SK's chairman warns of worst-ever memory imbalance while a U.S. fab stays under review

SK Group chairman Chey Tae-won has signalled that he intends to build a memory front-end wafer fabrication facility in the United States, and in the same breath forecast that the global memory market will face its worst-ever supply imbalance next year ^[1][3]. He also said he is worried about "chipflation," in which rising chip prices spread into the prices of finished products ^[4]. Read together, those two statements are a supplier telling buyers to expect scarcity and pass-through at once.

The interview was released by CNBC on the 13th, and the report carrying it is dated 17 August 2026, which places the imbalance Chey describes in 2027 ^[2][10][11]. Asked about further construction plans, Chey said he was willing to build but that finding a suitable site was truly difficult ^[2]. That is the part procurement teams should sit with. SK hynix is already putting $3.87bn, about 5.4 trillion won, into an advanced packaging plant in West Lafayette, Indiana ^[6]. Packaging is back-end work. The front-end investment, the part that actually adds wafer starts, is described only as under careful review, with surging construction costs and the infrastructure bottlenecks that TSMC and Micron have hit locally cited as the reasons for caution ^[7]. So the committed American capacity is assembly and test, and the capacity that would relieve a wafer shortage remains a study ^[13].

The demand story Chey offers is his own and worth labelling as such: he argues that in the AI era a structural shift is underway in which a single user runs more than 10 AI agents, which makes memory demand grow exponentially ^[5]. There is no audited number behind that, and the person making the claim sells the product. But it is at least directionally consistent with where the compute is heading. The alliance that LG chairman Koo Kwang-mo struck with Nvidia chief executive Jensen Huang covers robotics, AI factories and mobility, with a joint task force running from research through on-site validation to commercialisation, and is explicitly aimed at physical AI standards as the industry moves past the data centre ^[12].

The more consequential detail is commercial, not industrial. SK hynix plans to convert its Big Tech relationships from simple chip sales into long-term partnerships built around customised HBM, and to cut its own downturn exposure by expanding long-term agreements and introducing memory-as-a-service ^[8][9]. The purpose of an LTA is to move volume and price risk off the seller's book. If the largest buyers contract ahead, whatever is left for everyone else clears at the margin, and that margin is where a mid-sized hardware maker buys. A buyer who waits for spot to soften is betting against the supplier's stated strategy.

What to watch: whether the front-end review produces a named U.S. site, since Chey's own constraint is land and infrastructure rather than capital ^[2][7]; whether the Indiana packaging plant's schedule holds while front-end plans stay open ^[6][7]; and what shape memory-as-a-service actually takes when priced, because a service wrapper on a cyclical commodity is a way of selling the cycle back to the customer ^[9]. If you build anything with DRAM in it and your 2027 bill of materials still assumes flat memory pricing, that assumption is now contradicted by the seller ^[3][4].