After ECOC, what evidence matters for $POET, $COHR, $LITE and $AAOI?

05What it does in practice: six uses for photonics

Someone using an AI assistant does not normally buy a laser or transceiver. They buy a service; behind that service, someone else builds and operates the infrastructure. Interest in photonics arises from the need to move data between increasingly powerful machines without spending an excessive share of available energy on moving it.

Three recurring terms help explain the setting. GPUs are processors used as accelerators to perform large amounts of computation. Racks are the cabinets housing equipment. Switches route data between network connections. A hyperscaler operates cloud infrastructure on an enormous scale. An optical supplier may sell a small component inside a system that eventually occupies one of those racks.

Training a model means having many processors handle huge amounts of data and exchange intermediate results. Using the model, or inference, means generating a response, image or action. Inference can also require intensive communication when a model is distributed across processors. Services differ in size and networking needs, however: an AI function running on a phone does not necessarily require new optical connections in a data center.

On small screens, scroll the table horizontally.

The table maps functions; it does not certify interchangeable products. For AI networking, NVIDIA describes scale-out and scale-across applications of its platforms. NTT distinguishes connections between data centers from connections inside computers. NVIDIA, Vera Rubin platform and optical networks, May 31, 2026; NTT, optical data center connections and backbone, May 26, 2026.

An example helps separate the last two use cases. NTT and DOCOMO demonstrated AI video analysis using remote GPU resources connected through the IOWN optical network, an experiment relevant to distributed processing and robot control. NTT and MBRYONICS instead signed a memorandum to study modules for space optical communications. Both developments are documented, but their status and intended users differ. Neither establishes orders for $POET, $COHR, $LITE or $AAOI. NTT–DOCOMO demonstration, March 2, 2026; NTT–MBRYONICS memorandum, June 4, 2026.

Dedicated photonic platforms also exist for sensing, LiDAR and medical applications, such as those described by foundry SMART Photonics. The shared word is “photonics”; materials, processes, customers and sales cycles can differ substantially. Adding all these markets to an AI optics estimate would enlarge an opportunity that an individual business might not be equipped to serve. SMART Photonics, integrated photonics applications.

06How large could the market be, and how much can become revenue?

The first layer is infrastructure demand. In its April 2026 update, the International Energy Agency estimates data center electricity consumption rising from roughly 485 TWh in 2025 to 950 TWh in 2030, under its central scenario. TWh means terawatt-hours, a measure of energy. These are energy consumption figures, not optical component sales. They help explain why efficiency, density and power availability are real economic problems; they cannot be used to calculate the four companies’ future revenue. IEA, Key Questions on Energy and AI, 2026 executive summary.

The second layer is networking expenditure. A new data center requires land, buildings, power, cooling, processors, memory and connections. Only part of that spending goes to optics. Within optics, lasers, photonic circuits, modules, fibers, connectors and equipment must be distinguished. A module’s value includes its components: adding laser sales and module sales as though they were independent end markets can double-count the same value chain.

The third layer is the market a business can actually serve. Start with an application and product generation, identify compatible connections and assess which portion the company can address. Technology suited to a particular module does not automatically capture every 800G port worldwide. Qualification, capacity, pricing, support and competing suppliers narrow the opportunity.

A purely illustrative numerical example: if a program required 100,000 modules and a supplier delivered 20% at $400 each, sales would be $8 million. Doubling the total number of connections would not guarantee doubled revenue: the supplier’s share might change and unit prices could fall. These quantities and prices are not estimates for any of the four companies.

The useful question becomes: how many units can this company sell, to whom, when and at what margin? The total market describes the potential space. Revenue depends on programs won; shareholder value also depends on the capital required and the price paid for the shares.

07Who actually pays: end customer, system manufacturer and partner are different roles

A large cloud operator may define network requirements, buy complete systems or purchase some modules directly. A system manufacturer assembles equipment and integrates components. A transceiver manufacturer may buy lasers and optical engines from another supplier. A foundry manufactures chips for their designers; a packaging company assembles and tests them.

The customer generating a company’s revenue can therefore differ from the data center’s end user. A famous name in a presentation might denote a customer, investor, technical partner or simply another participant in the ecosystem. These relationships have different value and must be traced individually.

