Five Down Days, the Nikkei Below 60,000: Is It Valuations Falling, or Earnings?

What fell this week was, first of all, the price markets are willing to pay for the future speed of the AI buildout, and the terms on which that buildout is financed — not orders already booked. Seven to sixteen days before the crash, the four big US cloud companies raised or reaffirmed their capital-spending guidance, AMD guided a quarter up about 66% year on year, and Advantest raised its full-year profit outlook and announced a buyback. As of the November 6 close, no cloud company had cut capital spending, no regulatory law arising from this election had taken effect, and no permit halt had been confirmed.

But "the order book has not changed its story" is not "nothing has changed." Two of the spenders' free cash flow has turned negative, and both have reached toward stock and bond markets (a substantial part of the equity totals they announced has yet to be issued or received), and the most leveraged data-center operators saw their credit spreads at records back in October. Prices and financing fell; the shape of the financing changed — our call extends exactly that far, and no further. What would make us change it is graded in Section 5.

1. What actually happened in five days

Facts only, with the two markets' clocks kept separate.

United States, Tuesday November 3: midterm elections. Democrats retook majorities in both the Senate and the House — the result reported consistently by major media and wire services. The new Congress convenes in January 2027.

Tokyo, November 4: as the count proceeded, the Nikkei at one point fell more than 3,000 points intraday, nearing 60,000, then recovered much of it to close down 893.23 points (-1.42%). Advantest alone subtracted 258.9 points that day.

United States, November 4: all three major indices rose — the Dow +3.03% (its best day since April 2025), the S&P 500 +1.58%, the Nasdaq +1.44% — as markets took the settled result with relief. But the same day, data-center operators (IREN, Applied Digital and others) and the semiconductor index SOX fell. The repricing began inside the chain, a day before the index crash.

Tokyo, November 5: a split day. The Nikkei closed down 1,110.73 points (-1.79%) at 61,077.90, while TOPIX rose 1.16% and Prime advancers (1,130) outnumbered decliners (410) nearly three to one. Toyota, which reported in the afternoon, raised its outlook and strengthened shareholder returns, rose more than 16%.

United States, November 5: the worst day since the April 2025 tariff shock. Dow -3.44%, S&P 500 -4.36%, Nasdaq -5.52%; the Nasdaq closed 11.9% below its October 6 record close, in correction territory. The SOX fell 12.25%, its worst day since March 2020. Among the names: Nvidia -9.26%, AMD -17.26%, Micron -17.46%.

Tokyo, November 6: the split closed into a rout. The Nikkei closed at 57,434.77, down 3,643.13 points (-5.96%) — the third-largest point drop on record, a fifth straight down session, and the first close below 60,000 in about five and a half months. TOPIX fell 5.90%. Prime advancers: 21; decliners: 1,523. The intraday low was 56,745.81 — an intraday figure; this article's text uses closing prices throughout.

Across the five sessions the Nikkei fell 11.2%, from the October 29 close of 64,659.47 to 57,434.77.

2. Why a few stocks can move the whole index

The Nikkei is not weighted by company size. It is weighted by share price: the higher a stock's price, the more a 10% move in it moves the index. SoftBank Group closed November 6 at ¥3,549, down 10.83%, and by itself subtracted 348.70 points. Advantest closed at ¥28,015, down 4.99%, subtracting 317.15 points. Two stocks: 665.85 points. Add Fast Retailing, Tokyo Electron and Recruit Holdings, and the five biggest drags total about 1,264 points — roughly 35% of the day's 3,643.13-point fall, from five companies.

That does not make the fall an illusion. On November 6 more than 98% of Prime stocks declined; the breadth was genuinely bad. What matters is the sequence. On November 5 the index fell while seven in ten stocks rose: the selling was inside the AI chain only. On November 6 it spread to the whole market. Narrow first, broad second is the classic path of a panic broadening — and it is the first piece of evidence for telling a market-wide fundamental problem apart from the repricing of a single narrative, because if every company's business had worsened at once, TOPIX would not have risen on day one.

TOPIX, weighted by market value, is the complement: the Nikkei tells you how far the expensive stars fell; TOPIX tells you the average experience of the whole market. This week, reading only one of them means misreading the week.

3. What the market is actually afraid of

The selling has three layers, of very different evidentiary quality.

Layer one, confirmed fact: the election result itself. Democrats took both chambers, and in states with heavy data-center construction, governorships mostly went to the side more cautious about that construction (as reported by media). That is all. The new Congress does not sit until January 2027; as of November 6, no federal law arising from this election is in force, and no data-center permit suspension has been confirmed. Campaign pledges, local opposition and fights over electricity prices are real — as background, not as restrictions already in effect.

