Business Overview

Lumentum Holdings Inc.

NASDAQ: LITE

7 September 2026

Evidence base: Lumentum SEC filings FY2016–FY2026 held in the LITE research folder — annual reports through the FY2026 Form 10-K filed 17 August 2026 (fiscal year ended 27 June 2026), quarterly reports through Q3 FY2026, proxy statements and current reports — supplemented for industry structure by competitors’ own filings, standards bodies, named research firms and Lumentum’s FY2026 earnings releases and call transcripts, which are not held in the folder.

This is not a valuation and not a recommendation.

1. Executive Snapshot

2. What the Company Does

The problem being solved

An AI training cluster is thousands of accelerators that have to exchange data as if they were one machine. Copper stops working as a link medium beyond roughly two to three metres once each electrical lane runs at 200 gigabits per second, so above that distance every connection has to become a light path. Something has to make the light: a semiconductor laser, modulated at the line rate, launched down a fibre and recovered at the far end. Lumentum makes those lasers, and increasingly the assemblies and modules built around them.

The same physics has been selling into telecom networks for twenty-five years, where the distances are hundreds of kilometres rather than hundreds of metres and the buyer is a carrier rather than a hyperscaler. That older market is still there and still growing, but it is now roughly a quarter the size of the data-centre market (Cignal AI, 18 June 2026), and it is not where the change in this company came from.

The unit, traced end to end

The unit of economics is one indium-phosphide laser die and the wafer capacity behind it. A die is grown by epitaxy on an InP substrate in one of Lumentum’s wafer fabs, tested at wafer level, singulated, and packaged into a sub-assembly with a lens, an isolator and — for an electro-absorption modulated laser — an integrated modulator. From there it goes one of two ways: shipped to a module maker as a component, or built into a Lumentum transceiver and shipped as a finished system. Either way it is invoiced against a purchase order and collected in about 63 days (computed from FY2026 balance-sheet figures; 55.5 days a year earlier). There is no subscription, no recurring revenue and no contractual annuity anywhere in this business.

What the company will not tell you about that unit matters as much as the unit itself. Lumentum has never disclosed units shipped, average selling price, wafer starts, capacity or factory utilization — not in FY2026 and not in any prior year. The FY2026 revenue bridge is entirely qualitative: laser chip and laser assembly shipments “represent 78% of the total” Components growth, helped by “a slight increase in average selling prices of laser chip products driven primarily by a shift to 200G lane speeds,” while cloud transceivers grew “more than 173% due to an increase in shipment volume, partially offset by lower average selling prices” (FY2026 10-K, MD&A). That establishes direction and nothing more. An investor cannot compute revenue per die, cannot split volume from price in dollars, and cannot independently check the utilization claim that carries most of the margin story.

What it sells, and to whom

From FY2026 the company reports one segment and disaggregates revenue only two ways: by product type and by shipping geography. Components — semiconductor laser chips, laser sub-assemblies, line subsystems and wavelength management systems, sold to customers who integrate them into their own products — were $2,005.6m, 66.5% of revenue. Systems — complete stand-alone products including optical modules, optical circuit switches and industrial lasers — were $1,008.4m, 33.5% (FY2026 10-K, Note 18).

The position in the value chain is unusual and worth stating plainly: Lumentum sells at two levels of the same chain and therefore supplies its own competitors. The FY2026 10-K says so directly — “For data center interconnects, Lumentum offers both its own coherent pluggable transceivers and the underlying ultra-narrow linewidth laser and coherent components used by transceiver customers.” The Chinese module assemblers that compete with Lumentum’s Systems business are candidate buyers of its Components business, which is what makes the merchant chip position both an opportunity and a source of channel conflict.

Customer concentration is high and rising. Two customers were 26.6% and 15.0% of FY2026 revenue, 41.6% combined, against 31.4% a year earlier; one customer was 30.4% of gross receivables at year end (FY2026 10-K, Note 17). Neither is named. Nor is any competitor: the entire competition disclosure in the FY2026 10-K is two sentences saying the company competes with “various public and private companies” and that “some of these competitors are also our customers.”

