Micron - The Memory Bottleneck Behind the AI Supercycle

Rating: Buy / cyclical AI infrastructure leader

Style: Semiconductor memory supercycle, HBM, data-center infrastructure

Core debate: Has AI structurally changed the memory cycle, or is Micron showing peak-cycle earnings that the market should not capitalize too aggressively?

Executive View

Micron Technology has become one of the most important “picks and shovels” companies in the AI buildout. It does not make GPUs. It makes the memory and storage that keep those GPUs useful. Every AI accelerator needs high-bandwidth memory. Every inference system needs DRAM capacity. Every data center needs SSDs. And right now, memory is no longer a commodity afterthought. It is one of the biggest bottlenecks in AI infrastructure.

The latest quarter was extraordinary. Micron reported fiscal Q3 2026 revenue of $41.46B, up 346% year over year and 74% sequentially, with non-GAAP gross margin of 84.9%, non-GAAP EPS of $25.11, and $18.3B of adjusted free cash flow. Management then guided fiscal Q4 to $50B of revenue, approximately 86% gross margin, and $31 of EPS.

The bull case is simple: AI has pulled memory from the commodity bucket into the strategic infrastructure bucket. Micron has now signed 16 strategic customer agreements, covering roughly 20% of DRAM volume and around one-third of NAND volume over the contract period, with 14 of those agreements representing approximately $100B of minimum contracted revenue.

The bear case is also real. Memory has always looked best near the top of the cycle. Margins this high attract supply. Capex is rising. China is becoming more serious in memory. And the stock has already had a massive run. At around $920.90, Micron’s market cap is now approximately $1.05T.

So this is not a clean “cheap stock” thesis anymore. It is a question of whether the market is still underestimating how much AI has changed memory economics.

Why Now — HBM, AI Memory, and Strategic Customer Agreements

The reason Micron matters now is that AI demand has changed the role of memory.

Historically, memory was cyclical and mostly interchangeable. Customers bought quarter to quarter. Supply eventually caught up. Prices collapsed. Margins disappeared. Then the cycle started again.

That was the old Micron.

The new Micron is being pulled by HBM, DRAM, and data-center SSD demand at the same time. Management said data-center revenue exceeded $25B in fiscal Q3, equivalent to an annualized run rate above $100B. Data-center SSD revenue alone exceeded $5B and more than doubled sequentially.

The central change is HBM. High bandwidth memory is stacked DRAM designed to feed AI accelerators with enough bandwidth to keep them busy. It is more complex, more valuable, and more capacity-hungry than conventional DRAM. The baseline thesis framed this well: every HBM wafer allocated to AI pulls standard memory capacity away from the broader market, tightening the entire memory supply picture.

The second “why now” factor is the strategic customer agreements. These are not ordinary purchase orders. They include multi-year volume commitments, price floors, price ceilings, and customer deposits. That matters because Micron is trying to do something memory companies have rarely achieved: create forward visibility in an industry built on volatility.

The third factor is China. CXMT’s Shanghai debut reminded investors that memory is now a geopolitical industry. Reuters reported that CXMT shares surged 466% on their first trading day after raising $8.6B, briefly valuing the company at roughly $488B.

That does not break the Micron thesis today. CXMT is much more relevant in commodity DRAM than in leading-edge HBM. But it does remind us that the memory cycle is not dead. It has just changed shape.

What Micron Does

Micron makes memory and storage products. The two core technologies are DRAM and NAND.

DRAM is fast working memory. It is used in AI servers, PCs, smartphones, networking equipment, cars, industrial systems, and embedded devices.

NAND is flash storage. It is used in SSDs, smartphones, data centers, consumer electronics, and embedded applications.

In fiscal Q3 2026, DRAM represented 76% of Micron’s revenue, while NAND represented 24%. DRAM revenue was $31.3B, up 67% sequentially, while NAND revenue was $9.9B, up 99% sequentially.

The simplest way to understand Micron is this: Micron sells the memory bandwidth and storage capacity that AI systems cannot function without.

TAM - How Big Is The Opportunity?

The Micron TAM is no longer just “PC and smartphone memory.” That is the key shift.

The opportunity now has four layers.

