Issue 06: Who keeps the profits from AI's photonics buildout?

Google's Jupiter Evolving research, published at SIGCOMM 2022, describes production experience combining optical circuit switches, software control and a changed network topology.

The paper reports lower power and capital costs across that evolution. Those are results for an integrated network redesign over time, not a universal savings percentage that can be assigned to any OCS product.

The important lesson is the combination. Hardware enabled new connections; software and network design made those connections useful.

For an equipment supplier, a customer may need to change several parts of its operating model before adopting the product. A technically superior component can therefore face a slower sales cycle than its benchmark suggests.

• Faster pluggable module. What changes: More capacity in a removable form factor; Do not assume: Twice the bandwidth means twice the revenue.
• Near-packaged optics. What changes: Shorter electrical distance to the optical engine; Do not assume: Identical economics to CPO.
• Co-packaged optics. What changes: Optical engines integrate with the chip package; Do not assume: All lasers become inaccessible.
• External laser source. What changes: Light generation remains separate from the engine; Do not assume: One module is required for every port.
• Optical circuit switch. What changes: Optical paths are reconfigured; Do not assume: It replaces all packet-processing functions.
• Optical I/O chiplet. What changes: Optical connectivity attaches to a compute package; Do not assume: A demonstration establishes a production business.
• Coherent DCI module. What changes: Optical transmission serves links between sites; Do not assume: The product competes directly with every short-reach module.

Architecture distinctions are drawn from the technical sources linked in this section. The investment cautions are our analysis.

Power savings need a denominator

A claim that a component uses less power is meaningful only when the comparison specifies what was measured.

Was the benchmark a laser, an optical engine, a complete connection, a switch or an entire cluster? Those measurements cannot be substituted for one another.

A component can improve substantially while producing a modest percentage change in the facility's total electricity bill. Conversely, a modest power saving can be valuable if it frees capacity at a constrained site.

Our illustrative model below separates the electricity saving from the possible value of additional computing capacity.

A worked operating-cost example

Assume 100,000 comparable connection endpoints save 10 watts each. The total reduction is 1,000,000 watts, or one megawatt.

At continuous operation for a non-leap year, that saves 8,760 megawatt-hours of IT electricity. At an assumed $100 per megawatt-hour, the direct electricity saving is $876,000 per year.

If the incremental facility power overhead is represented by a 1.2 multiplier, the illustrative saving becomes $1,051,200. This assumes overhead scales with the saved IT load; an actual facility may behave differently.

Each 5 watts saved per endpoint is worth about $438,000 a year

5 W saved, 0.5 MW$438,000

10 W saved, 1.0 MW$876,000

15 W saved, 1.5 MW$1,314,000

20 W saved, 2.0 MW$1,752,000

If the more efficient installation costs an extra $5 million, the direct electricity payback at 10 watts saved is about 5.7 years. Including the assumed overhead benefit reduces it to about 4.8 years.

Counting facility overhead cuts the payback to about 4.8 years

5.7 yearsElectricity only$5 million extra at 10 watts saved

4.8 yearsWith 1.2× facility overheadAssumes overhead scales with IT load

That is enough to show why “lower power” alone cannot settle a purchasing decision. Equipment life, replacement timing and the cost of capital still matter.

A different electricity tariff changes the answer. So do partial utilization, standby power and the number of endpoints actually replaced.

Power capacity can be worth more than electricity

Suppose the facility cannot obtain additional electrical capacity promptly. Saving a megawatt may allow the operator to install more computing equipment within its existing allocation.

The additional contribution depends on equipment availability, cooling, customer demand and the value of the workloads. An unused power saving does not generate this second benefit.

This creates two separate values: avoided operating cost and the economic contribution from redeployed capacity. A supplier may capture some of either through pricing.

Do not add the full value of both if they describe mutually exclusive uses of the same saved electricity. If the freed megawatt is immediately consumed by new GPUs, the original electricity saving is no longer retained in full.

A saved megawatt can be banked or redeployed, not both

Bank the saving

• Avoided operating cost
• A lower utility bill
• Value set by tariff and utilization

vs

Redeploy the capacity

• More compute inside the existing power allocation
• Value set by workloads and customer demand
• Needs equipment, cooling and demand

A disciplined customer model has separate cases for banking the saving and redeploying the capacity.

• Is the constraint electricity capacity or electricity price? Why it changes the result: Capacity relief can have value beyond the utility bill.
• Are power savings measured at both endpoints? Why it changes the result: Counting the same saving twice inflates the benefit.
• Is the workload continuously active? Why it changes the result: A peak-power comparison may overstate annual savings.
• Does cooling scale down with IT power? Why it changes the result: Facility overhead may not fall proportionally.
• Can the operator use the freed capacity? Why it changes the result: Additional compute revenue requires other available resources.
• Does the new design require extra spares? Why it changes the result: Reliability provisions can increase the upfront cost.
• How long will this hardware stay installed? Why it changes the result: A payback beyond the expected replacement date is less useful.

This is our purchasing framework, not a vendor forecast. It explains why different customers can make different choices about the same optical product.

Reliability is an economic variable

A network fault can waste more than the cost of its replacement part. Depending on the workload, it can interrupt a job, require recovery or leave expensive hardware underused.

That makes the number of failures, their consequences and the time needed to restore service relevant to the investment thesis.

Broadcom's October 2025 reliability release reports one million cumulative 400G-equivalent port device-hours without link flaps in Meta's high-temperature laboratory characterization environment.

The setting matters. Cumulative device-hours are not a claim that one switch operated for a million hours, and laboratory characterization is not the same as fleet-wide production experience.

A link flap is a connectivity interruption. Absence of link flaps in a specified test does not establish the absence of every possible hardware or software failure.

The result is useful evidence. It should be recorded with the test conditions and failure definition attached.

Translate reliability into expected cost

Consider a hypothetical deployment where failures cause 20 expensive job interruptions each year. Assume each interruption costs $50,000 in lost work and recovery expense.

Under these assumptions, the annual expected cost is $1 million. Reducing the frequency to five interruptions would reduce that estimate to $250,000, producing a $750,000 benefit.

Cutting interruptions from 20 to 5 saves $750,000 a year

20 interruptions$1,000,000

15 fewer at $50,000 each−$750,000= $250,000

5 interruptions$250,000

These are deliberately illustrative assumptions. Actual losses depend on checkpoints, redundancy, scheduling, fault isolation and whether work can continue elsewhere.

A design that saves electricity but raises recovery costs can be unattractive. A design with a higher purchase price may be preferable if it materially improves useful operating time.

