How Lumentum ($LITE) Became AI's Most Critical Supplier

In a warehouse the size of an aircraft hangar, tens of thousands of graphics chips are training the next version of ChatGPT. Each chip can perform trillions of math operations per second. But every chip in the building has to finish each round before any chip can start the next one. The training job moves at the speed of the slowest handshake between them.

That handshake used to travel over copper, but it cannot anymore. At the data rates modern AI training requires, copper makes the signal get hot and noisy. The industry is replacing every copper link inside its newest data centers with pulses of laser light shooting through glass fiber. The company that makes most of that light is Lumentum Holdings, ticker LITE. The stock is up roughly 1,600% in twelve months. Nvidia put $2 billion into the company in March. The sell-side cannot agree on whether the fair price is $600 or $1,000.

Both numbers are defensible, but the more interesting question is not the price target but the physics underneath it.

Lumentum controls the physics bottleneck that every AI data center must pass through, and the bottleneck is getting tighter.

What AI Training Actually Does to a Network

A CPU (central processing unit) is the general-purpose brain of a computer. It handles instructions one at a time, or a small handful at a time. That worked for decades because most software is sequential: do step one, then step two, then step three. A GPU (graphics processing unit) was originally designed for video games, where the machine has to do the same simple math on millions of pixels at once.

AI training turns out to need the same thing. Training a model means multiplying enormous grids of numbers (called matrices) across billions of parameters, all at once. One GPU can do thousands of those operations in parallel where a CPU would do them one by one. That is why AI runs on GPUs.

A modern AI model is far too large to fit on a single GPU. The work gets split across hundreds or thousands of chips, each computing a piece of the answer. They cannot work independently. After every round of computation, every GPU has to share its results with every other GPU before the next round can begin. The full training job moves only as fast as the slowest exchange. Which means the network connecting the GPUs matters roughly as much as the GPUs themselves.

Each link between GPUs has to carry an enormous amount of data. The industry measures this in gigabits per second, or Gbps. Today’s lanes run at 100 to 200 Gbps and get bundled together. An 800G link uses four lanes at 200G each, which is roughly the equivalent of transferring a hundred feature-length movies every second. The next generation, 1.6T, doubles that with eight lanes.

Nvidia’s current racks hold 72 GPUs, the next generation holds 576, and planned clusters stretch past 100,000. Every doubling in GPU count multiplies the total traffic the network has to carry.

Copper has moved data around inside computers for thirty years, and at moderate speeds it is fine. But above roughly 100 Gbps per lane, electrical signals on copper start to degrade within inches. Signal regenerators can stretch it a bit, but they burn power. Past ten meters at 200 Gbps, no amount of power keeps copper alive.

A photon (a particle of light) traveling through a glass fiber does not heat the medium or crosstalk with photons in adjacent fibers. Optical links today consume an order of magnitude less power than copper at the same speed. The further the signal has to travel and the faster it has to go, the larger optical’s advantage gets.

Inside a data center, light is replacing copper at three scales. Within a compute tray (the shelf that holds GPUs, distances of a few inches), the wiring was copper and is going optical now. Rack to rack (tens to hundreds of feet across the data center floor) has been optical for years and is the layer growing fastest in absolute dollar terms. Within a rack (the refrigerator-sized cabinet that holds multiple trays, distances of a few feet), the wiring is still copper today, but will start going optical in the next generation. Each transition adds lasers, receivers, and switches. Every signal that used to be a voltage on a wire becomes a pulse of laser light inside a glass fiber.

The physics in one table

Power consumption per bit transmitted is the single cleanest way to compare interconnect technologies, because it captures distance, speed, and efficiency in one figure. Copper carries a clear penalty that grows with every speed doubling. Optical starts flat and keeps dropping as the industry moves from pluggable modules to co-packaged optics.

The Only Material That Can Do This

Making a laser requires a semiconductor material with a very specific property. When you run current through it, it has to emit light efficiently. This property is called a direct bandgap.

Silicon, the material that makes virtually every modern chip, is an indirect bandgap material. It is terrible at turning electricity into light. For AI-class speeds (200G per lane today, 400G per lane coming), the only material that works is indium phosphide. InP is a compound of two elements, indium (a soft, rare metal) and phosphorus, fused into a single crystal structure.

