly shipments expected to rise sequentially."

That statement is checkable, and the subsequent Q1 2026 result - the strongest quarter in the company's history at the time - is consistent with it. Management also repeatedly claimed the industry lacked "vicious competition" and that pricing was disciplined enough to hold margins; the reported gross margin staying broadly stable at a high level into Q1 2026 supports that claim rather than contradicting it.

Where to be sceptical: management's forward statements have grown noticeably more confident and more specific over the six briefings, culminating in the July 2026 "five signals" (2027 orders locked with monthly delivery, 2028 roadmaps, 1.6T scarcity through 2027, higher-margin 2.4T/NPO, expanding GPU:module ratio). These are strong, datable claims, but most of them concern 2027-2028 and cannot yet be verified. The track record on the near-term promises (1.6T ramp, silicon-photonics majority, capacity completion at Tongling) is genuinely good, which lends credibility; but the newer claims are further out and made into a demand environment that has, so far, only surprised to the upside - a condition that can flatter any management team.

Net read: management does roughly what it says on the near-term operational promises, and has been more conservative than reality on demand. The 2027-2028 claims are the ones to hold them to as they come due.

10. Shareholder Friendliness Index

Dividends. Zhongji Innolight pays a rising cash dividend and has been increasing it as profits have surged. Total cash distributed has climbed each year: roughly RMB 157m paid in 2023 (FY2022), ~RMB 355m in 2024 (FY2023), ~RMB 556m as the FY2024 final in April 2025, and ~RMB 1,111m as the FY2025 final paid in early 2026. The company also initiated an interim dividend for the first time in 2025 (~RMB 444m paid August 2025), a shift from annual-only to semi-annual distribution that signals more shareholder-friendly intent. Even so, the payout ratio is modest relative to the explosive earnings growth - the company is retaining the large majority of a rapidly rising profit to fund capacity, which is appropriate for a business in a demand-scarce ramp. (Dividend figures from exchange dividend filings; FY2025 final approved at RMB 11.20 per 10 shares, i.e. ~RMB 1.12/share.)

Buybacks and dilution. In the last ~90 days, the chairman proposed (board-approved July 28-29 2026) a share buyback of RMB 40-80 million for cancellation, priced up to 150% of the trailing 30-day average - a symbolic, tiny program relative to the company's scale, more a confidence signal than a capital return. Over the last three years there was no large repurchase program; the more important fact is dilution running the other way. The share count is growing, not shrinking: employee stock-incentive exercises have added shares (and diluted the founders passively), and the July 30, 2026 Hong Kong H-share listing issued a large new tranche of stock raising ~US$6.8bn. So on a three-year view the count has expanded materially, with the H-share issuance the dominant driver.

Verdict: Neutral - a rising and now semi-annual dividend and a token cancellation buyback show shareholder awareness, but net share count is growing (H-share issuance plus incentive dilution), so capital is being raised and reinvested, not returned.

11. Insider Activities

Insider and major-shareholder transactions for a Shenzhen-listed company are disclosed via Shenzhen Stock Exchange filings and cninfo. The most material recent disclosure is the controlling-shareholder group's ownership change.

Breaking down that group change: Zhongji Holding itself sold ~5.5m shares (stake 11.45% → 10.38%); Wang Weixiu's stake fell from 6.29% to 5.96% by passive dilution (not active selling); and Wang Xiaodong sold ~0.71m shares (0.26% → 0.19%). The company was explicit that the aggregate >1% ownership decline stemmed from three causes: the H-share listing (which dilutes existing A-share holders), stock-incentive exercises, and secondary-market reductions.

Reading the sells. This is not a distress signal, but it is not nothing. The active selling (~6.2m shares) is modest relative to the group's ~193m-share holding, and a meaningful part of the reported "reduction" is passive dilution from the H-share issuance rather than insiders dumping stock. There is no disclosed footnote attributing the secondary-market sales to a specific reason (e.g., a pre-set plan), so for the actively sold portion the reason is best described as founder/controlling-shareholder diversification during a period of extraordinary share-price appreciation. That is the most common and least alarming reason for insider selling in a stock that has re-rated this hard.

Buys. No open-market insider purchases were identified in the last 12 months. The nearest bullish insider signal is the chairman's July 2026 proposal for a (small) buyback-for-cancellation, which is a company-level confidence gesture rather than a personal open-market purchase.