NVIDIA provides a documented example for $COHR and $LITE. On March 2, 2026, it announced non-exclusive multiyear agreements with both. The Coherent arrangement includes multibillion-dollar purchase commitments and access rights to future capacity for advanced lasers and optical networking products; separately, NVIDIA announced a $2 billion equity investment. The Lumentum arrangement covers advanced laser components, multibillion-dollar purchase commitments and future capacity, with another, separate $2 billion equity investment. NVIDIA–Coherent agreement; NVIDIA–Lumentum agreement.

Here, the buyer also confirms the commercial relationship. The $2 billion invested in each company is not product sales revenue. Nor does the general March announcement automatically allocate commitments to the new PhotonLink platform, September’s DWDM ELSFP module or a particular application. The agreements improve visibility into the industrial relationship, but individual supply schedules still require verification. Coherent’s 10-K documents the shares issued for the investment; Lumentum also publishes the agreement on its own site. Coherent, 2026 Form 10-K, equity note; Lumentum’s March 2 announcement.

$AAOI has a public relationship with Amazon. The agreement filed in March 2025 refers to manufacturing and development relationships and a warrant granted to an Amazon subsidiary. The relationship and warrant are still described in the June 30, 2026 10-Q. The $4 billion purchasing threshold over ten years tied to vesting of part of the warrant is an incentive condition: by itself, it is not a guaranteed order of that size. Nor does it automatically identify Amazon as the unnamed customer behind the more-than-$200 million 1.6T order. AOI–Amazon agreement and exhibits; AOI, Q2 2026 10-Q, Customer Warrant.

For $POET, public names must be attached to the correct program. Lumilens is the counterparty to the commercial agreement discussed here; LITEON is a module development partner. Neither relationship proves that a particular hyperscaler has qualified and purchased the final product. When the buyer remains unnamed, the analysis must keep it unnamed.

There is also the end-user layer: NVIDIA lists CoreWeave, Lambda, Meta, Microsoft and Oracle Cloud Infrastructure among early adopters of Spectrum-X Photonics. This provides concrete evidence of end demand for NVIDIA’s platform. It is not a list of direct customers for each of the four suppliers, nor does it prove shipments of an identical configuration to every operator. NVIDIA, official Silicon Photonics page.

Commercial confidentiality is common. An unnamed customer is not necessarily a poor customer; it simply leaves the market with less verifiable information. Conversely, a prestigious name does not remove concentration risk, buyer bargaining power or the possibility of qualifying a second supplier.

08Which countries are ahead? A map of capabilities

There is no single meaningful league table for all of photonics. Strength in materials research differs from high-volume transceiver production or influence over which components enter data centers. This map selects centers with documented activity. It does not assign national market shares or claim to cover every country.

On small screens, scroll the table horizontally.

The evidence gives substance to the map. In April 2026, TSMC indicated a 2026 production start for its on-substrate COUPE solution; NVIDIA separately describes TSMC, SPIL and Foxconn’s contributions to its platform. These support Taiwan’s role, without certifying that every project using that supply chain is already in mass production. TSMC, 2026 Technology Symposium; NVIDIA, Spectrum-X production ecosystem, May 2026.

For China, the industrial reference is Suzhou-based InnoLight’s official data center portfolio. For Japan, NTT describes IOWN’s development and photonic device expansion in 2026. Network operation, customer evaluation and future programs must again be read separately. InnoLight, AI and data center applications; InnoLight locations; NTT, strategy and investor questions, May 8, 2026.

In Europe, PIXEurope involves 20 institutions in 11 countries and a €400 million budget. This is investment in shared industrial infrastructure, not a component manufacturer’s revenue. Imec contributes platforms, integration and packaging; its New Origin collaboration concerns transferring and industrializing a silicon nitride platform. Italy participates through Politecnico di Milano and FBK. Imec, PIXEurope, June 2026; New Origin–imec, March 2026; Politecnico di Milano, Italian contribution.

In March 2026, Singapore announced an S$800 million program to strengthen semiconductors, including advanced photonics and packaging. Currency and scope matter: this is not US$800 million allocated solely to optical components. A*STAR, Committee of Supply 2026.

