Last week I assigned homework. This week I turned it in.
Issue 020 ended in an uncomfortable place.
Nuclear already had a lot of my capital.
Memory had earned another look.
Optical networking had become too important to leave sitting politely on a research board.
The easy response was obvious:
That is the seductive version of portfolio management. Every good idea gets a chair. Nobody ever has to leave the table.
The harder question was better:
It is easy to love Tower Semiconductor without selling anything. It is easy to say AI networking matters while keeping every old space proxy, fintech position, chip-IP name and flying-car experiment exactly where it was.
A watchlist can expand without consequence.
A portfolio cannot.
So Friday, after the September jobs report landed and the first wave of market volatility settled, the Frontier File portfolio changed.
Not cosmetically.
Actually changed.
Seven positions out. Six new names funded. Several conviction positions resized. The purpose was not to own more stories. It was to force every dollar to compete for its seat.
And the interesting part is what did not happen.
I did not make the portfolio more aggressive by finding worse companies.
I made it more aggressive by concentrating harder in bottlenecks I can explain.
That distinction is this issue.
Friday looked easy. The week was not.
September payrolls increased by only 29,000. Unemployment moved to 4.2%. July and August were revised down by a combined 60,000 jobs, and average hourly earnings were up 3.0% year over year.[1]
Markets liked the jobs report because another immediate Fed hike suddenly looked less necessary. The S&P gained 0.73% Friday and the Nasdaq rose 1.19%. For the full week, though, the S&P still lost 0.27%, the Dow fell 1.26%, and the Nasdaq gained only 0.45%.[2]
That rally also came one day after the 10-year Treasury touched roughly 5.34%, its highest level since 2002.[3]
Soft jobs changed the expected Fed path.
They did not make capital cheap again.
The future is still the trade.
The future also has rent.
Micron answered the question.
The memory trade stopped being hypothetical.
Last issue, I wrote down the question before the earnings report:
Then Micron reported $54.23 billion of quarterly revenue.
Previous quarter: $41.46 billion.
A year earlier: $11.32 billion.
Operating cash flow: $43.97 billion.
Non-GAAP gross margin: 87%.
Then management guided the next quarter to approximately $61.5 billion of revenue with non-GAAP gross margin around 86.25%.[4]
The numbers are ridiculous.
Usefully ridiculous.
The important part is that I did not see them Wednesday night and invent a thesis Thursday morning.
The thesis already existed.
The question already existed.
The report answered it.
MU target: 10% → 14%. Enough capital to reflect stronger evidence. Not enough capital to require extraordinary memory conditions forever.
Not 25%.
Not “this time is different.”
Not “memory only goes up now.”
Fourteen.
Enough to acknowledge that the evidence improved.
Not enough to demand perfection.
That is the kind of aggression I want more of.
I broke my own BWXT ceiling.
So I owe the explanation.
Earlier issues treated roughly 18% as the ceiling for BWXT.
Friday's redistribution moved the target to 22%.
That deserves more than a line in a portfolio table.
At its September 29 Investor Day, BWXT showed total backlog growing from roughly $4.0 billion at the end of 2023 to $8.4 billion by Q2 2026. Trailing-twelve-month book-to-bill reached 1.7x.[5]
That is contracted work and customer commitments — not merely a market-size slide.
Then Thursday brought another receipt: approximately $189 million to manufacture and deliver naval nuclear reactor fuel, with work scheduled through August 2027.[6]
The same week, Amazon and Constellation announced a 20-year power agreement at Calvert Cliffs covering 690 megawatts, including roughly 190 megawatts of new uprated capacity expected between 2030 and 2032.[7]
That contract does not pay BWXT.
It does something useful anyway.
It keeps proving that hyperscalers no longer consider electricity somebody else's problem.
AI keeps walking backward through the technology stack.
GPUs.
Memory.
Networking.
Power.
