The Bounce Bought a Day. Seoul Took It Back.

What Needs Your Attention Today

Watching means normal, warning means look closer, danger means act. The word always carries the meaning — never the colour alone.

The 60-Second Picture

Two Mondays in one: in New York, buyers returned to exactly the stocks they had dumped — memory and storage first, up 3–6%, with Nvidia up 1.38% to $217.09. In Seoul the next morning, the same trade was marked down hard. Nothing about demand for computing changed overnight in either city. What changed is what investors are willing to pay, today, for a promise that pays back over years — while the cost of waiting, the 10-year yield at 5.15%, keeps rising.

The Macro Vector: A Bounce Before the Proof

Here is the claim: Monday was positioning, not conviction. The mechanism is timing. The results investors say will settle the argument — SpaceX and Netflix today, Meta and Tesla on Wednesday — had not been published when the buying happened, so the buying cannot have been a response to new information. It was a response to price: after the worst week since late June, the stocks were cheaper, and some buyers stepped in ahead of the evidence.

The evidence for narrowness is in the cross-section. Only four of the eleven market sectors rose. Technology added 1.19% in its first gain after three falls, and shoppers' stocks (consumer discretionary) led at +2.73% — but utilities fell 1.97% and property fell 1.17%, the two groups whose profits are most sensitive to borrowing costs, with banks down 0.79% behind them. At one point in the session roughly 70% of the S&P 500's members were lower on the day even as the index rose — a handful of giants carrying the average. Meanwhile the bond market moved the other way: yields rose at nearly every maturity, the two-year to 4.77% and the ten-year by six basis points to 5.15% on the Treasury's official daily table (the traded note itself closed at 5.135% — same story, two rulers). Stocks said "cheaper now." Bonds said "still expensive to wait."

Nor was credit the only crack the bounce papered over. The chip index's eight-session slide — its longest since October 2008 — began with reports of advances by Chinese memory maker CXMT, which knocked 14.6% off Micron and 24.3% off SK Hynix last week, and unverified reports of slowing revenue growth at AI developer Anthropic struck at a load-bearing assumption: if frontier-lab growth is decelerating, the end-user demand underwriting the entire build-out's circular financing has to be repriced. And part of the build-out's financing is circular, as Northwestern Mutual's Brent Schutte put it this week: "One company's debt issuance becomes another company's revenue." A bounce can heal a price. It does not, by itself, heal a funding loop.

Seoul then showed what the new-issue market thinks. Elice Group — an AI company that began in education and now builds small, fast-erected data centers — priced its first public share sale at 90,500 won, the top of its range, raising 201.1 billion won (about $150 million). Institutional investors had bid for 388 times the shares on offer; retail investors, 408 times. On Tuesday it opened at 75,700 won, fell as low as 61,000 won, and traded near 65,300 won — about 28% under water — while Seoul's main index lost nearly 4%. Elice is not a broken company: it earned an operating profit of 10.3 billion won on revenue of 34.4 billion won in the first half of this year, more profit than in all of last year. Massive oversubscription at the offer price and a 30% discount the next morning is not a verdict on one firm. It is the market repricing the whole category — anyone whose plan is "raise money, build data centers, earn it back later" — in the week after CoreWeave's bonds hit distress levels.

What is priced in: the five giant spenders lifting combined capital spending to about $1.2 trillion next year, per Goldman Sachs estimates reported by Reuters — and the revenue to justify it arriving on schedule. What would prove this wrong: Meta trimming its spending guide on Wednesday, or one more new AI listing pricing below its range. Either turns a narrow bounce into a second leg down.

Winners and Losers Split: Memory Got the Bid, Borrowers Got the Bill

Why did memory and storage lead? Bank of America's analysts argue the sell-off overshot: after a correction of about three and a half months, the big memory makers trade at just 2–3 times expected 2027–28 earnings — investors are paying only $2–3 for each $1 of yearly profit those firms are expected to make — with SK Hynix and Samsung down more than 50% from their June peaks, pricing that, in the bank's telling, belongs to a slump the industry is not actually in. (Western Digital and Seagate, Monday's two biggest risers, make disk storage rather than memory chips — the broader point stands: all five sell the capacity every AI server must buy.) Whether or not the bank is right, the logic of Monday's pecking order is sound: memory is a toll booth. Every AI server needs it, the suppliers collect in cash, and none of them needs to borrow at 15% to make it.

CoreWeave, at the other end, is the toll payer. Its shares slipped 0.44% to $61.12 on the market's best tech day in weeks, after a fall of about 25% last week. Its filings show $35.6 billion of total borrowing at June 30; its newest unsecured bonds — loans not backed by any specific asset a lender could seize — traded at yields between 15% and 16% in mid-October, and the market price of insuring its debt against default hit a record 1,194.5 basis points on October 15. The analyst split tells the same story in miniature: Goldman Sachs upgraded the stock to Buy on Monday — and in the same note cut its price target from $139 to $106 and the profit multiple it is willing to pay from 22× to 18× expected earnings. "Cheaper" and "safe" are different words, and the market is using both at once.

An analogy: Monday was the first sunny day after a storm warning — the beach refilled, but nobody cancelled their insurance. It breaks in one place: beach weather is free to enjoy, while holding a leveraged AI stock at a 5.15% risk-free rate costs real money every day it fails to prove itself. That carrying cost is why the bounce needs this week's earnings to be good, not merely fine.

One longer lens, from the research shelf: McKinsey and the Global Semiconductor Alliance estimate that by 2030 about 60% of AI computing demand will be inference — answering questions — rather than training models, and argue the winners will be judged on cost per useful answer, not on spec sheets. That favours exactly the layers Monday rewarded — memory, networking, and whoever lowers the cost of each answer — over whoever simply builds the most capacity. See McKinsey, “The cost of intelligence: How to scale AI inference” and, on the hidden fragilities underneath, McKinsey on the supply chains behind US semiconductors.

What to Watch This Week

• Meta, Wednesday — the week's main event. Meta is one of five giant spenders expected to lift combined spending to about $1.2 trillion next year. A reaffirmed or raised guide, with advertising revenue to match, is the cleanest way the bounce becomes a recovery. A trim hits every supplier, every memory maker, and every lender on this page at once.
• SpaceX and Netflix, today. SpaceX reports today in only its second results as a public company — a live read on what investors will pay for AI-adjacent growth at today's prices. Netflix, expected to show slower revenue growth, tests the consumer side of the same story.
• The 10-year yield at 5.15%. It rose on a day stocks rallied — the opposite of relief. Below 5% loosens every financing on this page; a push back toward the early-October high near 5.31% tightens them all, and hits utilities and property first, as it did on Monday.
• The next new listing, anywhere. Elice's debut is now the template fear: huge demand at the offer, a discount on day one. Watch whether the next AI infrastructure listing prices at all — a pulled deal would say the window for new builders is shut, not just repriced.
• If you work at — or sell to — a builder: the question clients and finance teams are now asking is the one State Street's Ninghui Liu put plainly this morning: whether every dollar of AI spending earns more than it costs to finance. Start with the one-line discipline of knowing what your team spends on AI tools and computing, then build the measurement that actually matters before someone asks for it in a budget review: cost per finished task, tracked over time and benchmarked against the workflow it replaced. Teams that can show a falling cost per answer keep their funding when money gets picky.