Micron Technology (MU)

The #3 HBM supplier betting on a memory-manufacturing comeback — Micron's HBM3E is qualified with NVIDIA for mainstream inference and some training, but the company faces a steep uphill battle against SK Hynix's commanding HBM4 lead and Samsung's scale.

1. Core Product / Service

Micron Technology (NASDAQ: MU) is an American memory and storage semiconductor company headquartered in Boise, Idaho. It is the third-largest HBM (High Bandwidth Memory) supplier behind sk-hynix and samsung-hbm, and the largest US-based memory manufacturer.

AI-relevant products:

• HBM3E — Micron's 8-Hi and 12-Hi HBM3E (High Bandwidth Memory Gen 3 Enhanced) products. Micron achieved NVIDIA qualification for HBM3E in mid-2024, making it the second HBM supplier cleared by NVIDIA after SK Hynix. Bandwidth per stack up to 1.2 TB/s at 9.6 Gbps/pin. Volume shipments began in second half of 2024.

• HBM4 — Micron is developing HBM4 (16-Hi stacks, >2 TB/s bandwidth per stack) targeting mass production in 2026. Micron's HBM4 strategy uses TSMC's 12nm base die technology — the same approach as sk-hynix — giving it a potential path to competitive HBM4 performance.

• DDR5 / LPDDR5X — Conventional DRAM used in AI servers for CPU system memory. Micron has strong positions in DDR5. AI server demand for DDR5 is growing because each GPU node requires substantial system DRAM.

• NAND (SSD) — Data center SSDs for AI training data storage and checkpointing. Micron is a top-3 NAND manufacturer alongside Samsung and Kioxia/WD.

• HMC (Hybrid Memory Cube) — legacy technology that predated HBM but laid the groundwork; no longer commercially relevant.

Key differentiator: Micron is the only US-based HBM manufacturer and the only HBM supplier outside of South Korea. This geographic diversification matters for US government AI chip supply chain security policy.

2. Target Users & Pain Points

• nvidia — Micron's HBM3E qualified for NVIDIA's H200 and B200 GPUs for mainstream inference and some training workloads. Micron is seeking HBM4 qualification for NVIDIA's Rubin GPU generation (2H 2026).

• broadcom custom ASIC customers — Broadcom uses Micron HBM alongside SK Hynix and Samsung for Google TPU and other custom AI accelerators.

• AMD — AMD's MI300X and MI325X use HBM3e; Micron supplies some portion.

• Hyperscalers (Microsoft, Amazon, Google) — for AI server DDR5 and enterprise SSD storage.

• Traditional data center and PC OEMs — non-AI DRAM customers (still >50% of Micron's revenue).

Pain solved: HBM is the bandwidth bottleneck for AI inference. Without enough HBM bandwidth, GPU compute cycles idle waiting for data. Micron provides an alternative to SK Hynix's dominant supply — particularly important for customers who want dual-source risk mitigation. For DDR5 and NAND, Micron supplies the broader memory required to build complete AI server systems.

3. Competitive Landscape

Micron's competitive position: pragmatic #3 — it is not competing head-to-head for HBM volume leadership, but rather positioning as a stable, qualified second/third source for AI GPU makers. Micron's HBM3E is already qualified for NVIDIA, and its HBM4 effort is progressing. The question is whether Micron can achieve the yield and capacity needed to meaningfully increase its HBM market share from the current ~10-15% to 20-25% by 2027.

4. Unique Observations

• Aschenbrenner 13F Q1 2026: PUT $584M (1.728M shares, 4.27% of book) + CALL $422M (3.09% of book) — VOLATILITY STRADDLE — This is NOT a directional short. Aschenbrenner holds both PUT and CALL options on MU simultaneously — the classic signature of a volatility straddle (betting on large price movement regardless of direction). Micron is the only position in the entire 13F with significant PUT + CALL exposure. This suggests Aschenbrenner anticipates a major catalyst for MU: HBM4 qualification outcome, memory cycle inflection, or an earnings surprise. The PUT is larger than the CALL ($584M vs $422M), giving a slight downside tilt, but the structure is fundamentally a volatility play, not a directional bet. Notional implied price: ~$338/share (both sides).

• HBM market share trajectory: Micron's HBM market share was estimated at ~10% in 2024, projected to reach perhaps ~15-20% by 2027 if HBM4 qualification succeeds. This is constrained by: (a) SK Hynix's first-mover advantage with HBM4, (b) Samsung's scale and vertical integration, and (c) Micron's own capacity allocation between HBM and conventional DRAM.