What has been entered and what has been exited

The product-line history explains the current economics better than any strategy statement. In FY2019 Lumentum left the datacom transceiver module business, selling product lines to Cambridge Industries Group for $25.5m net and taking $30.7m and $4.3m of long-lived-asset impairments in FY2019 and FY2020. The stated reason was explicit: “the market at the transceiver level is gross margin challenged due to extreme competition” (FY2019 10-K). It also wound down lithium niobate modulators, selling those assets for $17.0m, on the view that indium-phosphide photonic integrated circuits would replace them.

Then in November 2023 it paid $728.5m for Cloud Light Technology and re-entered the very transceiver-module business it had exited four years earlier, and in FY2024 it wrote off $29.1m of in-process R&D and stopped developing coherent DSPs and RFICs in-house — a step back up the chain, ceding that layer to Broadcom and Marvell. The reversal is the clearest single statement of how the economics moved. Modules were unattractive when the buyer was a telecom equipment vendor grinding on price; they became attractive when the buyer was a hyperscaler paying for supply.

The other half of the company has been shrinking throughout. Industrial lasers and consumer 3D sensing, reported as the Industrial Tech segment until FY2025, fell from $703.9m of revenue in FY2022 to $234.2m in FY2025, with segment profit collapsing from $373.5m to $12.1m. In FY2022 that business produced more segment profit than the networking business did ($373.5m against $266.9m). The FY2026 10-K gives no figure for it at all, because the segment no longer exists in the reporting.

3. Industry, Competitive Position and Moat

What the industry sells, and why prices fall

This industry converts scarce compound-semiconductor process capability into bandwidth, and it is priced off cost per bit, which falls every generation by construction. An IEEE 802.3df working-group contribution (Chang and Yu, November 2022) shows that moving from a 400G LR4-10 module to an 800G LR4 module on the same architecture raises component cost only about 1.2 times — a doubling of capacity for a fifth more component cost. Structural price decline is therefore the industry’s normal state, and each new switch-ASIC generation makes the prior optics generation uneconomic per bit almost immediately. Cignal AI’s June 2026 observation that “component shortages are slowing the usual price declines” confirms both halves of that: decline is the norm, and the present is an exception created by supply rather than by product differentiation.

Where the profit actually sits

The value chain runs substrate to epitaxy to laser chip or silicon-photonics circuit, then to optical sub-assembly, transceiver module, line card and finally the network operator. The digital signal processor arrives from an entirely separate CMOS supply chain. Capital intensity is concentrated at the substrate and fab end; module assembly is labour and working-capital intensive rather than capital intensive. The reported margins at each layer are the most useful outside evidence available:

Company-reported figures from each company’s own results release for the period shown. Chinese figures are from financial data aggregators rather than Shenzhen-filed reports and should be treated as directional. Lumentum FY2026 GAAP equivalents are 41.7% gross margin and 17.4% operating margin.

Two things fall out of that table. The first is that assembly on its own is worth very little: Fabrinet earns 12.2% gross margin doing exactly this at $4.6bn of revenue, which is the market price of building modules without owning the chips, and therefore the level a module-only strategy converges toward. The second contradicts the standard narrative. The Chinese module assemblers currently out-earn the Western vertically integrated component makers on both gross and operating margin — Eoptolink at 47.0% and 42.3% against Coherent’s 39.4% and 20.5%. The claim that owning the chip layer means capturing the profit is a forward expectation, not a present fact. The most likely explanation is that assemblers are earning scarcity rent on finished modules while their own input costs are contractually fixed, in which case those margins compress when supply loosens; the direct test is whether Innolight’s and Eoptolink’s gross margins hold through 2027.