There is a competing approach that tries to get around InP using silicon, but in the end every silicon photonics device ends up with an InP laser bonded on top. You do not escape the InP dependency, you can only relocate it. At the next speed tier, 3.2 terabits per second using 400G per lane, silicon photonics breaks down entirely because silicon’s material properties cannot handle the modulation speeds. Intel tried to make silicon photonics the dominant industry path and eventually sold its transceiver business to Coherent. That sale is as clean a signal as the market gives.

The issue is that InP has chokepoints at every layer of its supply. Raw indium metal is a byproduct of zinc mining, and roughly 70% of the world’s refined supply comes from China. In February 2025, China added indium to its export licensing list, which means every shipment now needs a government permit. The United States imports 100% of its refined indium.

The metal then has to be turned into polished crystalline discs of indium phosphide, a few inches across, that serve as the foundation for laser chips. Three companies make nearly all of them. Sumitomo Electric in Japan holds 50% to 60% of the market. AXT, through its Tongmei subsidiary in China, holds 25% to 35%. JX Nippon Mining is the smaller third. AXT alone has a backlog of more than $60 million, with customer forecasts stretching past 2030. But the deepest bottleneck is not the wafers themselves. It is what happens next.

From Wafer to Laser: The Manufacturing Bottleneck

A raw InP substrate wafer is just a blank disc. To turn it into a functional laser, the manufacturer has to deposit ultra-thin layers of crystalline material on top of it, one atomic layer at a time, lined up precisely with the crystal structure below.

A working InP laser has 20 to 40 such layers, each with a specific chemical composition that determines how the finished chip handles light. This process is called epitaxial growth, or epi for short. It happens inside Lumentum’s and Coherent’s own facilities, and it is currently the binding constraint in the entire optical supply chain.

You can import substrate wafers faster than you can build the cleanroom capacity to grow epi layers on them.

Epi growth happens inside a device called an MOCVD reactor, short for Metal-Organic Chemical Vapor Deposition. Picture a sealed chamber with the wafer inside, heated to 600 to 800 degrees Celsius. Carefully controlled gases (compounds of indium, phosphorus, and other elements) flow in and deposit atoms on the wafer surface in the exact crystal pattern the engineer wants, one layer at a time. Each reactor costs several million dollars. Each wafer takes hours to process.

Only two companies make these machines at scale. Aixtron, a German company, has roughly 90% of the market for optoelectronics. Veeco, an American company, won its first production InP contract in March 2026. Lead times for a new reactor run 12 to 18 months. Both companies are sold out for 2026 and taking orders into 2027.

Even once a new reactor is installed, qualifying the epi process for production takes one to two years. Laser chip customers, Lumentum included, have to prove that every layer is within tolerance, that the finished lasers hit reliability specs, and that the output is consistent wafer after wafer. Aixtron’s management calls laser tool qualification “the most sticky and the most difficult to re-qualify” of all its product segments.

Put the numbers together and you get a three-to-five-year timeline from order to qualified laser production. Twelve to eighteen months to build and install the reactor, plus one to two years to qualify the epi process, plus another one to two years for the end customer to qualify the finished chip. A new competitor that started spending today would reach scale somewhere around 2029.

Global demand for InP devices is estimated at two million pieces in 2025 but production capacity is roughly six hundred thousand. That is a 70% gap between what the market wants and what the industry can make. Lumentum is undershipping customer demand by approximately 30%, and the gap is widening. The company is sold out through 2028.

What does Lumentum Sell?

Lumentum is vertically integrated from wafer to laser and chip production. The company buys raw InP wafers, grows the epi layers itself, fabricates the chips, packages them into finished modules, and ships to customers like Nvidia and Google. Broadly, their products can be divided into 6 categories: EML, OCS, UHP Lasers, Transceivers, Telecom/deep sea cables, and Industrial Tech/other. 

EMLs (Electro-absorption Modulated Lasers)

An EML is a tiny chip, a few millimeters across, that turns electrical data into pulses of light. It contains two parts built out of the same piece of indium phosphide. A laser section generates a continuous beam of infrared light, and a modulator section rapidly blocks and unblocks that beam to encode ones and zeros. The modulator can switch the beam on and off hundreds of millions of times per second.

