Snapshot — the group at a glance

An optical transceiver is the small plug at each end of a fibre-optic cable that turns electrical signals into pulses of laser light (and back again) so data can travel down the glass at very high speed. Inside a modern AI data centre, the thousands of GPUs (the chips that do the actual AI computation) have to talk to each other constantly while they train and run a model, and they do it over fibre — which means every link needs a transceiver at both ends. The unit this group sells is one transceiver module, sold in generations named by their speed: 400G → 800G → 1.6T ("G" = gigabits per second, "T" = terabits per second; each generation roughly doubles the data per port). The companies here either build the finished module (Coherent, Lumentum, Applied Optoelectronics), build the laser chips and optical parts that go inside it (Lumentum, Coherent again), or assemble it under contract for others (Fabrinet). This is the "networking" layer of the AI/AGI build-out: the plumbing that lets a cluster of chips behave like one giant computer.

Demand currently runs ahead of supply at the leading edge: 800G modules have been on allocation (rationed to customers) since 2025, lead times stretch to 30-50 weeks, and the 800G-to-1.6T jump is being pulled forward by AI buildouts. What limits supply is not the module assembly but the laser chips inside it — high-power EMLs (electro-absorption modulated lasers — the chips that drive a fibre link) and silicon-photonics engines — plus the 18-24 months it takes to bring new capacity online. The Price-of-exposure section below sets out, in neutral arithmetic, what an owner pays today per dollar of each company's revenue; the reader draws the conclusion.

Source: 500-stocks scan — Networking & Connectivity, sub-sector 1 "Optical Transceivers & Components"; market-size figure is general-knowledge est., not live-verified.

The product & how money is made

The product is a transceiver module: a sealed package, often the size of a thick USB stick, that plugs into a switch or a server's network port. Inside it are a tiny laser that fires light pulses, a photodetector that reads incoming pulses, and the optics and electronics that shape and decode the signal. The fibre cable carries the light; the transceiver is the translator at each end. Because every fibre link needs one at both ends, transceivers are sold by the millions of units, and the count scales directly with how many connections a network has.

Money is made in three layers, and different companies in this group sit at different layers:

• Selling the finished module (price per unit × volume). A finished 800G module sells for roughly $500-$1,500 est. depending on type and reach; 1.6T modules sell for more. The maker's profit is the gap between that price and the cost of the parts and labour inside. Newer, faster generations have carried higher prices because they are harder to make and have been in shorter supply, so the first makers to a new generation have historically earned wider margins early, before competing supply arrives.
• Selling the parts inside the module (lasers and optical chips). The single most valuable component is the laser chip — especially the high-power EML and the silicon-photonics engine. A company that makes these chips can sell them to many module assemblers, so it captures value across the whole industry rather than just its own modules. This is the scarce, higher-margin layer.
• Assembling modules under contract (a fee for manufacturing). Some firms do not own the design or the brand; they build modules for other companies and earn a manufacturing margin. This is steadier and lower-margin — closer to renting a skilled factory than owning the scarce technology.

The key money word for an owner is gross margin (revenue left after the direct cost of making each unit). Owning the scarce laser/optics technology has tended to earn the highest gross margin; assembling finished modules from bought-in parts the thinnest. After that comes capex (capital expenditure — cash spent on factories and equipment) and finally free cash flow (the cash actually left for the owner). This group is less capital-heavy than chip foundries, but ramping laser-chip capacity still takes real money and 18-24 months of lead time.

Source: 500-stocks scan sub-sector 1 (component list: lasers/EMLs/VCSELs, photodetectors, modulators, silicon photonics; 18-24 month capacity lag); unit prices and margin characterisation are general-knowledge est., not live-verified.

Demand — how much the world will want this

Demand is driven almost entirely by AI data centres, and the scan is blunt about the mechanism: "Every GPU in a training cluster needs multiple high-speed optical links. A 100k-GPU cluster can require 200k+ transceivers." Because the number of links grows faster than the number of chips (each GPU connects to several switches), transceiver demand scales super-linearly with cluster size — bigger clusters need disproportionately more optics.

