Micron (MU) explained - DRAM, HBM, NAND, and the only US-listed pure-play memory bet

The standard $MU bear thesis writes itself: Micron is the smallest of the three HBM producers, SK Hynix has 60% market share and just secured ~70% of $NVDA's HBM4 orders for the Vera Rubin platform, and memory is the most cyclical sub-sector in semis. All three of those facts are correct. The thesis still missed Micron crossing $1 trillion in market cap on May 26, 2026 - the first memory company in history to do so.

The reason it missed is the structural read that's now playing out: for US-retail and US-institutional investors, MU is the only pure-play memory exposure available on the US tape. SK Hynix and Samsung are Korean (no clean US ADR for retail; institutional access via KRX is constrained and high-cost). Kioxia is a Japanese OTC pink-sheet with thin liquidity. SanDisk is NAND-only post-spinoff. Micron is the only ticker on a US exchange where you can express the DRAM + HBM + NAND thesis in a single line item. Those three products are also three of the four rungs on the memory difficulty ladder, so one name carries the whole manufacturing-moat spectrum. That scarcity premium is real, durable, and structural - and it's compounded by Micron actually being good at its job. This piece walks through what Micron does, where it sits in QA's DRAM / HBM Memory bubble (weight 1.0 - the cluster anchor), the customer mix, the cycle vs structural-break framing, and what would tell you whether the $1T mark holds.

Updated 2026-07-18. Micron has since reported FQ3-26 (2026-06-24): revenue $41.46B (+346% YoY), a record 84.9% gross margin, and data-center revenue above $25B in the quarter (>$100B annualized run-rate). The "~$70B FY26" consensus this piece originally carried is superseded - FY26 actuals-plus-guidance now sum to ~$129B. Figures throughout refreshed to the print; the structural read is unchanged.

The TL;DR. Micron is the only US-listed pure-play DRAM + HBM + NAND producer. HBM3E shipped in volume to NVIDIA in late 2024-2025 with "preferred supplier" status on the most advanced nodes; HBM4 share is the open question (SK Hynix appears to have won the lion's share of the Rubin platform). FQ3-26 reset the scale: revenue $41.46B (+346% YoY) at a record 84.9% gross margin, with FY26 actuals-plus-guide summing to ~$129B vs FY24 ~$38B. Market cap crossed $1T on 2026-05-26 (~1.13B shares outstanding). The bull is structural scarcity + a demand step-change; the bear is cycle reversal once HBM supply catches up - possibly 2027-2028.

What does Micron do?

Micron Technology designs and manufactures three categories of semiconductor memory:

• DRAM (dynamic random-access memory) - server, PC, and mobile. The bulk of the historical business; the most commoditized memory market; the most cyclical pricing dynamics.
• HBM (high-bandwidth memory) - stacked DRAM dies on a silicon interposer, used in AI accelerators (NVIDIA H100/H200/B100/B200/Blackwell, AMD MI300/MI350, custom hyperscaler ASICs). The single tightest constraint in AI compute scaling - see HBM is the tightest bottleneck in the AI cycle for the supply-chain map.
• NAND (flash storage) - SSDs, mobile, embedded. Diversified market with structural cyclicality similar to DRAM. Micron is one of five global NAND producers (alongside Samsung, SK Hynix, Kioxia/SanDisk, YMTC); see SanDisk (SNDK) explained for the NAND-specific pure-play angle.

The HBM business is the part that re-rated the stock. Historically, memory traded on cycle dynamics - boom-bust pricing of fungible commodity products. HBM is functionally different: long-cycle qualification with hyperscaler customers, capacity commitments rather than spot pricing, and a customer concentration on NVIDIA + a handful of frontier AI ASICs that gives suppliers pricing power they don't have on commodity DRAM.

Micron's HBM3E ramp started in late 2024 and accelerated through 2025-2026. As of recent disclosures, the company holds approximately 25-26% of the global HBM market by revenue - third-place behind SK Hynix (roughly 34-60% depending on which research house's methodology you trust) and Samsung (roughly 15-33%). The two consensus facts across analyst sources: Micron has captured "preferred supplier" status on NVIDIA's most advanced HBM3E nodes, and SK Hynix appears to have secured the lion's share of HBM4 orders for the Rubin generation.