• Failure frequency. Useful unit: Events per operating period; Verify: A consistent definition of a failure.
• Fault scope. Useful unit: Affected links or devices; Verify: Whether faults remain local or propagate.
• Recovery time. Useful unit: Minutes or hours; Verify: Detection, diagnosis, replacement and restart.
• Work lost. Useful unit: Compute-hours or completed jobs; Verify: How much work must actually be repeated.
• Spare requirement. Useful unit: Parts and capital; Verify: Inventory needed to meet recovery targets.
• Repair complexity. Useful unit: Technician time and equipment; Verify: Whether service requires removing a larger assembly.
• Customer outcome. Useful unit: Cost per completed workload; Verify: Whether component improvements reach the application.

The customer-outcome measure is the destination. A product benchmark is more persuasive when the operator can connect it to completed work at an acceptable cost.

Serviceable lasers solve only part of the problem

The OIF external laser source agreement announcement describes front-panel placement, hot-swap replacement and common mechanical, thermal, electrical and optical parameters.

It also notes that one external source can supply one or more optical engines. That is why multiplying every port by a fixed external-module price can overstate a supplier's opportunity.

Lumentum's external laser reliability white paper explains its rationale for keeping lasers away from the switch's thermal environment and making them field-serviceable.

This is a manufacturer's technical argument for its approach. We would still want customer evidence on complete-system operation before treating every claimed benefit as a realized fleet outcome.

Replacing a laser does not repair a damaged optical engine, a package defect or a firmware issue. Different failure modes need different remedies.

A replaceable laser solves only part of the reliability problem

A serviceable laser helps with

• Light-source failures
• Hot-swap replacement at the front panel
• Keeping lasers away from the switch's heat

vs

It does not repair

• A damaged optical engine
• A package defect
• A firmware issue

Our investment inference is that serviceability can support adoption without eliminating integration risk. The companies that diagnose and contain failures effectively may have an advantage that is absent from a simple bandwidth comparison.

Moving the optics changes who gets paid

A network connection contains several things someone can sell: the switching chip, signal-processing electronics, optical engine, laser, connectors and fiber. The customer also pays for packaging, testing and assembly.

Changing the design changes the bill of materials. Some components gain value; others can lose their place or become less valuable per connection.

Moving the optics reshuffles who gets paid per connection

1. Switching chip
2. Signal-processing electronics
3. Optical engineMoves beside the chip in CPO
4. LaserCan stay accessible at the front plate
5. Connectors and fiber
6. Packaging, testing and assembly

What a network connection contains

The laser illustrates why the details matter. NVIDIA's June report from Coherent's Sherman facility identifies Coherent as a supplier of external laser modules for its photonics switches. The light source can remain accessible at the switch's front plate even as the optical engine moves beside the chip. CPO does not mean every optical component disappears into one sealed package.

That report also describes expansion of indium phosphide manufacturing. Indium phosphide is a compound semiconductor used in these lasers. Producing qualified parts at scale gives the supplier something more defensible than a photonics label.

Our inference is that laser manufacturing deserves particular attention, but manufacturing capacity is not automatically pricing power. A successful expansion increases supply. Buyers can qualify alternatives. Equipment spending and poor early yields can absorb the cash that investors expected to receive.

For every company, ask: what does it sell in the old design, what does it sell in the new one, and how much gross profit remains per connection?

From a laboratory result to a profitable factory

Manufacturing is where a promising device becomes either a repeatable business or a recurring capital request.

The important quantity is qualified output: parts that meet the customer's requirements, pass the relevant tests and can be delivered consistently.

A company can increase nominal capacity faster than qualified output. Investors should avoid treating those two measures as interchangeable.

Follow the production sequence

The following is a simplified diligence map. Actual manufacturing routes differ by product and supplier; the map identifies questions to ask rather than prescribing a universal process.

• Device design. Establishes: Intended performance and physical structure; Risk: The design may not tolerate manufacturing variation.
• Wafer processing. Establishes: Devices are formed on the substrate; Risk: Defects or process variation reduce usable output.
• Wafer-level testing. Establishes: Early screening identifies working devices; Risk: Passing devices may fail later packaging tests.
• Dicing and handling. Establishes: Individual parts become available for assembly; Risk: Handling damage can reduce yield.
• Packaging. Establishes: Devices gain usable electrical, optical and thermal connections; Risk: Alignment and assembly can introduce losses or defects.
• Burn-in and reliability testing. Establishes: Early failures and degradation are screened; Risk: Test conditions may not capture every field stress.
• Customer qualification. Establishes: A product is accepted for a specific application; Risk: Approval may not transfer to another customer or design.
• Volume delivery. Establishes: Output becomes repeatable shipments; Risk: Forecast changes can leave excess inventory.
• Field operation. Establishes: The product works outside the factory; Risk: Returns, repair costs or redesigns can erode margins.

This sequence explains why additional wafer starts do not immediately become additional sales. Work can accumulate before the bottleneck rather than pass through it.

If packaging capacity is constrained, spending only on upstream wafer output can increase inventory without increasing customer deliveries.

If customer qualification is the slowest step, even a technically successful factory expansion may take time to influence revenue.

Spending upstream of the bottleneck builds inventory, not deliveries

1. Wafer output (Spending added) to Packaging (Constrained)
2. Packaging (Constrained) to Qualification (Customer approval)
3. Qualification (Customer approval) to Deliveries (Becomes revenue)

Yield is a multiplier, not a slogan

Imagine a hypothetical batch of 10,000 potential devices. Assume 80% survive fabrication, 90% survive packaging and 95% pass final qualification testing.

Make a guess

Out of 10,000 potential devices, how many qualify after those three yields?

Tap your guess

5,000Your guess: 5,0005,500Your guess: 5,5006,000Your guess: 6,0006,500Your guess: 6,5007,000Your guess: 7,0007,500Your guess: 7,5008,000Your guess: 8,0008,500Your guess: 8,5009,000Your guess: 9,0009,500Your guess: 9,50010,000Your guess: 10,000

0123456789,012345678901234567890123456789Qualified output

Spot on.Close, a little low.Close, a little high.Way off, too low.Way off, too high. It was 6,840.

6,840: 10,000 × 0.80 × 0.90 × 0.95. Three yields that each sound high compound to lose almost a third of the batch.

The resulting qualified output is 6,840 devices: 10,000 × 0.80 × 0.90 × 0.95.

Three respectable yields leave 6,840 of 10,000 devices

Potential devices10,000

Fabrication, 80% pass−2,000= 8,000

Packaging, 90% pass−800= 7,200

Final test, 95% pass−360= 6,840

Qualified output6,840

If packaging yield improves from 90% to 97%, qualified output rises to 7,372 devices without adding wafer starts. That is approximately 7.8% more saleable output.

