August 2026 · Positioning

The Month the Site Went Stale

August 17, 2026 · Prices as of the August 14, 2026 close · Benchmark: SPY

The most useful thing that happened this month wasn’t a trade. It was an audit of this site that found three of my own pages publishing statements that had quietly stopped being true — not typos, but claims that had inverted while I wasn’t looking. That is the subject of this letter, and it belongs ahead of the performance numbers, because a research site whose pages drift out of date is failing at the one thing it exists to do.

The book itself is worth $34,629.23 as of the August 14 close, up $14,400.53 (+71.19%) against a cost basis of $20,228.70. Since the July letter that is a decline of 0.56% — essentially flat. I want to sit on that word for a second, because the recent daily sessions have looked anything but flat.

Flat month, dramatic days

On August 14 alone, NBIL rose 17.77%, Nebius rose 8.88%, SanDisk rose 7.39%, and AMD rose 6.50%. Read those four numbers on their own and you would assume the book had a tremendous month. It did not. Over the same stretch since the July letter, SanDisk’s return on cost fell from +261.01% to +239.63%, AMD’s fell from +176.94% to +158.05%, and Micron’s slipped from +202.23% to +198.21%. A violent up-day inside a drawdown is still a drawdown.

The same session ran the other way for other names: Broadcom fell 5.94%, CrowdStrike fell 3.80%, Tempus fell 4.03%. This is the dispersion I have written about in the Opinion notes all month — the AI complex has stopped moving as one block, and the split is increasingly between the names selling physical capacity (memory, compute, power) and the software names priced off the same theme. I do not think one session establishes that as a regime. I do think it is the most interesting thing in the book right now.

Palo Alto is the quiet exception worth naming: it is the one large position that is meaningfully higher on cost than it was in July, +123.01% to +139.69%. It got there without a single headline I felt compelled to write up.

Portfolio composition

All 32 positions, sorted by weight. Average price is my real per-share cost basis; return is against that basis at the August 14 close. Same figures as the Holdings table and every individual security page — if any of those three disagree, that is a bug, and this month it was.

By bucket

• Core / quality compounders — 59.31%. The anchor. Sixteen names across semiconductors, cybersecurity, cloud software, power and consumer.
• Index hedge — 18.33%. SPY, VOO, QQQ, SCHD. SPY and VOO together are 12.91% of the book, which is more index duplication than I would design from scratch.
• Emerging growth — 13.20%. Smaller operating companies with real revenue.
• Speculative sleeve — 5.72%. Well under its 25% cap.
• Leveraged / derivative — 3.43%. NBIL alone, under its 5% cap but up sharply as Nebius rallied.

Six positions caught up on earnings

Coming into this month the Earnings page had six held positions whose most recent quarter had come and gone without a written report: SanDisk, Datadog, Talen, Vistra, Vertiv and Intel. All six now have one. The two that changed my thinking:

SanDisk’s fiscal Q4 was the largest print in the book — revenue of $8.97 billion, up 371.6% year over year, with non-GAAP EPS of $39.25 beating consensus by roughly 15%. The number that matters for the thesis is not the year-over-year figure, which mostly says how depressed NAND pricing was a year ago; it is the 51% sequential growth, roughly two-thirds of it price rather than volume. I wrote that up separately in The NAND Supercycle, and the honest reading is that this is a commodity cycle going my way, not a company out-executing its peers. Those deserve different levels of confidence.

Vertiv’s Q2 is the one I got least comfortable with. Adjusted EPS grew 60% and beat, guidance went up across every metric, and the stock fell anyway on a 3.4% revenue miss that management attributed to supply-chain timing. That explanation is plausible and completely unverifiable from outside. VRT is my only position underwater, at −3.63% on cost, and it recovered a long way from the −11.41% it was sitting at when I wrote the report. I am holding, but the “timing, not demand” claim needs to be settled by the next print rather than by management’s framing of this one.

What I got wrong

This is the section that actually matters this month. I ran a full audit across all 238 pages of this site — every date claim, every hardcoded figure, checked against the underlying data rather than read by eye. It found four categories of failure, and three of them are mine.

1. The site published prices under the wrong date. The Watchlist and all fifteen of its deep-dive pages carried the line “Prices as of August 17, 2026 close.” There was no August 17 close — the numbers shown were August 14 closes. The update script stamped the date it ran rather than the date the data came from. This is the second time that exact bug has published a wrong date here. The first time, in early August, I corrected the label and did not fix the script that generated it, which is the actual mistake: I treated a systematic error as a typo. The script now takes the data date explicitly and falls back to today only when nothing is passed.