How they make money

The actuals through fiscal Q3 dwarf what consensus modeled a quarter earlier. FQ3-26 (reported 2026-06-24) landed at $41.46B revenue, +346% year-over-year, at a record 84.9% gross margin - operating income $33.68B (81% of revenue), net income $28.86B (~70% margin), EPS $25.11, all above the high end of guidance. Stack the first three fiscal quarters ($78.96B) on the FQ4 guide midpoint ($50B) and FY26 lands near $129B - not the ~$70B consensus this piece originally carried, and a step-function 3.4x above FY24 (~$38B). The driver is HBM mix shift: HBM ASPs run several multiples of commodity DRAM per bit, and Micron converted fab capacity that previously made DDR5 into much higher-margin HBM3E. Gross margin by business unit in FQ3 tells the story - Cloud Memory 83%, Core Data Center 87%, Mobile & Client 87%, Automotive & Embedded 79%.

The structural shifts in the revenue line:

• DRAM: still the largest revenue segment in absolute terms, but margin profile is improving as the commodity DDR5 supply that flooded the market through 2023-2024 has been converted to HBM (industry-wide, not just Micron). Spot DDR5 pricing is firming.
• HBM3E: shipping in volume to NVIDIA + hyperscalers + AMD MI350. Management now expects demand to exceed supply beyond calendar 2027, and has signed 16 strategic customer agreements (SCAs) - including its first five-year SCA - that put multi-year price floors under what used to be spot-cycle revenue. Pricing is contracted, not spot; the cycle dynamics that historically defined DRAM don't apply to this slice.
• HBM4: in qualification across the industry; Micron's share appears to be lower on Rubin (SK Hynix dominant) but recovering on the post-Rubin generation. The HBM4 share read is the single most important forward variable.
• NAND: smaller revenue contribution; less re-rated than DRAM/HBM; cyclical dynamics still apply. SanDisk's $42B contracted-backlog approach is a cycle-break attempt that Micron has not (yet) replicated at the same scale.

Customer concentration is concentrated but not dangerously so: NVIDIA is the single largest customer (HBM3E + DDR5 module sales), with hyperscalers (Amazon, Microsoft, Google, Meta) and AMD next in the mix. The customer concentration risk is downstream - if NVIDIA's GPU shipment rate cools, Micron's HBM revenue cools with it on a 6-12 month delay.

Where it sits in DRAM / HBM Memory

Micron is the cluster anchor of the DRAM / HBM Memory bubble at weight 1.0 - the only memory name to carry the full anchor weight in QA's taxonomy. The bubble structure:

• $MU - weight 1.0, the US-listed pure-play anchor
• $SNDK - weight 0.85, US-listed pure-play NAND
• SK Hynix (000660.KS) - Korean, dominant HBM3E + HBM4, no clean US ADR
• Samsung Electronics (005930.KS) - Korean, broad memory + non-memory, no clean US ADR
• Kioxia (KIOXY OTC) - Japanese, SanDisk's JV partner, thin US pink-sheet liquidity

For US-retail investors, the access asymmetry is the entire story. SK Hynix is the strongest pure-play HBM exposure globally, but accessing it from a US retail account requires either a broker with KRX access (a short list including Fidelity and Interactive Brokers with specific account structures) or settling for diluted exposure via Korea ETFs (EWY, FLKR). Most US-retail flow targeting the memory cycle ends up in MU by default.

The bubble correlation is high - when memory moves, MU moves with it. The idiosyncratic upside is on HBM4 share announcements: any disclosure that Micron is recapturing HBM4 wallet share against SK Hynix on the post-Rubin generation would be the next discrete re-rating catalyst.

The numbers

The margin line is what re-rated the name and what most bears did not model. An 84.9% non-GAAP gross margin for a company that operates its own fabs runs above fabless designers like NVIDIA - a structural anomaly the SCA price floors and the HBM shortage produced, not a number memory has ever printed cyclically. The question the multiple is asking: is a fab operator sustainably out-margining fabless peers, or is this the top of the steepest cycle in the sector's history? Memory has historically traded 3-6x sales at cycle peaks and 1-2x at troughs; the current tape prices memory as no longer fully cyclical. That bet is the whole thesis.