The company may therefore grow shipments through better execution rather than a proportionate increase in factory size.

• Early ramp. Fabrication yield: 70%; Packaging yield: 85%; Final-test yield: 90%; Qualified output: 5,355.
• Base process. Fabrication yield: 80%; Packaging yield: 90%; Final-test yield: 95%; Qualified output: 6,840.
• Better packaging. Fabrication yield: 80%; Packaging yield: 97%; Final-test yield: 95%; Qualified output: 7,372.
• Broad improvement. Fabrication yield: 90%; Packaging yield: 97%; Final-test yield: 98%; Qualified output: 8,555.

Each case starts with 10,000 hypothetical potential devices. The broad-improvement case is rounded to the nearest device. These yields are not estimates for a named manufacturer.

Higher yields can improve unit costs because more usable products share part of the same production expense. The exact saving depends on where failures occur and which costs are incurred before rejection.

A failure identified late in assembly can destroy more accumulated value than one screened out early. This makes testing strategy financially important.

Yield improvement is particularly valuable when demand is strong enough to absorb the extra output. Otherwise, the factory may simply run fewer starts.

Larger wafers do not guarantee proportionate economics

Wafer area grows with the square of diameter. Moving from a three-inch circle to a six-inch circle multiplies gross geometric area by four.

That arithmetic does not establish a fourfold increase in good devices. Edge exclusions, layout, process maturity and defect rates still determine what can be sold.

A larger wafer process can initially have lower yield than a smaller mature process. Its economic advantage depends on improving that yield while controlling processing cost.

For Coherent's capacity story, the useful disclosure is the progression from nominal capacity to qualified output and then to cash generation. We should not assign all three milestones to the date a building opens.

Packaging can capture value that the laser does not

TSMC's 2025 annual report describes COUPE as integrating silicon photonics and electrical control chips, alongside work on co-packaging optical I/O with advanced computing or networking hardware.

That creates another important economic layer: the supplier that integrates and packages the parts can participate even when it does not sell the final network system.

Our inference is that scarce integration capacity may influence customer choice. A component with excellent specifications is less useful if the complete assembly cannot be delivered reliably.

For stock selection, however, technical participation is only the first question. The activity must also be large enough to matter within the company's total profits.

TSMC is therefore relevant to the supply-chain map without automatically becoming a focused photonics investment. A narrow product opportunity can be financially diluted inside a much larger business.

• New building announced. Shows: Management intends to add space; Does not establish: Equipment is installed or qualified.
• Equipment ordered. Shows: Capital has been committed; Does not establish: Delivery and commissioning are complete.
• Line commissioned. Shows: Equipment can operate; Does not establish: Customer-qualified yields are attractive.
• Capacity doubled. Shows: A defined capacity measure increased; Does not establish: Saleable output or revenue doubled.
• Customer approval. Shows: A stated requirement was met; Does not establish: Orders are large or noncancelable.
• Shipments increased. Shows: More products left the supplier; Does not establish: Cash was collected or margins improved.
• Gross margin increased. Shows: Reported production economics improved; Does not establish: Capital spending has been recovered.
• Free cash flow improved. Shows: Less cash was absorbed or more generated; Does not establish: The improvement will persist after working-capital timing reverses.

Standards can grow demand and reduce bargaining power

An interface standard can make a product easier to adopt. It can also make it easier for customers to evaluate another supplier.

Both effects can happen at once. A larger market with more competition may produce lower profit per unit than a smaller proprietary market.

An open standard can enlarge the market and shrink the margin

Standards grow demand

• Products become easier to adopt
• Integration uncertainty falls
• The market can get larger

vs

…and reduce bargaining power

• Customers can evaluate another supplier
• More competition in that larger market
• Profit per unit can fall

The OCI MSA describes work toward an open optical architecture for scale-up connectivity. OIF's external laser work similarly defines interfaces that support interoperability.

The economic implication is conditional: standardization can reduce integration uncertainty, but the distribution of profits depends on which parts remain difficult to reproduce.

A standard connector does not make every laser equivalent. Reliability, manufacturing quality, optical performance and delivered cost can still distinguish suppliers.

Equally, a supplier should not be credited with permanent customer lock-in merely because its current product requires qualification.

Ask what remains proprietary

A useful competitive analysis separates the open interface from the specialized work behind it.

• Device. Possible source of differentiation: Performance within the customer's requirements; Evidence to seek: Comparable tests under stated conditions.
• Process. Possible source of differentiation: Cost and yield at commercial volume; Evidence to seek: Margin and output progression during a ramp.
• Package. Possible source of differentiation: Efficient, reliable integration; Evidence to seek: Customer-qualified assemblies and repeat production.
• Control electronics. Possible source of differentiation: Useful diagnostics and power management; Evidence to seek: Operational integration and service data.
• Software. Possible source of differentiation: Deployment and fault-management capability; Evidence to seek: Customer adoption beyond a laboratory trial.
• Supply agreement. Possible source of differentiation: Delivery certainty and allocation rights; Evidence to seek: Disclosed terms and actual shipment conversion.
• Support. Possible source of differentiation: Faster diagnosis and replacement; Evidence to seek: Field-service capabilities and customer retention.

Our preference is for differentiation that remains valuable after buyers understand the product. Novelty alone is a weak basis for a long-term margin assumption.

A company can also choose to give customers more flexibility in exchange for faster adoption. Lower switching costs may be part of a successful volume strategy rather than evidence of failure.

The valuation model needs to reflect that strategy. A volume supplier with competitive pricing should not be valued as though it collects a monopoly royalty.

Model customer commitments without counting them twice

The NVIDIA investments and purchase commitments validate commercial interest. They do not provide enough public detail to construct an exact annual revenue schedule.

We do not know the complete product mix, price adjustments, shipment conditions or timing from the announcement alone. Those unknowns should stay visible in an investment model.

Different payments do different jobs

• Equity investment. Primary economic role: Finances the business in exchange for ownership; Common analytical mistake: Recording it as product revenue.
• Customer prepayment. Primary economic role: Provides cash before delivery under agreed terms; Common analytical mistake: Treating cash received as completed revenue.
• Purchase commitment. Primary economic role: Describes future purchasing obligations or intentions under a contract; Common analytical mistake: Assuming the full amount arrives in one year.
• Capacity access right. Primary economic role: Gives access to specified future production; Common analytical mistake: Treating access as proof of full utilization.
• Design selection. Primary economic role: Identifies a product intended for a program; Common analytical mistake: Treating the program as an irrevocable volume order.
• Backlog. Primary economic role: Captures orders under the company's definition; Common analytical mistake: Assuming every order has the same cancellation protection.
• Shipment. Primary economic role: Moves goods toward a customer; Common analytical mistake: Ignoring acceptance terms and collection risk.
• Recognized revenue. Primary economic role: Records sales under the applicable accounting policy; Common analytical mistake: Equating sales with cash collected.

