Will AI permanently change memory from a cyclical commodity into a structurally scarce strategic asset?

HBM makes the same wafer base supply fewer bits

HBM brings memory close to the accelerator in stacks that must satisfy demanding bandwidth, power, thermal and reliability requirements. The production chain runs from DRAM wafers through good dies, stacking and testing to customer approval. A wafer entering a fab is several steps away from memory that an accelerator customer can deploy. Micron says successive HBM generations are increasing the manufacturing trade-off against conventional DRAM. [3]

That trade-off gives a relatively small product category an outsized effect on supply. TrendForce estimates that HBM absorbed 18% of leading suppliers’ DRAM wafer input at the end of 2025, rising to 22% at the end of 2026 and 30% at the end of 2027. Its corresponding estimates for HBM’s share of DRAM bits are 8%, 9% and 13%. [1]

The simplest arithmetic is also the most revealing. If HBM’s wafer allocation rises from 22% to 30%, conventional DRAM’s allocation falls from 78% to 70%. Total wafer starts must therefore rise by 78 divided by 70, less one: 11.4%, just to preserve the number of wafers allocated to conventional DRAM. Productivity improvements can add bits beyond that, but the first portion of the expansion is already spoken for. [1]

Demand is pushing from the other direction. TrendForce forecasts nearly 31% growth in AI-server shipments in 2026. NVIDIA’s published GB300 NVL72 configuration contains 20 TB of HBM, compared with 13.5 TB for GB200, a 48% increase calculated from those specifications. More systems and richer configurations both create demand, although the product comparison and industry shipment forecast cover different populations. [4] [5]

The effect extends into the server’s main memory. HBM and RDIMMs together are estimated to account for 51% of DRAM bit supply in 2026, a category that includes conventional servers as well as AI infrastructure. Our view is that AI tightens conventional DRAM through both direct demand and competition for manufacturing resources. [6]

This is the strongest physical reason to expect a prolonged upcycle. Whether it creates durable pricing power depends on how difficult those resources are to replace.

Qualification creates strategic value that must be earned again

The relevant scarce asset is the ability to deliver the right memory on the customer’s schedule. A supplier with available stacks still needs an approved combination of speed, stack height, thermals and reliability. Micron’s HBM4 disclosures distinguish volume shipments for a lead platform from samples sent to additional customers for qualification. That separation is where nominal capacity becomes commercially useful capacity. [3]

The leading suppliers are advancing through that process. SK hynix began mass shipments of HBM4 in the second quarter of 2026 and planned to increase production in the second half. Samsung reported scaling HBM4 sales while shipping HBM4E samples to major customers. Micron had shipped more than $1 billion of cumulative HBM4 revenue by its June results and expected HBM4E volume production in 2027. [7] [8] [3]

These milestones cut both ways. They show how much development and qualification stand between a new design and useful supply. They also show competitors moving through the barriers. Our view is that HBM supports recurring shortages at the technological frontier: each generation can renew differentiation, and each requires suppliers to earn their position again.

Customers are adapting too. In August, TrendForce reported that NVIDIA was evaluating several HBM4 and HBM4E stack configurations for Rubin Ultra beyond its original design. The final specification remained unsettled. That is a concrete response to supply risk: the accelerator’s memory configuration itself becomes a variable. [9]

Even the economic ranking of products can change. TrendForce estimated that revenue per wafer from DDR5 64GB RDIMMs overtook HBM in the first quarter of 2026. Higher conventional-memory prices can alter where suppliers prefer to allocate capacity. We are skeptical of an argument that assumes HBM will always command the best return on a wafer simply because it is the most technically demanding product. [1]

Qualification explains why customers care about particular suppliers. Their willingness to commit money and volumes explains how much of that strategic value suppliers can keep.

Longer contracts are the clearest structural change

Micron’s customer agreements are stronger evidence of a changed business model than another declaration that capacity is sold out. In June, the company described 16 strategic agreements as take-or-pay, generally spanning 2026–2030. Together they covered roughly 20% of expected DRAM volume and one-third of expected NAND volume over that period. Fourteen agreements carried approximately $100 billion of minimum contractual revenue over their remaining terms. [3]

The contractual minimum gives future capacity a firmer revenue foundation. Spread evenly over an assumed four to five remaining years, $100 billion is equivalent to roughly $20–25 billion annually. Actual deliveries will follow the agreements’ schedules. The investment significance is the shift of some demand risk toward customers willing to secure supply years ahead. [3]

Customers are also helping to fund that supply. Micron expected about $18 billion of cash deposits within approximately $22 billion of deposits and related commitments. The refundable deposits provide financing; their value to the producer is the cash available to support expansion under the agreed terms. [3]

The change reaches beyond HBM. Sandisk described eight new business-model customer agreements, covering approximately half of expected bits in FY2027 and two-thirds in FY2028. SK hynix reported long-term agreements with around ten customers. These arrangements differ in their provisions, but collectively they show customers attaching greater value to assured supply. [10] [7]

We give this evidence substantial weight. Longer commitments can improve investment planning, redistribute financing risk and support better returns. They can do so even when physical supply eventually catches demand.

The coverage also sets the limit of the claim. Micron’s cited strategic agreements leave roughly 80% of its DRAM volume and two-thirds of its NAND volume outside those particular commitments. Other contracts may cover some of that production. The cited agreements provide a meaningful foundation, while much of the business remains exposed to future commercial conditions. [3]

For shareholders, the next question is whether better commercial terms translate into cash after the factories have been paid for.

