AI is swallowing memory, and phones and PCs are paying the bill

The noise is a crash, the signal is scarcity

The noise says the AI bubble has burst. The Kospi lost 454.59 points on 15 October; an intraday fall of as much as 9.29% triggered a 20-minute circuit breaker, and the index has shed 21% since 6 October. The Nikkei 225 fell 4.90% and the Taiex 3.41%. The headline reading is that AI demand is over.

The signal comes from the production line. On the same day TSMC raised its 2026 dollar-denominated sales growth forecast to 44%, from just above 40% in July; the midpoint of its fourth-quarter guidance implies 58% growth year on year. Chief executive C.C. Wei said demand still exceeds supply. ASML, in quarterly results reported by the Taipei Times on 15 October, moved its 2026 sales forecast to €44–45 billion and warned of a supply gap over the next two years. Demand has not collapsed; the scarcity is eating into another market.

Memory scarcity is eroding consumer electronics

The measurable break is in consumer devices. Preliminary IDC data show third-quarter smartphone shipments down 22% year on year at roughly 255.3 million units; Omdia measures the decline at 20%. IDC says memory costs rose by about 350% in a year. The pace of increase has slowed but not stopped: the quarter-on-quarter rise in the third quarter was in the low double digits, against 50–100% in earlier quarters.

PCs tell the same story. IDC put third-quarter shipments at 62.7 million units, down 20.1%; Omdia put them at 58.1 million, down 21.2%. According to Omdia, DRAM and SSDs now make up about 40% of a PC's bill of materials, up from around 15% in the past. Omdia expects another 24% decline in the fourth quarter.

The mechanism is simple. Memory makers are steering capacity towards high-margin AI server memory, and the price of standard memory is pushing up the price of consumer devices. Manufacturers are raising prices, cutting cheaper models and switching to cheaper displays and sensors. Omdia says the combined share of Samsung and Apple has risen to about 45%, a gain of 8 points in a year; scarcity is concentrating the market among the largest players.

What the Kospi priced in

Korea's sell-off is pricing funding costs, not a collapse in demand. US consumer inflation came in at 3.7% for September against expectations of 3.6%, and the Korea Times reports that the 10-year US Treasury yield rose above 5.3%. The night before, CoreWeave had fallen 15.46% and Oracle 8.83% in the US; both are among the companies building out AI infrastructure on borrowed money.

The second pressure is doubt over returns. A Gartner projection cited by Herald Business puts AI infrastructure spending at about $1 trillion in 2026 and a further $6 trillion in 2027–2030. Data-centre revenue, by contrast, may reach only $3.6 trillion by 2030, while $8.5 trillion would be needed to justify the spending. As rates rise, the present value of that gap grows.

Korea's fragility comes from concentration. Samsung and SK hynix account for roughly half of the Kospi's market capitalisation, and the two stocks lost 24% and 31% respectively over 7 sessions. The biggest winners of the memory shortage were the first assets sold once funding costs rose.

The constraints are equipment and interest rates

Equipment is the constraint that sets supply. ASML plans to lift low-NA EUV capacity by 30% next year to about 85 machines. Lead times for advanced tools exceed a year, and some customers want to place orders for 2029–2030. TSMC held its 2026 capital budget at $60–64 billion and said it would more than double 2-nanometre capacity in 2027.

That timetable delays relief on memory. Omdia does not expect the memory crisis to ease before 2028 and projects the phone market to shrink 12% in 2026 and 6% in 2027. Interest rates are the constraint that sets demand. As long as AI investment is debt-financed, every rise in US yields tests the spending decisions of cloud companies, and that test shows up fastest on the Seoul and Taipei exchanges.

The shortage is spreading to mature-node chips, and governments are stepping in with capital. On 14 October GlobalFoundries was chosen to operate two New York fabs that Japan will finance with up to $21 billion; the US Commerce Department is taking a stake of about 4% in the company. Its chief executive said the whole industry is short of supply. The fabs will make automotive and industrial chips, but the first will not come online before 2029. This is a policy written for supply security in the 2030s, not for today's shortage.

What comes next

The first indicator is capital spending guidance in cloud company earnings at the end of October. If guidance holds, the scarcity reading of TSMC and ASML remains valid and the Kospi sell-off stays contained as a funding scare. The second indicator is fourth-quarter device data from IDC and Omdia. If Omdia's 24% PC decline materialises, memory costs will have turned into a lasting contraction of the consumer market.

Our reading is that the most likely path is a divergence in which AI hardware demand stays strong while consumer electronics shrink through 2027. The development that would falsify this thesis is at least two of the large cloud companies cutting their 2027 capital spending guidance, together with TSMC coming in below its guidance range in the fourth quarter.