• First, HBM for AI accelerators. This is the highest-value part of the market. Micron previously forecast the HBM market growing from roughly $35B in 2025 to around $100B by 2028, implying approximately 40% annual growth.
• Second, data-center DRAM. AI infrastructure needs far more memory capacity than traditional servers. Micron said industry data-center DRAM and NAND bit shipments in calendar 2026 are expected to more than double from two years earlier.
• Third, data-center SSDs. As AI models, agents, and inference workloads scale, context storage becomes a larger bottleneck. Micron specifically highlighted AI context memory storage and HDD displacement as drivers of the SSD addressable market.
• Fourth, edge, automotive, robotics, and physical AI. Micron said L2+ and above vehicles carry more than five times the memory and storage content of an average vehicle, and that this mix is expected to exceed 40% of vehicles by 2030. It also highlighted humanoid robots as a potential multi-decade memory demand cycle.

So the TAM story is not only “AI servers need HBM.” The broader point is that memory content per device is rising across data centers, cars, robots, smartphones, PCs, and industrial systems.

That is why Micron’s market opportunity looks structurally larger than in prior cycles.

How Micron Wins

Micron wins because memory is a concentrated, capital-intensive, technically difficult industry.

At the highest end, only a few companies matter: Samsung, SK Hynix, and Micron. Building leading-edge memory capacity requires huge capital investment, process expertise, advanced equipment, customer qualification, and packaging capability. It is not a market where new entrants can quickly appear and compete at the high end.

Micron’s specific edge today is execution.

The company says its 1-gamma DRAM and G9 NAND nodes are ramping well and are on track to become the highest-volume nodes in company history. It has already shipped more than $1B of HBM4 revenue, and management says the HBM4 12-high volume ramp is tracking twice as fast as HBM3E 12-high.

Micron also benefits from geography. It is the only U.S.-based DRAM maker at scale, and that matters more in a world where AI infrastructure has become strategically important. The company is investing heavily in U.S. manufacturing, with Idaho and New York projects underway. ID1 is expected to produce first wafer output in mid-calendar 2027, while ID2 is expected in late calendar 2028.

In other words, Micron is not just selling memory. It is selling scarce, qualified, strategically important memory supply.

Business Model

Micron makes money by selling bits of memory and storage.

Revenue is driven by five variables:

• bit shipments
• average selling prices
• product mix
• customer agreements
• capacity allocation

The operating leverage is extreme. Fabs are expensive to build and run, but once fixed costs are covered, higher prices can flow very quickly into profit. That is why Micron’s income statement can move so violently in both directions.

In this cycle, pricing has done most of the work. In fiscal Q3, DRAM bit shipments rose only in the low-single-digit range sequentially, but DRAM ASPs rose in the low-60% range. NAND bit shipments rose in the mid-single-digit range, while NAND ASPs rose in the mid-80% range.

The new part of the model is the SCA layer.

Micron’s strategic customer agreements are structured as take-or-pay agreements with binding purchase commitments over multi-year terms. The largest agreements generally have price ceilings around current calendar Q2 market prices, but also include floor prices through the contract term. For the agreements with price bands, management says the floor price still enables gross margins well above prior-cycle peak margins.

This is the most important business-model change at Micron. The company is trying to turn part of memory from a pure spot-price business into a contract-backed strategic supply business.

That does not eliminate cyclicality. It reduces it.

Business Units / Segments

Micron reports four main business units.

• Cloud Memory Business Unit (CMBU) is the hyperscaler and HBM-heavy business. It reported $13.8B of revenue in fiscal Q3, representing 33% of company revenue, with 83% gross margin and 78% operating margin. This is the crown jewel.
• Core Data Center Business Unit (CDBU) covers enterprise data center demand and data-center SSDs. It reported $11.5B of revenue, up 653% year over year, with 87% gross margin and 83% operating margin.
• Mobile and Client Business Unit (MCBU) serves smartphones and PCs. It reported $11.5B of revenue, with 87% gross margin and 86% operating margin.
• Automotive and Embedded Business Unit (AEBU) reported $4.6B of revenue, with 79% gross margin and 75% operating margin.

The striking detail is that every segment is now producing margins that would have looked impossible in earlier memory cycles.

That is why investors are struggling to value the company. Either Micron has structurally changed, or these are peak-cycle numbers.

By The Numbers

Fiscal Q3 2026 was one of the most extreme quarters I have seen in semiconductors.