This is an analytical classification, not a claim that either NVIDIA contract contains every item listed here.

The most important questions are who carries excess-capacity risk and who benefits when manufacturing costs fall.

A customer may fund expansion because supply is strategically important. It may also negotiate favorable pricing or allocation in return.

That can lower the supplier's financing risk while limiting its upside per unit. The trade can still be attractive; its terms matter more than the size of the headline.

Avoid supply-chain double counting

Suppose a hypothetical laser manufacturer sells a $100 component to a module maker, which sells a $500 module to a systems company.

The systems company then sells a $2,000 assembly that includes the module. Adding all three sales gives $2,600 of company revenue across the chain, but only $2,000 of final customer spending.

Summing the chain gives $2,600; the customer spent $2,000

1. Laser maker (Sells a component) to Module maker (Sells a module): $100
2. Module maker (Sells a module) to Systems company (Sells an assembly): $500
3. Systems company (Sells an assembly) to Final customer (Spends $2,000): $2,000

Both views can be useful. The problem is calling the $2,600 total an independent end market and applying a market share to it without specifying the layer.

A supplier revenue model should follow that supplier's invoice. An industry spending model should define whether it measures components, modules or final systems.

The same caution applies to counting both an internal component transfer and the eventual external module sale inside a vertically integrated company.

• Choose the product. Required definition: Laser, optical engine, module, chiplet or complete system.
• Choose the customer. Required definition: Direct purchaser rather than only the ultimate operator.
• Choose the unit. Required definition: Device, endpoint, port, module, switch or rack.
• Choose the price. Required definition: Realized supplier price rather than retail system value.
• Choose the period. Required definition: Quarter, fiscal year or calendar year.
• Choose the recognition point. Required definition: Contract, shipment, acceptance or recognized revenue.
• Remove overlap. Required definition: Exclude internal transfers and duplicated value-chain layers.

The model is only useful if its units remain consistent from the physical deployment forecast to the supplier revenue estimate.

Five companies, five different underwriting problems

The five names from our photonics watchlist belong in different categories. The comparison describes business exposure rather than ranking expected returns.

• NVIDIA, NVDA. Role in the thesis: Integrates photonics into the AI networking platform; Evidence that would strengthen the case: Deployed systems improve customer economics and support networking demand.
• Broadcom, AVGO. Role in the thesis: Supplies networking silicon and optical connectivity technology; Evidence that would strengthen the case: Customers adopt its optical designs alongside its switching products.
• Marvell, MRVL. Role in the thesis: Existing connectivity business plus future Photonic Fabric revenue; Evidence that would strengthen the case: New optical programs progress from selection to recognized sales.
• Coherent, COHR. Role in the thesis: Lasers, optical components and manufacturing; Evidence that would strengthen the case: Qualified capacity produces higher cash generation after capital spending.
• Lumentum, LITE. Role in the thesis: Optical components and systems, with emerging CPO products; Evidence that would strengthen the case: New products add profitable sales without weakening the existing business.

NVIDIA: optics supports the economics of the system

NVIDIA says Spectrum-X Ethernet Photonics entered production in its Vera Rubin production announcement. That is a stronger milestone than a demonstration, although production status does not disclose installed volume or a standalone CPO profit figure.

The investment logic extends beyond selling optical hardware. If networking helps customers get more useful computation from a cluster, it can support demand for the wider system. That is our interpretation of the platform economics, not a separately reported revenue stream.

NVDA therefore offers broad AI infrastructure exposure. A successful photonics rollout could matter operationally while remaining difficult to isolate in the stock's earnings. Its valuation must still be supported by the rest of the business.

Broadcom: check the optical contribution inside the AI total

Broadcom's OFC release includes Tomahawk 6-Davisson CPO alongside switching chips and optical signal-processing products. It has multiple ways to participate in network spending.

Its September 2 earnings release reports $16.7 billion of third-quarter AI semiconductor revenue and attributes demand to custom accelerators and networking. That figure is much broader than photonics. Treating it as optical revenue would inflate the evidence for this particular thesis.

Our preference here is for evidence of customer adoption and durable networking economics. A large AI revenue number cannot establish which optical architecture wins or how much incremental profit it adds. Broadcom's software business also needs its own valuation; a photonics multiple applied to the whole company would obscure what shareholders actually own.

Marvell: the acquisition closed before the revenue arrives

Marvell completed the Celestial AI acquisition on February 2. At closing, management expected initial revenue in the second half of fiscal 2028, reaching a $500 million annualized rate in that year's fourth quarter. Those were forward-looking targets, not booked sales.

A $500 million annualized quarterly rate means roughly $125 million in that quarter. It does not mean $500 million of revenue over the fiscal year. Fiscal years also differ from calendar years; keep the company's fiscal labels when tracking the target.

A $500 million run rate is about $125 million in a quarter

$500MAnnualized revenue rate targetFourth quarter of fiscal 2028

$125MImplied revenue that quarterThe run rate divided by four

The latest quarterly release, August 27, reports $2.739 billion of company revenue and 46% growth in the data center business. Those results support the existing business, but do not establish that Celestial's previously announced ramp has happened.

Our approach would value the operating business first, then assign a probability-weighted value to the future optical program. Paying as though a product has already reached volume removes much of the protection against delays.

Coherent: the factory has to earn its cost

Coherent's appeal is its position in the physical supply chain. The NVIDIA agreement names advanced laser and optical networking products; the Sherman expansion provides a concrete manufacturing project to follow.

The question is how much productive capacity the investment buys. A larger wafer has more area, but usable output also depends on defects, device size, processing and customer qualification. Factory floor space is an input to the thesis, not its financial result.

We would track capital expenditures against operating cash flow, the ramp in qualified output where disclosed, and whether margins hold as shipments grow. Faster sales accompanied by persistent cash absorption would weaken the claim that a scarce manufacturing capability is becoming a strong shareholder return.

Coherent also serves communications and industrial markets, as its partnership release describes. Its whole-company earnings cannot be treated as a pure AI laser business.

Lumentum: distinguish today's results from the next product cycle

Lumentum's August 11 results show quarterly revenue of $1.0063 billion and a 47.4% GAAP gross margin. Management also described an initial external laser source module order and increasing demand for CPO lasers. Those statements describe different stages of commercial maturity.

Spot the red flag

Lumentum reported a profit and a $7.2 billion loss for the same quarter. Which line explains the gap?

