Micron at $940: The AI Memory Trade Stopped Being a Thesis and Became a P&L

The three memory markets AI is building simultaneously

1. HBM for GPU compute. Every H100 has 80GB of HBM2e. Every H200 has 141GB of HBM3. The Blackwell successor will carry more HBM4 than any prior generation. You cannot run a frontier AI cluster without HBM — there is no substitution. Micron is one of three companies on Earth that makes it, and demand exceeds the combined supply of all three. The HBM market was roughly $4B in 2023. It crossed $15B in 2025. Industry estimates put it at $30-40B by 2027.

2. DDR5 for AI server CPUs. Every GPU needs a host CPU to coordinate the workload. Agentic AI — where models run persistent, multi-step reasoning chains — requires vast DDR5 memory pools attached to the CPU to hold context, manage state, and orchestrate tool calls. The shift from GPU-heavy training to inference-heavy deployment dramatically expands the DDR5 TAM because inference clusters have 4-8x more CPU memory per GPU node than training clusters.

3. LPDDR5X for edge AI. Apple Intelligence. Samsung Galaxy AI. The entire on-device AI category requires the next generation of low-power DRAM. Micron's LPDDR5X is the leading node for this market. Every AI-capable smartphone shipped in 2026 and beyond needs it. Micron is designed in at Apple, Samsung, and virtually every major handset OEM.

The memory bandwidth problem is getting worse, not better

As AI models grow in capability, the memory bandwidth requirements grow faster than compute requirements. Larger context windows mean more data moving between memory and compute on every token generated. More parameters means larger model weights to load and unload. More concurrent inference sessions means more parallel memory reads. The trend that drove Micron's Q2 print isn't reverting — it's accelerating.

The inference shift matters most here. Training is capital-intensive but runs for months in batch jobs. Inference runs continuously, at scale, in real time, for billions of end users. Inference per user query requires 10-50x the memory bandwidth of training per parameter. As AI moves from "research clusters" to "always-on consumer infrastructure," the memory intensity per dollar of compute revenue goes up, not down.

Where each player stands

SK Hynix is the clear HBM market leader. First to yield HBM3E at volume. Sole supplier of HBM for Nvidia's H100 and primary supplier for H200. Estimated ~50% HBM market share in 2025. Their lead is real and their relationships are locked in for current generation products.

Samsung has struggled with HBM3E yield rates. Reports of yield issues delayed their entry into the Nvidia supply chain. They remain the world's largest memory company by capacity but have lost significant share in the highest-value AI memory segment. Their HBM4 ramp is the critical risk to watch — if Samsung fixes yield in H2 2026, it floods the HBM market and compresses pricing.

Micron is the comeback story. Shipping HBM3E to hyperscaler customers. CEO Sanjay Mehrotra confirmed HBM4 samples are shipping to lead customers. Estimated ~15-20% HBM market share in 2025, targeting meaningful expansion in 2026. Micron's U.S. manufacturing footprint gives it a geopolitical advantage that neither competitor possesses.

Why HBM4 is the swing factor

HBM4 offers roughly 2.5x the bandwidth of HBM3E. For AI training and dense inference workloads, that bandwidth delta is not incremental — it's the difference between being able to run the next generation of frontier models in cluster and not. Every major AI compute platform that ships in 2027 will be built around HBM4. The supply agreements being written right now will define who captures the margin upside of that cycle.

Micron's HBM4 competitive position comes down to one question: can they ramp yield faster than Samsung, and can they protect against SK Hynix locking up all of Nvidia's allocation? The CHIPS Act fab investments are relevant here — a domestic HBM4 supply chain has strategic value to Nvidia, hyperscalers, and the U.S. government that goes beyond economics. That's not a guarantee. But it's a real factor in the supply agreement calculus.

Micron's manufacturing footprint

Micron operates major DRAM and NAND manufacturing facilities across five geographies:

• Boise, Idaho (USA) — primary DRAM R&D and advanced production; CHIPS Act expansion underway
• Clay, New York (USA) — new 1-million sq. ft. greenfield fab under construction; targeting leading-edge DRAM by 2026-2027
• Manassas, Virginia (USA) — 3D NAND production
• Hiroshima, Japan — advanced DRAM (acquired from Intel's NAND business); key HBM production node
• Singapore — NAND Flash and assembly/test operations
• Taiwan — DRAM production

The CHIPS Act advantage

In July 2023, the U.S. Department of Commerce announced approximately $6.1 billion in CHIPS Act grants for Micron — the largest semiconductor investment commitment in U.S. history at the time. This covers the Idaho expansion and the New York greenfield fab. The New York fab alone represents a $100B+ long-term commitment and is intended to bring advanced DRAM production — including HBM — back to U.S. soil.

Why does this matter for the investment thesis? Three reasons. First, the subsidy represents roughly one full quarter of Micron's CapEx reimbursed by the government. That reduces the capital intensity of the expansion and improves the long-term return on investment. Second, it creates a domestic supply chain preference from U.S. hyperscalers and government customers who increasingly need to demonstrate "Made in America" AI infrastructure — and who are willing to pay a premium for it. Third, it generates goodwill with the federal government that translates into regulatory protection and policy alignment that no Korean competitor can replicate.

Fab ownership in a supply-constrained world

In 2020-2022, being fabless was the fashionable business model. Asset-light, high-multiple, capital-efficient. Then COVID hit supply chains, then AI hit demand, and suddenly every hyperscaler and government in the world was writing blank checks to anyone who could reliably manufacture leading-edge chips. Micron didn't become a great business by going fabless — it's a great business because it never did. The vertical integration from wafer to finished memory product means Micron controls yield improvement, capacity allocation between commodity and premium products (HBM vs. standard DRAM), and technology roadmap timing. In a market where HBM4 allocation will be negotiated by CEOs at the board level, owning the factory is the moat.

