Korean Chips Crashed Harder. Wall Street Is Buying Them Anyway.

On Monday, August 4, four Wall Street firms — Bank of America, Stifel, Wolfe Research, and RBC Capital — initiated bullish coverage on SK Hynix, with BofA setting a Buy rating and a W3,000,000 price target ^[1]. The same day, Bank of America named MLCC manufacturers Murata, Samsung Electro-Mechanics, and Yageo as top picks for AI growth, structuring a trade with up to 23x upside potential ^[2]. Morgan Stanley had already upgraded South Korean stocks to overweight two days earlier, forecasting 36% upside and a KOSPI target of 9,000 ^[3]. Also on August 4, US AI infrastructure stocks were selling off as investors reassessed the sustainability of compute demand and capex returns, with analysts expecting consolidation through September ^[4]. The week before, the EY US AI Pulse Survey had landed with a blunt finding: 98% of executives using token-based AI tools are reconsidering their strategies over mounting costs ^[5]. The bullishness is striking not because Korean stocks were spared in July. They were not. SK Hynix fell 46% over the month, Samsung dropped 35%, both worse than the Philadelphia Semiconductor Index's 25% decline ^[6]. Korean firms had spiked short-term debt in the first half of the year — commercial paper issuance rose 19% to W282.8 trillion and short-term bond issuance surged 90.4% to W990.1 trillion — driven by brokerages' capital needs during the AI-fueled KOSPI rally ^[7]. When the leverage unwound, it unwound everywhere. The NASDAQ's semiconductor sector posted its worst month since 2008, and the sell-off was explicitly linked to "leverage-driven failures including a crash of the Korea Kospi" ^[8]. Everyone crashed together. Korea crashed harder. The story is not in the crash. It is in what each layer must prove to recover. The Korean hardware case rests on supply, not sentiment. Micron has said its entire supply of HBM chips for 2026 is sold out and shortages will persist beyond 2027 ^[9]. SK Hynix, which holds 55 to 60% of the HBM market, reported record second-quarter results: W79.3 trillion in sales and a 76% operating margin ^[1]. South Korea's July exports surged 62.8% year-on-year to $98.89 billion, with semiconductor exports up 179% and computer equipment sales up 404% ^[10]. Samsung reported a 250-fold increase in Q2 chip profit and secured multi-year data center supply agreements ^[10]. The government has committed $518 billion to a Samsung-SK Hynix chip-making hub, targeting 80% of the global HBM market by 2028 ^[11]. Even as the KOSPI was crashing in July, South Korea's second-quarter GDP grew 0.6%, beating the Bank of Korea's 0.2% forecast, and real gross domestic income surged 15.6% year-on-year — the highest since 1988 — driven by higher chip prices and improved terms of trade ^[12]. The US platform layer faces a different kind of question. Amazon, Google, and Microsoft are projected to spend $900 billion on AI infrastructure in 2026, up from $450 billion in 2025, rising to $1.4 trillion in 2027 — funded by more than $400 billion in borrowing ^[13]. The returns on that spending, as one report put it, "remain deeply uncertain" ^[13]. ServiceNow's 2026 Enterprise AI Maturity Index found that corporate AI spending rose 110% while the maturity score reached only 51 out of 100 ^[14]. The EY survey captured the pivot plainly.

‘AI saves time’ is no longer a sufficient business case when the costs are mounting and difficult to ascertain over the long run. — Dan Diasio

Coupang's director of AI identified what he called the "Request-To-Silicon Gap" — enterprises cannot trace a single business transaction through agentic loops and model calls down to the specific compute, memory, and energy consumed.

If you can't trace a request to the silicon that served it, you aren't operating AI infrastructure. You're guessing. — Srikanta Datta

The silicon layer is the one posting record profits. The layer above it cannot yet say what a single transaction costs. The asymmetry is not in who crashed. It is in what each layer needs to prove to recover. Korean chipmakers need only to keep selling out — and a supply-constrained market does that work for them. The US AI platforms need to prove the spending was worth it, a question the market still cannot answer and may not be able to for years. The line between the two layers is not clean. Nvidia itself is investing more than $500 billion in Korean sovereign AI infrastructure — a partnership with SK Group for AI factories and HBM4 optimization, a $1 billion stake in Naver, and a 2-gigawatt SK Telecom AI cloud targeted for 2027 ^[15]. The company whose chips are the reason Korean fabs are sold out is also a major investor in the same hardware layer Wall Street is now bullish on. The two layers are entangled even as the market is pricing their recovery on different timelines.