For investors, a company’s legal domicile or listing market does not fully describe geographic exposure. Production locations, critical suppliers, customers, alternative capacity and the time required to transfer a process must be mapped. Diversifying plants can reduce some dependencies, but requires capital and new qualifications. That economic consequence of geography is more useful than a generic national ranking.

09Why copper does not disappear, and better technology can take years

The choice depends on distance, speed, power use, cost and maintenance. A short electrical connection can be simple and inexpensive; converting a signal into light still requires devices, power and assembly. Optics becomes attractive when the system-level benefit outweighs those costs. There is no universal replacement threshold for every network.

Broadcom provides a concrete example: the Tomahawk 6 family offers both copper connectivity and CPO options. In October 2025, it announced shipments of the Davisson CPO version. These alternatives show why an optical future does not imply the immediate replacement of every electrical connection. Broadcom, Tomahawk 6, June 2025; Broadcom, Davisson, October 2025.

Reliability also involves more than making a prototype work. A customer wants to know how to replace a failed source, cool the system, identify a fault and use alternative suppliers. Shared interfaces help, but meeting a specification does not automatically mean passing the customer’s tests. This is one reason Lumentum’s new module refers to an OCI MSA interface and distinguishes the underlying laser from the complete configuration.

Finally, a constraint exists outside photonics: a data center must be built and supplied with electricity. In its 2026 update, the IEA identifies bottlenecks in transformers, turbines, advanced components and grid connections. An order can therefore shift in time without any defect in the optical product. IEA, demand and physical constraints update, April 16, 2026.

10From demonstration to revenue: what changes at each step

The following sequence is an analytical tool, not an obligatory path shared by every program. An order may precede final qualification; a sample may be paid for; a development contract may generate revenue before serial production.

1. Product introduction or demonstration. This shows what a supplier intends to offer and, when supported by appropriate measurements, some of its performance. It does not necessarily establish manufacturing cost, long-term reliability or volume availability.

2. Sampling. The customer receives units to evaluate. It matters whether they represent the final product and the factory process that will be used. A sample built with extensive manual work may function without demonstrating sustainable costs at high volumes.

3. Qualification. The product undergoes checks required for a particular use: performance, compatibility, reliability and the supplier’s ability to sustain them. The useful question is “qualified by whom, for which product and which application?” Component qualification does not automatically certify a complete module; module qualification at one customer does not automatically apply to every other customer.

4. Order. There is a more concrete commercial request. Investors still need to read conditions, timing, any modification or cancellation rights, and outstanding steps. A trial order also serves learning purposes; a production order has a different objective. The headline value does not remove contractual conditions.

5. Delivery and revenue. The business must satisfy the conditions for recognizing a sale. In simplified terms, product revenue depends on transfer of control; some development services can generate revenue over time. Cash collection may come before or after revenue recognition: sales and money entering a bank account are not synonyms. POET explicitly distinguishes engineering services recognized over time from product sales recognized at a point in time. POET financial statements, Note 3 on revenue.

6. Repeat orders and repeatable production. The customer returns, shipments recur and the supplier sustains quality and costs. This is richer evidence than a single announcement, especially if margins improve and the customer base broadens.

The practical distinction is substantial. An order reduces uncertainty about customer interest. Completed qualification reduces part of the technical risk. Recurring revenue with sound economics reduces part of the industrial risk. None of these steps eliminates all the others.

11$POET: follow separate programs without adding them up as cash

At least three disclosures should be separated for POET.

The first concerns Infinity 800G. The company’s July 14 FAQ refers to an order exceeding $5 million, announced in October 2025, with shipments expected in the second half of 2026. Useful confirmation would concern execution of that program: quantities shipped, qualification status and revenue contribution. The announced delivery window does not establish that deliveries have already occurred. POET’s official commercial FAQ.

The second is Lumilens. The program uses the Electrical-Optical Interposer, or EOI, a platform integrating electrical and optical functions. Its roadmap starts with 800G and 1.6T pluggable transceivers and extends toward NPO and CPO. The agreement announced in May includes a $50 million initial order within a potential purchasing opportunity exceeding $500 million over five years: not a guaranteed $550 million. Fulfillment and revenue depend on successful development, qualification and manufacturing scale-up. Engineering samples are expected in late 2026 and the production ramp in 2027. Official POET–Lumilens agreement.

