HBMX ETF Review 2026: Tuttle Capital Concentrated Memory Stack ETF

HBMX at a Glance 2026

What is HBMX?

HBMX is the Tuttle Capital Concentrated Memory Stack ETF, launched in June 2026 on the Cboe exchange. Tuttle Capital Management — known for thematic concentrated ETFs including FOTO (photonics) — designed HBMX as a strict, pure-play vehicle focused exclusively on the memory semiconductor ecosystem.

The fund's defining rule: every holding must derive at least 25% of revenue from — or have substantial strategic focus on — one or more of the following memory-stack segments:

• DRAM (Dynamic Random-Access Memory) — volatile memory used in servers, PCs, mobile, and AI accelerators
• NAND flash — non-volatile storage underpinning SSDs, enterprise storage, and AI inference caching
• HBM (High Bandwidth Memory) — stacked DRAM with through-silicon vias; the defining memory technology of the AI era
• Advanced packaging — CoWoS, HBM-on-interposer, 2.5D/3D IC integration that physically connects HBM to GPUs
• Semiconductor test and inspection equipment — wafer-level testing required for every HBM die before shipping

Critically, HBMX is NOT a broad semiconductor ETF. It does not hold NVIDIA (GPU/compute), Intel in its processor business, or Qualcomm (mobile chips). If it's not memory or memory-enabling, it doesn't belong in the portfolio. This is intentional concentration — the point is maximum exposure to the memory cycle, not diversification across the whole chip sector.

Why Memory Matters for AI

Every AI training run is fundamentally memory-bandwidth-constrained. You cannot simply add more GPU cores without proportionally more memory bandwidth — the data pipelines starve the compute. The industry's answer is High Bandwidth Memory (HBM): DRAM dies stacked 8–12 layers high using through-silicon vias (TSVs), delivering roughly 10× the bandwidth of standard DDR5 per watt.

NVIDIA's H100 GPU uses 80GB of HBM3; the H200 uses 141GB of HBM3E; the Blackwell B200 uses 192GB of HBM3E. Each new GPU generation consumes more HBM, and hyperscalers order tens of thousands of these accelerators per quarter. The three HBM suppliers — SK Hynix, Samsung, and Micron — represent essentially the entire global supply. There is no alternative to HBM for frontier AI training. Demand for HBM is projected to grow approximately 5× by 2028 as model sizes scale and inference deployments multiply. DRAM and NAND more broadly are also entering what analysts are calling a structural supercycle driven by AI — not a typical cyclical recovery.

Full Holdings Breakdown — Top 10

HBMX spans the complete memory value chain: from the equipment that manufactures memory dies, to the companies that make them, to the packagers and testers that deliver finished product to GPU makers.

Remaining positions (11–35) include UMC (mature node foundry for memory controllers), Camtek (inspection for advanced packaging), ASE Technology (OSAT packaging), and smaller equipment and materials companies. Holdings are rebalanced quarterly. Weights as of estimated June 2026 inception.

HBMX vs SMH vs SOXX vs MU Direct

The key question for any thematic ETF: what does it give you that cheaper alternatives don't? Here's how HBMX stacks up against the most common alternatives.

HBMX's unique value proposition: it gives you the memory semiconductor ecosystem — including packaging and test equipment companies that SMH/SOXX barely touch — without the non-memory dilution. You're not buying NVIDIA's GPU business when you buy HBMX. The tradeoff is a 0.60 percentage-point higher expense ratio versus the big semi ETFs.

HBM Technology Deep Dive

What is High Bandwidth Memory? HBM is a type of DRAM (Dynamic Random-Access Memory) where multiple memory dies are stacked vertically using through-silicon vias (TSVs) — microscopic copper pillars that pass electrical signals through each die. The stack sits on a silicon interposer alongside the GPU or AI chip, connected by an extremely wide data bus (1,024 bits in HBM3 vs 64 bits for a standard DDR5 DIMM). The result: dramatically higher bandwidth in a smaller physical footprint with much lower power consumption per bit transferred.

SK Hynix first-mover advantage: SK Hynix supplied essentially all of NVIDIA's HBM3 and early HBM3E while Samsung struggled with qualification. This gave SK Hynix extraordinary pricing power and margin expansion in 2024–2025. Micron is now aggressively ramping its HBM3E and is the second qualified supplier for NVIDIA. Samsung is working to close the gap. The HBM4 transition — expected in 2026–2027 — will reset supplier dynamics as all three compete for next-generation design wins. HBMX's holdings across Micron, AMAT, ASML, and packaging companies give exposure to this transition across the stack, not just the chip manufacturers.