Lumentum fiscal fourth quarter, as reported$ millions

Quarterly revenue1,006.3Revenue was solid. It doesn't explain why the two profit measures diverge.Non-GAAP net income326.3One side of the gap, not its cause.Noncash debt-extinguishment loss−7,800Converting certain note amounts into equity created a $7.8 billion noncash charge. Not a cash operating loss, but it changes the shareholder claim.GAAP net loss−7,200The result, not the cause.

Tap the line you think is the tell.

Spotted it.Not that line. A $7.8 billion noncash debt-extinguishment loss sits between $326.3 million of non-GAAP profit and a $7.2 billion GAAP loss. Neither figure alone describes the business; the next question is what happened to the share count.

Explain every lineHide the other explanations

The same release reports a $7.2 billion GAAP net loss, driven by a $7.8 billion noncash debt-extinguishment loss from converting certain note amounts into equity. Non-GAAP net income was $326.3 million. Using either profit figure without explaining the reconciliation would distort the business.

Our focus would be the bridge between operating performance and the fully diluted shareholder claim. A noncash financing charge does not equal a cash operating loss. Equally, excluding it from adjusted earnings does not remove the need to examine the capital structure and share count.

What the financial statements add to the technology story

The technical case tells us why customers might buy. The financial statements tell us what the supplier keeps and how much capital it needs along the way.

A useful review reconciles reported growth, margins, cash flow and the number of shares participating in the outcome.

Coherent: a concrete cash-conversion checkpoint

Coherent's August earnings summary reports fourth-quarter revenue of $2.05 billion and a 38.5% GAAP gross margin.

Make a guess

In fiscal 2026, how did Coherent's operating cash flow compare with its spending on property and equipment?

Tap your guess

Short by over $1 billionYour guessShort by under $1 billionYour guessSurplus under $500 millionYour guessSurplus over $500 millionYour guess

−$0123456789,012345678901234567890123456789.0123456789MFiscal 2026

Spot on.Close, a little low.Close, a little high.Way off, too low.Way off, too high. It was −$1,023.4M.

Short by $1,023.4 million: operating cash flow of $79.5 million against $1,102.9 million of additions, after a $192.8 million surplus in fiscal 2025. Our simple subtraction, not a company-defined measure.

Fiscal 2026−$1,023.4M

Fiscal 2025$192.8M

Its full fiscal 2026 release, cash-flow table, supplies a different perspective:

Coherent's factory spending outran operating cash by $1.02 billion in fiscal 2026

2025 operating cash flow$633.6M

2025 equipment additions$440.8M

2025 difference$192.8M

2026 operating cash flow$79.5M

2026 equipment additions$1,102.9M

2026 difference−$1,023.4M

This does not prove that expansion is uneconomic. It establishes that the current investment case requires a future improvement in cash conversion, rather than assuming growth has already funded itself.

The next analysis should reconcile working capital, exceptional cash items and capital expenditure timing using the detailed filings. We do not attribute the entire shortfall to a single cause from this summary.

Separate operating performance from financing

A growing cash balance does not necessarily mean a business generated cash. It can also reflect equity issuance, debt issuance, asset sales or customer funding.

Those sources may be sensible. They simply answer a different question from whether customers are paying enough to cover operations and reinvestment.

The reverse is also true. A declining cash balance can reflect deliberate repayment of debt rather than operating deterioration.

For each company, reconstruct the movement in cash before drawing a conclusion from the ending balance.

• Customer collections. Classification for this analysis: Operating; Interpretation: Evidence of cash received from the business.
• Supplier and employee payments. Classification for this analysis: Operating; Interpretation: Cost of running the business.
• Inventory build. Classification for this analysis: Working capital; Interpretation: Cash committed before the related sale or collection.
• Receivables growth. Classification for this analysis: Working capital; Interpretation: Revenue that has not yet become cash.
• Factory equipment. Classification for this analysis: Investing; Interpretation: Capacity that must earn a return over time.
• Asset sale. Classification for this analysis: Investing; Interpretation: Proceeds that may not recur.
• New debt. Classification for this analysis: Financing; Interpretation: Cash raised with a repayment claim.
• New shares. Classification for this analysis: Financing; Interpretation: Cash raised by sharing future ownership.
• Debt repayment. Classification for this analysis: Financing; Interpretation: Cash used to reduce a creditor claim.
• Share repurchase. Classification for this analysis: Financing; Interpretation: Cash used to reduce the outstanding share base.

These categories describe the economic distinctions we inspect. Exact financial-statement classification can depend on the transaction and accounting policy.

Working capital can hide inside a successful ramp

Consider a hypothetical supplier whose annual sales rise from $1 billion to $1.5 billion. If customers take 60 days to pay in both periods, receivables rise by roughly $82 million using a simple 365-day approximation.

That additional receivable is an asset, but it is not yet spendable cash. The company may have paid employees and suppliers before collecting it.

If inventory must also grow to support the higher sales rate, the business needs more funding even while its income statement improves.

Supplier payment terms can offset some of this need. Longer payable periods help cash flow temporarily, but they should not be mistaken for a permanent improvement in product profitability.

The practical question is whether working capital is growing in proportion to the business or becoming less efficient.

• Receivable days. Simple calculation: Average receivables divided by annual sales, multiplied by 365; Question it helps answer: Are collections slowing?
• Inventory days. Simple calculation: Average inventory divided by annual cost of sales, multiplied by 365; Question it helps answer: Is stock accumulating faster than production needs?
• Payable days. Simple calculation: Average payables divided by annual cost of sales, multiplied by 365; Question it helps answer: Is supplier financing supporting reported cash flow?
• Cash conversion cycle. Simple calculation: Receivable days plus inventory days minus payable days; Question it helps answer: How long is cash tied up in operations?
• Capital expenditure intensity. Simple calculation: Capital spending divided by revenue; Question it helps answer: How much reinvestment accompanies the growth?
• Cash conversion. Simple calculation: Operating cash flow compared with operating performance; Question it helps answer: Does accounting profit translate into cash over time?

These are approximate screening measures. Use consistent periods, averages where available and adjustments for acquisitions or major changes in the business.

A single quarter can be distorted by shipment timing. A persistent deterioration across several periods deserves a stronger response.

Adjusted earnings need an ownership check

Adjusted earnings can help isolate ongoing operations from unusual accounting charges. They are less useful when treated as a substitute for understanding the excluded items.

For a manufacturer, acquisition-related amortization may differ from current cash spending. But acquiring a business still required someone to provide cash or shares.

Stock-based compensation can be noncash in the current period while reducing existing shareholders' participation. Ignoring the share count would miss that economic cost.