Overlap with other suppliers therefore concerns some end applications. A POET engine may form part of a module developed with Lumilens; AOI’s transceiver is a product at another level of integration, while PhotonLink encompasses components and complete assemblies. Sharing an application does not prove technical equivalence, common customers or matching adoption schedules.

The agreement also includes a warrant: the right to purchase up to 22,921,408 shares at $8.25, partly linked to customer payments. The maximum does not represent shares already issued; potential exercise can dilute existing shareholders. POET financial statements, customer-related warrants.

The third disclosure is the August 13 quarterly update: POET reports a new $2.4 million order from an existing customer, received after June 30. This commercial fact belongs in the comparison without naming the customer or arbitrarily allocating it to Infinity or Lumilens. In the same update, expectations for the remaining 2026 quarters concern production units to be sent for qualification. That wording does not state that final qualification has already been completed. POET results and commercial update.

The quarter ended June 30 recorded $569,925 in combined engineering and product revenue and approximately $796.3 million in cash and short-term investments. Liquidity provides resources to execute programs; it does not demonstrate product adoption. Nor can all reported revenue be labeled sales of mass-produced optical engines. Interim financial statements filed with the SEC.

The post-ECOC test is therefore concrete: identify the next step actually completed for each program. A presentation adds visibility; delivery of a conforming batch adds a different piece of information. Company details are covered in the dedicated $POET Stock Hub.

12$LITE: one company can have mature products and demonstration projects

Lumentum’s September 21 announcements illustrate the distinction clearly.

The first concerns an eight-wavelength external laser source in a pluggable ELSFP format. In plain language, it uses several “colors” of light to feed compact optical connections. The underlying laser platform is already in production; initial availability of the new DWDM ELSFP configuration with an OCI MSA optical interface is expected in the first half of calendar 2027. A mature core technology can therefore coexist with a new module yet to be introduced. Lumentum DWDM ELSFP announcement.

The second announcement presents a demonstration with Qualcomm and Corning for short-reach optical connections between chips, called D2D (die-to-die), based on small 1060-nanometer VCSEL lasers. The company describes it as a proof of concept. It is evidence of technical collaboration, not an announcement of orders from Qualcomm or Corning. Participating in a demonstration does not automatically make the partners customers of the finished product. Joint ECOC demonstration announcement.

The ECOC Exhibition Industry Awards add another reference point, and show the limits of this kind of recognition. In the Optical Component Innovation (Package/OSA/Fibre) category, the 2026 award went to Lumentum’s VCSEL-based FOWLP optical engine for AI scale-up networks; POET Blazar was among the shortlisted nominees in the same category. An industry panel’s award signals technical recognition. For either company, it is not a customer qualification, an order or revenue. ECOC Exhibition Industry Awards 2026.

The August update had already mentioned a first order for ELS modules. It would nevertheless be incorrect to assign that order to the specific DWDM configuration introduced in September without a disclosure linking them. This is the same mistake as assigning company revenue to its newest product. Lumentum results, August 11, 2026.

The confirmations to watch for $LITE differ by program: actual availability and qualification of the new module, any specific orders, and progression of the chip-to-chip demonstration into a customer product. The company’s industrial experience matters, but does not replace these checks.

13$AAOI: orders exist, but capacity must become saleable output

Applied Optoelectronics provides a useful comparison because commercial growth is already visible. For the quarter ended June 30, 2026, it reports $191.9 million in revenue, including roughly $107.7 million from data centers. The remainder is primarily CATV business associated with cable networks. Total revenue is therefore not a measure solely of AI optics sales. AOI Form 10-Q, revenue breakdown.

On March 9, the company announced a first volume order for 1.6T transceivers exceeding $200 million, from an unnamed long-standing hyperscale customer. Shipments were expected from the beginning of the third quarter through the fourth quarter of 2026, after product qualification. An expected date must be distinguished from subsequent shipment confirmation: the March release alone does not establish how much was delivered by September. Official 1.6T order announcement.

