The Pipes
Are Full.
Who Pays?

Last week was the allocation. This week was the exam.

Issue 021 moved capital toward companies collecting real revenue from technological bottlenecks. Seven positions out. A 23% bandwidth sleeve in. Larger allocations to memory and nuclear. I wanted the companies collecting from the constraint, not the companies with the most extravagant predictions about it.

This week did not argue with that thesis. It argued with the price of it.

The businesses supplying AI infrastructure may be getting stronger while the financial environment underneath their valuations gets more hostile. Higher yields, $100-plus oil and a sudden argument over how AI companies count revenue can all coexist with enormous infrastructure demand. That matters more for this book than for most, because I deliberately concentrated it in memory, optics, nuclear and power.

The rebalance still makes sense. It just met a harder test sooner than I expected. I spent last week deciding which toll booths I wanted to own. This week reminded me that somebody still has to finance the highway.

The invoice went up the week the data went soft.

Friday, October 2, brought a +29,000 jobs report and a market that decided the Fed could relax. Monday, ISM's services prices index printed 74.0, the highest since July 2022, and the 10-year touched about 5.35%, its highest since April 2002.

Wednesday supplied the clearest receipt. The Treasury sold $39 billion of 10-year notes at 5.300%, the highest 10-year auction yield since 2000. Demand was strong anyway: bid-to-cover reached 2.77; indirect bidders took roughly 80%; dealers were left with about 2.5%. That matters. This was not an auction nobody wanted. Investors were willing to lend for a decade. They simply demanded a substantial yield to do it.

An hour later, the September FOMC minutes landed. Most officials thought another hike would likely be appropriate by year-end. They did not pick October or December.

Then the line that matters most for this book. A few participants named increased expectations for AI-related borrowing among the factors possibly contributing to higher long-term yields. The Desk also noted commentary about heavy private-debt issuance financing AI infrastructure. That does not prove AI financing caused the rise. It establishes that the scale of the buildout is large enough to enter the bond-market discussion.

The Fed staff attributed most of the rise in longer-term yields to real rates, and only a few officials raised the AI point. AI borrowing has entered the Fed’s explanation; it has not been established as the driving cause. Broadcom is reportedly in early-stage discussions about financing exceeding $50 billion for OpenAI-related custom chips. Those are talks, not completed funding. But the scale makes the tension easy to see.

The buildout is now borrowing from the same pool of capital that prices the buildout's stocks.

By Friday the 10-year had eased to about 5.24%, slightly below its roughly 5.28% starting level. The S&P 500 rose 1.2%, the Nasdaq 0.6%, and the Dow 0.9%. The Russell 2000 fell 0.9%. Brent finished near $105. A high auction yield and a lower week-ending Treasury yield are not contradictory. They describe different points in a volatile week.

The market is still willing to pay for established earnings and credible growth. It is getting less generous toward anything that needs cheap financing or several more years before the cash shows up. That line runs straight through my Core, Growth, Speculative and Moonshot buckets.

The pipes are full.

The strongest receipt of the week came from a company I don't own.

Lumentum CEO Michael Hurlston told Bloomberg on Friday that its optical-component capacity was fully committed through early 2029. He described an inability to meet about 70% of demand for certain products through 2027, and 30% for others through 2028. Six months ago, he was discussing capacity selling out through 2028. That is a notable change in the demand horizon.

TSMC’s September revenue increased 54.6% year over year. At its investor day, Marvell described a market opportunity in optical interconnect expanding at roughly 60–70% annually. One company is reporting revenue, another is forecasting an industry, and a third is describing capacity constraints. They are useful signals, but they are not the same kind of evidence.

The bottleneck is not imaginary. Demand for bandwidth is running years ahead of the capacity to supply it.

And yet my 23% bandwidth sleeve returned roughly +0.08% for the week. Ciena rose almost 15%. Coherent fell 7.3%, Celestica 6.5% and Tower 5.0%. Three of four constituents finished lower in the week the industry's loudest supplier said it had nothing left to sell. Lumentum itself ended the week up 1.7%.