Do not subtract stock compensation twice by fully charging it and independently assuming the same dilution again without reconciling the model. The objective is consistent treatment of the shareholder claim.

• Acquisition amortization. Useful question: What cash or shares were paid to obtain the acquired earnings?
• Stock compensation. Useful question: How does the diluted share count evolve?
• Restructuring. Useful question: Is this exceptional, or does a similar charge recur regularly?
• Debt conversion charge. Useful question: What changed in the creditor and shareholder claims?
• Asset impairment. Useful question: Which earlier investment failed to earn its expected return?
• Disposal gain. Useful question: How much of current profit will disappear next period?
• Tax adjustment. Useful question: What cash tax rate is reasonable for future operations?

Lumentum's reconciliation makes this discipline especially relevant. The earlier section gives the reported figures; the next step is understanding the share and debt consequences, not choosing whichever earnings measure makes the valuation look cheaper.

Build the revenue model from physical units

The most useful market model begins with the product a supplier invoices. It does not begin by assigning a percentage of total AI capital expenditure to the company.

For an illustrative optical component business, use the following chain:

Deployed systems × eligible connections per system × optical adoption × supplier share × content value per connection = potential annual revenue.

Every term needs a unit and a time period. If the product is sold per module, convert the connection count into modules before applying the price.

A complete hypothetical example

Assume an annual deployment of 10,000 systems, each with 100 eligible connection endpoints. That produces one million eligible endpoints.

Assume 30% use the optical architecture being modeled. The relevant optical endpoint count is then 300,000.

Assume the supplier wins 25% of that content and receives $200 per winning endpoint. The resulting annual revenue estimate is $15 million.

That is far smaller than one million endpoints multiplied by $200, because adoption and supplier share both matter.

One million eligible endpoints become $15 million of revenue

1. 1,000,000 (10,000 systems × 100 endpoints) to 300,000 (Optical, at 30% adoption)
2. 300,000 (Optical, at 30% adoption) to 75,000 (Supplied, at 25% share)
3. 75,000 (Supplied, at 25% share) to $15 million (At $200 per endpoint)

If one module serves several endpoints, the module price must be allocated accordingly. The OIF external-source architecture shows why this is more than a theoretical concern.

If the supplier sells both a component and an assembled module, count only the external sale appropriate to each unit. Do not add the internal component value again.

Separate adoption from market share

A company can gain market share in a slowly adopting architecture. It can also lose market share in a rapidly expanding market and still report revenue growth.

Those outcomes imply different competitive positions. A revenue chart alone cannot distinguish them.

Two assumptions swing revenue from $2 million to $48 million

Pick an adoption rate and a supplier share

Optical adoption

10%30%60%

Supplier share

10%25%40%

Annual revenue$2 million$5 million$8 million$6 million$15 million$24 million$12 million$30 million$48 million

$2 million$48 million

This sensitivity analysis shows which assumptions deserve the most research. A small change to a dominant input can matter more than a precise forecast for a minor cost.

The model also needs replacement sales, spares or additional applications if those are material. Add them explicitly rather than hiding them inside an unexplained growth rate.

Define the calendar

A platform can be announced, sampled, qualified and deployed in different fiscal years. Revenue recognition may follow a different date from the customer's capital announcement.

Build a timeline that keeps these events separate. For an emerging product, the gap between milestones is part of the investment risk.

• Research publication. Evidence strength for near-term revenue: Shows a technical idea or result.
• Live demonstration. Evidence strength for near-term revenue: Shows operation under specified conditions.
• Product sampling. Evidence strength for near-term revenue: Shows customers can begin evaluation.
• Design selection. Evidence strength for near-term revenue: Shows intended use in a program.
• Qualification complete. Evidence strength for near-term revenue: Removes a stated adoption hurdle.
• Initial production shipment. Evidence strength for near-term revenue: Establishes early commercial delivery.
• Repeat volume orders. Evidence strength for near-term revenue: Gives stronger evidence of durable demand.
• Collected cash. Evidence strength for near-term revenue: Establishes payment for the commercial activity.

The sequence is not rigid for every product. Its purpose is to prevent a model from treating an early milestone as though all later ones are complete.

Intel's June 2024 optical I/O announcement explicitly described the demonstrated chiplet as a prototype at that time. A working optical link was real evidence, but the announcement did not establish a mature sales stream.

Marvell's March 2026 coherent interconnect release likewise specified future sampling for newly announced products. Those products should not inherit the commercial status of older products sold by the same company.

A larger factory needs an incremental return

Revenue growth is not sufficient if every additional dollar of sales requires too much new capital.

An incremental-return model asks what the next expansion earns relative to the money required to build and operate it.

The following model is hypothetical and uses annual operating figures. It is intended to reveal the assumptions, not to estimate any of the five companies.

• Incremental annual revenue. Hypothetical amount: $500 million.
• Incremental operating margin. Hypothetical amount: 20%.
• Incremental operating profit. Hypothetical amount: $100 million.
• Assumed tax rate. Hypothetical amount: 25%.
• Incremental after-tax operating profit. Hypothetical amount: $75 million.
• New plant and equipment investment. Hypothetical amount: $600 million.
• Additional working capital. Hypothetical amount: $150 million.
• Total incremental invested capital. Hypothetical amount: $750 million.
• Simple incremental return. Hypothetical amount: 10%.

The final result is $75 million divided by $750 million. This simplified calculation does not replace a discounted cash-flow model or capture the full timing of the ramp.

If the operating margin is only 12%, after-tax operating profit falls to $45 million and the same capital earns 6%.

If capital required rises to $1 billion while the original operating assumptions hold, the simple return falls to 7.5%.

Every 4 points of margin moves the return on new capital by 2 points

12% operating margin6.0%

16% operating margin8.0%

20% operating margin10.0%

24% operating margin12.0%

28% operating margin14.0%

This makes the central manufacturing question measurable: can the company sustain margins and throughput high enough to justify the capacity being added?

A subsidy or customer investment can reduce the supplier's own funding burden, but the project still consumes resources. Analyze the economic return and the financing structure separately.

Equity funding also shares the outcome with new owners. A strong project return does not automatically produce the same percentage gain in earnings per existing share.

A valuation test that survives a good earnings report

More revenue can coexist with a disappointing stock return. Here is a hypothetical supplier, with annual revenue starting at $1 billion and a gross margin of 40%.

Assume shipments rise 40%, but average selling prices fall 15%. Revenue rises only 19%:

1.40 × 0.85 = 1.19.

If gross margin then falls to 35%, gross profit rises from $400 million to $416.5 million, an increase of about 4.1%. Research costs, factory spending and working capital still need funding. This is an illustrative sensitivity, not a forecast for any company above.