The DRAM/NAND Cycle — Why This Time Is Different

Memory semiconductors are historically among the most cyclical industries in the market. DRAM prices have experienced peak-to-trough collapses of 70–80% across multiple cycles (2001, 2008, 2015–2016, 2022–2023). The 2022–2023 DRAM glut saw manufacturers cut prices to the bone as PC and smartphone demand cratered post-COVID. Micron reported multi-billion dollar losses. Equipment companies cut forecasts.

Why bulls argue this cycle is structurally different:

• AI data center demand for HBM is not cyclical demand — it scales with hyperscaler capex, which is guided up through 2027+
• HBM supply is capacity-constrained (requires specialized packaging and TSV infrastructure), not just foundry capacity
• DRAM ASPs (average selling prices) are rising due to HBM premium mix shift — even commodity DRAM benefits from tighter overall DRAM supply
• NAND: AI inference requires fast NVMe SSDs for KV-cache retrieval at scale — enterprise SSD ASPs holding up better than consumer NAND
• Memory inventory has normalized post-2023 glut; days of inventory are back to historical averages across the supply chain

The bear case is that AI capex eventually normalizes or gets disrupted by more compute-efficient models (e.g., sparse architectures, model compression), causing a demand air pocket. Historical precedent suggests these cycles end badly for memory stocks. HBMX holders need to be comfortable with that cyclical risk.

Advanced Packaging — The New Semiconductor Bottleneck

For most of semiconductor history, performance improvement came from making transistors smaller (Moore's Law). As transistor scaling has slowed, the industry has turned to packaging innovation — physically integrating multiple chips together in increasingly sophisticated ways. This is where HBMX's packaging holdings (AMKR, KLA, ONTO, Camtek) become critical.

CoWoS (Chip-on-Wafer-on-Substrate) is TSMC's 2.5D packaging technology that mounts an HBM stack and a GPU die side-by-side on a silicon interposer, connecting them with thousands of micro-bumps. Every NVIDIA H100, H200, and B200 ships in a CoWoS package. CoWoS capacity was the binding constraint on NVIDIA GPU shipments in 2023–2024 — not the GPU die itself.

OSAT (Outsourced Semiconductor Assembly and Test) companies — primarily Amkor (AMKR) and ASE Technology — are building dedicated HBM packaging capacity. As HBM4 transitions to even more complex packaging (hybrid bonding, backside power delivery), OSATs with advanced capability command significant pricing power. Amkor is the largest US-listed OSAT and a direct HBMX holding.

Advanced packaging is now a competitive moat, not a commodity service. It is why HBMX extends beyond just Micron and adds packaging names that traditional semiconductor ETFs underweight.

Bull Case for HBMX

• AI HBM supercycle: hyperscaler capex (Microsoft, Google, Meta, Amazon, xAI) guided north through 2028 with memory as the binding constraint
• Memory oligopoly: SK Hynix + Samsung + Micron control 95%+ of global DRAM supply — pricing discipline is structurally easier than most industries
• No substitute for HBM in AI training: the physics of bandwidth-per-watt makes HBM irreplaceable at the frontier for the next 3–5 years
• HBM3E → HBM4 transition creates upgrade cycles across the entire value chain — equipment, packaging, test all benefit from capex renewal
• HBMX provides pure, concentrated exposure: instead of owning 0.5% Micron through SMH, you own 8% Micron plus all the enabling stack companies
• Packaging bottleneck gives AMKR and advanced packagers pricing leverage not seen in prior memory cycles

Bear Case for HBMX

• 0.95% expense ratio is expensive: over 10 years at 8% annual return, you pay ~$85 per $1,000 invested vs ~$32 for SMH — a 73% higher fee drag
• Extreme concentration: 8 positions = roughly 60% of the fund; one bad earnings from ONTO or MU materially impacts NAV
• Memory cyclicality is real and severe: 2022–2023 saw commodity DRAM fall 50%+ and memory stocks crater — HBM did not exist at scale but a demand air pocket could repeat
• Customer concentration risk: NVIDIA is the majority buyer of HBM globally — a slowdown in NVIDIA shipments would hit every HBMX holding simultaneously
• Small and new ETF: HBMX launched June 2026 with limited AUM; thin trading volume = wide bid-ask spreads and closure risk if AUM doesn't scale
• China risk: ASML, AMAT, and LRCX face ongoing US export control restrictions on selling to Chinese memory fabs; escalation would hurt equipment holdings

Key Remaining Holdings — Teradyne, ACLS, MKSI, KLA, UMC

Beyond the top four or five positions, HBMX fills out its 20–35 stock portfolio with companies that play specific, non-obvious roles in the memory ecosystem:

Key Risk Factors to Monitor as an HBMX Holder

Owning a concentrated thematic ETF means tracking the specific drivers that could break the thesis. For HBMX, watch these quarterly indicators:

Who Should Own HBMX?