EMLs are the workhorse component inside optical transceivers, the pluggable modules that sit on the front panel of data center switches and convert electrical signals into light signals that travel through fiber. An 800G transceiver contains four EML chips, each running at 100G per lane. A 1.6T transceiver contains eight. Every speed doubling means more EMLs per link.

Lumentum holds roughly 50% share of high-speed EMLs per Bank of America, citing LightCounting and Dell’Oro. The company is undershipping demand by about 30% and is sold out through 2028. BofA projects total Datacom laser revenue of $2.17 billion in CY2027.

OCS (Optical Circuit Switches)

Traditional data center networking converts light to electricity at every switch. A light signal arrives through a fiber, the switch chip reads the destination address using electronics, then converts the signal back to light and sends it out the correct fiber. Every conversion wastes power and adds delay.

An Optical Circuit Switch skips all of that. It routes light signals directly from one fiber to another without ever converting to electricity. Lumentum’s OCS uses a technology called MEMS, short for Micro-Electro-Mechanical Systems, which is a grid of hundreds of tiny silicon mirrors, each a few hundred microns across (roughly the width of two human hairs), suspended on microscopic hinges etched from the same silicon wafer. When voltage is applied, a mirror tilts by a precise angle. Light coming in from one fiber hits the mirror, bounces off, and lands on a different output fiber. 

Google built MEMS-based OCS in-house for ten years to manage traffic in its AI training clusters. It is now outsourcing a large part of that program to Lumentum. Citi calls Lumentum “heir to Google’s internal OCS program.” Current share is roughly 50% of OCS deployments (Citi), projected to reach 59% of a $4.33 billion total addressable market by 2029 (BofA). Backlog exceeds $400 million. Citi forecasts OCS revenue of $1.1 billion in FY2027 and $3.4 billion in FY2028.

UHP (Ultra-High-power Pump) Lasers (400mW+)

A pump laser is a high-power InP laser that boosts the strength of an optical signal, like an amplifier for light. The technology was originally developed for undersea fiber optic cables, where signals have to travel thousands of miles and need periodic boosting. Those same lasers now feed into optical engines for co-packaged optics and AI networking equipment. The 400 milliwatt tier requires extreme precision in the epi process and years of reliability testing.

This is the closest Lumentum comes to a monopoly. Per Citi in March 2026, Lumentum holds roughly 80% share, Coherent holds 20%, and nobody else has started the 15-month qualification process required to compete. Even if a challenger began today, Lumentum would hold its lead for more than a year. Each pump laser carries an average selling price of $50 or more, roughly four times the price of a standard 200G EML.

The biggest demand driver for these lasers is Co-Packaged Optics, or CPO. Lumentum does not build CPO engines, but CPO is a major driver of their future growth because of what goes inside them. The engine is the package that integrates the lasers with the switch chip, and the competition to build them involves Broadcom, Marvell, TSMC with Ayar Labs, and Coherent.

In a traditional data center, optical transceivers plug into the front panel of a switch, connected to the chip by several inches of copper on a circuit board. At high speeds, copper hits its limit. CPO fixes the problem by mounting the optical components directly onto the same package, putting the light source millimeters from the processor instead of inches away.

Lumentum builds the ultra-high-power lasers that sit inside every one of those engines. They are the picks-and-shovels play on CPO, and as the demand for CPO increases, it becomes the next phase of growth.

Nvidia announced CPO-based switches in March 2026 with Lumentum as a named laser supplier, backed by a $2 billion strategic investment. BofA projects CPO-related laser revenue of $483 million in CY2027, expanding roughly 240% to $1.63 billion in CY2028.

Transceivers (Cloud Light)

Every fiber link in a data center needs a transceiver at each end: one to turn the outgoing signal into pulses of light, and one at the far side to read those pulses and turn them back into electricity that the next switch chip can process. They’re roughly the size of a candy bar and inside sit four EMLs (which generate and modulate the light), matching photodetectors (which read incoming light), and a small circuit board that handles the electrical interface to the switch. 

