Korea's memory premium now prices in Chinese chip exports

The technology gap is noise; the trade balance is the signal

Noise: Korea's memory leadership collapsed on 13 October. The Kospi lost 348.89 points (5.46%) that day and dipped below 6,000 intraday; SK hynix fell 10.09% and Samsung Electronics 7.39%. The trigger was CXMT chairman Cao Kanyu saying in Shanghai on 10 October that the first 4F²-cell product would be a DDR5 server module due at year-end.

The signal lies elsewhere. CXMT gave no speed, no capacity and no firm date. Industry sources in Korea say Samsung and SK hynix already know the 4F² structure but are leaving mass production until around 2028 because of yield and cost. What has changed measurably is China's place in the chip trade: customs data released on 14 October show September integrated-circuit exports up 136% at $44.95 billion.

This report asks a narrow question: what did $3.35 billion of foreign selling price in, and how quickly is the constraint that would erode Korea's memory premium loosening? The answer lies not in a technology announcement but in China's shift from buyer to seller in its chip balance.

What the sell-off priced

The size of the 13 October sell-off was out of proportion to the technical news. A day earlier, on 12 October, the same CXMT news had knocked 3.6% off Samsung and 3.4% off SK hynix. The next day the losses doubled. Three other pressures piled on: a 2.67% fall in the Philadelphia Semiconductor Index the night before, a rumoured delay to Nvidia's Vera Rubin shipments, and the end of domestic chipmakers' share buybacks.

Foreigners did the selling and domestic investors absorbed it. Institutions bought a net 1.69 trillion won and retail investors 2.23 trillion; the won weakened by 10 won to 1,345.1 per dollar. The same day in Taiwan foreigners made their third-largest net sale on record at NT$165.34 billion, and in Japan Kioxia fell more than 13% and Tokyo Electron more than 7%. The sell-off repriced not a single company but memory and the regional chain that supplies it with equipment.

China's chip balance shifts from buyer to seller

Chinese customs data show both directions at once. In January–September integrated-circuit exports rose 99.1% and imports 56.6%, while imports of storage components grew 1.5-fold. China is still a big buyer of memory, but its exports are growing almost twice as fast as its imports.

CXMT is the most visible part of this shift. Counterpoint data cited by the Seoul Economic Daily on 3 September show the company's share of global DRAM revenue rising from under 1% in 2023 to 4% in the second quarter of 2025. It reached 8% in the first quarter of 2026 and 10% in the second. More important, 63.8% of its revenue in the first half of 2026 came from outside mainland China, a rise of about 27 points in a year.

The customs data have a blind spot. They do not separate how much of September's $44.95 billion came from higher prices, how much from volume and how much from foreign-owned plants in China; that breakdown could not be verified. Reading 136% on its own as a domestic technology leap would therefore be wrong, but CXMT's non-China revenue share independently confirms that the direction has turned towards exports.

Korea's concentration constraint

On the Korean side, fragility comes from concentration. Korean customs data show chip exports for 1–10 October up 170.9% at $12.1 billion, about 46% of total exports of $26.1 billion; exports to China rose 167% to $7.59 billion. Adjusted for calendar effects, export growth per working day is only 28.5%.

This layers two risks on top of each other. Korea's growth depends on a single product, and that product's largest buyer is the country growing substitution investment in memory fastest. As CXMT's share grows by two points a quarter, the commodity memory Korea sells to China begins competing with memory China makes itself.

Washington is building a second squeeze. Commerce Secretary Lutnick said on 2 September that chipmakers not building plants in the US would pay a price for market access. According to the Seoul Economic Daily, as of 3 September neither Samsung nor SK hynix had announced a new US chip plant under the Trump administration. Korean memory has to defend its pricing power between Chinese substitution on one side and a possible US tariff on the other.

The constraints that set the pace of substitution

The first constraint is yield. SemiAnalysis estimated in June that CXMT's overall yield on its previous-generation 8-layer HBM3 was about 25%, and no yield data have been disclosed for 4F². One industry source estimates CXMT cannot bring 4F² into mass production before 2030; that estimate could not be verified.

The second constraint is customer qualification. Reliability and compatibility testing for server memory takes months, and CXMT is placing its first 4F² product in exactly that segment. The third constraint is price. China's nine-month integrated-circuit exports are up 99.1% while memory prices are high; as long as prices stay high, every module CXMT can sell enters the market without directly eroding Korea's premium.

What comes next

The first indicator is Korean customs data for the first 20 days of October: if the 167% rise in exports to China holds, the 13 October sell-off will stand as a mispricing. The second is any CXMT disclosure, by year-end, of speed, capacity and customer names for its DDR5 RDIMM. The third is Chinese customs' October data in mid-November and the MOFCOM suspension that expires on 10 November.

Our reading: the likeliest path is that the selling eases within weeks, but Korea's memory premium is retested with greater volatility at every CXMT announcement. What would refute this thesis is documented evidence that CXMT's 4F² product passed qualification at a major Chinese cloud customer by the end of 2026.