Which barriers actually bind

Three barriers are evidenced rather than asserted. Indium-phosphide epitaxy capacity and yield is the binding one: Coherent’s management called InP “our primary constraint” on its August 2026 call, and Lumentum’s chief executive put its own supply-demand gap at “somewhere greater than 30%” in May 2026. Substrate access is the second — three merchant InP substrate suppliers globally, China added indium phosphide to its export control list in February 2025, and China holds roughly 70% of refined indium production. Third is hyperscaler qualification, and the strongest evidence for it is behavioural: NVIDIA paid $2.0bn each to Lumentum and Coherent in March 2026, with purchase commitments and capacity access rights attached, rather than qualify a new supplier.

Two barriers that sound impressive do not bind. Assembly skill is one, for the reason given above. Fab ownership as such is the other: Applied Optoelectronics owns fabs, ran 29.8% non-GAAP gross margin in the June 2026 quarter, and still posted a GAAP operating loss in the strongest demand environment this industry has seen. Fab ownership below scale is a fixed-cost liability, not a moat.

Lumentum’s specific position

Lumentum owns five indium-phosphide wafer fabs and is converting a sixth. It is expanding capacity across its two Japanese fabs and qualifying both continuous-wave and EML process flows there; in March 2026 it bought an operating fab in Greensboro, North Carolina for $38.0m and is converting it from gallium arsenide to six-inch InP, with first revenue guided to early 2028 and further investment described as “hundreds of millions of dollars over the next several years” (Lumentum press release, 26 March 2026; Q4 FY2026 call, August 2026).

The most valuable thing about that position right now is not the capacity itself but who else has it and what they do with it. Coherent stated in August 2026 that “every bit of capacity… we have and then some” goes to its own internal datacom modules — it is not a merchant EML supplier. Lumentum is. That leaves it as the merchant supplier of record for laser chips at the exact moment the merchant market is short, which is what a well-funded competitor would find hardest to reproduce: not the design, but qualified epitaxy capacity plus the years of hyperscaler qualification standing behind it.

The position could weaken from three directions, and none is speculative. NVIDIA wrote the identical $2.0bn cheque to Coherent on the same day — Lumentum is one of two, not the one. Coherent is 2.4 times larger ($7,118.2m of FY2026 revenue against $3,014.0m) and, on stated plans, is expanding InP faster: doubling internal output by the end of its fiscal first quarter 2027, a quarter ahead of schedule, then more than doubling again by the end of calendar 2027, with six-inch lines already running in Texas and Sweden. And Broadcom’s stated optical portfolio now spans DSPs, silicon photonics, co-packaged optics, EMLs, VCSELs and continuous-wave lasers — the switch-silicon monopolist, targeting roughly 66% non-GAAP operating margin, has entered the laser layer.

The outside evidence points to a fairly specific winner profile in this industry: own III-V fab capacity, at enough scale to absorb fab fixed cost through a downturn, and sell merchant chips rather than only finished modules. Lumentum matches on asset type and on merchant position, which is why FY2026 looks the way it does. It does not match on scale, and scale is precisely the clause that determines what happens on the way down.

4. Growth Engine

FY2026 revenue was $3,014.0m, up 83.2% from $1,645.0m (FY2026 10-K). The first question with a number like that is how much of it was bought.

Reported against organic

For once the answer is clean, and the history makes the point better than the current year does. The last material acquisition, Cloud Light, closed in November 2023 and was fully lapped by FY2025. The only FY2026 acquisition was the $38.0m Greensboro fab in March 2026, whose revenue contribution is not disclosed and whose attached transitional supply agreement is recognized net, as agent, contributing $2.3m. FY2026 growth is therefore essentially all organic.

That was not true of the preceding decade. FY2023 reported revenue growth of 3.2% contained $340.4m of acquired NeoPhotonics revenue, which means the underlying business shrank by roughly 16%. FY2024 reported a 23.1% decline while including $199.5m from Cloud Light, so the organic decline was closer to 45%. The company does not disclose acquired revenue separately for FY2020, FY2025 or FY2026, so those years cannot be decomposed at all. The pattern through FY2019–FY2024 was acquisitions masking organic contraction; FY2026 is the first year in which the reported number and the underlying number are the same number.

