Micron HBM3e and HBM4

Learning Objectives

• Understand why high-bandwidth memory (HBM) is the dominant constraint on AI accelerator performance
• Identify Micron's HBM3e and HBM4 generation specs and their roles in NVIDIA accelerators
• Evaluate what the completed HBM4 ramp means for AI compute capacity

What Is Micron HBM?

High Bandwidth Memory (HBM) is a stacked-DRAM technology designed for the extreme memory bandwidth that AI accelerators require. Micron is one of three companies in the world (alongside SK hynix and Samsung) that manufactures HBM at scale, and it now ships two generations in parallel.

For AI specifically, HBM is the gating constraint on accelerator performance more often than raw compute. NVIDIA H200, B200, AMD MI300X, and most other modern AI accelerators are bounded by how fast they can move weights and activations between HBM and the compute cores. Micron's HBM3e ships in NVIDIA H200 GPUs (introduced 2024) and is also qualified for NVIDIA B100/B200. Its successor, HBM4, entered high-volume production in March 2026 and is designed for NVIDIA's Vera Rubin platform.

Pricing & Access

HBM is sold to a small number of large customers (chip designers and system integrators) under long-term supply agreements. Pricing is not publicly disclosed.

HBM supply has been the rate-limiter on AI accelerator output for the past two years. Customer queues and allocations matter as much as nominal pricing.

Core Capabilities

HBM3e 8-Hi 24 GB Cubes

The mainstream production tier as of 2025-2026. Stacks 8 DRAM dies to deliver 24 GB per cube at over 1.2 TB/s of bandwidth per cube. Used in NVIDIA H200 GPUs (which contain 6 HBM3e cubes for 141 GB total at 4.8 TB/s).

HBM3e 12-Hi 36 GB Cubes

Higher-density variant stacking 12 dies for 36 GB per cube. Targets the highest-tier AI accelerators including NVIDIA B100/B200 (Blackwell). NVIDIA B200 specs 192 GB of HBM3e at 8 TB/s aggregate bandwidth — using these higher-density cubes.

Power Efficiency Lead

Micron has differentiated on power efficiency — early HBM3e benchmarks showed Micron's products consuming meaningfully less power per GB-throughput than competitors, which matters at AI-data-center scale where HBM contributes substantially to total accelerator power.

Market Share Trajectory

Micron entered HBM late — at approximately 5% market share in 2023, with SK hynix and Samsung dominating. By end of 2025 Micron targeted 20-25% market share through aggressive capacity expansion. For NVIDIA's H200 and B100/B200, Micron has secured material design wins, breaking the SK hynix near-monopoly that prevailed earlier.

Through mid-2026 the bottleneck has only tightened: Micron has reported its HBM output effectively sold out for the year, with memory emerging as the single scarcest component in an AI server and HBM now its fastest-growing product line. That demand has made memory makers — not just GPU designers — central to how fast the industry can add AI capacity.

HBM4 — In Production Since March 2026

HBM4 is the largest architectural change since HBM's introduction, and Micron completed the transition rather than merely announcing it. On March 16, 2026 the company said its HBM4 36 GB 12-high stacks had entered high-volume production, designed for NVIDIA's Vera Rubin platform:

• Interface width doubled from 1,024 to 2,048 bits per stack
• Pin speeds above 11 Gb/s, well beyond the 8 Gb/s the specification originally targeted
• Over 2.8 TB/s per stack — roughly 2.3 times HBM3e's bandwidth
• More than 20 percent better power efficiency than HBM3e
• New packaging required, meaning real re-tooling for both DRAM makers and accelerator vendors

Micron has also shipped HBM4 48 GB 16-high samples to customers, stacking 16 dies for a 33 percent capacity increase per HBM placement. The next step, HBM4E, moves base-die manufacturing to foundry partner TSMC rather than staying in-house.