• HBM4 strategy parity: Micron is adopting the same TSMC 12nm base die approach as SK Hynix for HBM4, potentially leveling the playing field. However, SK Hynix has a 2+ year head start on HBM4 process learning and will already be in volume production when Micron samples.

• US government supply chain angle: As the only US-headquartered memory manufacturer, Micron benefits from CHIPS Act funding (~$6.1B in direct grants for US fab expansion in New York and Idaho). This provides a structural cost advantage for domestic memory manufacturing.

• Memory cycle risk: MU's non-AI business (PC DRAM, mobile DRAM, NAND SSDs) represents >50% of revenue and is subject to the notoriously cyclical memory market. AI-driven HBM demand provides a growth buffer but does not fully insulate MU from DRAM/NAND price cycles.

• AI token supply chain role: HBM is the memory layer in L1 (Physical Compute) of the ai-token-supply-chain. Every AI GPU needs HBM — it's the bandwidth bridge between compute and data. Without HBM3E/HBM4 capacity, token throughput is limited. Micron's HBM capacity incrementally adds to the global HBM supply that AI chips consume.

• Capacity buildout: Micron is building a $15B DRAM fab in Boise (2025-2028), a $15B expansion in Clay, New York, and expanding its Singapore and Taiwan facilities. HBM-specific capacity is being added at its Taichung, Taiwan facility. These investments are backstopped by CHIPS Act grants and tax credits.

5. Financials / Funding

• Listed: NASDAQ: MU; market cap ~$140-170B (2026, fluctuates with memory cycle)
• FY2024 (Aug 2024) revenue: $25.1B (+61% YoY from FY2023 trough of $15.5B), reflecting the memory upcycle driven by AI HBM demand
• FY2024 net income: $2.8B (recovered from FY2023 loss of -$5.8B)
• FY2025 estimated revenue: ~$35-40B, driven by HBM3E volume ramp and DDR5 pricing recovery
• HBM revenue: ~$5-8B estimated for FY2025 (growing rapidly but still a minority of total revenue)
• Gross margin: ~36% (FY2024); recovering toward normalized 40-50% range
• Capex: ~$8-12B annually (2024-2026), heavily weighted toward HBM-capable DRAM capacity
• CHIPS Act grants: $6.1B in direct funding + $7.5B in loan guarantees announced December 2024 for US fab expansion (Boise, ID + Clay, NY)
• R&D spending: ~$4B annually

Aschenbrenner / Situational Awareness LP — Q1 2026 13F Position: PUT leg:

• Security: Micron Technology (MU) — Put Options
• Value: $584,000,000 (4.27% of 13F book)
• Shares: 1,728,000 shares
• Type: PUT
• Voting Authority: Sole
• Status: NEW position (not held in Q4 2025) CALL leg:
• Security: Micron Technology (MU) — Call Options
• Value: $422,000,000 (3.09% of 13F book)
• Shares: 1,246,000 shares (estimated from CALL notional)
• Type: CALL
• Voting Authority: Sole
• Status: NEW position (not held in Q4 2025)
• Notional implied price: ~$338/share
• Interpretation: Volatility straddle — bet on large MU movement regardless of direction. Slight downside tilt (PUT larger than CALL). Only position in the 13F with significant bidirectional options exposure.

6. People & Relationships

• CEO: Sanjay Mehrotra (since 2017) — co-founder of SanDisk; led MU through HBM pivot
• Chairman: Lynn Dugle
• History: Founded 1978 in Boise, Idaho; acquired Elpida Memory (2013) and Inotera Memories (2016) to gain DRAM scale
• Key partnerships:
  • nvidia — HBM3E qualified; HBM4 qualification in process
  • tsmc — base die partner for HBM4 (same TSMC 12nm approach as SK Hynix)
  • broadcom — HBM supply for custom AI ASIC programs (Google TPU, etc.)
  • AMD — HBM3E supply for MI300X/MI325X
• Government: CHIPS Act beneficiary; strategic US memory supplier
• Competitors: sk-hynix (dominant HBM supplier), samsung-hbm (scaled competitor), conventional memory manufacturers for non-AI business
• Equipment partners: be-semiconductor (TC bonders for HBM stacking), asmt (TCB for HBM assembly)