Where the FY2026 growth came from

Components grew $889.3m, or 79.7%, to $2,005.6m; Systems grew $479.7m, or 90.7%, to $1,008.4m. Within Components, laser chips and laser assemblies were 78% of the increase — roughly $694m — driven by volume with a “slight” ASP increase from the shift to 200G lane speeds; the remaining 22%, roughly $196m, was shipment volume in long-haul terrestrial line subsystems and undersea pump products. Within Systems, cloud transceiver product lines grew more than 173% on volume with falling ASPs, and optical circuit switches contributed more than $90.0m in what management called the initial phase of shipments (FY2026 10-K, MD&A).

Ranked most to least impactful, with what each one actually is:

• Laser chip and assembly volume into AI clusters — structural, with a cyclical overlay. Roughly $694m of the FY2026 increase. Structural because AI cluster architecture converts copper links to optical above two to three metres and every speed generation adds channels; cyclical because the volume is a direct function of hyperscaler capital budgets, which are set annually.
• Factory utilization and capacity ramp — management-driven. Not a revenue driver but the largest earnings driver: management attributes roughly 54% of the gross-margin dollar increase to “higher internal factory utilization.” Headcount went from 7,257 in FY2024 to 13,757 in FY2026, of which 11,916 in manufacturing.
• Cloud transceiver volume — cyclical. More than 173% growth on units, with ASPs falling. This is the layer where Fabrinet earns 12.2% gross margin and Chinese assemblers compete directly; the growth is real and the pricing is not defensible on its own.
• Optical circuit switches — structural and management-driven. More than $90.0m in year one, against a multi-year, multi-billion-dollar purchase agreement described on the May 2026 call, with the first triple-digit revenue quarter guided for Q1 FY2027. Management also noted the customer retains an internal version of the product, which caps how much of it Lumentum can own.
• Scarcity pricing — temporary. Management said on the August 2026 call that it “did reprice a little bit,” with “some benefit… on the gross margin line on pricing” that “still has some room to play through.” Price increases in a structurally deflationary industry are rent on a shortage, and they last as long as the shortage.
• Telecom and network-equipment recovery — cyclical. Long-haul and undersea data transport volume grew; independently, Dell’Oro reported that carrier inventory correction completed in late 2025 after the worst telecom equipment downturn in 22 years, and forecasts optical transport up 16% in 2026.
• Industrial and consumer lasers — structural decline. The only driver working against the total. Down from $703.9m in FY2022 to $234.2m in FY2025; not disclosed for FY2026.

One caution on the geography table: shipping destination is where a contract manufacturer’s factory is, not where the demand originates. The FY2026 10-K says so explicitly. Mexico rising from 9.0% to 14.7% of shipments and Thailand from 17.7% to 20.8% tells you where modules are assembled, not who bought them.

5. Margin, Cash and Capital Allocation

How the margin was made

Gross margin went 18.5% in FY2024, 28.0% in FY2025, 41.7% in FY2026. Management’s own attribution of the FY2026 step is unusually specific and should be taken at face value: approximately 54% of the gross-margin dollar increase came from “lower manufacturing costs as a percentage of revenue, primarily due to higher internal factory utilization,” 29% from a mix shift to higher-margin products, and the remaining 17% from lower amortization of acquired intangibles (FY2026 10-K, MD&A).

That composition matters more than the level. The largest single component of the improvement is operating leverage on a fixed manufacturing base, and operating leverage is symmetric. The 17% from intangible amortization is durable and will continue shrinking as the Oclaro, NeoPhotonics and Cloud Light intangibles run off — acquired developed-technology amortization inside cost of sales fell to $77.6m, or 2.6% of revenue, from 5.0% a year earlier. Only the 29% from mix is both durable and demand-dependent.