HBMX is a satellite position — not a core holding. Here's the ideal investor profile:

• AI infrastructure bulls who already hold broad semiconductor ETFs (SMH, SOXX) and want a memory-specific tilt on top
• Investors who want exposure beyond owning just Micron (MU) as a single stock — HBMX diversifies across the memory value chain
• Portfolio allocators comfortable with 5–10% in a concentrated thematic bet; not 20–30%
• Active investors who monitor quarterly earnings from ONTO, MU, AMAT, and AMKR and can respond to changing HBM demand signals

NOT right for:

• Passive, low-cost, buy-and-hold-forever investors — the 0.95% fee and single-sector concentration are a poor fit for a core holding
• Investors with less than 3–5 year horizon — memory cycles can crater in 12 months
• Anyone who cannot tolerate a 50%+ drawdown in a down cycle
• Investors who don't already own diversified equity exposure — HBMX is not a standalone portfolio

The 0.95% Fee — Does It Kill the Return?

HBMX's 0.95% annual expense ratio is 2.7× higher than SMH (0.35%) and dramatically higher than direct stock ownership. Over long holding periods, fees compound into meaningful performance drag. Here's how the math works on a $10,000 investment at 10% gross annual return:

The fee drag is real but not fatal for a tactical position. Over 1–3 years, the difference is small relative to HBMX's potential to outperform if the memory thesis plays out — a 10% outperformance vs SMH in a single year more than offsets the fee gap. The fee becomes a serious problem only if HBMX is a large core holding held for 10+ years without differentiated return. Used as a 5–10% satellite for 3–5 years, the fee is a reasonable cost for targeted exposure.

What HBM4 Means for HBMX Holdings

The HBM3E to HBM4 transition — expected to begin volume production in 2026–2027 — is not just an incremental spec bump. HBM4 involves significant architectural changes that affect nearly every company in the HBMX portfolio:

• Die stacking increases from 8–12 layers (HBM3E) to potentially 16 layers — more TSV drilling, more packaging complexity, more test coverage required per stack
• Microbump pitch shrinks further, requiring tighter lithography — ASML's EUV tools become even more critical; KLAC's inspection tools must adapt to finer defect detection
• Logic die integration changes: HBM4 may integrate more functions into the base die, potentially benefiting AMAT's advanced deposition tools
• New interposer designs for CoWoS-L (liquid-cooling compatible packaging) create additional advanced packaging demand for Amkor
• Test time per device increases as stack height grows — more revenue per tool for Teradyne and Onto Innovation
• First customer: NVIDIA's next-generation GPU platform after Blackwell is expected to be the launch customer for HBM4 in high volumes

The HBM4 transition is a catalyst that could drive incremental capital expenditure across the memory stack in 2026–2028, benefiting HBMX holdings at multiple layers of the value chain simultaneously. It is one of the key arguments for owning the full stack rather than just Micron alone.

Bottom Line Verdict

HBMX is a well-constructed, deliberately concentrated expression of one of the most compelling structural themes in technology investing: the memory semiconductor stack that makes AI training possible. The holdings cover the full value chain — from ASML's EUV lithography through Micron's DRAM fabrication through Amkor's advanced packaging through Onto Innovation's quality inspection. That is genuinely differentiated from anything in SMH or SOXX.

The risks are real: 0.95% annual fee, limited track record (launched June 2026), extreme single-sector concentration, and exposure to one of the most cyclical industries in markets. These are not dealbreakers — they are terms of the trade.

For investors who believe the AI infrastructure buildout sustains HBM demand through at least 2028, and who want concentrated, full-stack exposure rather than the diluted memory tilt of a broad semiconductor ETF, HBMX deserves a 5–10% satellite position. For everyone else, MU direct or the memory weighting inside SMH is probably sufficient.

Analyze memory semiconductor stocks

Read Next