Stable Supply Agreements

Micron's HBM business runs on multi-quarter (sometimes multi-year) supply agreements with NVIDIA and other accelerator vendors. Customers lock in capacity early; spot purchases are rare.

Long-Horizon Research: Micron Research Labs

On August 20, 2026 Micron committed $10 billion over the next decade to Micron Research Labs, anchored on a flagship campus in Boise, Idaho, with groundbreaking expected in calendar 2027. The company describes it as the first dedicated memory research hub of its kind in the United States.

What makes it unusual is the time horizon. The lab is explicitly aimed at memory and compute architectures, packaging and semiconductor manufacturing more than ten years out rather than the next product cycle — a long bet in an industry that plans in generational ramps like the HBM3e-to-HBM4 transition above. Named partners include Stanford University, the University of Texas at Austin, Applied Materials, Lam Research and imec. Chief executive Sanjay Mehrotra framed the reasoning as a matter of scale: the future of memory and compute is too complex to invent alone.

The commitment sits separately from the more than $250 billion Micron has already pledged to US manufacturing and research.

Strengths

• Third major HBM supplier: One of only three companies globally producing HBM at scale — NVIDIA, AMD, and other accelerator makers benefit from supply diversity
• NVIDIA H200 + B200 design wins: Material share of NVIDIA's flagship accelerator memory now sources from Micron, breaking SK hynix near-monopoly
• Power efficiency: Micron HBM3e benchmarks favorably on power-per-GB-throughput, meaningful at AI-data-center scale
• Capacity expansion: Aggressive investment in HBM production capacity targeting 20-25% market share
• HBM4 shipping in volume: Reached high-volume HBM4 production in March 2026 for NVIDIA Vera Rubin, with 16-high 48 GB stacks already sampling
• Stable customer base: Multi-quarter supply agreements with major AI accelerator vendors

Limitations & Considerations

• Three-supplier oligopoly: Micron, SK hynix, Samsung control all HBM supply — pricing and allocation dynamics are concentrated and opaque
• Long lead times: HBM capacity additions take years; customers cannot hedge near-term shortages by buying spot capacity
• Geopolitical exposure: Most HBM is produced in South Korea (SK hynix, Samsung) and the US (Micron); export-control regimes affect cross-border supply
• Dual-generation complexity: With HBM3e and HBM4 both in production, customers running mixed accelerator fleets carry two memory architectures and two supply queues at once
• Customer concentration: NVIDIA is the dominant HBM3e buyer; demand cyclicality at NVIDIA flows directly to Micron HBM revenue
• Not a finished product: HBM is sold as memory cubes to chip designers — end customers buy GPUs, not HBM directly

Best Use Cases

When to choose alternatives:

• HBM is the dominant memory for AI accelerators — there are no direct alternatives at the same bandwidth tier
• For lower-cost inference at smaller scale, GDDR6X (consumer GPUs) is meaningfully cheaper but bandwidth-limited
• For specialty AI ASICs (Cerebras, Groq), on-chip SRAM partially replaces HBM for very specific workloads

Key Takeaways

• Micron HBM3e is the high-bandwidth memory powering NVIDIA H200 (141GB at 4.8 TB/s) and NVIDIA B100/B200 (192GB at 8 TB/s aggregate)
• HBM is the dominant constraint on AI accelerator real-world performance — bigger and faster HBM materially outperforms equivalent compute with smaller memory
• Micron is one of three HBM suppliers globally (alongside SK hynix and Samsung), targeting 20-25 percent market share by end of 2025 vs ~5 percent in 2023
• HBM4 reached high-volume production in March 2026 for NVIDIA Vera Rubin — doubled interface width (2,048 bits), over 2.8 TB/s per stack, and more than 20 percent better power efficiency than HBM3e
• Micron committed $10 billion over a decade to Micron Research Labs in Boise, a memory research hub aimed more than ten years ahead rather than at the next product cycle
• HBM supply is the rate-limiter on global AI accelerator output; customer allocations matter as much as nominal pricing