Below gross margin the leverage is starker still. R&D fell from 22.2% of revenue in FY2024 to 11.8% in FY2026 while rising only 17.3% in dollars; SG&A fell from 22.9% to 12.1% while rising 4.3%. Operating income was $524.8m, a 17.4% GAAP operating margin, against losses of $180.1m and $434.0m in the two prior years. On management’s non-GAAP basis, which excludes $170.2m of stock compensation and $138.2m of acquired-intangible amortization, FY2026 operating margin was 29.8% and the June quarter alone was 36.6%.

The GAAP net loss is not what it looks like

FY2026 GAAP net loss was $6,935.1m, or $92.96 per share. It is an accounting consequence of the share price, not of the business. In April and May 2026 the company issued roughly 10.6 million shares to retire $1,124.9m of convertible note principal, and the applicable standard required the conversion value in excess of principal — $7,755.1m — to be recognized as a loss on extinguishment. It is non-cash, substantially non-deductible for tax, and will not recur. Operating income of $524.8m is the meaningful line for FY2026, and the same year also contains two one-off credits that will not repeat: a $236.3m release of the US deferred-tax valuation allowance in the June quarter, and a $27.5m escrow settlement from the Cloud Light acquisition recognized in other income.

Cash conversion

Operating cash flow was $751.4m against $524.8m of operating income, and free cash flow was $300.1m after two consecutive negative years. Working capital absorbed $292.0m of that: receivables rose 108.1% with days sales outstanding going from 55.5 to 63.0, and inventory rose 47.1%. Capex of $451.3m understates the true commitment — $181.4m of unpaid property and equipment sat in payables at year end against $43.4m a year earlier, and construction in progress rose from $152.3m to $377.4m, all of which management expects to place in service within twelve months. Depreciation was $128.8m against $451.3m of capex, a ratio of 3.5 times, which means the depreciation line steps up materially in FY2027 whether or not revenue does.

Source: FY2022, FY2024, FY2025 and FY2026 Forms 10-K. Net cash is cash plus short-term investments less total debt principal. Comparability breaks across these columns: reportable segments were redefined in FY2024 and eliminated entirely in FY2026, so no continuous segment series exists; ASU 2020-06 was adopted at the start of FY2023 on a modified retrospective basis, which breaks the interest-expense and debt-carrying-value series before that year; and the company refined its non-GAAP methodology in the first quarter of FY2025, which breaks the non-GAAP series before FY2024. FY2022 operating income excludes no unusual items; FY2025 operating income includes a $34.9m gain on sale of a facility.

Where the cash went

Ranked over the five years FY2022 to FY2026, using cash flow statement figures:

• Debt repayment and settlement — about $1,848.8m, of which $1,377.1m in FY2026 alone. Most of the FY2026 figure is cash settlement of convertible notes that were deep in the money, not deleveraging by choice.
• Acquisitions net of cash — about $1,600.5m, being NeoPhotonics and the IPG telecom lines in FY2023, Cloud Light in FY2024 and Greensboro in FY2026.
• Capital expenditure — $1,035.0m, of which 43.6% fell in FY2026.
• Share repurchases — $719.5m, all of it in FY2022 and FY2023 at an average of $81.66 per share. Nothing was repurchased in FY2024, FY2025 or FY2026, and the $1.2bn authorization expired in May 2025 with $569.6m unused.
• Withholding tax on net share settlement of restricted stock — $422.9m, economically a share retirement, $281.0m of it in FY2026.
• Dividends — none. Lumentum has never paid one on its common stock and states it does not expect to.

The behaviour that ranking describes is a company that bought assets aggressively into a downturn — NeoPhotonics and Cloud Light for $1.6bn while its own revenue was falling 23% and its gross margin was at 18.5% — bought back stock while it was leveraged, then stopped entirely and let the authorization lapse. It then funded the capacity build with $2.0bn of convertible preferred sold to NVIDIA at $695.31 a share in March 2026, rather than with debt or its own cash flow. Across the whole period the convertible structure turned coupons of 0.375% to 1.50% into substantial equity: shares outstanding went from 69.8m at June 2025 to 89.7m by August 2026, plus 2.9m of preferred.