NAND Flash: the recovered patient

NAND was the worst segment in the FY2023 bust — gross margins went negative as oversupply crushed prices. The recovery since then has been real: enterprise SSD demand (for AI model storage, checkpoint I/O, and hyperscaler cold storage) has driven pricing higher, and Micron's 9400 Pro enterprise NVMe SSD has been well-received. But NAND is inherently more commoditized than DRAM, more Chinese-competitor-exposed (YMTC), and more vulnerable to digestion periods when enterprise storage buyers catch up to their needs.

The key NAND question for 2026-2027: does AI storage demand (massive model weights, training checkpoints, inference logs) continue to absorb the NAND supply Samsung and SK Hynix are building? If yes, pricing holds. If not, NAND becomes a drag that partially offsets DRAM's upside. Right now, NAND is recovering nicely. It's not the risk it was in FY2023. But it's the segment most likely to show weakness first if the cycle turns.

Mobile DRAM: the silent beneficiary

On-device AI is real and it requires memory. Apple Intelligence on iPhone 16 and beyond requires LPDDR5X memory that runs at speeds commodity LPDDR5 can't match. Samsung's Galaxy AI series requires the same. Every smartphone OEM competing on AI features is racing to spec more and faster memory. Micron is designed in as a primary LPDDR5X supplier at multiple top-5 OEMs. The AI phone upgrade supercycle is early innings — the market replacement cycle for AI-capable handsets hasn't fully kicked in — and it drives meaningful DRAM ASP improvement on every unit shipped.

Embedded & auto: steady, not spectacular

Micron's embedded segment (automotive, industrial, IoT) generates roughly $1-1.5B per quarter and grows at mid-single-digit rates. Automotive is the interesting long-term story — autonomous vehicles and ADAS systems require more memory per vehicle than any prior automotive electronics generation. But this is a FY2028-2030 revenue inflection story, not a FY2026 driver. It's ballast, not catalyst.

The valuation paradox at 16x forward earnings

Here's the thing that makes Micron genuinely hard to analyze: on a strict forward P/E basis, the stock looks cheap. FY2026 EPS is tracking toward ~$58 (Q1 $4.60 + Q2 $12.08 + Q3 estimated ~$19 + Q4 estimated ~$22). At $940 per share, that's roughly 16x FY2026 earnings. A $1T company growing revenue 197% year-over-year at 75%+ gross margins trading at 16x forward earnings is not an obviously expensive stock.

The bear case challenge to that framing: those are peak-cycle earnings. FY2023 EPS was -$5.34. If the cycle turns in FY2027 and earnings revert to $20-25 — which is not a catastrophic scenario, just a normalization — then 16x current earnings becomes 40-50x normalized earnings. That's not cheap. That's the multiple compression risk embedded in buying a cyclical at the peak.

The bull case challenge to the bear framing: AI may have permanently broken the memory cycle. HBM is not commodity DRAM. If inference demand keeps growing and Micron captures HBM4 supply agreements, the "trough" EPS may never go below $40-50. Under that scenario, 16x forward earnings is genuinely inexpensive for a $1T AI infrastructure company.

Both framings are internally coherent. The scenario you believe determines the trade.

How the math shifts at different entry prices

The same three scenarios produce dramatically different risk/reward ratios depending on where you enter. A hypothetical investor with no current position would find the math at $700-800 substantially more attractive than at $940. That's not a forecast that the stock reaches $700-800 — it may not. The June 24 print could gap it to $1,100. But the asymmetry math is the math, and it's worth knowing.

The cyclical discount problem

Memory stocks have always traded at discounted multiples relative to growth because the market embeds a cyclical discount — it knows the trough will come and it refuses to pay fully for peak earnings. The question of whether AI has permanently broken that discount is the most important analytical question in the MU thesis. If AI memory is structurally different — if HBM is more like Nvidia's datacenter GPU than like commodity DRAM — then the cyclical discount should shrink and the multiple should expand. If it's not, the discount will reassert itself with brutal efficiency the moment any hyperscaler signals demand softness.

The honest answer is that we don't know yet. FY2023 was three years ago. The structural AI demand thesis has been validated by actual revenue — not just projections. But the memory cycle is typically 4-5 years peak to trough, and the current cycle started in earnest around FY2025. There is no precedent for AI-driven HBM demand at this scale, which means both the bulls and the bears are extrapolating from incomplete data.

Position framing for existing holders

For a hypothetical investor who held MU through the year — through the April 2025 tariff trough at $65, through the Q2 FY2026 sell-the-news pullback to $338, to today at $940 — the math behind trimming into strength has two pieces. First, the position has likely grown to 3-5x its original portfolio weight simply from appreciation. Concentration rebalancing is rational regardless of fundamental view. Second, with June 24 earnings 27 days away and a $33.5B guide in the price, the near-term binary risk is elevated. Trimming 20-30% of an oversized position ahead of a known catalyst is not bearish — it's portfolio management.

For a hypothetical investor with no current position, the math is less compelling at $940 than it would be at $750-800. Patience costs nothing if the thesis is right. If Q3 beats and Q4 guides $42B+, you may pay $1,100 to get in. That's a real cost. But you'll have much better information than you have today about whether the cycle is extending or peaking.