In August, AOI reported that 800G volumes more than doubled sequentially. It indicated capacity approaching 200,000 units per month and a target of roughly 650,000 combined 800G/1.6T units per month by the end of 2026. The second figure remains a target, not production already achieved; neither automatically means units sold. Operating update, August 6.

There is also recent evidence of industrial commitments: September 8-K filings describe additional properties and leases for manufacturing, assembly and logistics. These transactions document investment and obligations; they do not directly establish how many transceivers have left the production lines. September 1 filing, property purchase disclosed September 10, lease disclosed September 15.

Financial risk accompanies production risk. On August 21, AOI entered into a program allowing market sales of new shares up to $600 million. The authorized ceiling is not money already raised and does not establish that all shares have been issued, but it signals a potential financing and dilution channel. Form 8-K on the equity program.

For investors, the question becomes how much announced capacity turns into conforming, delivered and profitable products, and how much capital is required to achieve that result.

14The product–customer–qualification–order–revenue matrix

The matrix must be completed program by program. “Not specified” means the cited documents do not provide that detail; it does not mean the contract or technical work does not exist.

On small screens, scroll the table horizontally.

Matrix sources: official releases and documents linked in the four company sections and the customer section, including the March 2, 2026 NVIDIA agreements. The cells summarize the cited public evidence and its limits; they do not assign undisclosed technical or contractual status.

15Revenue helps explain maturity, but the figures measure different businesses

On small screens, scroll the table horizontally.

Sources: POET financial statements, AOI 10-Q, Lumentum results, Coherent results.

These figures help explain company scale. They cannot assign each company a market share in the same niche because their product scope differs. Nor do they show which stock offers a more attractive valuation.

Revenue, gross margin and net income must also be distinguished. GAAP refers here to US accounting principles. Non-GAAP measures are company adjustments excluding particular items and must be read alongside their reconciliation. In the second quarter, AOI reported a $22.8 million GAAP net loss and $5.5 million in non-GAAP net income. These are not interchangeable ways of saying the company is profitable. AOI reconciliation.

Lumentum also illustrates why the distinction matters: in the June fiscal quarter, it recorded a GAAP loss of approximately $7.2 billion, affected by a $7.8 billion non-cash charge related to convertible debt extinguishment, while non-GAAP net income was $326.3 million. Those figures do not describe an equivalent cash outflow in the quarter and cannot simply be replaced by the adjusted number. Lumentum release and reconciliations; Form 10-K, debt extinguishment explanation.

16High production is not enough: the units must be saleable

Manufacturing yield is the proportion of units passing inspection. A fast line with high scrap rates may be less useful than a slower, stable one. Testing time, rework, component availability and customer acceptance also matter.

A purely illustrative example helps. A factory starts 100,000 units per month and 80% pass inspection. That produces 80,000 conforming units before other delivery constraints. If the batch costs $8 million, the cost is $100 per conforming unit. With the same cost and a 60% yield, it rises to approximately $133. These are teaching numbers, not data from the four companies.

That is why “we have room to triple production” requires further answers. Have machines been ordered or installed? Is the process qualified? Does capacity refer to an optical engine, laser source or complete module? Under which shifts and product mix? Are there orders to absorb the output?

Inventory growth does not resolve the question either: it may prepare for future deliveries or indicate products awaiting testing or customers. Explanations and corroboration in sales, margins and cash flow are needed. The interpretation depends on context, rather than a positive change in one line item.

17When a competitor’s news matters for the others

A competitor’s success may confirm that customers care about a technical problem. Turning that into a conclusion about another company requires an additional connection.

Shared demand. If customers invest more in connectivity, opportunities may expand for several suppliers. That does not yet establish which product will be selected.

Supply relationship. One company’s success can benefit another if the latter actually sells a component to the winner. Evidence is needed about the relationship, the product generation and its economic relevance. Technical similarity cannot identify that relationship.

Competition. A complete solution may reduce customers’ need to integrate parts from other suppliers. Alternatively, a customer may prefer multiple sources to limit dependence on a single business. These are plausible scenarios to verify, not automatic consequences of the PhotonLink launch.

Substitution between architectures. A connection designed for a particular distance, fiber or system position may not replace another. Equal nominal speed does not eliminate differences in power, cooling, maintenance and total cost.