What post-dates the reported figures

Three things sit after the FY2026 numbers. As of 14 August 2026 the company had received early conversion requests for $757.8m of note principal, which must be settled in cash; the remaining converts carry $1,544.6m at cost against an estimated fair value of $7,607.2m, an equity claim of roughly $6.1bn that does not appear on the balance sheet. On 11 August 2026 management guided fiscal Q1 2027 to revenue of $1.225bn to $1.275bn and a non-GAAP operating margin of 39.5% to 40.5%, and said the guide meant “reaching our target model more than a quarter ahead of schedule” — the target model itself is not quantified in the release. And no acquisition or divestiture is announced and unclosed as of the FY2026 10-K.

6. Cyclicality, Constraints and What to Monitor

Exposure, and what is no longer disclosed

The end-market split — cloud and networking against industrial and consumer — was disclosed for the last time in FY2025, at 85.8% and 14.2%. It is not disclosed for FY2026 at all, because the single-segment reorganization in the first quarter of FY2026 removed it. What remains is geography, and it is a shipping-destination table rather than a demand table: the Americas 35.9%, Asia-Pacific 58.2%, EMEA 5.9%, with China and Hong Kong together 26.6% of shipments. Business with Huawei, historically the largest networking customer in China, is now zero, and the FY2026 risk factors add a sentence that was not there before: “we may never recover this demand even if such export restrictions ease.”

Where the business sits in its own cycle

This is the single most useful framing available for FY2026, and it is not the one the headline suggests.

Source: FY2016–FY2026 Forms 10-K. FY2021 reported a 30.2% operating margin but included a $207.5m merger termination fee from the abandoned Coherent transaction; excluding it, FY2021 was 18.3%, and FY2022’s 17.7% is the cleanest prior peak.

So: revenue is at an emphatic all-time high, 70.6% above anything the company has produced before, while gross margin and operating margin have still not reclaimed the levels they reached in FY2022 on 76% less revenue. The margin peak arrived only in the final quarter — the June 2026 quarter ran a 47.4% GAAP gross margin, above the prior full-year peak, and 50.4% on the non-GAAP basis. A record margin at a cycle peak and a record margin at a cycle trough are opposite facts; this one is the former, and it has been in place for one quarter.

The downside, as a mechanism

The fixed-cost base has grown faster than either revenue or margin. Manufacturing headcount went from a company total of 7,257 in FY2024 to 13,757 in FY2026, 11,916 of them in manufacturing. Net property and equipment rose 60% to $1,159.1m, with Thailand doubling to $450.6m. Construction in progress of $377.4m enters the depreciable base within twelve months. Purchase obligations stand at $2,354.4m, of which $2,112.5m falls due within a year — 1.3 times FY2026 cost of sales — and while the terms allow rescheduling, they are described as legally binding. Against that, roughly 54% of the gross-margin improvement came from utilization.

The mechanism, therefore, is not a demand forecast but an arithmetic one. Every announced capacity expansion in this industry lands in the same two-year window: Coherent more than doubling InP output again by the end of calendar 2027, Lumentum’s Greensboro fab producing first revenue in early 2028 and reaching full pitch by 2029, and JX Advanced Metals expanding substrate capacity seven to ten times over four years. If AI order rates flatten while that capacity arrives, the utilization that produced 54% of the margin gain reverses through exactly the same channel, and the reversal hits a cost base that is 90% larger than it was two years ago.

There are two precedents, one of them the company’s own. In FY2024 revenue fell 23.1%, gross margin fell to 18.5%, operating margin to negative 31.9%, and the FY2024 gross margin carried $20.7m of excess-capacity charges. Further back, the corporate ancestor is the cleaner warning: JDS Uniphase’s quarterly revenue fell from $920.1m in the March 2001 quarter to $261.8m in the March 2002 quarter, a 71.5% decline, with gross margin guided to 15–18%. Management’s own explanation at the time is the sentence to keep: “much of the forecasted revenue decline is in components, typically among the Company’s highest margin product lines.” The fabs that are the moat at the top of the cycle are the fixed cost at the bottom of it.