Current demand (known facts from the scan): the AI segment alone is "growing 30-40% annually," 800G modules have been "in allocation since 2025" (meaning customers are rationed because makers cannot fill every order), and "the 800G-to-1.6T transition is being pulled forward 12-18 months by hyperscaler AI buildouts" (hyperscalers = the largest cloud operators, e.g. Microsoft, Amazon, Google, Meta). In plain terms, the buyers want the next generation faster than the industry planned to deliver it.

Forward demand (forecast — given AGI is arriving): if recursive self-improvement and broad AI deployment proceed, the demand driver is not one cluster cycle but a sustained, compounding need to wire ever-larger groups of chips together. Each new frontier-model training run, each new inference fleet (the fleet of chips that serve a finished model to users), and the coming wave of physical AI (robots, autonomous systems running large models) all require more chips talking to each other — and therefore more transceivers. As clusters scale from ~100k toward 1M+ GPUs, and as speeds climb 800G → 1.6T → 3.2T, both the unit count and the price per unit can rise together. Industry forecasts commonly put AI optical-transceiver demand growing at high-double-digit percentages per year for several years est. These are forecasts, not contracted orders; the binding question for an owner is whether supply can keep up, addressed next.

Who the buyers are: a concentrated set of very large, very well-funded customers — the hyperscalers and the big AI-cluster builders, mostly buying through the switch and system vendors (Nvidia, Arista, Cisco) who integrate the optics. That concentration makes near-term demand visible, but it also means a slowdown in AI capex by a handful of giant buyers would be felt across this whole group quickly.

Source: 500-stocks scan sub-sector 1 ("200k+ transceivers," "30-40% annually," "in allocation since 2025," "pulled forward 12-18 months"); forward growth figures are general-knowledge est. and forecasts.

Supply — how much can be made, and what limits it

Supply is the constraint that defines this group right now. The scan answers "Supply Constrained?" with a flat "Yes," and names the choke points: "Laser chip supply (especially high-power EMLs and silicon photonics co-packaged optics) is the bottleneck. Lead times stretched to 30-50 weeks. Capacity expansions take 18-24 months to come online."

Current capacity & expansion (known/announced): the finished-module side can be scaled relatively fast — assembly lines can be added — but the laser chips and silicon-photonics engines that go inside cannot. Those are made in specialised semiconductor fabs (chip factories) by a small number of companies (Coherent and Lumentum are among the leading Western laser-chip suppliers; several Asian makers also produce). Players are expanding, but a new laser-chip line takes 18-24 months and deep process know-how to bring to acceptable yield (the share of chips that come out working). That lag is why 800G has been rationed even as demand surged.

The bottleneck is a stack of hard-to-replicate inputs, in roughly this order:

• High-power laser chips (EMLs). The lasers that drive 800G/1.6T links are the single tightest input named in the scan. Making them at high volume and good yield is a specialised, capacity-limited process.
• Silicon photonics / co-packaged optics (CPO — moving the optics right next to the switch chip). This next architecture promises lower power and higher density but is harder to manufacture, and capacity for it is still being built — another gating input rather than a solved one.
• Capacity lead time. 18-24 months to add a meaningful new line means supply physically cannot respond to a demand spike inside a year. Laser-chip capacity cannot be rushed.
• Know-how / yield. Getting a high fraction of working laser chips and well-aligned modules is an accumulated skill. Being able to fund a line is necessary but not sufficient, which is one reason incumbents with proven processes (Coherent, Lumentum) have an established position relative to new entrants.

Market-share structure (who controls supply): the finished-module market is more fragmented than chip foundries — it includes Western makers (Coherent, Lumentum, Applied Optoelectronics) and several large Chinese module makers (such as Innolight and Eoptolink, not US-listed) that are reported to hold significant global module share est. The laser-chip layer underneath is more concentrated, and that is where the named bottleneck sits. So "who controls AI-relevant supply" depends on which layer you mean: modules are contested, the scarce laser/optics components less so.

Source: 500-stocks scan sub-sector 1 (EML/silicon-photonics bottleneck, 30-50 week lead times, 18-24 month capacity lag); module-share and Chinese-maker characterisation are general-knowledge est., not live-verified.