Toll booths can be in exactly the right place and still be priced for more traffic than arrives this quarter. A shortage may support pricing. It may also provoke supply additions, customer redesigns, and expensive capacity expansion. What counts is how much durable cash flow reaches each shareholder.

Decision: Hold CIEN, COHR, CLS and TSEM. No automatic addition, and no fifth optical position. Celestica’s October 27 results should give the first direct operating test of this sleeve.

Price is not a receipt.

Two of my names had huge Tuesdays. Neither announcement represented a new, signed revenue-producing customer order.

BWXT climbed nearly 8% Tuesday after Prodigy Clean Energy and its First Nations partners selected the BANR microreactor for a proposed transportable plant in Belledune, New Brunswick. A provincial letter of intent supports prospective power purchases, but it is not a definitive electricity contract. The plant targets the early 2030s. That is potentially valuable commercial optionality. It is not revenue recognized today.

Ciena rallied about 14% Tuesday alongside Marvell’s expansive market-size presentation. That is a third party describing a bigger pond, not a customer handing Ciena a purchase order. The rally can be rational without being independent proof that Ciena’s margins or bookings improved that day.

I’m happy to own both moves. I’m not willing to count the stock-price gains themselves as new operating evidence. When the 10-year is around 5.3%, the market paying 8% and 14% for prospective opportunities tells me how much future success may already be priced in.

BWXT is still the kind of business where showing up with a pitch deck does not qualify you to manufacture naval reactor fuel. That is a conclusion about competitive position. Buying it at any price would be a conclusion about expected returns, and those are not interchangeable.

One regulatory milestone I missed in Issue 021 deserves to be added: on September 28, the NRC issued TVA a construction permit for a GE Vernova Hitachi BWRX-300 at Clinch River, Tennessee. It was the first U.S. construction permit for that reactor design. A permit is not a new sales order, and an operating license is still needed before the unit can run. But it is a concrete step toward deployment.

Decision: Hold BWXT at its 22% October target, with no addition after the rally. Hold Baker Hughes and GE Vernova. Physics has apparently declined the invitation to become a software problem.

Who pays?

Wednesday the minutes asked who finances the buildout. Thursday the market asked who is paying for it.

Reports placed OpenAI’s annualized revenue near $50 billion at the end of September, rather than the roughly $70 billion figure that had circulated earlier. Axios and Bloomberg explained that partner-sales accounting and attempts to compare OpenAI with Anthropic were central to the confusion. Bloomberg separately reported that OpenAI expects to reach or exceed a $70 billion annualized run rate by year-end. Those are different dates and potentially different presentation methods; they are not evidence of an overnight $20 billion collapse in sales.

The coverage did not establish that actual revenue had declined. The market sold the uncertainty anyway. Thursday, Oracle fell about 6%, CoreWeave about 8% and Nvidia about 3%. In my book, Coherent fell 9.6%, Tower 7.9%, Micron 4.8%, Ciena 4.6% and Celestica 4.1%.

That is what the "three or four enormous stories" warning in Issue 021 looks like on a single afternoon. A reporting argument at one AI lab hit optics, memory and manufacturing simultaneously.

My suppliers do not require a precise OpenAI revenue headline to be correct this quarter. They require customers willing and able to fund sustained procurement at acceptable economics. AI demand could continue growing while higher financing costs compress supplier valuations. That is the uncomfortable math: demand growth is not the same thing as a shareholder return.

Memory: record profits, falling stocks.

Samsung’s preliminary third-quarter guidance projected about ₩107.4 trillion in operating profit and ₩195 trillion in revenue, both extraordinary absolute numbers. Different analyst surveys produced different expectations for the quarter; I’m not going to call the entire release a clean beat or miss. Shares fell roughly 2.4% on the day as investors questioned how durable the boom could be.

Micron fell 4.3% for the week. Two items there were costs, not catastrophes.

First, Micron reached an agreement with Netlist: $600 million over five years, paid at $30 million per quarter, for a license that includes high-bandwidth memory intellectual property and settles the identified legal proceedings. That reduces litigation uncertainty. It also introduces a real recurring cost. The HBM profit pool has more claimants than memory shareholders.

