Opinion · August 2026
A Fund Manager Trimmed Memory in Q2. The Market Gave It a New Reason in Q3.
A Form 13F disclosed Monday that Stanley Druckenmiller’s Duquesne Family Office sold its Micron and Intel positions in the second quarter, after both stocks had already tripled and roughly tripled in price — textbook profit-taking on names that had run hard. Hours later, on the same day that filing became public, Micron rose 4.13% and SanDisk rose 8.88% on a genuinely new input: a reported Trump administration policy signal favoring domestic and allied memory suppliers over Chinese sourcing, plus Elon Musk publicly arguing that memory, not compute, is now the binding constraint on agentic AI. Those are two different information sets pointing at the same names from two different points in time, and I don’t think either one contradicts the other.
A 13F told me what a very good investor thought about memory on June 30. Six hours later the market repriced the same names on information that did not exist in June. Both things can be true, and the interesting part of Monday is neither of them — it is how few of the day’s largest moves I can trace to a dated cause at all.
A trim on old information, a rally on new information
A Form 13F disclosed Monday that Stanley Druckenmiller’s Duquesne Family Office sold its Micron and Intel positions in the second quarter, after both stocks had already tripled — textbook profit-taking on names that had run hard. Hours later, on the same day the filing became public, Micron rose 4.13% and SanDisk rose 8.88% on a genuinely new input: a reported Trump administration policy signal favoring domestic and allied memory suppliers over Chinese sourcing, plus Musk publicly arguing that memory is now the binding constraint on agentic AI.
The 13F reports positions as of June 30 and is published up to 45 days later. Monday’s catalyst is dated Monday. A stock can be a legitimate trim in Q2 and a legitimate rally in Q3 without either investor being wrong; the mistake is treating a quarter-old filing as a verdict on today’s price. What it is evidence of is that a very good investor found the risk/reward unattractive at June prices — which is worth knowing, and is not the same claim.
The part of Monday I actually can’t explain
Figure 1 · Attribution audit
Seven of the day’s ten biggest moves had no name-specific cause
The ten largest moves among names this site tracks, August 17, 2026, classified by whether I could find a dated, company-specific reason for the move. “Sector weakness” and “rates rose” count as macro, not name-specific. Source: my own sourcing check against that day’s reporting.
ServiceNow, Datadog, Nebius, Microsoft and Meta all fell between 3% and 5% with no comparable dated catalyst. The 30-year Treasury yield hitting its highest level since 2007 is real and sourced, but it explains a rate-sensitive rotation out of long-duration growth in general — not why ServiceNow specifically fell 5.08% while Datadog fell 3.19%. This is the same pattern this site has logged repeatedly this month: a real move sitting next to an unexplained one, and a rotation that looks like conviction until you check the sourcing.
The discipline is not deciding which story is right. It is being honest about which moves I can trace to a dated cause and which I cannot — and then not letting the untraceable ones quietly become part of a thesis.
A projection: what a share shift is actually worth
“Reshoring” is a share-shift argument, not a demand argument — a policy preference redirects existing memory purchasing, it does not create new memory demand. So the honest way to size it is to ask what a given amount of redirected spend is worth against these companies’ actual revenue bases.
| If policy redirects | Annual revenue | As % of SanDisk FY26 revenue | As % of Micron annualised |
|---|---|---|---|
| 2 points of share redirected | $0.24B | 1.2% | 0.1% |
| 5 points redirected | $0.60B | 3.0% | 0.4% |
| 10 points redirected | $1.20B | 5.9% | 0.7% |
| 20 points redirected | $2.40B | 11.9% | 1.4% |
Assumptions, stated plainly. The $12 billion annual figure I am applying the share shift to is my own rough proxy for Apple’s addressable memory spend, built from public estimates rather than any disclosure, and it could easily be wrong by a wide margin — treat it as an order of magnitude, not a number. Micron’s comparison base annualises a single quarter ($41.46B × 4), which overstates a full year if the current price level does not hold. Nothing here assumes any policy is actually enacted; no rule has been written down. What the table is for is proportion: even a fairly aggressive ten-point share shift is worth low-single-digit percentages of these revenue bases, which is a long way from justifying an 8.88% single-session move on its own. That gap is the part of Monday I would want explained before treating the rally as durable.
Where this connects to my own thesis
Musk’s framing — memory as the binding constraint on AI rather than compute — is a live version of the capacity question the Physical Limits of Compute whitepaper raised about the buildout generally: every layer of this stack eventually meets a physical or geopolitical bottleneck, and Monday it happened to be memory sourcing. One policy signal and one CEO comment do not settle a multi-year question. They are a dated data point in favour of memory demand staying structurally tight rather than cyclically peaking — which is a different and much narrower claim than saying Monday’s 8.88% was rational.