For $POET, growth at $COHR or $LITE may therefore support the relevance of the problem being addressed. It does not demonstrate adoption of the Optical Interposer. For $AAOI, CPO development does not prove that pluggable transceivers immediately disappear: investors must see which networks and customer programs actually change.

18The updates that could change the comparison

Events do not all carry the same information. A practical way to read upcoming releases is to identify the fact that resolves an outstanding question.

On small screens, scroll the table horizontally.

Company timelines must remain attached to their subject: late 2026 for Lumilens engineering samples does not mean late 2026 for all agreement revenue; first-half 2027 availability of the new ELSFP is not already-qualified volume production; a PhotonLink revenue forecast does not apply equally to every architecture.

The same applies to missed windows. If a window passes without public confirmation, the appropriate status is “outcome not confirmed in the available sources.” Silence proves neither success nor failure. An explicit schedule revision or loss of an order would provide different and stronger information.

19Other questions investors ask

Does AI photonics mean quantum computing?

No. In this article, light’s main function is to carry data between elements of computing infrastructure. Photonic computing and photonic quantum computing are distinct fields with their own devices and requirements. A data center transceiver order does not establish a commercial position in quantum computing. NTT itself presents networking, photonics–electronics convergence and quantum systems as distinct areas. NTT, functions of photonics–electronics convergence; NTT, 2026 strategy.

Does 1.6T mean an AI service is twice as fast as 800G?

It means twice the nominal transmission capacity of the connection concerned, using the same definition. The speed experienced by users also depends on processors, memory, software, congestion and network utilization. Widening a road helps most when that road was the bottleneck; it does not automatically double the speed of the entire journey.

Does halving optics power use halve the data center’s electricity bill?

No. If, in a purely arithmetic example, the affected part represented 10% of total consumption, halving it would reduce the total by 5%, with everything else unchanged. The benefit can still matter, including by freeing power for other tasks. Read the boundary of the comparison: laser, module, network and entire facility consumption are different measurements. The example’s 10% is not an estimate of a real data center’s consumption mix.

Can these companies work together and compete?

Yes. A supplier may sell components to a module manufacturer while competing with it in another product line. The supply chain should therefore not be presented as four separate teams always offering the same product. Any particular relationship must nevertheless be documented: the technical possibility of collaboration does not prove an agreement between companies.

Does a large cash balance solve industrial risk?

It funds development, staff, equipment and waiting time. It does not replace customer acceptance. Conversely, a major order without enough capital to fulfill it can create financial strain. Analysis must connect demand, liquidity, investment and potential dilution.

Does a small company automatically have more upside than a large one?

A new contract can have a large impact on a small revenue base, but the same sensitivity applies to delays or lost programs. A larger business may have more resources and customers, without making any share price attractive. Commercial comparison does not replace valuation analysis or assessment of per-share risk.

What is the most useful check after an announcement?

Identify what changed since the previous disclosure: a completed test, available product, accepted order, shipment, payment or repeat order. Then ask whether that fact changes timing, quantities, margins or required capital. The connection between industrial progress and economic results gives the announcement substance.

20The role of social media and the investor’s final question

The September 23 POET discussion thread includes expectations linked to ECOC and possible further announcements. The accessible portion illustrates community attention; it is neither a representative sentiment measure nor verification of orders or customers. These are retail users’ comments, not institutional analysis. September 23 Reddit discussion.

That attention can move expectations ahead of results. Even good news can be followed by a negative share-price reaction if investors expected more, or expected it sooner. Conversely, a positive reaction does not certify the industrial quality of an announcement.

Commercial analysis makes assumptions more precise. Share valuation requires another step: how much future success is already priced in, at which margins, with how much capital and how many shares? A business can sell more products without generating an adequate return for someone buying its shares at any price.

After ECOC, the useful question is which concrete step each program has completed since its last update. Coherent needs to make execution of the PhotonLink timetable visible. POET needs to demonstrate conversion of programs into qualified products and recurring revenue. Lumentum needs to distinguish established operations from the maturation of new configurations. AOI needs to turn expansion and orders into saleable output with sustainable economics.

The matrix remains open: a qualification, accepted delivery or repeat order can change it. Their value lies in the question they answer, not how often an announcement uses the words “AI” or “photonics.”