Updated: 2 days ago
Data through the October 14 close
We read the AI-infrastructure sell-off as concentrated deleveraging, not broad liquidation, into the second half of October. Selling that began in data-centre financing in mid-September has spread to compute and memory, and by October 14 SPY had logged its third heavy selling day of the month. Financing cracked first, CXMT's memory news added a second shock, and herding is tightening inside AI while Health Care, Staples and Utilities still draw buyers. Converging, not yet contagious.
Data-centre credit stress has turned into a cross-border deleveraging trade. Korea, Japan and Taiwan sold off, and an Australian data-centre IPO was pulled. In memory, Chinese chipmaker CXMT added pressure with plans to enter the server DDR5 RDIMM market by year-end, challenging Samsung, SK Hynix and Micron.
On October 14, the US joined. Every AI-infrastructure group fell, from −4.8% (memory) to −9.6% (financing/neocloud), and SPY lost 1.1%.
The flows show where the selling started, and where it is herding.
Financing cracked first, chips followed
Institutional selling started in data-centre financing in mid-September and spread to AI compute and memory after October 7.
Financing Names Were the Early Warning
We track 28 US-traded AI-infrastructure names in four groups using XTech US Equity Flow powered by BMLL; constituents are in the Method.
Financing and neocloud names bled steadily from mid-September, shedding $9.6B of institutional flows by October 14.
Meanwhile, institutions were still adding AI compute, peaking at +$11.1B on September 21, and memory, at +$5.4B on September 22. Both reversed in late September, as Exhibit 1 shows.
Memory Took a Second Shock
Selling accelerated after October 7 and intensified in memory after CXMT's October 10 announcement.
CXMT unveiled plans to roll out DDR5 RDIMM server modules by year-end, nearly two years ahead of Samsung and SK Hynix's 2028 mass-production targets. By October 14, compute stood at −$8.9B and memory at −$8.4B, joining financing at −$9.6B. Power saw little selling (−$2.0B).
Since the September 30 close, memory names are down 18%, the worst price drawdown of any AI group, as Chinese DDR5 supply enters server and consumer channels. Financing names are down 16.5%.
Chips and Memory Led October 14's Selling
October 14's heaviest abnormal sellers were SMCI (z −2.5), AMD (−1.9), MU (−1.4), VRT (−1.1), NBIS (−1.0), SNDK (−0.9) and ORCL (−0.9).
Within memory, outflows concentrated in US-listed Micron (MU) and SanDisk (SNDK), as CXMT's 8,000 MT/s DDR5 ramp and 4F² server roadmap threatened incumbents' pricing power.
Korea is mixed. Samsung and SK Hynix tumbled in Seoul on the CXMT news, yet US-listed SK Hynix (SKHY) saw only light selling (z −0.4), and EWY attracted +$224M over October 1–14.
If CXMT rather than credit stress explains most of the memory selling, the spillover from financing is weaker than this sequence suggests.
The Defensive Bid Is Selective
Institutions are selling Technology and buying Health Care, Staples and Utilities, but the defensive bid is narrow and recent.
Tech Sells While Defensives Buy
Information Technology is the persistent seller among the 11 S&P 500 GICS sectors, at z −1.6 on October 8 and −1.7 on October 14.
Health Care registered +2.1 on October 12 and +1.4 on October 14, Consumer Staples +1.7 and +1.3, and Utilities +0.9 on October 14.
Below normal means a negative z-score, not net selling. On October 14, only Tech, Communication Services and Materials were below normal, yet 8 of 11 sectors were net sold in dollars, led by Tech (−$9.2B). Health Care (+$3.3B), Staples (+$0.9B) and Utilities (+$0.4B) were net bought.
In the late-September shocks, on September 16, 23 and 30, all 11 sectors were both net sold and below normal. Exhibit 2 shows the split day by day.
Health Care Is the Only Net Buyer
Over September 15–October 14, Health Care was the only sector with positive cumulative institutional net flow: +$10.7B, with net buying on 15 of 22 sessions.
Energy was not a destination, at −$12.7B. Across all 11 sectors, institutions net sold $128.3B over the window and $44.8B October 1–14.
The defensive bid broadened in October. Over October 1–14, Health Care drew +$8.0B, Utilities +$1.6B and Consumer Staples +$0.3B. Utilities and Staples remained net sold across the full window, so their buying is a reversal, not accumulation
Index-Level Selling Is Building
SPY recorded z −1.6 and −$1.2B on October 14, its third heavy selling day this month after October 5 (−2.0) and October 7 (−1.7).
That is consistent with index-level risk reduction, although the flows do not establish changes in investors' total equity exposure.
Aggregate institutional flows remain consistent with a risk-off market. If net selling persists, further downside remains our working scenario.
These aggregates cannot show that the same investors sold Technology to buy Health Care or Utilities, and if the defensives slip back below normal, the selective bid is gone.
Herding Is Concentrated, Not Yet Broad
Herding peaks inside AI as Tech’s flows decouple from the rest of the market.
Herding Is Tightening Inside AI
In four of the last eight sessions, 71–82% of the 28 AI names showed below-normal institutional flows.
Those sessions were October 5 (82%), October 8 (71%), October 12 (75%) and October 14 (82%). In September, this happened three times, all before mid-month. Exhibit 4 shows both herding gauges.
Across Sectors, Tech Is Decoupling
The average 10-day correlation between Tech's flows and the other ten sectors' fell from +0.42 on October 1 to −0.16 on October 14.
Technology selling coincides with selective defensive buying, not market-wide liquidation. Sector aggregates do not reveal changes in cash or Treasury holdings.
A move in the correlation back above zero, combined with negative z-scores in the defensives, would signal broad liquidation.
Bottom Line
Selling pressure has converged across AI infrastructure and is now visible in SPY, while selected defensive sectors retain a buying bid. What the flows cannot yet tell us is whether capital is leaving equities overall.
Three triggers would mark the shift from rotation to liquidation: Tech's correlation turning positive while the defensives' z-scores go negative; SPY selling persisting below z −1.5 as below-normal sectors climb from 3 of 11 toward 9–11; and the Health Care and Staples bid fading. A positive correlation alone could mean both sides are being bought, and ETF selling alone does not prove transmission to the underlying stocks. A fading defensive bid is usually closer to capitulation than to the start of a sell-off.
Until the defensive bid fades or index selling spreads into the underlying stocks, we read this as concentrated AI deleveraging, not broad liquidation.
Converging, not yet contagious.
Method
Sources: XTech US Equity Flow powered by BMLL, US Equity Sector Flow S&P 500 and SIP daily bars, via XTech Unifier. Institutional flows use XTech's institutional classification: net buy minus sell dollars, divided by dollar volume. Each name and sector is z-scored as (today's flow ratio − mean of the prior 60 sessions) / sample standard deviation, current session excluded. SKHY (US-listed since July 10) uses a shorter baseline, with a 30-session minimum, until it reaches 60. Stock flows include extended hours, and returns use regular-session closes. Group returns are equal-weighted. Institutional flows are trade-based buying-minus-selling estimates, not fund subscriptions or redemptions. They do not track investors' portfolio allocations or the destination of sale proceeds.
Universe
Data-centre financing and neocloud: ORCL, CRWV, NBIS, IREN, CORZ, APLD, CIFR, EQIX, DLR. Memory: MU, SKHY, SNDK, WDC, STX, EWY. AI compute: NVDA, AVGO, AMD, TSM, MRVL, SMCI, DELL, ANET. Power: VRT, CEG, VST, ETN, GEV.
Watch this video highlighting XTech US Equity Flow powered by BMLL.
Comments