Lumentum has historically been the chip supplier, selling EMLs into other vendors' transceivers. In November 2023, the company bought Cloud Light, a Hong Kong-based transceiver maker, for $705 million in cash. The acquisition moved Lumentum up the stack: instead of only selling chips into other vendors' transceivers, Lumentum now sells finished transceivers alongside them. Specifically, these are the pluggable modules that clip into the front panel of data center switches, so the module can be replaced or upgraded without opening up the switch.

Cloud Light is roughly $719 million of Lumentum’s FY26 revenue, about 25% of the company. Morgan Stanley projects $1.1 billion in FY27, and BofA models the finished transceiver business at $1.6 billion in CY27 growing to $2 billion in CY28. Backlog extends more than a year. 

Management previously self-imposed a $1 billion annual ceiling on the business because transceivers drag on corporate margin, but walked that cap back during the Q2 FY26 earnings call as volumes accelerated. BofA estimates Lumentum now holds about 3% share of the 800G finished transceiver market and will reach 8% in the next generation 1.6T devices. Microsoft, Google, and Meta are all shipping customers, and a fourth hyperscaler is committed for calendar 2026.

However, margins remain a point of tension. Cloud Light runs at roughly 31% to 32% gross margin against Lumentum’s corporate target in the low 40s. The drag comes from manufacturing execution, CW laser sourcing (Cloud Light currently buys continuous-wave lasers from third-party foundries while Lumentum directs its own InP capacity to higher-margin external EML sales), and the tail end of the 800G transition. 

CFO Wajid Ali said “1.6T margins are significantly better than 800G” on the Q2 FY26 call. As the mix shifts to 1.6T and Lumentum integrates its own CW lasers into the transceivers, the margin profile improves, even if the business remains dilutive to the corporate blend.

The legacy business

The rest of the business splits into two buckets. Industrial Tech (laser processing, 3D sensing, precision optics for manufacturing) was $234 million of FY25 revenue, 14.6% of the company, and it is shrinking. Consumer 3D sensing is on its way to under 5% of revenue. These are legacy businesses Lumentum inherited and does not prioritize.

The subsea, telecom, and data center interconnect business is a different story. The subsea pump laser business is up 90% year over year. Data center interconnect (DCI) is the same kind of long-haul optics but on land, and sales have grown 70%+ year over year for eight quarters straight. This is highly tied to the hyperscaler story as the subsea cables connect their regions and DCI connects their campuses. This growth sits inside the Cloud and Networking segment alongside the AI products, which makes a clean dollar breakout hard to pull.

Revenue and profit mix

LITE does not publish revenue or gross margin by product line. The 10-K shows two segments (Cloud and Networking, Industrial Tech) and nothing below that. What follows is constructed from sell-side broker research with estimates rather than reported figures.

Profit breakdowns were generated by multiplying revenue estimates by profit margins. The vertically integrated component businesses (EML chips and ultra-high-power lasers) run at roughly 55% gross margin because Lumentum owns the fab and sells the raw silicon. Finished transceivers run at roughly 30% because the assembly step adds cost without adding much differentiation. As CPO ramps, the product mix shifts further toward the high-margin components and away from the lower-margin finished systems, which means the CPO transition is accretive to margins before it’s even fully accretive to volume.

MEMS vs. Liquid Crystal: The OCS Bet

There are two ways to avoid converting light to electricity: MEMS, with its tiny rotating mirrors, and liquid crystal. And while liquid crystal may sound like the more high-tech on its face, the market is settling on MEMS.

Liquid crystal uses a cell filled with the same material that powers an LCD television, which bends light when voltage is applied and has no moving parts. But the reason MEMS is winning comes down to light loss. Every switch loses some of the light passing through it, and MEMS loses far less than liquid crystal does. At hyperscale, where a buyer might deploy thousands of these switches, that cost compounds quickly. Google placed its MEMS bet ten years ago, and the reliability record since has been clean.

Coherent makes a sincere counter-argument: no moving parts means no mechanical wear. Coherent’s CTO Julie Eng put it plainly in March 2026: “Usually you don’t want something that moves inside the data center if you can help it.” CEO Jim Anderson added: “I believe that the better technology is one that’s non-mechanical.”