Constraints

The binding constraint today is on the supply side, which is why the margin is where it is. Lumentum’s chief executive put the supply-demand gap at “somewhere greater than 30%” in May 2026 and said in August 2026 that “we are way behind… on high-powered lasers,” with EML unit output targeted to grow more than 50% year on year by the December 2026 quarter and pump laser shipments to rise fourfold over several quarters. Regulatory constraints compound the physical one: China added indium phosphide to its export control list in February 2025 and holds roughly 70% of refined indium production, and the FY2026 10-K states that “China restricted exports to Japan, which affected our substrate supply chain globally.” Export licensing is named by management as one of its own capacity bottlenecks. Tariffs run through cost of sales as a period cost and are nowhere quantified.

Durable against borrowed

Leading indicators, and where each is published

7. Risks, Unknowns and Questions for Deeper Work

Cyclicality is covered in Section 6 and is not repeated here. What follows are the exposures that cyclicality does not capture, ordered by how badly they compound.

• Two customers, 41.6% of revenue, and almost no contractual commitment behind it. Concentration rose from 31.4% to 41.6% in a single year, and one customer is 30.4% of gross receivables. The FY2026 10-K states that the majority of customers purchase under orders “that do not contain volume or long-term purchase commitments,” and then spells out the transmission mechanism: if forecast orders do not materialize the company may “fail to optimize our manufacturing capacity and incur charges for such underutilization” and “incur liabilities with our suppliers for reimbursement of capital expenditures.” Against $2.35bn of purchase obligations and 11,916 manufacturing employees, one customer changing its build plan is a fixed-cost event rather than a revenue event. Management said in August 2026 that it holds “multiple long-term customer agreements” with three-year take-or-pay terms that offset planned capex; no counterparty, volume or term appears in any filing.
• Architecture, not competition, is the long-run threat to laser die volume. Silicon photonics does not remove the laser — silicon cannot emit light — but it changes the count: a 1.6T module built from discrete electro-absorption modulated lasers uses eight modulated die, while the silicon-photonics equivalent uses two continuous-wave sources. Co-packaged optics goes further and replaces the pluggable module entirely with an external light source; Cignal AI forecasts that external-laser module market above $1.5bn by 2030 against a datacom optical component market already running $7.7bn per quarter. Lumentum participates in both architectures — it has shipped ultra-high-power CPO lasers and taken a first order for external light source modules, and says those carry meaningfully higher selling prices than the lasers alone — but the revenue pool it participates in shrinks per port. The offsetting case, that co-packaged optics converts copper intra-rack links to optical and multiplies total port count, is real and currently unproven.
• The switch-silicon monopolist has entered the laser layer. Broadcom’s stated optical portfolio now spans optical DSPs, silicon photonics, co-packaged optics, EMLs, VCSELs and continuous-wave lasers, funded from a business targeting roughly 66% non-GAAP operating margin. A competitor able to price lasers as an attachment to the switch it already sells is a structurally different opponent from Coherent, and the constraint that protects Lumentum — fab capacity — is one Broadcom can buy.
• The export-control investigation is an operational exposure, not a financial one. Lumentum made a voluntary self-disclosure to the Bureau of Industry and Security in December 2023, supplemented in April 2024, and received a BIS administrative subpoena and a related Department of Justice subpoena in August 2024. Both remain open; the company states it is “unable to predict the likely outcome” and carries no accrual. The named potential consequences include “denial of export privileges.” For a company that already lists export licensing among its capacity bottlenecks and ships 79.2% of revenue outside the United States, that is a supply-chain risk with a legal trigger, and it compounds with the China substrate restrictions rather than sitting beside them.
• The convertible overhang competes with the capacity programme for the same cash. All four note series were convertible at holder option entering FY2027 and are classified as current at $1,596.9m; principal must be settled in cash. By 14 August 2026 the company had received early conversion requests for $757.8m of principal. New language appears in the FY2026 risk factors that is absent from FY2025: if holders convert a significant portion within a short period, liquidity “could adversely impact our ability to continue as a going concern.” With $2,738.4m of cash and short-term investments this is a sequencing problem rather than a solvency one — but the same dollars are earmarked for a fab conversion and a capacity ramp.
• Disclosure moved backwards while the business became more complex. In the year revenue nearly doubled, the company collapsed two reportable segments into one, stopped disclosing segment profitability, stopped disclosing the end-market split, and continued to disclose no units, no ASPs, no utilization rate, no backlog figure, no customer names and no competitor names. The utilization claim that carries 54% of the margin improvement is therefore unverifiable from outside, and so is the durability of the mix shift that carries another 29%.