Second, Micron’s Taoyuan union secured authorization to strike, with about 99% of votes cast favoring the action. No strike date has been set, and the report did not establish any production stoppage. The union has sought an employee profit-sharing arrangement pegged to 15% of operating profit and has discussed a possible October 19 demonstration. A demonstration is not a strike, and authorization is not an actual shutdown.

The stock market doesn't pay you for a good quarter. It pays you for what the quarter implies relative to what everyone already expected. When everyone expects extraordinary, extraordinary becomes the admission price.

When Micron reports non-GAAP gross margins around 87%, workers notice. So do patent holders, competitors and customers. That is not automatically a thesis break. It is a margin question with a calendar. I’m watching contract prices, HBM capacity additions, labor negotiations and the rate at which unusually high margins might normalize.

Decision: Maintain Micron’s 14% target. No automatic dip purchase. The burden of proof is rising.

The scarce link moved down a layer.

In Issue 021, Seagate punched me the day I wrote it down. This week the fight moved underneath it.

Bloomberg reported that Seagate and Toshiba are both bidding for TDK's hard-drive head business. TDK is the only large independent head supplier, and Toshiba depends on it. Toshiba denied any such talks, and TDK said no decision had been made. Seagate fell 7.8% for the week.

Whether or not a transaction happens, the logic is useful. Expansion in finished drives requires the specialized components inside them. If a small number of suppliers control those components, capacity economics can be decided one layer below the brand name on the drive. This is a research lead, not confirmation that the reported bidding contest will produce a deal.

The Micron labor dispute is a reminder that bottlenecks are not always pieces of silicon. The scarce input in a scarce supply chain can also be the people.

Decision: Hold Seagate’s 3% target. No panic sale on a transaction report Toshiba has denied.

ASTS goes to the front of the line.

AST SpaceMobile fell 10.5% Friday and 12.8% for the week, the worst move in the book.

SpaceX agreed to acquire Grain Management’s nationwide 800 MHz spectrum portfolio, encompassing up to 14 MHz of low-band spectrum, subject to FCC approval. The companies did not disclose financial terms; the approximately $8 billion figure is attributed to a Wall Street Journal report. SpaceX framed the acquisition as strengthening Starlink Mobile’s ambition to become a larger mobile carrier. Its stock exposure in this book rose roughly 2.3% over the week.

I own both, which is the interesting part. Owning two companies in the same revolution doesn't mean their interests are aligned. Sometimes one company's progress is the other's problem. Rocket Lab is a different exposure: launches and space systems, not a fight over phone spectrum. Its 7.7% weekly drop doesn't undo the contract behind its October seat.

ASTS is 2% of the portfolio, so the dollar damage was small. The thesis damage could be larger. The question isn't whether AST has spectrum; it has its own holdings and its carrier partners'. The question is whether a better-funded competitor with satellites, a launch business and a proposed new spectrum position changes what AST's partners are willing to pay for.

So ASTS becomes the first position to re-underwrite from scratch: spectrum access, carrier economics, launch and deployment timing, financing requirements, and what a stronger rival could mean for pricing.

Decision: No panic sale, but no averaging down. Research first. A 2% weight limits portfolio damage, not business-model risk.

The book.

The exact share-level ledger has not been reconciled for October 9. What follows is a price-only estimate, not a recorded closing value.

Taking Issue 021's October target weights and applying each stock's October 2-to-9 closing-price change, the book would be about $8,480, roughly flat from $8,463. On $5,270 of contributed capital, that would be about +60.9%. This is a price-only estimate with no share counts, fees or same-day drift. It is not a closing NAV, and I will not treat it as one.

The estimate still says something true about structure. BWXT, at 22%, added about $124. The other 21 positions combined subtracted about $110. One position turned a down week into a flat one.

That is what concentration looks like when it works. Issue 021 provided the other half of the math: a 30% BWXT drawdown would cost about 6.6 percentage points of the portfolio, all else equal. Same sizing, opposite week.

And the exposure is broader than any one ticker. Twenty-two companies do not mean twenty-two independent economic bets. A slowdown in data-center capital spending could hit optics, memory, networking, equipment and power valuations together. My risk is not just single-name concentration. It is shared dependence on the same investment cycle.