Layer two, the market's fear: wire services and financial media attributed the selloff with unusual consistency — investors fear that Democratic control of Congress slows the siting, power supply and environmental approvals of data centers, that the buildout's pace falls short of what was assumed, and that assumptions about cloud capex and chip orders are then revised down. The fear has real fuel: over the past two years, disputes over data-center power use have heated up in several states, and power and grids were already among the buildout's bottlenecks. But note the grade of this evidence: it is the sellers' own account of why they sold, relayed by media — not any document that has been issued.

Layer three, our inference: what was cut first was the price paid for the future speed of construction, not the pricing of orders already placed. The timing supports this reading: on November 4, a day US stocks rallied hard, data-center operators and the SOX were already falling. The market was not reacting to the election result; it was repricing the variable called speed, and operator stocks are the most sensitive to that variable. Whether this inference holds is tested against the companies' books in the next section.

One more piece of background: this chain's valuations had been sliding well before November. The SOX's record close was in June, and by the time AMD reported on November 3 its stock was already about 22% below its October 6 high. This week was not the first brick to fall from the sky; it was the most violent stretch of a six-month repricing.

4. Has the companies' ability to make money actually changed?

Four ledgers decide this: capital spending, orders and guidance, cash flow, and financing terms. The first two answer no. The last two contain real deterioration. They are written separately, on purpose.

Capital spending: unchanged — and most recently revised upward. In the late-October earnings season, Meta (October 21) narrowed its 2026 capex range to $135–145 billion, raising the floor. Alphabet (October 28) spent $51.4 billion in the quarter, held full-year guidance at $195–205 billion, and said 2027 spending would increase significantly. Microsoft (October 28) guided calendar-year capex of roughly $170 billion, and its finance chief said customer demand continues to exceed available supply. Amazon (October 29) reaffirmed roughly $220 billion and said 2027 would be meaningfully higher. Combined guidance is roughly $720–740 billion (definitions differ — some figures include finance leases — so treat the total as approximate). These were the latest statements, seven to sixteen days before the crash. The direction was up or flat. None pointed down.

Orders and guidance: also no turn — and two of the samples are from this week. On November 3 AMD reported third-quarter revenue of $13.38 billion (up 45% year on year) and data-center revenue of $8.55 billion (up 97%), and guided the fourth quarter to about $17 billion, up roughly 66% year on year. Two days later the stock fell 17.26% in a day. In Japan, Advantest on October 28 raised its full-year operating-profit outlook by 14.3% to ¥967 billion and net profit to ¥741.5 billion — both records — and announced a buyback of up to ¥280 billion; eight days later it was the Nikkei's second-biggest drag. Read these two facts correctly: an upward revision and a buyback do not prove earnings are safe. They prove only that, as of the announcement, the company itself had no intention of guiding down. Treating a buyback as a certificate of sound fundamentals is a reading this piece refuses.

Cash flow: this is the ledger that genuinely changed. Alphabet's third-quarter free cash flow was about -$9.1 billion, a second consecutive negative quarter. Meta's was about -$2.1 billion, its first negative quarter in more than a decade, with no buyback in the quarter. Amazon's trailing twelve months were about -$42.4 billion. But draw the boundary precisely: the negatives come from the pace of spending, not from an inability to generate cash. Alphabet's operating cash flow in the quarter was still about $42.3 billion and Meta's about $32.8 billion; capital spending of $51.4 billion and $34.9 billion simply ran faster. At September 30 the two held about $265.1 billion and $82.45 billion in cash and marketable securities (Meta separately holds about $12.8 billion of restricted cash equivalents — escrow under multi-year infrastructure purchase agreements, not inside that balance and not freely usable). These companies have not stopped earning; the speed of investment is changing their cash needs and their financing arrangements, and both have begun reaching outside — but an announced total is not cash in hand. Alphabet's June package was announced at about $84.75 billion in total: the common-stock offering, the Berkshire Hathaway private placement and the mandatory convertible preferred were completed, and together with the at-the-market sales made since, about $61.9 billion had actually been received by September 30, with about $27.6 billion of ATM capacity still unissued. Nor is the whole package aimed directly at AI construction: the 10-Q states that the proceeds of the $40 billion ATM program are primarily intended to meet tax obligations associated with employee equity grants; AI infrastructure is an important use of the remaining parts of the financing, such as the mandatory convertible preferred. Meta's late-October framework runs to as much as $97.5 billion, but only $37.5 billion was priced — $12.5 billion of preferred stock issued and received, and $25 billion of common stock sold forward, with the cash arriving on settlement dates running to November 2027 — while the remaining $60 billion at-the-market program has not even been established. By the Bank of England's count, global AI-related debt issuance had reached about $450 billion by early September — more than double the whole of 2025.