What the sources could not answer

• Unit volumes, average selling prices, wafer starts, capacity or factory utilization rate — not disclosed in any fiscal year, so the central margin attribution cannot be checked.
• The FY2026 end-market split and any segment-level profitability — removed with the single-segment reorganization.
• Backlog or remaining performance obligation in dollars — defined in the filings, expressly disclaimed as an indicator, never quantified.
• The identity of Customer A (26.6%) and Customer B (15.0%), and whether NVIDIA is either of them.
• The size and term of NVIDIA’s “multibillion purchase commitment and future capacity access rights,” which appears only in NVIDIA’s press release and in neither company’s securities filing.
• The counterparties, volumes and terms of the three-year take-or-pay agreements management referenced in August 2026.
• FY2027 capital expenditure guidance, and any dollar figure for the capacity programme beyond “hundreds of millions of dollars over the next several years” for Greensboro.
• Quantified tariff cost, and quantified revenue lost to the Huawei restriction.
• Whether the Greensboro site is 148,000 square feet (FY2026 10-K, Item 2) or 240,000 square feet (Lumentum press release, 26 March 2026) — two of the company’s own primary sources conflict and neither states the basis.
• A published annual ASP erosion rate for optical transceivers — no reliable public figure was found, only qualitative confirmation from Cignal AI that price decline is the industry’s normal state.

What would need resolving before forming a thesis

• How FY2026 gross margin divides between volume-driven utilization, which persists while volume does, and scarcity pricing, which does not. Management gave the utilization share; it did not separate price.
• Whether the three-year take-or-pay agreements cover enough of the capital programme to change the shape of a downturn, or are too small to matter.
• Whether Coherent begins selling EMLs on the merchant market, which would remove the single clearest structural advantage Lumentum currently holds.
• Revenue per port under co-packaged optics against pluggable modules, which determines whether Lumentum’s participation in the next architecture is worth as much as its participation in this one.

8. Investor Takeaways

• What this business really is: a compound-semiconductor manufacturer that spent a decade assembling indium-phosphide fab capacity through acquisitions — Oclaro, NeoPhotonics, Cloud Light — and arrived at the moment that capacity became the scarcest input in the AI build-out.
• The core economic engine: laser die produced in owned fabs and sold both as merchant chips and inside its own modules. Roughly 54% of the FY2026 gross-margin improvement came from filling those fabs, which is operating leverage rather than price.
• The main growth lever: EML and high-power laser volume into AI clusters, running more than 30% behind demand as of May 2026, with unit output targeted to grow more than 50% year on year into December 2026 and a sixth fab producing from early 2028.
• What could break the story: the industry’s new capacity all lands in 2027–28, and the next architecture cuts laser die per port even where Lumentum participates. The predecessor company took its component gross margin from boom levels to 15–18% in four quarters when demand turned.
• What to monitor: non-GAAP gross margin against the 50.4% June-2026 quarter, top-two customer concentration against 41.6%, and whether Coherent starts selling EMLs on the merchant market.