But the market’s verdict so far is clear. MEMS holds over 70% of the optical switching market by deployment. Coherent is the sole liquid crystal OCS supplier. Morgan Stanley notes that “investors have given Lumentum more credit than Coherent when it comes to OCS.” Google’s decision to outsource its internal MEMS program to Lumentum rather than migrate to liquid crystal is the strongest signal available.

MEMS vs liquid crystal at a glance

The Numbers That Prove Scarcity

There is one obvious narrative that has driven LITE stock up 1,600% over the past twelve months: data center demand. Nvidia made a $2 billion strategic investment in March 2026 because the consensus is steady on the need for more and more data centers to power the AI wave.

The five large cloud and AI buyers (Microsoft, Google, Amazon, Meta, and Oracle) are on track to spend roughly $675 billion on AI infrastructure in 2026, up 63% year over year. BofA’s capex tracker models $748 billion in CY2026, growing to $869 billion in CY2027. Every one of those dollars fuels the optical component boom.

On EMLs, Lumentum holds roughly 50% share at high speed per Bank of America, citing LightCounting and Dell’Oro. The company is undershipping demand by 30% and sold out through 2028. BofA projects 200G EML revenue of $1.14 billion in CY2027 and $1.84 billion in CY2028.

On OCS, Lumentum holds roughly 50% of deployments today per Citi, rising to 59% of a $4.33 billion TAM by 2029 per BofA. Citi forecasts OCS revenue of $1.1 billion in FY27 and $3.4 billion in FY28. Backlog exceeds $400 million.

On ultra-high-power pump lasers, Lumentum holds approximately 80% share and Coherent holds 20%, per Citi, with a 15-month qualification moat against every other challenger.

On CPO, BofA projects $483 million in CY2027 and $1.63 billion in CY2028. Citi projects 4 million ultra-high-power laser units in FY2027. Lumentum supplies lasers into every competing engine architecture, which means the company wins regardless of which engine vendor ends up on top.

Taken together, BofA projects total revenue of $7.61 billion in CY2027 and $10.55 billion in CY2028. Morgan Stanley projects $4.98 billion in FY27 and $7.73 billion in FY28. The emerging AI businesses, which are currently about 30% of sales, are modeled to exceed 60% of the mix as the company breaks through a $2 billion quarterly run rate.

Street targets span from $595 (Morgan Stanley bear) to $1,040 (BNP bull), with BofA at $775, Citi at $800, and JPMorgan at $950. The stock trades at 85x forward P/E, which means investors are paying $85 today for every dollar of next year’s expected earnings, and 12x forward sales. It is priced for execution.

The person expected to deliver that execution is CEO Michael Hurlston. He joined in February 2025 from Synaptics. He was previously CEO of Finisar, the InP laser business now inside Coherent, so in his current job he competes against the operation he built. In his first year at Lumentum he expanded operating margins from roughly 10% to over 30%. The sell-side is unanimously bullish.

The Difference Between Now and the 2000–2001 Fiber Optics Story

We have seen a fiber optic buildout before. How did it end?

The shape rhymes in four places: supply is concentrated, demand is hyper-cyclical, multiples have expanded, and a lot of hot money is chasing the trade. JDS Uniphase was the iconic optical company of the dot-com era, the dominant maker of fiber optic components for the telecom buildout. The stock peaked at $153 in 2000 and bottomed at $1.58. In 2015 the company split into two parts. Viavi Solutions took the network test and measurement business, and appeared in our earlier PQC piece. Lumentum Holdings took the optical components business, the lasers and photonics that are the subject of this piece. The corporate DNA is as similar as the underlying physics, but now we’re in a new era.

The demand profile is different. The 1998 to 2001 fiber buildout was speculative. Telecom operators like MCI, Global Crossing, and 360networks laid fiber ahead of a demand curve that never showed up. All of them went bankrupt.

This cycle’s end customers are Microsoft, Google, Amazon, Meta, and Oracle, companies with more than $3 trillion in combined market capitalization and tens of billions in annual free cash flow. The capex is funding AI workloads already in production and already generating revenue, not speculative capacity ahead of a dream.

The revenue model is different, too. Fiber optic cable is infrastructure you install once and use for decades. Optical components are equipment that depreciates and gets replaced every few years as speed standards advance. That makes optical component revenue recurring instead of one-shot.

