The $1 Trillion Excuse

Micron reported the most straightforwardly bullish quarter in the memory industry's recent history on Wednesday night. Revenue beat. Margins beat. Q4 guidance beat. The stock surged 15% in a single session, its best day in years, and for about eighteen hours the AI trade looked like it had found its footing again after two weeks of ugly rotation out of semis.

Then the New York Times published a story about OpenAI possibly delaying its IPO until 2027, and the whole thing fell apart before Friday's close.

MU gave back 6.7% on Friday. Western Digital collapsed 13%. Seagate fell 12%. SanDisk dropped 10%. SoftBank, arguably the most financially leveraged node on the entire AI enthusiasm graph, plunged more than 12% in Tokyo. South Korea's Kospi shed 5.81%, the Kosdaq another 4.1%. Japan's Nikkei fell 4%. The Nasdaq posted its fifth consecutive losing session and is now down roughly 5.5% from its June 2 peak.

All of that, from a report about a company that has never had a profitable quarter, deciding not to list publicly at a valuation that would make it the most expensive single-product technology business in history.

Think about the causal chain implied there. A rumor that OpenAI might delay raising money it doesn't currently need caused a thirteen percent single-day collapse in Western Digital — a company whose business is selling storage hardware — because investors apparently priced that hardware demand as contingent on OpenAI's access to public capital markets. JPMorgan articulated it plainly in a note Friday morning: the delay raised concerns about the "sustainability of infrastructure spending given the delay in funding from the capital markets." Adam Crisafulli at Vital Knowledge called it directly: the IPO delay "could slow the pace of infrastructure spending."

That framing is worth sitting with. The AI capex supercycle — the $755 billion in hyperscaler spending projected for 2026, a figure representing 83% year-on-year growth — has been justified to equity markets as demand that will persist regardless of short-term rate or macro conditions. It was supposed to be structural. Inevitable. A new infrastructure buildout on the order of the 1990s fiber boom but with better unit economics and smarter capital allocation. That was the pitch.

The pitch did not account for the possibility that narrative confidence functions as a liquidity condition.

Because what the OpenAI IPO story actually revealed is that this entire trade has a sentiment dependency baked into it. The hyperscalers are still spending. Microsoft and Meta posted gains on Friday while the chip names got destroyed. The physical demand for HBM, for networking bandwidth, for co-packaged optics infrastructure — none of that disappeared because Sam Altman may prefer to float OPAI.PVT in 2027 instead of 2026. But the stocks moved as if it had, because a material portion of the multiple on semiconductor names was priced on the assumption that capital would keep flowing freely to AI-native companies, who would then spend it on compute, who would then buy chips, who would then justify the memory capex cycle.

Sever one link and you get WDC down 13%.

The memory efficiency angle makes this sharper. Qualcomm spent investor day this week promoting its "High-Bandwidth Compute" architecture — essentially a pitch for reducing reliance on HBM in AI workloads. Nvidia has been quietly adjusting Vera Rubin designs toward lower overall memory requirements. Cerebras, reporting its first post-IPO quarter, built its entire market positioning around being "HBM-free." None of these are coincidences. They are responses to a supply-chain constraint that has become a cost constraint, and cost constraints eventually become architectural constraints. The industry is beginning to engineer around the bottleneck rather than through it.

Apple, meanwhile, announced price increases on iPads and MacBooks this week, blaming the memory shortage directly. That is not a sign of healthy supply-demand equilibrium in NAND and DRAM markets — that is a sign that the shortage has become severe enough to get passed through to consumer hardware, which is the last place any tech company wants pricing friction to appear. Apple's stock fell 6% on Thursday.

The breadth picture in US equities is doing something interesting underneath all this. Even as the Nasdaq bled out, advancing shares outnumbered decliners on the S&P 500 on multiple sessions this week. Industrials were up 2.19% on Thursday. AbbVie surged 17% since June 18 to a record $253.35 after FDA expanded approval for Skyrizi. Visa and Walmart added over 2% Friday. The market isn't breaking — it's rotating. Money is leaving a trade that got overcrowded and moving into companies with earnings tied to things other than GPU rack density.

Alphabet replaced Verizon in the Dow on Friday, which was a notable index committee endorsement of the search-and-cloud business at exactly the moment the AI trade was getting clobbered. Alphabet itself was down 2.2%. The Dow barely moved, which was only possible because the components that gained — Microsoft 5.7%, Salesforce 5.5%, IBM 5.1% — happen to represent the mature, cash-generative enterprise software layer of AI deployment rather than the pure infrastructure buildout. That bifurcation is becoming visible in price action in a way it wasn't two months ago.

The University of Michigan consumer sentiment reading due Friday was expected to come in at 48.9%, near historic lows. At those levels, the consumer isn't really a stabilizing force for anything. Real average hourly earnings have declined as inflation outpaced wage growth for the first time in three years. The personal saving rate is at 2.6%. These are not recessionary numbers yet, but they are consistent with an economy that is running hotter than it should be while its households feel considerably colder than the equity indices imply.

Kevin Warsh hasn't moved rates yet. The bond market is doing the work for him, though. Treasury yields fell through the week as lower oil prices reduced the near-term inflation read — tankers are still crossing the Strait of Hormuz despite Trump's accusations that Iran has violated the ceasefire — but the base rate environment is not supportive of the multiple that AI infrastructure names were carrying at the June peak.

Micron's quarter was real. The HBM demand is real. The data center buildout is real. None of that is the question. The question is whether the equity prices on the companies enabling that buildout were priced for a continuation of the enthusiasm, the capital flows, and the narrative confidence that have characterized this trade since late 2024. This week suggested they were. And enthusiasm, it turns out, has a duration.

---

This newsletter is for informational purposes only and does not constitute investment advice.