2. Two pages had narratives that were factually backwards. The Compliance Ledger and the SanDisk security page both explained that SNDK had fallen back under its 10% position cap because the stock “continued to pull back in price.” By the time I read those sentences, SanDisk had rallied to +239.63% and its weight had climbed back to 9.48% — half a point from the cap. The pages were not merely out of date; they described the opposite of what was happening, in the section specifically meant to hold me accountable. The old text had even named this scenario as the thing to watch for: “if SNDK re-approaches 10% on a rally rather than a pullback, the same obligation applies again.” It did, and the page kept saying otherwise. Both are rewritten.

3. A tool claimed to do something it does not do. The Rebalance Visualizer stated that its bucket weights were “recomputed whenever Holdings updates — not a one-time snapshot.” They are hand-written into the file, and they were roughly four percentage points stale. I have corrected the weights and, more importantly, corrected the claim: it now says the numbers are entered by hand and that Holdings wins if the two disagree. An overstated methodology claim is worse than a stale number, because it tells a reader not to check.

4. Figures disagreed with each other across pages. The concentration index read 493 on three pages and 492 on a fourth. Thematic concentration read 60.93% site-wide when the live weights gave 60.37%. The Risk X-Ray’s aggregated Nebius exposure — the NBIL leveraged position counted at 2× plus the direct NBIS holding — still read 6.99% when it had grown to 8.46%. That last one is the one I would least want a reader to miss, because it moved up for the same reason the leaderboard looks good: Nebius rallied. Hidden exposure grows fastest exactly when the position is working.

One accounting note in the same spirit. This site quotes an all-time return of +72.14%, verified via Blossom on July 31. Running the same cost-basis arithmetic on the published average costs against the August 14 closes gives +71.19%. The gap is a different date and possibly a different method, not a contradiction — but I am not going to restate a Blossom-verified figure using my own spreadsheet and keep calling it Blossom-verified. The +72.14% stays labelled with its July 31 verification date until I have a new verified figure to replace it with.

Ongoing risks

SanDisk is approaching its cap again, and the trim still has not happened. At 9.48% the position is inside the 10% limit, but it got there and back by price alone. The trim I said was owed in June has never been executed. If the next leg higher puts it through 10%, that is a live breach and the Compliance Ledger will list it as one.

Effective Nebius exposure is 8.46%, not the 1.60% the holdings table implies. NBIL is a 2× daily-reset fund on NBIS, and I hold NBIS directly as well. Counted the way the IPS requires, that is a single-company exposure approaching the same 10% cap from a direction that no single line item in the table reveals.

Thematic concentration is 60.37% across 17 positions. Semiconductors, AI infrastructure, and the power that feeds them are one bet expressed many ways. The Thematic Exposure page keeps this checkable. It is disclosed rather than capped, deliberately, and it remains the largest structural risk in the book.

Concentration by position size still looks benign, and still is not the point. HHI of 492 implies roughly 20 equally-weighted positions. That measure cannot see that a third of the book moves on the same news, which is what The Concentration Premium exists to demonstrate.

The macro backdrop

Core CPI came in at 2.5% for July, down from 2.6% in June and still above the Fed’s 2% target. Unemployment is 4.1%. The FOMC held rates at 3.50–3.75% on July 29 in a 9–3 vote, with the three dissents wanting a hike on exactly that above-target inflation; the next decision is September 15–16. The two-year Treasury is at 4.17% and the ten-year at 4.68%, a +0.51 spread — an uninverted curve that has steepened slightly. High-yield spreads are 2.71%, which is not a market pricing credit stress.

None of that is a forecast, and none of it changed a position this month. It is the backdrop the book is sized against, and it is permissive. The risk in this portfolio is not the macro right now; it is the concentration described above.

No trades

No positions were opened, closed, added to or trimmed this month. That includes the SanDisk trim, which is now a stated obligation carried across three consecutive letters without action. I would rather write that sentence plainly than let it disappear into a page of otherwise good numbers.

How I’ll keep myself honest

The audit changed what I think the maintenance job actually is. I had been treating “update the site” as refreshing prices. Prices were the part that was already automated and mostly fine. What went wrong was prose — sentences written to be true at a moment, left standing after the moment passed, in a voice confident enough that a reader would have no reason to doubt them.

Three things change. The price script now records the data’s own date, so a label can no longer outrun the number it describes. Every figure that appears on more than one page — concentration, thematic exposure, aggregated leveraged exposure, position weights — is now checked against a single source rather than maintained separately in each place. And any sentence that explains why a number is what it is gets re-read whenever the number moves, because that is the class of error that produced the two backwards paragraphs above.

The point of publishing this is not that the site is now correct. It is that a site claiming to keep an honest record has to be willing to audit itself in public and report what the audit found, including the parts that make the record look worse. See the page for how these figures are sourced, and the Disclosures page for what this account is and is not.

Prices via Massive Market Data, August 14, 2026 close — the most recent session with published end-of-day data. Position weights and returns computed from published average costs. Macro series via Massive Market Data (Federal Reserve labor-market, inflation and Treasury-yield series); high-yield spread from the ICE BofA US High Yield Index. Personal views only, not investment advice.