Net assessment. Insiders are net sellers over the last 12 months, but the selling is concentrated in the controlling-shareholder group, is modest in size relative to their holding, and is substantially explained by mechanical dilution from the Hong Kong listing plus incentive exercises rather than by conviction-driven exits. There is no cluster of executives dumping stock and no CEO/CFO open-market selling flagged. Read: mild concern on the dilution and the absence of any insider buying, but not a red flag - consistent with founders trimming into strength rather than fleeing a deteriorating business.

12. Scenarios

Bull case. The AI build-out keeps compounding and the module-per-GPU ratio widens as clusters scale, so Innolight's shipment growth outruns even GPU growth. It stays first to volume on each new generation: 1.6T scarcity holds pricing through 2027, and 2.4T and NPO arrive on schedule at richer margins, so blended margins hold even as older generations deflate. The Thailand and Taiwan capacity lets it keep selling to US hyperscalers despite the geopolitical noise, effectively neutralising the 1260H overhang commercially. The in-rack "scale-up" optical market opens on top of the existing scale-out market from 2027, giving a second growth leg. Nvidia's insourcing progresses slower than feared because the volumes are simply too large for any one in-house line to absorb, so Innolight keeps its wallet share. The company uses its fresh Hong Kong war chest to integrate further down the supply chain and lock in scarce lasers and silicon-photonics capacity. Three years out it is a larger, more vertically integrated, geographically diversified franchise still sitting at the leading edge.

Base case. Demand stays strong but normalises from the vertical Q1 2026 growth rate. 1.6T ramps roughly as guided; 800G stays the volume backbone with steadily eroding price; 2.4T/NPO arrive in 2027-2028 broadly on management's timeline but as a small share initially. Nvidia's insourcing and multi-sourcing gradually pressure wallet share, so Innolight has to run hard on the value ladder just to hold blended margins - which it mostly does, given its execution record. The 1260H designation stays a headline risk that does not translate into an actual export ban, but it keeps a discount and a de-risking incentive hanging over the customer relationships, and pushes more production offshore. Capital return stays modest (rising dividend, token buyback) while the company reinvests. The business grows, remains the category leader, and delivers roughly what management has guided, with the share count somewhat higher after the H-share issuance.

Bear case. Two things go wrong together. First, AI-infrastructure capex digests: a couple of hyperscalers pause or re-phase, and because Innolight's demand is geared to a handful of buyers, orders that looked "locked" get pushed, inventory builds, and pricing cracks faster than the value-ladder can offset. Second, the geopolitical clamp tightens - the 1260H designation escalates toward export controls or explicit hyperscaler de-risking mandates, and Innolight starts losing qualified slots to Coherent, Eoptolink's non-China capacity, or Nvidia's own line, precisely in its 90%-of-revenue Western market. Layered on top, CPO adoption surprises to the fast side, led by Broadcom and Nvidia, shrinking the pluggable-module TAM just as competition intensifies. Margins compress from three directions at once - price erosion, lost wallet share, and technology substitution - and the founders' trimming looks, in hindsight, well-timed. The stock's entire premium was the assumption that this one company would keep sitting at the leading edge of an insatiable market; the bear case is that any one of demand, geopolitics, or CPO knocks it off that perch.

Sources: Zhongji Innolight - Wikipedia; BigGo Finance - trillion-yuan optical giant / FX rumor; BigGo - three structural dilemmas; STCN - Q1 2026 net profit +262%; cninfo IR record 2026-006 (2026-07-12); cninfo IR record 2026-005 (2026-05-15); dfcfw IR record 2026-004 (2026-04-24); JRJ - July 2026 investor call five signals; iccsz - 2027 demand for 800G/1.6T/2.4T/NPO; 9fzt - Q3 2025 results review; 21财经 - controlling shareholder reduction; Sina - chairman buyback proposal; 10jqka - dividend/financing history; China Daily - HK listing raises $6.81bn; CNBC - HK debut; DoD Section 1260H list (June 2026 PDF); Holland & Knight - 1260H update; Deep Dive: Optical Module Market (deepfundamental, 2024-09-27); ip-fiber - InnoLight & Eoptolink 60% of Nvidia 800G; Cignal AI - 800GbE shipments; LightCounting - 2025 transceiver sales $23.8bn