Shipments up 40%, gross profit up about 4%

−20−1010203040%0

+40.0%

−15.0%

+19.0%

+4.1%

ShipmentsSelling priceRevenueGross profit

Valuation can add a second disappointment. Suppose earnings per share grow 30% over a year, while the price-to-earnings multiple falls from 40 to 28. The implied price change, excluding dividends, is:

1.30 × (28 ÷ 40) − 1 = −9%.

Earnings grew 30% and the share price still fell 9%

Starting price100

EPS +30%+30= 130

P/E 40× to 28×−39= 91

New price91

The business improved and the share price fell. The initial price required more than the business delivered.

For a real valuation, build three cases and keep the assumptions visible:

• Adoption. Conservative case: Qualifications or deployments slip; Central case: Announced programs ramp gradually; Strong case: Multiple customers expand orders.
• Supplier economics. Conservative case: Price cuts offset much of the volume gain; Central case: Product mix supports margins; Strong case: Better yields and differentiation lift margins.
• Cash conversion. Conservative case: Capital spending and inventory absorb profit; Central case: Cash flow improves after the buildout; Strong case: Existing capacity supports profitable growth.
• Shareholder outcome. Conservative case: Dilution and a lower multiple reduce value; Central case: Returns depend on the purchase price; Strong case: Per-share earnings grow with limited dilution.

Use current filings to supply the numbers. Compare enterprise value, which adds debt and other financing claims to equity value and subtracts cash, with operating earnings or cash flow before debt payments. Compare equity value with earnings or cash flow available to shareholders. For a diversified company, value the other businesses separately before deciding what the market is charging for optics.

We are not assigning price targets in this issue. The evidence supports a differentiated watchlist; it does not establish that any of these five stocks is cheap at the research cutoff.

Reverse the valuation before choosing a stock

Instead of beginning with a target price, begin with the earnings a proposed purchase price requires.

This helps separate enthusiasm about photonics from the financial expectations embedded in a stock.

A required-earnings calculation

Suppose a hypothetical stock costs $100. An investor evaluating a three-year holding period wants a 10% annual price return and assumes no dividends.

The required price after three years is $133.10: $100 × 1.10³.

If the expected exit multiple is 25 times earnings, the business must earn approximately $5.324 per share in the final year to support that price.

If current earnings are $3 per share, the required growth to $5.324 is roughly 21.1% annually over the three years.

A lower exit multiple requires more earnings. The following cases hold the $133.10 required exit price constant.

A lower exit multiple demands much faster earnings growth

20× exit, needs $6.655 EPS30.4%

25× exit, needs $5.324 EPS21.1%

30× exit, needs $4.437 EPS13.9%

35× exit, needs $3.803 EPS8.2%

This does not tell us which multiple will prevail. It tells us how much the investment depends on a generous one.

A stock can look attractive at 35 times future earnings and unattractive at 20 times. That difference is an assumption to defend, not an error to hide in a spreadsheet.

For a photonics supplier, convert the required EPS into revenue, operating margin, interest expense, taxes and diluted shares. Then compare the resulting sales requirement with the physical-unit model.

If the implied volume exceeds the supplier's plausible capacity or market share, revise the valuation rather than automatically raising the growth forecast.

Dilution can consume operating progress

Assume a hypothetical company earns $300 million on 100 million shares. Earnings per share are $3.

If profit rises 30% to $390 million while the share count rises 10% to 110 million, earnings per share become about $3.545.

Per-share growth is approximately 18.2%, well below the 30% increase in company profit.

10% more shares turn 30% profit growth into 18% EPS growth

+30%Net income$300 million to $390 million

+10%Share count100 million to 110 million

+18.2%Earnings per share$3 to about $3.545

• Net income growth 20%. Share-count growth: 0%; Resulting EPS growth: 20.0%.
• Net income growth 20%. Share-count growth: 10%; Resulting EPS growth: 9.1%.
• Net income growth 30%. Share-count growth: 0%; Resulting EPS growth: 30.0%.
• Net income growth 30%. Share-count growth: 10%; Resulting EPS growth: 18.2%.
• Net income growth 30%. Share-count growth: 20%; Resulting EPS growth: 8.3%.
• Net income growth 40%. Share-count growth: 20%; Resulting EPS growth: 16.7%.

EPS growth equals one plus net income growth, divided by one plus share-count growth, minus one. These are hypothetical cases.

That is why manufacturing funding, acquisition consideration and stock compensation belong in the same investment model as the optical sales forecast.

A company can make rational financing decisions and still produce a lower per-share outcome than an investor expects. The purchase price must reflect the claim actually being acquired.

Five holdings can still be one concentrated trade

Owning five tickers does not necessarily diversify the underlying source of demand.

A portfolio of NVIDIA, Broadcom, Marvell, Coherent and Lumentum can have several different business models while remaining sensitive to the same AI infrastructure spending cycle.

It can also contain both sides of a purchasing relationship. Lower component prices may help a systems company while reducing a supplier's margins.

These relationships can create partial offsets, but they are not guaranteed hedges. A broad reduction in customer spending can hurt several layers simultaneously.

Five tickers, one spending cycle

1. AI infrastructure spending (The shared driver) to NVIDIA (Systems buyer of optics)
2. AI infrastructure spending (The shared driver) to Broadcom, Marvell (Networking silicon)
3. NVIDIA (Systems buyer of optics) to Coherent, Lumentum (Lasers and optics): Buys components

Map the common risks

• AI spending slows. Possible effect across the group: Fewer new systems and network upgrades; Why the effects differ: Revenue mix and customer contracts differ.
• Optical adoption is delayed. Possible effect across the group: Future optical programs shift out; Why the effects differ: Existing electrical and pluggable businesses may persist.
• Laser supply expands rapidly. Possible effect across the group: Prices may come under pressure; Why the effects differ: Buyers and manufacturers sit on different sides of the transaction.
• Packaging remains constrained. Possible effect across the group: Complete systems may ship more slowly; Why the effects differ: Some firms control more of the assembly process.
• A major customer changes design. Possible effect across the group: Existing content can be displaced; Why the effects differ: Alternative products and customer diversity matter.
• A broad valuation reset occurs. Possible effect across the group: Share prices can fall despite sales growth; Why the effects differ: Starting multiples and financing needs differ.
• Manufacturing yields improve. Possible effect across the group: More usable output reaches customers; Why the effects differ: Cost savings may be retained or passed through.

These are scenarios, not predictions about simultaneous company performance.

The portfolio question is which risks are being repeated. A second supplier with the same end customer and product cycle may add less diversification than its different ticker suggests.