We do not have to go all the way back to 2001 for the instructive precedent. In 2022 and 2023, hyperscalers briefly paused their capex expansion, flattening growth to roughly 4% after it had doubled between 2019 and 2022. The optical supply chain got hit hard. Lumentum’s total revenue dropped 23% year over year from FY2023 to FY2024. The Cloud and Networking segment fell 18%, shedding $238 million. The company took $20.7 million in excess capacity charges as factory utilization collapsed. Coherent saw inventory days spike past 160 and enacted $100 to $125 million of restructuring. The bullwhip effect, in which customers who had stockpiled during the shortage suddenly stopped ordering, hit merchant optical suppliers harder than proportionally. That is the risk baked into a capex-driven business with high fixed costs and a customer base that can turn on a dime.

What Breaks and What Compounds

At 85x forward earnings, the debate over Lumentum has moved past physics. Copper is losing to light, that is established. The open question is whether Lumentum’s revenue can keep growing even if hyperscaler spending slows. Which force is stronger: the potential slowing of data center demand, or the growth in optical content within each data center?

The content multiplier

Nvidia’s current Blackwell NVL72 rack holds 72 GPUs connected mostly over a copper backplane. The next generation, Rubin Ultra NVL576, holds 576 GPUs and runs past copper’s limits, forcing a shift to an optical mesh. The generation after that, Feynman NVL1152, holds 1,152 GPUs and requires practically every link inside the rack to be optical. The physics simply will not let copper carry that much data over those distances.

The multiplier translates to real dollars. At NVL72 today, Lumentum barely has content inside the rack because the GPUs talk over copper. Once CPO deploys at the NVSwitch level, Citi estimates optical content per GPU jumps to $500 to $1,000. A single External Laser Source (ELS) module packages eight ultra-high-power lasers at roughly $120 each, which adds another $960 per module on top. The laser content per rack can scale up to four times depending on the architecture.

The optical links between GPUs inside a rack are 9-16x larger than the rack-to-rack links Lumentum already sells into today. Add the generational jump from Rubin to Feynman and Citi models the addressable optical content per rack multiplying by ten. At ten times the content per rack, the architecture inside each rack ends up mattering more than how many new racks get built.

What the math says about a slowdown

If AI cluster deployments slow by 20 to 30%, the ten-times content jump per rack absorbs the volume hit and Lumentum still grows through the transition. On top of that, the company is already undershipping demand by 25 to 30% across its InP portfolio, which gives it a backlog cushion before any slowdown would touch reported revenue.

The sell-side has modeled the downside. Morgan Stanley’s bear case is $18 EPS at 20x, landing at a $360 stock price, if EML supply and demand hit equilibrium faster than expected and the CPO and OCS ramps slip. Citi’s bear case is $17 EPS at 30x, landing at $500, if Lumentum misses margin targets when aggregate volumes collapse. Citi raised that bear-case floor from $300 to $500 earlier this quarter, which says even the worst-case estimate has been rewritten upward.

Even that bear case is elevated by historical standards. A $17 to $18 EPS floor under a 20% capex contraction is dramatically higher than anything Lumentum earned in the old telecom cycle. The physics that forces optical integration does not go away in a downturn, which puts a floor under revenue even when spending slows.

Downside risks the content multiplier does not offset

Even when end demand holds up, inventory dynamics can crush merchant supplier revenue for two to four quarters. Customers stockpile during a shortage, then freeze orders while they burn through what they already have. Lumentum’s 23% revenue decline in FY2024 happened during a capex pause, not a capex collapse. A real downturn would be worse.

Silicon photonics might beat the physics assumption. Lumentum’s thesis rests on InP being the only material that works at 400G per lane. If silicon photonics figures out the higher speed tiers, and Tower Semiconductor and GlobalFoundries are ramping their silicon photonics capacity roughly five-fold, InP’s scarcity premium erodes. Management projects 80% EML share holds at 3.2T. 

Coherent could close the gap. Coherent is moving to six-inch InP wafers against the industry’s four-inch standard, which yields roughly four times more chips per wafer at lower cost. If Coherent hits high yields on six-inch first, it can undercut Lumentum on price. Nvidia invested $2 billion in Coherent too, hedging its supplier risk.

