South Korea selloff reflects forced deleveraging event

Tom Lee, the market strategist behind Fundstrat, has characterized the recent plunge in South Korean equities and AI-linked semiconductor shares as a leverage-driven liquidation rather than evidence that demand for artificial intelligence infrastructure has suddenly weakened. His view places the sell-off in a familiar late-cycle pattern: a rapid rise draws leveraged positioning, a reversal triggers margin calls, and forced selling pushes prices below levels that business fundamentals alone might justify.

Lee said the KOSPI reached roughly 9,300 on June 22, coinciding with a peak in the Philadelphia Semiconductor Index. The subsequent decline came after a near straight-line advance in AI-related shares, particularly companies exposed to memory chips and semiconductor equipment. Leveraged products widely used in South Korea amplified the reversal, according to Lee, turning an initial pullback into a broader deleveraging event.

Prime-broker data cited by Lee showed hedge funds reducing long technology exposure at their fastest pace in almost a decade. Such flows can matter more than earnings forecasts in the short term because funds facing risk limits or margin demands must sell quickly, regardless of their longer-term outlook. Lee argued that the speed of the unwind means many of the most price-insensitive sellers may already have exited the market.

Memory stocks face a more volatile demand cycle

The sell-off has reopened a familiar debate over whether the AI hardware trade has become too dependent on optimistic assumptions about data-center spending. Lee drew a distinction between Nvidia and companies supplying memory chips or semiconductor manufacturing equipment, arguing that the latter are more exposed to abrupt changes in ordering behavior.

He said Nvidia was trading at about 16 times forward earnings, while SK Hynix traded near seven times earnings at its recent high and around 4.5 times after the decline. Those figures reflect different business models, in Lee’s view. Nvidia benefits from the CUDA software ecosystem and a visible product-upgrade schedule that links its chips, networking products and software tools. Memory manufacturers sit further down the supply chain, where orders can be more vulnerable to inventory adjustments.

Cloud companies and other large data-center operators may order aggressively when they expect component prices to rise or supply to tighten. If those buyers later find they have accumulated too much inventory, suppliers can face a sharp drop in orders. This pattern, known as the bullwhip effect, often produces larger swings in chipmaker revenue than in the spending plans of the companies buying the equipment.

Lee’s comparison with Cisco during the late-1990s technology boom offers a warning against treating volatility as proof that a long-term theme has ended. He said Cisco rose from about $0.80 in 1993 to $80 in 2000, a 100-fold increase that included at least four declines exceeding 40%.

Cisco climbed to roughly $9 by 1997 before the Asian financial crisis drove a decline of about 40%, according to Lee’s account. It later rose to $18 in 1998, then dropped back to $9 amid Russia’s default and the Long-Term Capital Management crisis, before ultimately reaching $80 at the peak of the dot-com era. At that point, Lee said, Cisco traded near 200 times earnings as telecom companies used aggressive assumptions to justify enormous fiber-network spending.

The comparison does not imply that today’s AI market will follow the same path. It does show how a powerful capital-spending cycle can produce repeated, severe corrections long before the underlying technology stops expanding.

Software and digital networks gain relative strength

Lee said market leadership has begun shifting away from the most upstream AI names and toward businesses closer to software, platforms and digital-network activity. The discussion cited the GRNY ETF, described as a roughly $5 billion fund, as having outperformed the S&P 500 by about 120 basis points year to date and ranked in the top decile of comparable managers.

The release of Kimi K3, a Chinese open-source AI model said in the discussion to have 2.8 trillion parameters, has added to pressure on the idea that only the largest proprietary model developers can participate in the AI economy. Open-source models can lower the apparent cost of access for users, Lee said, although training, inference and maintaining AI services still require substantial computing expenditure.

That distinction may help explain why the market is reassessing which parts of the AI supply chain capture durable returns. Lower-cost models could expand AI usage while making it harder for hardware suppliers to rely on uninterrupted shortages and accelerating component prices.

Ethereum treasury model centers on staking income

Lee also pointed to Ethereum as a digital asset that can generate staking income, unlike a non-yielding treasury holding. He cited an Ethereum staking yield of about 3% and said BitMine Immersion Technologies held about 5.78 million ETH.

Based on the figures presented in the discussion, that position would generate approximately $6 million a week, or about $300 million annually, in staking income. If ETH reached $5,000, Lee said annual staking revenue could approach $1 billion. The company’s preferred-equity dividends were described as about $30 million a year, making staking income central to the argument that the capital structure could support its fixed obligations.

The analysis depends heavily on ETH’s market price, staking conditions and the company’s ability to stake assets at the projected rate. Yet it marks a meaningful contrast with bitcoin-treasury companies whose balance sheets are typically tied more directly to the price performance of their holdings and, in some cases, financing structures used to acquire them.

Lee said Ethereum had outperformed memory stocks by 72 percentage points since late June. He also cited tokenization activity and legislative movement around the CLARITY Act as potential supports for the Ethereum ecosystem, alongside claims that Robinhood Chain, built on Ethereum infrastructure, had surpassed $1 billion in daily transaction volume.

Inflation outlook remains part of the recovery case

Lee’s broader market view rests partly on his expectation that inflation will undershoot current forecasts. He pointed to an oil-price shock that he believes has peaked, along with softer housing and wage trends. If those forces reduce inflation pressure, the Federal Reserve could have more room to cut interest rates than current bond-market pricing suggests.

Gold presents the opposite setup in Lee’s framework. He said the metal’s three-year gain ranks near five standard deviations within a long-run history stretching roughly 12 centuries, while allowing that a further 10% decline remains possible. The discussion cited a 1% allocation as an appropriate scale for exposure.

For AI-linked equities, the immediate issue is whether liquidation pressure continues to override earnings and demand trends. Lee’s case is that the sell-off has primarily punished the most leveraged and cyclically exposed segment of the trade, leaving a sharper divide between companies selling components into a volatile supply chain and businesses positioned closer to recurring software, platform or network revenue.