Separate technology conviction from valuation conviction

A strong belief that more links will become optical is a technology view. A belief that a specific stock offers an attractive return at its current price is a valuation view.

The second requires more inputs: competitive share, margins, capital needs, timing and the purchase price.

Record those views separately so a successful technical milestone does not automatically override a disappointing financial result.

• A product passes qualification. Technology conclusion: A technical obstacle has been reduced; Valuation conclusion: Revenue and margin assumptions still need confirmation.
• Initial orders arrive. Technology conclusion: Commercial adoption has begun; Valuation conclusion: Order size may be too small to affect total earnings.
• Revenue beats expectations. Technology conclusion: Demand or execution was stronger; Valuation conclusion: The share price may already require more growth.
• Gross margin improves. Technology conclusion: Mix or production economics improved; Valuation conclusion: Capital spending and dilution still affect returns.
• A new competitor qualifies. Technology conclusion: The market has another supply source; Valuation conclusion: Pricing and share assumptions may need revision.
• A stock price falls. Technology conclusion: No technical conclusion by itself; Valuation conclusion: Expected returns improve only if business value holds.

This separation is especially useful after a dramatic market move. Price action can change prospective returns without settling the technical argument.

A quarterly research sheet for the five names

The aim of a monitoring sheet is to make changed evidence visible. It should not become a collection of favorable quotations.

For each company, keep the original assumption, the latest observation and the resulting model change on the same line.

“Management sounded positive” is not a model change. A revised shipment date, margin assumption or cash requirement is.

Every review line should end in a model input

1. Original assumption (Written before results) to Latest observation (What actually happened)
2. Latest observation (What actually happened) to Model change (A revised date, margin or cash need)

• NVIDIA. Record at each review: Photonics product status and customer deployment evidence; What would require revising the model: Availability without meaningful deployment progress.
• NVIDIA. Record at each review: Networking economics in the wider platform; What would require revising the model: Benefits that do not support customer adoption or pricing.
• Broadcom. Record at each review: Optical products separately from total AI semiconductor revenue; What would require revising the model: Treating accelerator growth as proof of optical growth.
• Broadcom. Record at each review: Customer adoption and competing connectivity designs; What would require revising the model: Loss of relevant content despite industry expansion.
• Marvell. Record at each review: Existing connectivity revenue separately from emerging programs; What would require revising the model: A future program being used to explain current reported sales.
• Marvell. Record at each review: Changes to the stated Photonic Fabric milestones; What would require revising the model: Delays that change discounted value or funding needs.
• Coherent. Record at each review: Operating cash flow and property and equipment additions; What would require revising the model: A larger or longer cash shortfall than the model allows.
• Coherent. Record at each review: Qualified output and manufacturing commentary; What would require revising the model: Capacity additions failing to translate into saleable output.
• Lumentum. Record at each review: Components, systems and new-product evidence; What would require revising the model: A new product offset by deterioration elsewhere.
• Lumentum. Record at each review: Earnings reconciliation and diluted ownership; What would require revising the model: Operating growth that fails to reach per-share results.

The sheet should retain negative observations. Deleting old assumptions after a miss prevents an honest assessment of the thesis.

A practical update can be brief: the original expected milestone, what actually happened, whether the difference is temporary, and which valuation input changes.

Predefine the response to mixed evidence

Several outcomes can coexist: revenue rises, gross margin improves, inventory grows and free cash flow deteriorates.

Do not force those facts into a single “good quarter” or “bad quarter” label. Decide which one is most important for the current stage of the thesis.

For an early manufacturing ramp, temporary cash absorption may be planned. For a mature operation valued on strong cash generation, the same absorption may be more concerning.

The response should depend on the assumptions written before the results arrived.

• Sales and cash improve together. Provisional interpretation: Stronger commercial and financial evidence; Next question: Is the improvement repeatable?
• Sales rise, inventory rises faster. Provisional interpretation: Growth may require more funding; Next question: Is the inventory qualified and linked to customer demand?
• Margins rise, capital spending surges. Provisional interpretation: Better operations with heavier reinvestment; Next question: What return will the new capacity earn?
• Orders rise, shipments slip. Provisional interpretation: Demand may exceed execution capacity; Next question: Which production stage is constraining delivery?
• A product delay accompanies stable existing sales. Provisional interpretation: Future value falls while current operations may hold; Next question: How much of the valuation depended on the delayed program?
• Earnings rise, diluted EPS lags. Provisional interpretation: Ownership expanded; Next question: Was the capital raised used productively?
• Cash flow improves after payables rise. Provisional interpretation: Timing may be supporting cash; Next question: Is supplier financing sustainable?

The purpose is to change the investment view when the evidence changes. A thesis that can explain every outcome after the fact cannot help allocate capital before the next one.

The evidence that would change our view

Our current judgment favors researching suppliers with named products, identifiable customers and a measurable route to cash generation. Coherent and Lumentum provide more direct optical exposure than NVIDIA or Broadcom; Marvell adds a distinct future adoption bet. Direct exposure increases sensitivity to the theme, including its disappointments.

The next review should test five things:

1. Deployment: Do product announcements become qualified shipments and repeat orders? Production availability alone cannot settle demand.
2. Content: Does the supplier earn more gross profit per deployed system after the architecture changes? Count displaced products as well as new ones.
3. Cash: Over several quarters, does operating cash flow cover the capacity buildout? A single inventory-heavy quarter is weaker evidence than persistent deterioration.
4. Pricing: Does added capacity preserve margins or transfer the benefit to customers? Lower prices help adoption while potentially hurting suppliers.
5. Ownership: Do earnings and cash flow improve per diluted share? Financing a successful expansion can still reduce each shareholder's participation.

Marvell's announced October 6 investor day is a specific checkpoint for its optical roadmap. For the laser suppliers, upcoming results should be checked against their own manufacturing and product commentary, rather than an assumed industry-wide adoption date.

We would reduce confidence if orders repeatedly slip while capital spending continues, if new optical sales merely replace higher-profit legacy content, or if healthy unit growth fails to reach cash flow per share. We would increase confidence when repeat customer orders, qualified output and cash conversion improve together.

What would move our view on the suppliers

Reduce confidence if

• Orders slip while capital spending continues
• New optical sales replace higher-profit legacy content
• Unit growth fails to reach cash flow per share

vs

Increase confidence when

• Repeat customer orders arrive
• Qualified output rises
• Cash conversion improves alongside both

AI's growing need for bandwidth gives photonics a credible demand case. The investment case rests on the price paid for the portion of that demand a company can profitably serve. NVIDIA's commitment makes the research worth doing. The supplier's cash generation determines whether the stock deserves the capital.