Eventually you reach turbines, uranium, fuel fabrication, grid hardware and concrete.
BWXT already lives in that world.
Now the uncomfortable part.
Twenty-two percent is enormous.
A procurement delay is no longer merely a BWXT problem.
An execution miss is no longer merely a BWXT problem.
A budget fight or valuation compression becomes a whole-book problem.
A 30% BWXT drawdown costs roughly 6.6 points of the portfolio. 22% position × 30% drawdown = approximately 6.6 percentage points.
I know that.
I sized it anyway.
This is not a safe position.
It is a Core business held at an aggressive size.
Those are different statements.
And from here, BWXT has to keep earning the privilege.
Twenty-two percent is not the beginning of another accumulation campaign.
It is where the burden of proof goes up.
Same traffic jam. Different toll booths.
Issue 020 gave Tower an interview.
Issue 021 gave the bottleneck money.
The largest structural change in the book is not BWXT.
It is this table:
| Company | Target | Role |
|---|---|---|
| CienaCIEN | 7% | Network systems and optical transport moving data inside and between facilities. |
| CoherentCOHR | 6% | Optical hardware that generates, guides and receives the light carrying that data. |
| CelesticaCLS | 6% | AI infrastructure hardware, systems design and manufacturing. |
| Tower SemiconductorTSEM | 4% | Specialty foundry manufacturing underneath silicon photonics. |
23% of the portfolio now targets the bandwidth bottleneck. Not four independent macro bets. Four different ways to collect from the same physical constraint.
That is not a watchlist anymore.
It is a thesis with consequences.
The common problem is simple.
AI processors can get faster faster than networks can move information between them.
Eventually the bottleneck changes.
Ciena sells the system.
Coherent sells the optics.
Celestica builds the hardware.
Tower manufactures pieces underneath it.
Tower stays the smallest on purpose.
On September 17, Tower and NewPhotonics announced high-volume shipments of laser-integrated optical-engine PICs designed for 800G-to-1.6T AI interconnect.[8]
Tower is also expanding silicon-photonics capacity in Japan, targeting full production readiness for the first expansion track in Q4 2027.[9]
That is much more interesting than:
“AI optics will be big.”
But huge demand and good technology do not automatically create great shareholder returns.
Tower still has to build capacity.
Customers still have alternatives.
Margins still matter.
Return on that expansion still matters.
So Tower gets 4%.
The interview became a position.
It did not become a religion.
Celestica gets the first major exam.
Bernstein initiated coverage this week around what it calls the AI “bandwidth wall” — compute growing faster than the networking infrastructure connecting it. Celestica was the firm's highest-conviction networking pick.[10]
Celestica reports and hosts its Investor and Analyst Day on October 27.[11]
A thesis.
A receipt.
A date.
You know I like those.
Physics is impatient.
So the power research finally got capital too.
Baker Hughes: 5%.
GE Vernova: 3%.
This is the newest branch of the AI infrastructure map.
Data centers need power.
The grid cannot always provide that power where it is needed, when it is needed.
And a billion-dollar data center does not become economically useful by politely waiting years in an interconnection queue.
Enverus Intelligence Research estimates roughly 29.6 gigawatts of behind-the-meter gas generation could be added in the United States through 2030, with data centers accounting for roughly 88% of that amount.[12]
Smaller turbines can arrive faster than large grid projects. Baker Hughes and GE Vernova are among the manufacturers positioned around that urgency.
This is not suddenly my grand ideological endorsement of natural gas.
It is much simpler.
A hyperscaler waiting for power has a financial incentive to buy speed.
Nuclear handles one portion of that problem.
Gas turbines handle another.
Grid equipment handles another.
Eventually transmission, storage and other generation handle more.
The common commodity is not merely electricity.
It is time-to-power.
Time has a price.
I want exposure to companies capable of charging it.
Rocket Lab earned the bigger seat.
Space needed a hierarchy.