The October allocation targets

These are October target weights, not a claim that Friday’s market close left every holding exactly on target.

What I'm not doing.

• No chasing strength. BWXT is already 22%; a letter of intent isn't cash flow. No fifth optics holding or Lumentum chase with 23% already in the theme.
• No automatic dip-buy. Micron's decline isn't proof of a bargain, and Seagate's disputed supply-chain story isn't a reason to panic-sell.
• No rushed ASTS trade. Neither sell in fear nor average down until I've reassessed SpaceX's competitive threat.
• No expanding the speculative bets. No added quantum exposure, and no return to Archer, Joby or SoFi merely because they look cheaper.
• No trading just to trade. The October rebalance holds. Rubrik below $95 means revisit the research, not place an order.

And no confusing a four-dollar share price with a cheap business. Cheap and distressed are not synonyms. Neither are low-priced and undervalued. The next dollar still has homework.

The calendar.

• Tuesday, October 13: JPMorgan, Citi, Wells Fargo and Goldman open bank earnings. Credit appetite and what a hiking Fed is doing to lending.
• Wednesday, October 14, 8:30 a.m. ET: September CPI. Also ASML's third-quarter results, the first read on chip-equipment orders.
• Thursday, October 15: TSMC's third-quarter results and capex. September PPI at 8:30 a.m. ET.
• Monday, October 19: Possible Taoyuan union rally in Taipei.
• Tuesday, October 27: Celestica's third-quarter results plus Investor and Analyst Day. Baker Hughes reports after the close. Direct tests for two parts of the rebalance.
• Wednesday, October 28: GE Vernova’s earnings webcast is scheduled for 7:30 a.m. ET, and the FOMC concludes its meeting. Power demand meets the cost of capital on the same day.
• Monday, November 2: BWXT's third-quarter results. The largest position gets its next exam.

No need to invent catalysts. There are enough real ones coming.

I wanted the toll booths. Now I need to study the traffic.

There is something reassuring about optical suppliers reporting demand years into the future. It tells me the bandwidth bottleneck is real. It does not tell me every stock exposed to that bottleneck is cheap.

There is something much less reassuring about watching AI stocks sell off over a revenue definition while the Treasury pays 5.3% to borrow for ten years and the Fed's own minutes mention AI debt. It tells me the market has started to interrogate the bill.

Both things can be true. The buildout can be enormous, the suppliers can be genuinely valuable, and investors can still overpay for them. Demand can remain extraordinary while the return earned on that demand disappoints.

The October rebalance was never supposed to end the research. It was the point where research began to have larger consequences. This week gave me more confidence in bandwidth demand, and more reason to scrutinize valuations, customer economics, and who is financing the customers.

The next move is not necessarily another trade. Sometimes the most aggressive thing you can do after a large allocation is refuse to touch it until the next real piece of evidence arrives.

The pipes are full. Now I want to know who pays for everything flowing through them.

See you next Sunday.

— Alex

Source desk

• U.S. Treasury — October 7 ten-year auction results
• Federal Reserve — September 15–16 FOMC minutes
• The Wall Street Journal — October 7 Treasury auction
• Bloomberg / Japan Times — Lumentum optical capacity
• TSMC — September 2026 revenue
• Marvell — October 6 Investor Day
• BWXT — BANR reactor selection announcement
• GE Vernova — Clinch River NRC construction permit
• Axios — OpenAI revenue-accounting differences
• Bloomberg — OpenAI year-end revenue-run-rate expectations
• Samsung — preliminary Q3 2026 guidance
• SEC / Netlist — Micron license and settlement
• Focus Taiwan — Micron Taoyuan strike vote
• Bloomberg — Seagate, Toshiba and TDK report
• Grain Management — Spectrum agreement with SpaceX
• GE Vernova — October 28 earnings webcast
• Issue 021 — The Homework Had an Answer

Portfolio accounting: Approximately $8,480 is an indicative application of October 2–9 closing-price changes to the published October 2 target weights and $8,463 opening valuation. It is not a share-level reconciliation, and it assumes no intervening trades, contributions, withdrawals or fees.