The Day a Model Became a Market Event

On 27 January 2025, the cap-weighted S&P 500 closed 1.46 per cent lower and the equal-weighted version of the same index closed 0.07 per cent higher. Two portfolios holding identical companies, differing only in how much of each they hold, spent the session moving in opposite directions. That gap is the whole event, and it is a fact about index construction rather than about artificial intelligence.

The session followed DeepSeek's release of R1, an open model whose reported training cost was a small fraction of the figures associated with frontier Western systems. What the tape did with that news is measurable, and Figure 1 measures it. Nvidia fell 16.97 per cent. A semiconductor ETF fell 9.83 per cent. The Nasdaq-100 fell 2.97 per cent. The broad index fell 1.46 per cent. The average S&P 500 constituent, which is what the equal-weighted series tracks, did not fall at all.

Section 01What the cross-section shows

The ordering in Figure 1 follows exposure to one premise: that frontier AI capability requires continually larger quantities of advanced silicon, and the power and cooling to run it. If R1 could be had cheaply, that premise was worth less than the price implied.

The single-name detail sharpens the point, because the losses did not track industry membership. Marvell fell 19.10 per cent and Broadcom 17.40 per cent, both roughly in line with Nvidia. But Vertiv, which makes data-centre power and cooling equipment and holds no semiconductor exposure at all, fell 29.88 per cent, and GE Vernova, an electrical-equipment manufacturer, fell 21.52 per cent. Meanwhile Microsoft, whose business would plausibly benefit from cheaper inference, fell 2.14 per cent, barely more than the index.

A sector classification would have grouped Microsoft with Nvidia and separated both from Vertiv. The market did the opposite. It sorted securities by their dependence on a single forecast about capital spending, and that ordering cut across industry lines in both directions. Whatever else the session was, it was a well-organised piece of price discovery about one assumption.

Section 02Why the index moved and the average constituent did not

An index return is a weighted sum of constituent returns. When the weights are close to equal, a large loss in one name is diluted by the four hundred and ninety-nine others. When a handful of weights dominate, the dilution largely stops working, and a single-name shock passes into the index at close to its full size.

The gap between the two series that day was about 1.5 percentage points. Measured like for like, comparing the two exchange-traded vehicles that hold the same constituents under different weights, it was 1.48 points, which Issue 005 places in the bottom half of one per cent of daily observations across the twenty-three years both have existed. Nothing about the constituent list changed that day. What differed was the weighting rule, and the weighting rule turned out to be carrying a position that no investor in a broad index had chosen to take.

r_index  =  sum_i ( w_i * r_i )          # w_i = market-cap weight, sum w_i = 1

equal weighting:   w_i = 1/N     a shock r* enters as  r*/N        (diluted)
cap weighting:     w* large      a shock r* enters as  w* * r*     (barely diluted)

r_cap - r_equal  =  sum_i ( w_i - 1/N ) * r_i

Section 03What the session does and does not establish

An event study of this kind establishes association inside a window. It cannot establish that the R1 release caused the move, and the design has an obvious confounder: the Federal Open Market Committee met on 28 and 29 January, inside the post-event window plotted in Figure 1. The cross-sectional evidence is stronger than the time-series evidence here, because a monetary-policy surprise would not sort securities by data-centre capital spending, and something clearly did.

What the session does establish is narrower and more durable. Counting securities overstated the diversification actually held. A broad index contained, by weight, a large undiversified bet on one forecast, and the bet was legible only once something priced it. Fifteen sessions later the gap had largely closed: Nvidia was 2.26 per cent below its pre-event level and the equal-weighted index 0.15 per cent above it. The exposure did not go away when the prices came back.

Supervisors had described the same geometry from the other side. In a 2024 Financial Stability Review feature, the European Central Bank considered what happens if many institutions come to rely on similar models from a small number of suppliers, and identified correlated decisions, herding and distorted asset prices as the consequences to watch. That is concentration in the supply of judgement. January 2025 was concentration in the demand for one outcome. The mechanism is the same in both: independent participants reach the same position, and the system finds out how alike they are only when a single question is put to all of them at once.

The index did not fall because the market learned something about models. It fell because so few premises were carrying so much of the weight.

This research is analysis and commentary for general information. It is not investment advice, an offer, or a solicitation, and it contains no price forecasts. Returns reported in the text are the author's own calculations from the sources cited; the interpretation is the author's.