Opinion · Positioning · August 28, 2026
The Concentration Risk I Named Thursday Showed Up Friday
Thursday’s piece on this page argued that the largest risk in this book is not the Fed — it’s that roughly 60% of the portfolio sits in one linked theme, unconstrained by any cap in my written policy. I did not expect that argument to get a live illustration within twenty-four hours, but Friday obliged. Marvell, Tempus AI, IonQ, Coinbase, Arm and Nvidia had each run up into Thursday’s close — several of them sharply — and gave a large share of that gain straight back Friday afternoon, on the same session a Fed speech that initially lifted the tape reversed into a selective repricing of the market’s most expensive names. This is not the durable, systemic version of the risk I described Thursday. It is a much smaller, faster, one-day version of the exact same mechanism: names that move together on the way up move together on the way down, and a policy that caps single-position size does nothing to cap that.
1. The pattern, shown directly
Twelve names I track split cleanly by what they did Wednesday-to-Thursday versus Thursday-to-Friday. Six of them — Marvell, Tempus AI, WhiteFiber, IonQ, Coinbase and Arm — rose into Thursday’s close and fell more than 4% Friday. Three more — Nvidia and Nebius among them — show the same shape at smaller magnitude. The names that avoided it were the ones that were never part of Thursday’s rally in the first place: ServiceNow, Amazon, Alphabet and Meta, all already-profitable large caps that rose Friday on no dated news at all, simply by not being part of the basket that ran.
X-axis: each name’s move from the Aug. 26 close to the Aug. 27 close. Y-axis: the same name’s move from Aug. 27 to Aug. 28. Marvell (top-left) is the exception — its decline traces to its own earnings guide, not the pattern.
| Ticker | Aug 26→27 | Aug 27→28 | Read |
|---|---|---|---|
| Nvidia (NVDA) | +8.74% | −4.57% | Partial give-back after a strong run |
| ServiceNow (NOW) | +10.04% | +4.54% | Kept rising — the exception to the pattern |
| IonQ (IONQ) | +6.07% | −7.68% | Gave back more than it gained |
| Coinbase (COIN) | +4.92% | −6.33% | Gave back more than it gained |
| Tempus AI (TEM) | +3.23% | −9.41% | Gave back nearly 3x the prior day’s gain |
| Nebius (NBIS) | +2.13% | −4.26% | Gave back roughly double the prior gain |
| Arm Holdings (ARM) | +1.65% | −6.33% | Gave back roughly 4x the prior gain |
| Marvell (MRVL) | −1.49% | −10.28% | The outlier — its own earnings guide, not the pattern |
| WhiteFiber (WYFI) | −3.10% | −8.41% | Down both days — no run-up to give back |
2. Why this matters more than the day’s dollar total
Friday’s net was −$430.79, a real but unremarkable number against a roughly $32,600 book — about 1.3%. The dollar figure is not the point. The point is that eight of the nine names in the table above sit inside, or immediately adjacent to, the semiconductor and AI-infrastructure theme Thursday’s piece flagged at 60.37% of the portfolio, and every one of them that had run up Thursday gave a large share of it back Friday, on the same session, for what reads as the same underlying reason: high-multiple, high-beta names being repriced together after a hawkish-sounding Fed remark. That is thematic clustering behaving exactly as described — not in a crisis, but in an ordinary Friday afternoon.
3. Sizing the mechanism, not predicting it
Thursday’s piece was explicit that the scenario worth watching is a durable repricing of AI-infrastructure spending, not a single hawkish afternoon. Friday is not that scenario — it round-tripped a few days’ gains, not a structural re-rating. But it is a convenient, low-cost way to size what the larger version would do, using Friday’s actual mechanism rather than a guess.
| If the concentrated sleeve fell an additional… | Dollar impact on the book |
|---|---|
| −5% from here | −$984 |
| −10% from here | −$1,969 |
| −15% from here | −$2,953 |
| −20% from here | −$3,938 |
Assumptions, stated plainly. This applies each hypothetical percentage decline to 60.37% of the book’s total value ($32,609.04 as of the August 28, 2026 close) — the concentration figure published in Thursday’s piece, measured as of the prior close. It assumes the entire thematic sleeve moves uniformly by the stated percentage, which no real basket of seventeen different names does; some would fall more, some less, and NBIL’s embedded 2x leverage means its contribution to any further decline compounds faster than the others. It ignores everything outside the sleeve, which could move in either direction and partially offset or add to the total. This is not a prediction that any of these declines will happen — it is Friday’s real, observed mechanism (correlated names giving back a shared gain together) scaled up to show what it would mean in dollars if the mechanism ran further than one afternoon.
What would change this read
WatchTwo things that would tell me this is more than a one-day pattern
1. Whether the give-back continues into next week. A single Friday reversal after a Fed speech is thin evidence of anything beyond ordinary volatility. A second consecutive session where the same names fall together, without a fresh dated catalyst each, would be a stronger signal. 2. Whether high-yield spreads move. Thursday’s piece named 2.71% high-yield spreads widening past roughly 4% as the trigger that would turn this from background risk into something the macro table can no longer ignore. Nothing in Friday’s session moved that number.
I am not trimming the sleeve on the strength of one Friday, and Thursday’s piece already explained why I haven’t trimmed it on the strength of anything else this month either. What changed is smaller: I now have a concrete, priced example of the exact mechanism I described in the abstract, twenty-four hours after describing it, which is a better argument for taking the concentration seriously than the abstract version was.