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 trim on old information, a rally on new information
Druckenmiller’s sale reflects what he knew as of June 30 — a large, profitable position after a historic run, the kind of thing a portfolio manager trims on valuation discipline alone. Monday’s move reflects something that didn’t exist in June: a specific policy tailwind for domestic and allied memory production, and a demand-side argument (Musk’s comment on agentic AI’s memory intensity) that wasn’t part of the public conversation two months ago. A stock can be a legitimate trim in Q2 and a legitimate rally in Q3 without either investor being wrong; the mistake would be treating a quarter-old 13F as a verdict on Monday’s price action.
What I think is more interesting is the shape of the rest of the day. 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 a real, sourced fact, but it explains a rate-sensitive rotation out of long-duration growth names in general, not why ServiceNow specifically fell harder than Datadog. 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 isn’t deciding which story is right; it’s being honest about which moves I can actually trace to a dated cause and which ones I can’t.
Where this connects to my own thesis
Musk’s framing — memory as the binding constraint on AI, not compute — is a live version of the capacity question my Physical Limits of Compute whitepaper raised about the AI buildout generally: every layer of this stack eventually runs into a physical or geopolitical bottleneck, and today it happened to be memory sourcing. I don’t think one policy signal and one CEO comment settle the multi-year question of where AI infrastructure spending actually constrains growth, but it’s a real, dated data point in favor of memory demand staying structurally tight rather than cyclically peaking — which is a different claim from saying today’s 8.88% move in SanDisk is fully rational or repeatable.