AI memory shortage persists into year two; HBM and DRAM supply still critically constrained.

The AI memory shortage that began squeezing hyperscaler supply chains in 2025 has now entered its second calendar year, with Micron Technology guiding fiscal Q4 2026 revenue to roughly $50 billion and HBM4 qualification samples still rationed among lead customers. For investors translating supply tightness into portfolio exposure, three exchange-traded funds capture different slices of the same trade: the VanEck Semiconductor ETF (NASDAQ: SMH), the Roundhill Memory ETF (DRAM), and the Invesco PHLX Semiconductor ETF (NASDAQ: SOXQ). Each attacks the theme from a different angle—SMH offers a blue-chip anchor with deep Micron and equipment weighting; DRAM is the only pure-play memory vehicle on US exchanges, concentrating the upstream pricing cycle into one ticker; SOXQ delivers similar broad-semiconductor exposure to SMH at a lower price of admission. The right pick depends on whether the investor wants to bet on memory specifically, on the entire chip stack, or on cost minimization.

Memory has historically been the most cyclical segment of the semiconductor industry. DRAM contract pricing fell more than 50% in 2022, and the industry carries a hard reputation going back four decades. What changed in 2025 was the structural demand floor created by AI inference workloads. A single Nvidia H200 GPU consumes 141GB of HBM3e, and hyperscaler buildouts have pulled HBM bookings into multi-year Strategic Customer Agreements that lock in pricing visibility.

Micron's most recent quarter illustrates the magnitude. Fiscal Q3 2026 revenue hit $41.4 billion, up 346% year-over-year, with GAAP gross margin expanding to 84.6% from 37.7%. CEO Sanjay Mehrotra told investors that "Micron's record fiscal Q3 financial results and even stronger outlook for Q4 reflect the strategic value of memory in the AI era." SK Hynix and Samsung have echoed similar tightness in their own commentary, with HBM4 ramping into 2027.

SMH is the most liquid way to play the semiconductor trade. Its weighting methodology allocates meaningful exposure to companies directly benefiting from the memory shortage. Micron accounts for almost 9% of the fund, making it the third-largest holding behind AMD at 10% and Broadcom at almost 10%—a trio anchoring the allocation in the part of the value chain that moves with memory pricing. NVIDIA sits at roughly 8%, capturing the demand side of the HBM equation, since every AI accelerator the company ships pulls more memory through the supply chain. The fund also owns the picks-and-shovels layer: ASML, Lam Research, and Applied Materials together represent about 19% of the portfolio, which matters because memory makers are spending record capex to add HBM capacity. Micron alone spent $7.8 billion on capex in fiscal Q3, with that money flowing directly to equipment vendors.

SMH's recent performance reflects the setup. The fund is up roughly 70% year-to-date and 121% over the trailing 12 months, with AUM at $65 billion. The 0.35% expense ratio is reasonable for the liquidity profile. The tradeoff is concentration: the top ten holdings represent almost 78% of assets, so the fund moves with mega-cap chip sentiment more than with the broader sector, with a beta of almost 2 telling the rest of the story.

DRAM offers the most focused way to target memory pricing rather than the broader semiconductor complex, building exposure centered on the companies that produce all of the world's HBM and most of its commodity DRAM. The fund holds Samsung Electronics at 25%, SK Hynix at 24%, and Micron at almost 24%—a trio anchoring exposure directly to the part of the value chain that moves with memory cycles. Those three names alone comprise 73% of the portfolio, a concentration that defines the character of memory pricing exposure and cycle sensitivity. The remaining weight rounds out the storage stack, with Kioxia, Sandisk, Western Digital, and Seagate each sitting between 4% and 5%, adding NAND flash and hard-disk exposure for investors wanting the full memory-and-storage picture. Nanya Technology and Winbond fill out the Taiwanese memory tail.

Volatility comes with the territory. DRAM gained about 18% in the trailing month and 159% since its early April reference price, while AUM has climbed to roughly $17 billion. The 0.65% expense ratio is the highest among the three funds covered here, reflecting the cost of access to a category no other US-listed ETF replicates. The structural risk worth flagging: memory has historically been the first segment to roll over when capex guidance softens, so the same concentration that powers upside in tight cycles works in reverse when supply catches demand.

SOXQ tracks the PHLX Semiconductor Sector Index, a benchmark distinct from the MVIS index used by SMH yet built with substantial overlap in constituents. The case for owning it instead of SMH rests entirely on cost and weighting methodology: the fund charges a 0.19% expense ratio, roughly half of SMH's and the lowest in the category, a fee gap that compounds meaningfully over a multi-year hold. The PHLX SOX index uses modified capitalization weighting, producing a slightly less top-heavy portfolio than SMH. For an investor wanting exposure to the same broad supply chain—NVIDIA, Broadcom, AMD, Micron, and the equipment vendors—without paying up for the liquidity premium of the largest semi ETF, SOXQ has emerged as the natural alternative. Multiple comparative analyses through June flagged SOXQ as the cost-efficient choice against SMH and SOXX.

SOXQ's recent performance has been strong, up about 86% year-to-date and 139% over the trailing year, with AUM around $2.6 billion. The tradeoff is liquidity: SOXQ trades meaningfully less volume than SMH, which can matter for larger orders or active rebalancing.

The decision sorts cleanly along two axes: how concentrated the memory bet should be, and how cost-sensitive the investor is. An investor viewing the HBM shortage as the dominant trade and wanting to express it directly finds the cleanest fit in DRAM. Those wanting AI semis broadly, with Micron and the equipment makers riding shotgun, should look to SMH. Cost-sensitive buyers wanting similar exposure to SMH but unwilling to pay 35 basis points should turn to SOXQ. The frameworks are not mutually exclusive. Some investors blend the three, weighting SMH heaviest for a broad-semiconductor core and adding DRAM for direct memory-cycle exposure, while others reverse that emphasis when leaning into the memory thesis. The funds layer rather than compete, which is unusual for a sector this concentrated.