The memory industry isn't cyclical anymore. AI killed that narrative.
At FMS 2026 in Santa Clara this week, Deutsche Bank sat down with Micron's top brass — including Jeremy Werner, SVP of the Core Data Center Business Unit, and Satya Kumar, VP of Investor Relations. What came out of that room wasn't just another supply-demand forecast. It was a declaration: storage has moved from being a commodity bolt-on to the central nervous system of AI infrastructure. Memory now accounts for nearly 50% of total system value, up from roughly 10% thirty years ago. If you're still thinking of DRAM and NAND as interchangeable parts, you're reading the wrong playbook.
Here's what's happening and why it matters.
Let's start with the financial reality. Micron just posted $41.46 billion in Q3 FY2026 revenue — up 347% year-over-year from $9.30 billion. Their adjusted operating margin hit 81.2%. For Q4, they guided to $50 billion. These aren't startup growth numbers on a tiny base; this is a mega-cap semiconductor company in a supposed "cyclical" industry printing margins that make software companies blush.
Deutsche Bank analyst Melissa Weathers called it a "luxury advantage" — Micron's ability to grow without sacrificing profitability. The bank maintains its Buy rating, and when you look at the supply dynamics, it's easy to see why.
Both DRAM and NAND are in structural shortage relative to demand. And unlike previous cycles, this isn't about a temporary spike in PC or smartphone sales. This is about AI fundamentally rewiring how data centers consume memory.
Here's the stat that should change how you think about AI hardware: memory's share of total system value has climbed from roughly 10% three decades ago to nearly 50% today. AI is accelerating a multi-decade revaluation that most investors and even many technologists have overlooked.
Think about what that means. When someone builds a $100 million AI cluster, roughly $50 million of that is going to memory and storage. The GPU gets the headlines, but the memory gets the money.
This shift isn't an accident. It's being driven by a fundamental technical challenge that every AI infrastructure team is grappling with right now: the KV cache problem.
When a large language model generates responses, it builds something called a key-value (KV) cache — essentially, the attention memory it needs to avoid recomputing previous tokens. A 70-billion-parameter model with a one-million-token context window generates approximately 320 GB of KV cache for a single user. That's four times the entire HBM capacity of an NVIDIA H100 GPU.
If the KV cache isn't managed properly, the GPU ends up recalculating results it already computed rather than doing new work. Your utilization metrics look great. Your actual output doesn't.
Micron's solution isn't a single product — it's a tiered memory architecture:
| Tier | Technology | Role |
|---|---|---|
| Hot Path | HBM4 (36GB 12-Hi stacks, 2.8+ TB/s bandwidth) | GPU-attached, highest bandwidth, smallest capacity |
| Warm Overflow | SOCAMM2 (256GB modules, 32Gb LPDDR5X dies) | CPU-adjacent, KV cache spillover, 1/3 the power and footprint of RDIMMs |
| Long Tail | DDR5 RDIMMs (up to 9,200 MT/s) | Decentralized resource pools for cold and less-frequent accesses |
| Persistent Storage | PCIe Gen 6 NVMe SSDs (Micron 9650, 64 GT/s per lane) | Model weights, vector databases, checkpoints |
The orchestration layer — software that manages what data lives where and when to evict or prefetch — is arguably as important as the hardware tiers themselves. Micron's FAMFS (fabric-attached memory file system) is one piece of this puzzle, connecting CXL-attached DRAM to NVIDIA's Dynamo inference stack with preliminary results showing a 5x to 10x speedup over traditional storage-backed KV cache implementations.
Here's the uncomfortable truth: HBM demand exceeds supply through at least calendar 2028. Micron has now pushed its HBM total addressable market forecast to cross $100 billion by calendar 2027 — a full year earlier than previous estimates. Customer requests for HBM3E, HBM4, and beyond exceed what Micron can manufacture.
This is why Strategic Customer Agreements (SCAs) matter. Micron now has SCAs covering roughly 40% of sales volume, with multi-year commitments, upfront cash deposits totaling more than $22 billion (including nearly $18 billion in actual cash), and pricing structured around quarterly resets within ceiling-and-floor bands.
Translation: the hyperscalers are so desperate for guaranteed memory supply that they're handing over billions in cash deposits for the privilege of being in line. This isn't how commodity markets work. This is how strategic infrastructure markets work.
One of the most overlooked signals from the Deutsche Bank meeting was Micron's commentary on CPU-side demand. Hyperscaler capex has been overwhelmingly GPU-focused, but as workloads shift from human-driven prompts to agent-driven computation (where one user request spawns dozens or hundreds of inference calls), the CPU side of the equation is dramatically under-provisioned.
Micron described this as "pre-season warmup" — agent adoption is currently limited to the most technically sophisticated enterprises, but the trajectory is clear. When agentic AI goes mainstream, CPU-side memory demand will explode. This isn't replacing GPU memory demand; it's additive.