For full fiscal 2026, Micron is tracking toward roughly $73 of non-GAAP EPS if it lands near Q4 guidance. That uses Q1 EPS of $4.78, Q2 EPS of $12.20, Q3 EPS of $25.11, and Q4 guidance of $31.

At around $920.90, that implies roughly 12.6x fiscal 2026 EPS power. The reason that multiple looks low is the entire debate: the market does not fully believe these earnings are sustainable.

Key Drivers

• The first driver is HBM demand. AI accelerators need high-bandwidth memory, and HBM is now one of the most valuable products in the semiconductor supply chain.
• The second driver is data-center SSD growth. Micron’s data-center SSD revenue exceeded $5B in fiscal Q3 and more than doubled sequentially. That shows the AI memory story is not only HBM. Storage is also becoming strategic.
• The third driver is strategic customer agreements. If more customers accept multi-year supply contracts with price floors and deposits, Micron’s earnings become more visible and less dependent on spot pricing.
• The fourth driver is supply discipline and capacity timing. Management expects tight DRAM and NAND conditions to persist beyond calendar 2027 because demand is growing quickly and new greenfield fab supply takes years to build.
• The fifth driver is U.S. manufacturing strategy. Micron’s U.S. fab investments give customers and policymakers a domestic strategic supply option in memory.
• The sixth driver is capital returns. Micron said it intends to increase capital returns after December 9, 2026, the second anniversary of its definitive CHIPS agreements, and over time expects to return 100% of excess cash to shareholders.

Risks

• The first risk is peak earnings risk. Gross margins near 85–86% are extraordinary. If pricing rolls over, earnings can fall much faster than revenue.
• The second risk is China. CXMT is becoming more relevant. Its IPO raised $8.6B, and Reuters reported that the stock surged 466% on its debut. Apple is also reportedly testing CXMT memory chips for China-market devices.
• The third risk is capex. Micron expects fiscal 2026 capital spending of approximately $27B, and quarterly capex in fiscal 2027 is expected to be above fiscal Q4 levels. That is rational if demand remains strong, but memory has a long history of overbuilding at the wrong moment.
• The fourth risk is customer concentration. AI memory demand is tied to a small number of hyperscalers, AI labs, accelerator vendors, and cloud customers. If AI capex slows, memory orders can move quickly.
• The fifth risk is uncontracted volume. The SCAs are a real structural improvement, but they do not cover the entire company. A large portion of DRAM and NAND volume remains exposed to market pricing.
• The sixth risk is valuation psychology. The stock may look cheap on near-term EPS, but that is exactly what happens near cyclical peaks. Investors need to decide whether Micron deserves a new multiple because AI has changed the cycle.

Bottom Line

Micron is having the best moment in its history.

The company has gone from cyclical memory supplier to strategic AI infrastructure bottleneck. HBM, data-center SSDs, tight supply, and customer agreements have created a profit profile that looks almost unrecognizable versus prior cycles.

The most important question is not whether the latest quarter was impressive. It clearly was.

The real question is whether the earnings power is durable enough to justify a higher valuation than investors have historically assigned to memory companies.I think the answer is partly yes.

The SCAs, HBM trade ratio, long fab lead times, AI demand, and U.S. manufacturing strategy all support a higher-quality Micron than the market was used to. But I would not pretend the old cycle is dead.

The better framing is this:

Micron is still cyclical, but the AI cycle is larger, longer, and more contract-backed than previous memory cycles.

That makes the business better than old Micron. But it still deserves disciplined position sizing.

Fair Value

My current fair value estimate for Micron is $1,250 per share.

Why $1,250?

If Micron earns roughly $73 of non-GAAP EPS in fiscal 2026, I would not put a high software-like multiple on those earnings because they are supercycle earnings. But I also would not value Micron like the old boom-and-bust memory stock anymore.

A reasonable valuation framework:

At around $920.90, the stock is not obviously cheap after the massive run, but it still offers upside if the market becomes more convinced that HBM demand, strategic customer agreements, and supply constraints create multi-year earnings durability.

My view: Micron is a Buy, but not a blind one. The best version of the thesis is not “memory is no longer cyclical.” The better version is:

AI has made memory strategically scarce, and Micron has finally found a way to lock in part of that scarcity through long-term customer agreements. That is a stronger business than old Micron but still a stock that needs respect for the cycle.