Remarks Prepared for No One in Particular
(delivered, in the writer's head, to an empty conference room somewhere between Sintra and Seongnam)
Thank you. I'll skip the pleasantries, because the numbers don't have time for them.
Next week, ASML and Taiwan Semiconductor report earnings, and I want to talk about why that matters more than anything Kevin Warsh has said or not said since May. The Fed Chair's silence gets treated as the dominant uncertainty of this cycle. It isn't. The dominant uncertainty is sitting inside two earnings reports from companies most retail investors couldn't locate on an org chart, and the reason is structural, not rhetorical.
Here's the shape of it. The Philadelphia Semiconductor Index closed back above its 50-day moving average this week, which sounds like recovery. It is not recovery. SOX still sits 11% below its June peak, even as the Nasdaq Composite grinds toward records on the back of exactly four or five names. That gap — mega-cap AI infrastructure ripping while the underlying semiconductor supply chain limps — is not a temporary dislocation. It is the market telling you, in the clearest language it has, that "AI trade" and "semiconductor trade" have quietly become two different assets wearing the same ticker soup.
Consider what just happened with SK Hynix. The company priced its Nasdaq ADR at $149, opened near $170, raised $26.5 billion. Every dollar of that goes toward memory capacity — new fabs, new equipment — aimed squarely at closing what Baillie Gifford's people are calling a persistent imbalance between AI compute demand and memory supply. That's not a bet on Nvidia's quarter. That's a multi-year capital commitment predicated on a demand curve nobody can actually verify yet, because the hyperscalers building the demand curve are themselves guessing at monetization timelines that are still, charitably, theoretical.
Now overlay geopolitics, because this is where the structural story stops being abstract. The Strait of Hormuz has been disrupted badly enough that the IEA is projecting the first annual decline in global oil demand since the pandemic — a supply-chain shock, not a demand collapse, and the distinction matters. Meanwhile Taiwan Semiconductor, the single most important company in the entire AI capex chain, sits ninety miles from a strait of its own that nobody in this market has priced for disruption in over a decade. We have just watched what happens to global energy logistics when one chokepoint goes hot. Nobody has run that scenario for the Taiwan Strait in a portfolio committee meeting since 2023, and I'd wager most haven't run it at all.
This is the pattern history keeps handing us and we keep declining to read. In 2000, the market conflated "internet company" with "profitable business" until the distinction became unavoidable and expensive. In 2007, it conflated "AAA-rated" with "safe" until a chain of counterparty assumptions broke all at once. The conflation this cycle is simpler and in some ways more dangerous: it's treating "AI infrastructure buildout" and "semiconductor cycle" as a single trade, when the first is a multi-decade capex thesis with genuinely uncertain payback periods, and the second is the same cyclical, capital-intensive, geopolitically exposed industry it has always been — the one that gave you a 47% SOX drawdown in 2022 on nothing more exotic than inventory correction and rate hikes.
Margin debt at a record $1.42 trillion is financing exposure to the first narrative while the underlying fundamentals still behave like the second. That is the actual risk architecture of this market right now, and it has nothing to do with whether Kevin Warsh says something hawkish in his next public appearance. Warsh's silence is a rounding error next to a supply chain running through the two most geopolitically exposed waterways on the planet, financed at record leverage, on the assumption that AI monetization arrives roughly on schedule and nothing between here and Hsinchu goes wrong in the meantime.
History doesn't offer comfort here. It offers pattern recognition. Every cycle that conflates a durable structural theme with a cyclical, leverage-financed vehicle for expressing it eventually forces a repricing of the vehicle, even when the theme survives intact. The internet survived 2000. Housing survived 2008, eventually. AI will very likely survive whatever correction eventually separates it from the semiconductor cycle currently carrying it. That's cold comfort if you're levered into the vehicle when the separation happens.
ASML and TSM report next week. Watch capex guidance, not headline earnings. That's where this thesis gets tested first.
Thank you.