The gap — demand vs supply

Putting the two together: at the leading edge of speed (800G now, 1.6T arriving) the product has been structurally short — more buyers than the industry can supply — since 2025. The evidence in the scan is direct and consistent: 800G "in allocation since 2025," lead times of "30-50 weeks," the 1.6T transition "pulled forward 12-18 months," and the AI segment "growing 30-40% annually" while new capacity takes 18-24 months to arrive. Older, slower modules (400G and below) are far less tight — that capacity is ample and competitive, and prices there have tended to decline as volume matures.

When could it flip to oversupply? Two paths, both forecasts. (1) Supply catches up: makers are expanding laser-chip and module capacity; if much of it lands on schedule into a demand pause, 800G could loosen as 1.6T becomes the new tight node — and historically the optical-component industry has been cyclical, with past gluts when telecom demand rolled over. (2) Demand stalls: because the buyer base is a handful of hyperscalers, a pullback in AI capex would cut orders fast, and unlike a fab a module line can be idled quickly. A third, slower risk is a technology shift — if co-packaged optics or linear-drive designs change which components are needed, today's scarce part could become tomorrow's surplus. Under the AGI-arriving premise, sustained and growing compute demand argues against a near-term broad glut, but the leading-edge tightness rotates from one speed to the next, and the timing of any flip should be treated as unknown, not contracted.

Source: 500-stocks scan sub-sector 1 (allocation, lead times, 18-24 month lag); pricing arrows and oversupply timing are general-knowledge est. and forecasts.

The players — who captures the money

The main US-listed names fall into three buckets — the integrated module-plus-component makers, the contract assembler, and the small pure-play — plus one adjacent system vendor. Figures below are approximate and not live-verified.

How the layers map for a deep-dive: Coherent and Lumentum make the scarce laser/optics layer that the scan names as the constraint. Fabrinet rides the same volume as a manufacturer-for-hire that does not own the scarce IP. Applied Optoelectronics is a small pure-play with high revenue variance. Ciena is adjacent — an optical-systems company more than an in-cluster transceiver maker.

Source: 500-stocks scan sub-sector 1 company list (COHR, LITE, AAOI, CIEN) plus FN (Fabrinet) per task scope; revenue and market-cap figures are general-knowledge est., not live-verified.

The price of exposure

In plain money terms, what does an owner pay today for $1 of these companies' annual revenue, and how much owner cash does that revenue actually throw off? The two questions matter together, because a high-revenue contract assembler can hand the owner less profit per dollar of sales than a smaller company that owns the scarce technology. Price-to-sales (market value divided by annual revenue) is used below only as a rough, like-for-like ratio; all inputs are estimates, not live-verified.

• Coherent (COHR). At roughly $5-6B revenue and a ~$10-15B market value est., that is very roughly $2-3 of market value per $1 of revenue est. (price-to-sales ~2-3x). It makes the laser-chip layer, and it also carries meaningful debt from the II-VI/Coherent merger, so the relevant denominator is enterprise value (market value plus net debt), not just market cap.
• Lumentum (LITE). Roughly $1.5-2B revenue and a ~$5-9B market value est. works out to price-to-sales around 3-4x est. The arithmetic: the price-to-revenue ratio is higher here than at Coherent or Fabrinet on these estimates, while legacy telecom revenue is shrinking underneath the AI-optics growth.
• Fabrinet (FN). Roughly $3-4B revenue and a ~$8-12B market value est. implies price-to-sales near 2-3x est. — but as a contract manufacturer its gross margins are thinner than the module/component makers, so each dollar of revenue carries less profit. It is also capital-light and has been consistently cash-generative, which is the opposite trade-off from the IP makers.
• Applied Optoelectronics (AAOI). Small revenue ($0.3-0.6B) against a ~$1-2B market value est. can imply a higher price-to-sales (often 3-5x+ est.) on an expected data-centre ramp, against a company history of losses. Here the ratio is set against future revenue, not today's cash.