Deutsche Bank calls this an "incremental pillar" of AI-driven memory growth that opens new addressable markets beyond the GPU ecosystem.
At FMS 2026, investors were buzzing about disruptive memory technologies that could challenge the HBM/DRAM/NAND hierarchy. Let's go through them:
SRAM: Currently handles roughly 5% of system workloads, mostly inference tasks with bandwidth needs but smaller memory footprints. SRAM doesn't scale the way DRAM does — it's inherently an on-chip solution. It's a new tier in the memory hierarchy, not an HBM replacement.
CXL (Compute Express Link): Heavily discussed at FMS, with Micron presenting the FAMFS + JBOM (Just a Bunch of Memory) architecture. Samsung's XGBoost-based hot-page placement framework and Meta's production hyperscale experiences with CXL suggest this is a real technology. But it's complementary to flash, not a replacement. CXL-attached DRAM and NVMe SSDs address different points on the latency-capacity curve.
Liquid-Cooled SSDs: Kioxia's CM10 — built on 332-layer BiCS10 NAND with direct cold-plate liquid cooling — made a splash at FMS. As flash storage gets co-located with accelerator racks running hundreds of watts per chip, thermal management becomes as important as bandwidth. But this is evolution, not disruption.
Samsung's zHBM: Samsung proposed stacking memory directly on the GPU die. Interesting concept, but it doesn't change the fundamental math: there's only so much silicon area on an accelerator package.
The bottom line: none of these technologies threaten Micron's core memory franchise in any meaningful timeframe. The moat is deep.
The storage industry is being permanently restructured around three realities:
1. Memory is no longer fungible. When hyperscalers sign five-year take-or-pay agreements with billions in deposits, they're not treating memory as a commodity they can source from anyone. They're treating it as strategic infrastructure.
2. The cycle is broken — in a good way. SCAs with price bands, persistent supply shortages, and customer cash deposits don't eliminate cyclicality entirely, but they fundamentally dampen it. The boom-bust memory cycles that defined the industry for decades are being replaced by something more predictable.
3. The value shift is permanent. Memory at 50% of system value isn't a temporary AI training spike. Inference — which will dominate AI compute over time — is even more memory-intensive than training. As context windows expand and agentic workloads multiply, memory intensity per GPU will continue rising, not falling.
Deutsche Bank's framing is worth repeating: Micron has positioned itself to grow without sacrificing profitability. With a full-stack portfolio spanning HBM4, SOCAMM2, DDR5, PCIe Gen 6 SSDs, LPDDR5X, GDDR7, and UFS 4.1, they cover every tier of AI infrastructure from the data center to the edge.
Their Q3 FY2026 numbers tell the story: $41.46 billion revenue, 81.2% operating margin, and Q4 guidance of $50 billion. They're investing $27 billion in capex this fiscal year, with greenfield capacity coming online in calendar 2028. Even with that additional supply, management doesn't expect it to meet demand.
The company that was once viewed as the "third player" in memory is now arguably the best-positioned semiconductor company for the AI era — precisely because they bet the farm on memory when everyone else was fixated on compute.
No analysis is complete without acknowledging what could go wrong:
The storage industry didn't just have a good quarter. It had its Copernican moment. The GPU used to be the sun that everything orbited. Now memory and storage are proving they're a star of equal magnitude — and the math says they're not done rising.
华尔街见闻 / Futunn: Micron Management Sends Strong Signal at FMS 2026 — Memory System Value Exceeds 50%, CPU-side AI Agent Demand in "Pre-Season Warmup" — Original Deutsche Bank research summary, August 7, 2026
TechTimes: FMS 2026 Opens Tuesday With Liquid-Cooled PCIe 6.0 SSDs and a Debate on AI Memory Tiers — FMS 2026 conference overview, August 2, 2026
StorageNewsletter: Microchip & Micron Demonstrate PCIe Gen 6 Storage Architecture for AI and Data Center — FMS 2026 Gen 6 demo details, August 5, 2026
Futurum Group: Micron Q3 FY 2026 — HBM and LPDRAM Drive the Next Phase of AI Memory Growth — Q3 FY2026 earnings analysis, June 25, 2026
Micron Investor Relations: Q3 FY2026 Earnings Press Release — Record $41.46B revenue, June 24, 2026
Igor's Lab: Micron at Computex 2026 — AI Needs Not Only Compute Power, But Memory with Throughput — HBM4, SOCAMM2, and PCIe Gen 6 portfolio overview, June 2026
NVIDIA NVMe Technology at FMS 2026 — NVMe technology in AI storage architecture
FMS 2026 Official Conference Page — Conference agenda and session details
Micron Data Center Memory Products — Official product specifications
Investing.com: Micron Q3 FY2026 Slides — Record $41.5B Revenue, 85% Margins — Financial data summary
