Investors Pour $25B Into Chip ETFs Amid 40% Memory Fund Plunge
Semiconductor ETFs have absorbed nearly $25 billion in new capital over the past month despite suffering steep losses, signaling that major investors still view the recent AI-driven selloff as a buying opportunity rather than a structural top.
Since peaking on June 22, major semiconductor ETFs have experienced a sharp reversal, but this has done little to deter capital inflows. Over the span of less than a month, four of the largest chip-focused funds took in a combined $24.7 billion. This aggressive buying occurred even as the underlying assets posted punishing declines.
The Roundhill Memory ETF (DRAM) illustrates the sheer scale of this dynamic. The fund plummeted nearly 40% from its June high of 80.72 to an intraday low of 48.64 on Friday. Yet, DRAM attracted $8.8 billion in fresh capital, pushing its total assets to $23.4 billion—just slightly below its late-June peak of $25.9 billion. Investor deposits effectively neutralized the market-driven destruction of value.
Broader funds saw similar patterns of heavy accumulation alongside painful drawdowns. The iShares Semiconductor ETF (SOXX) surrendered 24% while taking in $8.5 billion. The VanEck Semiconductor ETF (SMH) fell 20% and added $2.3 billion. The leveraged Direxion Daily Semiconductor Bull 3X Shares (SOXL) suffered a brutal 61% drop but still pulled in $5.1 billion.
The immediate catalyst for the pullback is straightforward profit-taking following an exceptional run. Prior to the retreat, SOXX was up 118% year to date, SMH had gained 86%, and the newly launched DRAM had surged 191%. The 3x leveraged SOXL had skyrocketed 616%. High-flying individual names like Micron Technology, Marvell Technology, and Applied Materials surrendered a large portion of those gains, though all four ETFs remain comfortably higher on the year.
Beyond simple profit-taking, the selloff coincides with renewed skepticism regarding the longevity of AI infrastructure spending. The release of Kimi K3, a Chinese open-source model reported to rival offerings from Anthropic and OpenAI, revived memories of the DeepSeek episode from early 2025. That previous event temporarily convinced markets that cheaper models would curtail massive capital expenditure budgets at hyperscalers.
Analysts note that while Kimi K3 appears competitive on capability, it currently lags on cost per task. Regardless, the uncertainty surrounding open-source threats to closed models has not shaken the fundamental bull thesis. The massive inflows suggest market professionals are interpreting the third-quarter drawdown as a routine correction in a continuing AI capital expenditure cycle, not the end of the semiconductor boom.
