Another ticker. Same wallet.
I have been wondering whether I leaned too heavily into nuclear while memory kept moving. Then Tower Semiconductor landed on the reading list, which is a very efficient way to turn one allocation question into three.
There is a useful version of this restlessness: revisit the portfolio, find what has changed, and discover a better opportunity. There is also the version where I buy whatever would have made last week’s account look smarter.
I would prefer to know which one I am doing before placing the order.
Mizuho initiated Tower with an Outperform rating on Friday, September 25.[1] That prompted the closer look. What keeps my attention is the underlying business: a specialty chip manufacturer with a role in moving data through AI infrastructure using light.
Issue 007 followed the physical constraints underneath AI. Issue 013 asked how much of that progress becomes shareholder value. Optics brings those questions together neatly.
The interesting work is finding where the economics leave room for me.
Big indexes up. Small caps down.
Index returns for the week ending September 25, 2026.[2]
The large indexes recovered while small caps finished lower. My read: investors remained willing to pay for growth, but the rally was uneven. Four index returns cannot tell me which business deserves more of my money.
Fast chips still have to talk.
An AI system needs processors to do the work, memory to keep information close, and connections that let the parts exchange data. A processor waiting for information is expensive equipment spending part of its day waiting.
Optical connections use light to move that information. Silicon photonics puts functions that handle light onto chips. The appeal is more bandwidth within practical power and space limits. Making those parts work together reliably—and manufacturing them economically—is the difficult bit.[4][5]
Tower is a foundry: it manufactures chips for customers. That puts it in a different position from a company selling network equipment or a supplier assembling optical components into a finished product.
| Company | What it sells into the theme | What needs to pay off |
|---|---|---|
| CienaCIEN[3] | Network systems, optical transport, and interconnects that move data within and between facilities. | Customer spending must turn into profitable deliveries. Watch the mix of products sold and the margin retained. |
| CoherentCOHR[4] | Optical components and integrated assemblies that generate, guide, and receive light. | Selling more of the finished connection must improve profit, rather than simply add work and manufacturing costs. |
| TowerTSEM[5] | Specialized chip manufacturing, including silicon photonics. | Factories must stay productively busy and earn enough to justify the cost of expanding them. |
The theme tells me why customers might spend. It does not tell me which supplier keeps the most attractive share of the profit—or which stock already charges me for it.
Coherent has a nearer checkpoint.
On September 21, Coherent introduced PhotonLink, its integrated optics platform for AI infrastructure. Management expects revenue from the platform to begin ramping in the fourth quarter of calendar 2026.[4]
That gives the story a relatively near date to test. The announcement is evidence of a product plan; the expected ramp is still a forecast.
The next questions are whether the revenue arrives, how much it contributes, and what Coherent earns on it. Becoming more useful to a customer is promising. Getting paid adequately for the extra work is where the shareholder joins the celebration.
Tower has shipments and a construction bill.
On September 17, Tower and NewPhotonics announced high-volume shipments of laser-integrated optical-engine chips for high-speed AI connections.[6] Tower’s existing business is already shipping products. The whole case does not wait for a future factory.
Its second-quarter report, released in August, showed $460 million in revenue, up 24% year over year. The cash figures are just as useful.[7]
Quarter ended June 30, 2026. Calculation uses Tower’s non-GAAP free-cash-flow definition; it is not a net-income figure.[7]
Substantial cash came in from operations. A little more went back into equipment. That can be a sensible trade when the investment earns attractive future returns.
The other half of the picture: first-half free cash flow was positive $344 million, calculated from $687 million of operating cash flow less $343 million of net equipment investment. Operating cash flow benefited from a $283 million net increase in customer advances.[7] Those payments help fund the build, but the associated customer commitments still have to be served.
Both periods matter. One negative quarter does not establish a broken business; advance payments do not establish recurring annual cash generation.
The next expansion is substantial. In July, Tower outlined a two-track Japanese capacity plan with approximately $3 billion of company investment after $1 billion in grants. The first track targets full production readiness in Q4 2027; the second includes a new facility subject to completing related agreements.[8]
These are plans and target dates. The return depends on the cost of getting capacity ready, the demand that shows up, and the economics of the work performed there.
Give the opportunity a price.
The strongest version of the Tower case is appealing: specialized manufacturing becomes harder to replace as customers qualify products, demand fills the new capacity, and profit grows faster than the money needed to support it.
The less flattering version is plausible too. Spending arrives on schedule, customer demand arrives late, and the factory earns less than the stock price assumed.
Those are scenarios to test. I have not completed a valuation that tells me which one today’s price is paying for.
So the next step is a side-by-side estimate for TSEM, COHR, and CIEN: revenue that can realistically arrive, margins after production costs, cash required for growth, and the debt and share count that determine what remains for each shareholder.
Then make the assumptions less comfortable. If slower demand or weaker pricing wipes out the prospective return, the headline opportunity is doing too much of the work.
Evidence that customer commitments support profitable capacity use, with funding terms and a share price that leave room for delays.
An expansion whose returns disappoint under modestly weaker assumptions, or a valuation that requires almost everything to go right.
Tower deserves the comparison. It has not yet earned a buy conclusion. The test includes another dollar in memory or nuclear, and keeping the money available. An existing holding gets no exemption just because I know its story.
Two dates that can change the math.
BWXT and Micron were already on last issue’s calendar. Now the useful preparation is deciding which answers would change an estimate.
8:30 a.m. Eastern
Connect capacity spending to delivery schedules and cash generation. A stronger plan could support the business case; position size still needs a separate check against total nuclear exposure.
4:30 p.m. Eastern
Test high-bandwidth-memory commitments, pricing, margins, and required spending. Estimate what earning power remains if supply grows or pricing becomes less generous.
A compelling memory opportunity may still exist at a higher price than the one I remember. An encouraging nuclear presentation may still leave the position too large. The new evidence and the price have to be considered together.
The thesis board, heading into the week.
Compare TSEM, COHR, and CIEN on economics and valuation. More tickers alone do not diversify a shared dependence on AI spending.
Use BWXT’s updated plan to revisit both expected returns and portfolio concentration.
Use Micron’s report to distinguish durable earning power from unusually favorable conditions.
I may need a better allocation.
I do not need a busier watchlist.
The worry behind this issue is real: maybe I have leaned too heavily into a familiar theme while better opportunities developed elsewhere.
That deserves work. But adding three optical stocks could leave me with more company names and more dependence on the same AI infrastructure spending. A longer holdings list can still have a very short list of things that must go right.
Then there is the feeling that I need to catch up because I have not watched the market constantly.
I understand that feeling. I would not put it in charge of the brokerage account.
Nor should “be patient” become the answer before I have done the comparison. There may be a better allocation available. It very well may involve buying something that has already risen.
I want enough evidence to change my mind—and enough honesty to admit when the main thing that changed was my mood.
Portfolio accounting: no refreshed brokerage snapshot accompanies this edition. Current weights, account returns, and new transactions are not reported here. The thesis board describes research priorities.
The connection is real.
The return needs work.
Optics gives this issue a useful new question: which part of moving AI’s data can earn attractive returns for shareholders after the cost of serving that demand?
Tower brings manufacturing into the comparison. Coherent supplies a nearer product-revenue checkpoint. Ciena keeps the network-equipment economics in view. None gets selected by the theme alone.
The next dollar has homework: compare those businesses at their prices, incorporate what BWXT and Micron report, and decide whether the portfolio can actually be improved.
More homework is useful only if it eventually produces an answer.
— Alex