China might disrupt indium supplies or replace the supply chain. 70% of refined indium flows through China, and export permits currently take about sixty business days to clear. Separately, Lightelligence in Shanghai ($1.1 billion valuation, backed by Tencent and Baidu) started mass-producing OCS modules in December 2025. If Chinese hyperscalers default to domestic suppliers, Lumentum loses a chunk of addressable market that would be hard to win back.

LITE is effectively a one-end-market company. Industrial tech is down to 14.6% of FY2025 sales and shrinking, and consumer 3D sensing is on its way to under 5% of revenue. Telecom and subsea is growing fast, but that growth is adjacent to data center demand rather than independent of it. There is no meaningful non-AI-optical buffer, and if the AI cycle pauses there is no other business to fall back on.

Upside forces

Optical scale-up is likely to eventually replace copper inside the rack. Lumentum’s management calls it the company’s fourth growth driver. As GPU clusters grow from 72 to 576 to 1,152, copper backplanes run out of reach and every intra-rack link goes optical. There's nine to sixteen times more optical content inside a rack than between racks. The first shipments would land in late CY2027, and that is the force that makes the ten-times content multiplier real.

CPO could pull forward higher-margin content. Every switch that moves from pluggable transceivers to co-packaged optics raises Lumentum’s dollar content per port and tilts the product mix toward pump lasers at $50-plus ASPs. CPO also saves enough power per port that hyperscalers cannot economically defer the transition.

However, as CPO ramps in 2027 and beyond, it gradually replaces the pluggable transceivers that Cloud Light makes. This could be seen as a cannibalization risk, but Lumentum is the laser supplier into every major CPO engine. Revenue that eventually leaves Cloud Light flows back into ultra-high-power pump lasers at higher margins. The product mix inside the company shifts from finished systems to components, and they continue to make money from the higher-margin transceivers that remain in demand during that transition period.

The refresh cycle also makes revenue recurring. As speed standards climb from 800G to 1.6T to 3.2T, hyperscalers have to replace transceivers and upgrade switch optics every two to three years. That ties Lumentum’s revenue to the upgrade cycle rather than a one-time buildout.

Long-term agreements may yet lock in volume. Lumentum has shifted from spot sales to multi-year, take-or-pay contracts worth billions, which changes the revenue profile entirely. Management says the agreements have killed the quarterly price-down negotiation, with prices “holding or increasing.” That is a very different business than the transceiver commodity cycle of 2022 and 2023.

What to watch

FQ3 2026 earnings in mid-May should bring updated OCS backlog, CY2027 guidance, and the first CPO shipment timeline.

Google’s CPO deployment decisions in the back half of 2026 will show whether the search giant keeps outsourcing to Lumentum or builds an alternative.

OFC 2027 next March will refresh the industry roadmap for 3.2T and scale-up CPO.

Coherent’s six-inch InP yield data is the single most important competitive variable to track. Any move in hyperscaler capex guidance, especially from Microsoft or Google, will move the stock. Inventory days at Lumentum and Coherent are the canary in the coal mine, the first number to move if customers get over-stocked and start freezing orders.

The bottom line

The historical rhyme with JDS Uniphase and the early 2000’s fiber optics boom can’t be ignored. Supply is concentrated, demand is hyper-cyclical, the multiple has expanded, and hot money is chasing the trade. 

However, the end customer this time is internal hyperscaler capex backed by live cash flows, not a speculative telecom buildout chasing a demand curve that never materialized. The revenue recurs through the upgrade cycle instead of landing once. The bottleneck sits deeper in the stack, closer to the raw material, which makes it harder to route around.

The stock is priced for execution, and at 85x forward earnings any miss will sting. Execution is exactly what CEO Michael Hurlston has delivered since February 2025: operating margins tripled in his first year, and backlog stretches through 2028 on products that are already sold out. The physics argument that underpins the whole bull case is the kind of argument that does not get relitigated once it is settled. Light is replacing copper inside the biggest infrastructure buildout in the history of computing, and Lumentum makes the light.

The question for the next twelve months is not whether any of that holds, but how much of the future the stock has already priced in.

Want to build your own thesis? Launch Terminal X