The bull case

• Structural scarcity premium for US-retail. The only US-listed pure-play memory exposure. As long as the AI memory thesis is in retail conversation, MU is the default ticker. This is not going away.
• HBM contracts decouple from spot pricing. Contracted hyperscaler revenue at multi-year ASPs is structurally different from commodity DRAM cycle dynamics. As HBM mix grows toward 50%+ of memory revenue, the cycle volatility compresses.
• DDR5 supply rebalance. Industry-wide capacity conversion from commodity DDR5 to HBM3E has tightened DDR5 supply. Spot DDR5 pricing is firming after the 2023-2024 oversupply.
• Sold out through 2026 and into 2027. HBM3E capacity is fully contracted. Each new GPU generation requires more HBM stacks per accelerator (H100: 5 stacks → B100: 8 stacks → Rubin: more). Demand growth is structurally above industry capex ramp speed.
• Cluster-anchor flow. MU benefits from passive ETF flows targeting the memory theme (DRAM ETF, SOXX, SMH, SOXQ) as the largest US-listed name.

The bear case

• HBM4 share appears soft. SK Hynix reportedly secured ~70% of NVIDIA's HBM4 orders for the Rubin generation. If that holds, Micron's HBM4 revenue compresses materially in CY27-28.
• Memory cycle has not been repealed. Five complete cycles since 2003, each ending in 50-80% drawdowns from peak. Crossing $1T does not mean the cycle is broken - it may mean the next cycle is bigger in absolute dollars.
• Concentration on NVIDIA. A single customer accounts for an outsized share of HBM revenue. NVIDIA's own pace of GPU shipments is the gating variable.
• Multiple compression risk. Memory historically trades at 3-6× sales at cycle peaks. The current ~15× sales multiple requires the HBM-contract-floor thesis to continue compounding. A single guidance miss compresses the multiple before it compresses earnings.
• HBM supply expansion. SK Hynix, Samsung, and Micron are all aggressively expanding HBM capacity. The 2026-2027 ramp brings online substantial new bits. If demand growth slows for any reason - NVIDIA shipment pause, frontier-model training plateau - the new supply has nowhere to go.

How to access

Direct stock. $MU trades on NASDAQ in USD. Any US-retail brokerage supports it directly. For US-resident retail, the IBKR-via-LLC structure is the cleanest path - see /stack/ibkr for the access mechanics.

Through thematic ETFs:

• $DRAM (Roundhill Thematic Memory ETF) - MU is a top holding, the cleanest single-vehicle memory cycle exposure.
• $SOXX + $SMH (semiconductor ETFs) - broader semi exposure; MU is in the indices but diluted by NVDA / AVGO / TSM.
• $SOXQ (Invesco PHLX Semiconductor ETF) - alternative semis vehicle with similar dilution.

QA tracks the full ETF holdings table for MU on /stocks/mu.

Through bubble-themed watchlists:

• AI Infrastructure Stack - MU as the memory bottleneck (Layer 3, the tightest bottleneck #1).
• Top performers - 2026 YTD - MU at +90% YTD.

What to watch

• FY26 Q4 earnings (typically late September 2026) - the first print against the record FQ3 bar and the $50B FQ4 guide. FQ3 (2026-06-24) already cleared the high end; watch whether the SCA-backed demand commentary holds and how FY27 capex ($27B+ trajectory) frames next-year supply.
• The robotics read-through. Management flagged humanoid robots needing roughly 10x the memory content of a vehicle as a next demand leg - an early, unquantified signal, but the kind of structural-demand framing that would extend the "beyond 2027" tightness if it materializes.
• HBM4 share disclosures. Any concrete read on Micron's Rubin and post-Rubin HBM4 wallet share. Recovery vs SK Hynix is the single most material forward variable.
• NVIDIA capex cadence and Rubin shipment timeline. Each Rubin slip pushes Micron's HBM4 ramp window with it.
• Spot DRAM and DDR5 pricing (DRAMeXchange, ContrFLM). The legacy DRAM line still moves meaningfully on spot pricing; firming spot prices flow through to gross margin.
• Competitor capex announcements - SK Hynix, Samsung HBM fab capex. Sustained industry capex growth is the leading indicator of the supply-side cycle response.
• Bubble correlation. If MU breaks correlation with the broader memory bubble - either decoupling on idiosyncratic HBM4 news or being dragged down disproportionately on cycle fears - the structural-anchor read shifts.

Live data on this ticker: /stocks/mu - price, ETF holdings (DRAM, SOXX, SMH+), bubble correlation, bot positions.

Bubble context: /bubbles/memory - the DRAM / HBM Memory cluster MU anchors and how it's moving.

Adjacent reading: HBM is the tightest bottleneck in the AI cycle for the broader memory supply-chain context. SanDisk (SNDK) explained for the NAND-specific pure-play.

QuantAbundance is educational research. Nothing here is investment advice. See /disclosures.