Rocket Lab signed its largest commercial Electron agreement yet:
That takes Synspective's total contracted Electron missions to 47, the most of any Rocket Lab launch customer, and pushes Rocket Lab's launch backlog above 100 missions.[13]
Twenty launches is not a vibe.
It is not a partnership announcement.
It is not somebody standing beside a rendering of a future factory.
It is work.
So RKLB moves to a 6% target.
SPCX stays at 3%.
AST SpaceMobile stays at 2%.
Planet and MDA leave.
Not because either became a bad company.
Because the portfolio finally admitted it did not need five ways to say:
Redundancy is not diversification just because the tickers are different.
Rocket Lab currently has the operating receipts I trust most.
So it gets the larger seat.
This is one of my favorite changes in the rebalance.
The portfolio got more aggressive while becoming less cluttered.
That is progress.
Seagate slapped me immediately.
Good.
I kept Seagate at a 3% target because data storage remains one of the physical constraints underneath the AI build.
Then Friday, a Nikkei report said Toshiba plans to invest roughly $380 million to double HDD production capacity for AI data centers by fiscal 2027.[14]
Seagate and Western Digital both sold off roughly 10% on Friday.[2]
Excellent.
Nothing tests a freshly written thesis quite like getting punched before the ink dries.
Demand can stay enormous while additional capacity changes the economics investors are willing to capitalize.
That is why STX is 3%.
Micron is 14%.
Putting both under “AI storage” does not make them the same investment.
No averaging because it fell.
No panic sale because Toshiba may build.
Research first.
Then capital.
I have apparently heard that somewhere before.
The homework changed the book.
The percentages below are the October execution targets. Friday's closing prices caused only minor same-day drift; I am not pretending the close landed on a mathematically perfect allocation.
BWXT 22% · MU 14% · NVDA 3% · STX 3% · GOOGL 2% · MSFT 2% · AMZN 2%
CIEN 7% · COHR 6% · CLS 6% · RKLB 6% · BKR 5% · TSEM 4% · SPCX 3% · GEV 3% · RDDT 2%
ASTS 2% · OKLO 2% · AEVA 1% · HONA 1%
SERV 1% · QBTS 1%
USD 2%
Seven positions got fired.
Selling is part of research too.
Real progress. Real customers. Rocket Lab simply earned the space capital more convincingly.
Useful space exposure. Redundant exposure. There is a difference.
Interesting company. No longer central to the bottleneck map I am actually investing around.
Excellent asset. I prefer the current risk/reward farther down the AI infrastructure chain.
The company got more strategically interesting — and more complicated, capital-intensive and execution-heavy.
The future may still include electric aircraft. At five-percent-plus yields, that future has a much higher hurdle.
Aluminum was a useful expression of the grid-build thesis. BKR and GEV give me a cleaner path to the paid power bottleneck.
Accounting should be boring.
The Frontier File ledger closed Friday at approximately:
Two violent weeks.
A Treasury shock.
Several major company events.
A near-total redistribution of the portfolio.
And the headline number barely moved.
Very boring result.
Very different book.
The Frontier File portfolio began with a $5,000 seed. Another $270 was added around the SpaceX IPO, taking total contributed capital to $5,270.
$5,270 → $8,463 = +60.6%. The original $5,000 seed is useful history. Total contributed capital is the honest performance denominator.
No denominator shopping.
No making the percentage prettier because I prefer one version.
The stock picking can be weird.
The arithmetic should not be.
Portfolio accounting: Performance figures are based on recorded Frontier File executions, trade decisions and reconciliations.
Twenty-two tickers do not equal twenty-two independent bets.
The book is less cluttered now. It is not necessarily more diversified.
23% sits directly in CIEN, COHR, CLS and TSEM. Four businesses, one broad dependence on AI networking demand.
22% sits in one company. Stronger evidence does not abolish single-name concentration risk.