The money-in / money-out shape of the group: this is moderately capital-intensive — heavier than a pure software business, far lighter than a chip foundry. Money goes in as capacity expansion for laser chips and modules (18-24 months ahead of revenue) plus inventory to feed long order backlogs; money comes out as module and component sales when the lines fill. The owner-cash spread across the group is wide: the IP makers (Coherent, Lumentum) have earned the highest gross margins but reinvest heavily and, in Coherent's case, service debt; the contract assembler (Fabrinet) earns thin margins but is capital-light and steadily cash-positive; the small pure-play (Applied Optoelectronics) is priced on an expected ramp rather than current cash. The arithmetic the reader can sit with: on these estimates, a higher price-to-revenue ratio in this group corresponds either to the scarce laser/optics layer or to a forecast-growth story, and a lower ratio corresponds to the contract-manufacturing margin or the legacy-telecom decline. The reader draws any conclusion.

Source: revenue/market-cap and all multiples are general-knowledge est. (cutoff ~early 2026), not live-verified; capex/margin characterisation grounded in scan sub-sector 1 (capacity lead times) and well-known company structure.

What to deep-dive next

Where a company-level deep-dive would add the most factual detail:

• Makers of the named bottleneck: Coherent (COHR) and Lumentum (LITE). These two make the scarce laser-chip / optics layer that the scan names as the constraint. A deep-dive should verify the live AI/optics share of revenue, EML and silicon-photonics capacity-expansion plans and timing, 800G-vs-1.6T mix, gross margin trajectory, backlog, and — for Coherent specifically — net debt and interest cost against current market cap.
• Most volume-exposed but IP-light: Fabrinet (FN). The factual questions are the durability and margin of a contract manufacturer building others' designs. A deep-dive should check customer concentration (how much revenue rides on one or two customers), the AI-optics share of its book, margin stability, and free cash flow.
• Smallest, highest-variance pure-play: Applied Optoelectronics (AAOI). The factual question is whether it converts data-centre design opportunities into sustained, profitable revenue. A deep-dive should look at design wins, gross-margin progress toward profitability, cash burn and dilution history.
• Adjacent system vendor, smaller direct slice: Ciena (CIEN). Worth a deep-dive to understand the optical-systems and data-centre-interconnect angle, but it is NOT mainly an in-cluster transceiver maker — useful for the broader optical-networking picture rather than the transceiver gap itself.

Source: 500-stocks scan sub-sector 1 company list and supply/demand notes; FN added per task scope.

Sources & confidence

What was used:

• 500-stocks AGI sector scan — Networking & Connectivity module, sub-sector 1 "Optical Transceivers & Components": /Users/ravf/projects/work/.claude/worktrees/sector-hub/research/investments/500-stocks/03-networking-connectivity.html. This is the source for the company list (COHR, LITE, AAOI, CIEN), the 400G→800G→1.6T generations and component list (VCSELs, EMLs, CW lasers, photodetectors, modulators, silicon photonics), the "200k+ transceivers per 100k-GPU cluster" figure, the "30-40% annually" AI growth, the "800G in allocation since 2025" and "30-50 week lead time" facts, the "laser chip / EML / silicon-photonics bottleneck," and the "18-24 months to add capacity" lag. Fabrinet (FN) was added per the task's stated scope; it is not in this sub-sector's named list.
• Prior group report: none existed for this group at build time, so no V1 dated numbers could be folded in. A separate internal sector note ("optical-networking") referenced elsewhere in the project (with a "$29B TAM by 2031" figure) was NOT accessible in this worktree; if located, its hard figures should be merged on the next revision.
• General knowledge (cutoff ~early 2026): all market-size, share, revenue, market-cap, unit-price, margin and valuation-multiple figures, used because live web retrieval was unavailable.

Hard vs approximate: Treated as relatively hard (from the scan / well-known structure): the company list and what each makes; that a 100k-GPU cluster can need 200k+ transceivers; that the AI segment grows ~30-40%/yr; that 800G has been on allocation since 2025 with 30-50 week lead times; that high-power EML lasers and silicon photonics are the named bottleneck; that new capacity takes 18-24 months. Treated as approximate and NOT live-verified (all tagged est.): every dollar/percentage market figure — the ~$15-20B market size, the per-unit module prices ($500-$1,500 for 800G), the per-company revenue and market-cap numbers, all price-to-sales multiples, and the module-market-share characterisation. Confirm these against current 10-K/10-Q filings (annual/quarterly SEC reports) and live quotes before relying on them. All forward demand-growth and oversupply-timing statements are forecasts, not contracted orders.

Source: as listed above. Live retrieval unavailable; estimate-tagged figures are not live-verified.