Today’s cash generation is real; its durability is untested

Micron is retaining substantial cash even after investment. In its fiscal third quarter ended May 28, 2026, revenue reached $41.46 billion, operating cash flow was $25.39 billion and net capital expenditure was $7.1 billion. Subtracting the latter from operating cash flow gives $18.29 billion, consistent with reported adjusted free cash flow of $18.3 billion. That is roughly 44% of quarterly revenue. [11]

This is a tangible benefit from current conditions. It also comes with a continuing obligation to invest: Micron guided to approximately $27 billion of capital expenditure for fiscal 2026. Strong present cash flow and substantial future capacity spending can coexist. [3]

SK hynix illustrates both the strength of the present and the importance of the cycle. Its reported companywide operating margin reached 76% in the second quarter of 2026. In fiscal 2023, it recorded an operating loss of KRW7.73 trillion. Micron’s own 2023 downturn included a $1.83 billion inventory write-down. [7] [12] [13]

Our judgment is that leading manufacturers deserve recognition for greater technological differentiation and better contractual visibility. They have yet to demonstrate a higher earnings floor through a full downturn. One excellent quarter can establish that scarcity is profitable; only weaker conditions can reveal how much protection the new business model provides.

The economic exposure is uneven. SK hynix, Micron and Samsung participate in qualified HBM and server memory, while Samsung’s broader operations make consolidated cash generation a less direct measure of memory returns. Sandisk and Kioxia need to be assessed on NAND economics. On the purchasing side, NVIDIA, AMD and hyperscalers bear higher component or procurement costs, but avoiding delayed deployments can make assured supply worth paying for. The contracting evidence shows customers accepting part of that burden. [7] [8] [3] [10] [14]

How much value remains with the memory producer ultimately depends on the supply response. Here the distinction between memory categories becomes decisive.

NAND can loosen while advanced DRAM stays tight

NAND provides the clearest challenge to an all-memory scarcity thesis. TrendForce forecasts a 4–5% NAND supply deficit in 2026, followed by a positive supply balance in the second half of 2027. Its separate DRAM outlook anticipates substantial new-fab output contributions in 2028. The markets have different production technologies, capacity additions and demand mixes. [15] [2]

AI is important to storage demand without determining the whole market. Servers account for more than 40% of NAND bits, while smartphones and notebooks together still represent nearly 40%. Weak consumer demand can therefore help new NAND supply catch up even while enterprise storage grows. Our view is that stronger AI-related SSD demand and better customer contracts can coexist with renewed NAND oversupply. [15]

DRAM’s supply response takes longer, but it is already visible. Micron expected additional Singapore HBM packaging capacity to contribute in the first half of 2027 and first wafers from its initial Idaho fab in mid-2027. SK hynix planned a Yongin cleanroom opening in early 2027 while accelerating M15X. These are successive milestones on the route to qualified volume. [3] [7]

Process improvements also matter. TrendForce projects a 27% increase in combined server-DRAM and HBM bit supply in 2027 from migrations and capacity ramps. Chinese DRAM and NAND expansion adds another potential source of commodity-market relief, although reported capacity and domestic demand must be considered together. [6] [16]

Nor does memory control the timing of every AI deployment. TSMC chief executive C. C. Wei told investors in July that its packaging capacity was “so tight that now it limits my customers’ growth.” When another component delays completed accelerators, memory orders and actual deployment can diverge. [17]

Software gives buyers another way to respond. Google’s TurboQuant research reported at least a sixfold reduction in the KV-cache component of memory in its benchmarks. The economic outcome depends on how broadly such methods are deployed and how much cheaper inference expands usage. We see a credible demand-adaptation mechanism, with an unsettled net effect on purchased memory. [18]

These forces explain why we expect recurring frontier shortages while remaining skeptical of permanent industry-wide scarcity. The size of the initial deficit matters less than whether supply can grow a little faster than demand for long enough.

A prolonged shortage can still be repaired

TrendForce estimates the 2026 DRAM supply shortfall at 1–2%. A small imbalance can produce a large price response when customers urgently need compatible components. It can also disappear if effective supply gains a sustained growth advantage. [2]

Our conditional model makes that distinction explicit. It combines the forecast HBM allocation shift with a starting supply-to-demand ratio of 98.5%, the midpoint of that estimated deficit. The assumptions below are sensitivity inputs chosen to test how difficult catch-up would be; they are neutral scenarios rather than a bullish or bearish forecast. [1] [2]

Under those assumptions, a 25% improvement in pre-mix capacity and productivity still leaves an approximately 3.6% DRAM deficit in 2027. Balancing the market would require about 30% pre-mix growth; across demand-growth assumptions of 15–25%, the requirement ranges from roughly 24% to 35%. This capacity measure includes manufacturing productivity as well as wafer availability. [1] [2]

The longer-term arithmetic is less forbidding. Starting from that hypothetical 3.6% deficit, effective saleable-bit supply growing about 21.5% annually against 20% demand growth would restore aggregate balance over three years. The supply measure includes the effects of mix, yield and qualification. Specific products could remain scarce within a balanced market. [1] [2]

That is why we place the major supply-response test in 2028–2030. A shortage can survive several years of impressive output growth and still end. Historical memory markets have crossed that boundary quickly: TrendForce reported server-DRAM supplier fulfillment rising from 90% in the fourth quarter of 2018 to 120% in the first quarter of 2019. The episode is a reminder of how inventory and demand shifts can amplify a supply response. [19]