MU, NVDA, STX, CIEN, COHR, CLS, TSEM, GEV and BKR look different on a sector chart. A serious data-center capex slowdown would introduce them to one another very quickly.
A 10-year yield above 5% keeps raising the hurdle for every dollar of future cash flow in the book.
Three things now have to stay broadly true.
- AI infrastructure spending stays enormous.
- Scarcity lasts long enough for the suppliers to capture attractive economics.
- Five-percent-plus yields do not compress valuations faster than the businesses grow into them.
That is fine.
As long as I remember it.
New portfolio. Same ability to do something stupid.
No Rubrik above $95 just because I still like Rubrik.
No buying Archer or Joby back because softer jobs helped duration for one day.
No new quantum size.
No turning OKLO into BWXT because both involve nuclear reactors.
No Tower add because I enjoyed writing the Tower section.
No Seagate panic sale because Toshiba may add capacity.
No automatic BWXT add from 22%.
No pretending Friday's jobs report solved the rate problem.
The 10-year is still above five.
The landlord is still home.
The new book gets tested quickly.
2:00 p.m. ET
I want the internal debate behind September's decision — especially how much appetite existed for continuing to tighten before the labor data weakened.
8:30 a.m. ET
Weak jobs plus cooler inflation gives duration room. Weak jobs plus sticky inflation gives everybody a headache.
The first major company-specific exam for the new bandwidth sleeve. At 6%, “AI networking is huge” is not enough. I want economics.
Friday's jobs report changed the rate conversation. Inflation gets to speak before the meeting.
after close
At a 22% target weight, the company earns more scrutiny, not less.
The aggressive move was not buying more risk. It was deleting weaker expressions of conviction.
There is a childish version of aggressive investing.
Buy more rockets.
Buy the smaller company.
Buy the thing with no earnings because the upside can technically be enormous.
Find a low-priced stock and confuse a small share price with asymmetric risk.
I have done versions of that.
There is another kind of aggression that is less fun to describe.
Sell seven perfectly defensible ideas.
Put 7% into the networking company.
Put 6% into the optics company.
Put 6% into the manufacturer.
Put 4% into the foundry.
Increase memory after the cash register proves itself.
Increase nuclear after the backlog proves itself.
Then accept the uncomfortable part:
If the thesis is wrong, the consequences are now larger.
That is aggression too.
Probably the adult version.
The portfolio was not too safe.
It was too comfortable.
Last week I worried that the portfolio had become too safe.
I think that diagnosis was wrong.
It was too comfortable.
Those are different problems.
BWXT felt familiar because I had researched it for months.
Micron felt like something I had already missed because the stock had moved without waiting for me.
Some positions had been in the portfolio long enough to become furniture.
Planet was there.
MDA was there.
SoFi was there.
CENX had become an old expression of the power thesis.
Archer and Joby were stories I had spent enough time thinking about that keeping them somehow felt easier than asking whether I would buy them again today.
Then the receipts arrived.
Micron printed.
BWXT laid out the backlog.
Rocket Lab sold twenty launches.
Optics stopped looking like a research side quest.
The power bottleneck got harder to ignore.
And suddenly the uncomfortable question from Issue 020 had an answer.
The correct response was not to buy some tiny stock so the portfolio felt dangerous again.
It was to look at every dollar already deployed and ask:
For seven positions, the answer was no.
For several others, the answer became more.
That is the work.
Issue 019 · Five percent is a competitor. / Issue 020 · The next dollar has homework. / Issue 021 · The homework produced an allocation.
Do the homework.
Then move the damn dollar.
Research is not useful because it makes a watchlist longer.
It is useful when it changes what I am willing to own, how much I am willing to own, or what I am finally willing to sell.
This week it did all three.
Now the new portfolio gets the same treatment the old one did:
No loyalty points.
No permanent residents.
No free pass because I wrote a convincing paragraph about it.
The next receipt can strengthen the allocation.
The next receipt can also fire it.
— Alex