Financing terms: stress is concentrated at the most leveraged end, and it predates the election. The five-year credit default swaps of data-center operator CoreWeave rose to a record area around 1,000 basis points in mid-October and set another record on November 5; its unsecured bonds yielded around 15% in October. SoftBank Group's CDS reached a four-year high on November 6 (Reuters market data). These are real price signals, but they mark a small group — the companies that build data centers with borrowed money — not the credit market as a whole: on November 5 the US 10-year Treasury yield fell, to about 4.84% (Reuters market data), a flight-to-safety bid rather than a rate shock. On high-yield spreads, we could not obtain an official reading for November 5 itself; the most recent reading we could find (late October, around 340 basis points) sits far below past stress episodes — but without same-day data, "credit markets are fine" is a conclusion we will not print. That cell stays empty rather than filled with a guess.

The contrary evidence, laid out as found: Arm, reporting November 4, guided roughly in line, but royalty revenue grew only 8% and missed estimates, and its royalty growth outlook for the next quarter was low single digits — the stock fell 13.08% the next day. That is one of this week's few genuine data points of slowing demand — except that the slowing was in the smartphone chain (company and media accounts point to memory prices squeezing handset shipments), while data-center royalties doubled. Among equipment makers, Lasertec's July–September operating profit fell 41% year on year: one company's weak print. Qualcomm's smartphone-chip miss is the same species. All three go into the "weakens" column of the next section, not the "overturns" column — because none of them has spread into a pattern.

5. OUR READING: repricing, or a fundamental turn?

Our call is in the thesis: this week was first a repricing — the price layer and the financing layer moved first, the expectations layer followed, and the fundamental layer (orders, capital spending, sector-wide guidance) had not moved as of November 6. We state the strength of that evidence ourselves.

Supporting it: the direction of the four cloud companies' guidance, the timing of AMD's and Advantest's guidance, Tokyo's split breadth on November 5, and yields falling rather than spiking. Making it uncomfortable: three negative free-cash-flow ledgers (all three sit on the investment side — both companies' operating cash flow for the quarter was still strongly positive), the credit prices of leveraged operators, and Arm's royalty growth. Together those three say the buildout's speed has come to depend more heavily on outside financing — and if the financing window narrows, speed gets cut before orders do. So our complete sentence is: prices and financing terms fell; fundamentals have not yet turned, but the cost of their financing premise has risen.

What would change our mind is fixed here, in three grades:

Directly overturns (any one of these, and the thesis is withdrawn and rewritten): a major cloud company formally cuts its capex guidance; regulation turns into enacted law, permit restrictions, or the actual delay of a named project; or this earnings season produces a broad pattern of order and outlook cuts across semiconductor and equipment companies.

Weakens the call (accumulating evidence downgrades us to "repricing and fundamental deterioration together"): cloud free cash flow stays negative and reliance on outside funding deepens; Arm-style growth slowing spreads to a second and third company; weak equipment earnings multiply; credit stress at leveraged operators spreads to investment-grade borrowers.

Insufficient on its own: one company's miss (Qualcomm), one local bill or campaign pledge, one day of margin-call rumors, or macro data such as CPI — CPI moves the environment of rates and valuations; however hot it prints, it proves nothing about whether AI orders improved or worsened.

The checkpoints are scheduled. November 10: SoftBank Group reports its interim results (disclosure at 15:30 JST) — simultaneously this week's biggest Nikkei drag and one of the largest holders of AI assets, its outlook and asset-valuation commentary are the nearest hard test. The same day, the US publishes October CPI (Bureau of Labor Statistics schedule) — read it only for what it does to the rate environment. Tonight, US time November 6, comes the October jobs report. Later in November, Nvidia's earnings are the chain's next full audit. As for the rumored margin-call cascade: as of the November 6 close, all we found was analyst speculation quoted by media, and no confirmed evidence. Absent evidence is reported as absent.

A final discipline: the temporary absence of bad news is not proof that fundamentals will never deteriorate. This section bets on the evidence as of today — it is not a prophecy that AI investment will never run into trouble.

6. What an investor can actually control

On the market's side, an individual controls nothing. What can be controlled is one's own position inside a selloff. Three questions, in order:

First: is your risk concentrated in the same chain of AI names? The portfolios hurt worst this week were not hurt by picking one wrong stock; they held what looked like four or five positions — chips, equipment, operators, SoftBank — that were in fact a single variable. Before the next question, see that clearly.

Second: do you carry financing, leverage, or margin-call risk? If you do, this piece does not advise you to "do nothing tonight" — the opposite. Verify three things tonight: how far you are from a margin call; in the worst case, which of your money gets sold by force; and whether any large, unavoidable payment falls due in the next three months. Leverage turns a wrong judgment into an immediate exit, and that is its fundamental difference from an ordinary paper loss. Only people without leverage have the standing to "wait and see."

Third: what is this money's horizon? Money needed in three months and money that can sit for three years should not take the same action in front of the same -5.96% candle. In selloffs, what actually wounds people is usually not a wrong direction call but a wrong maturity match: the forced seller at the low lost not on judgment, but on their calendar.

Answer these three first; only then talk about what to buy, sell, or add. Do not reverse the order. What follows is the author's own practice — the habits one person changed over the past half year. Not a law, and not advice.