Photos: Audun Wickstrand-Iversen © DNB
By Audun Wickstrand-Iversen, manager of the DNB Fund Disruptive Opportunities
The market continues to treat memory chip manufacturers as classic cyclical stocks. Yet long-term supply contracts, constrained capacity and AI’s growing appetite for memory could reshape the sector’s profitability profile.
For a long time, the equity-market logic applied to memory chips was fairly straightforward: when prices and earnings rise sharply, caution is warranted. High margins prompt higher investment, new capacity comes on stream and the cycle eventually turns. What initially looks like excellent results can therefore already signal the end of the expansion phase. That is precisely the scepticism now shaping views of the AI-driven memory boom: despite exceptionally strong demand, the market assumes the pattern will repeat. Yet the key question for investors is this: what if it is not just demand that has changed, but the structure of the memory market itself?
Artificial intelligence is turning memory into an increasingly critical bottleneck. More powerful processors are of little use if data cannot be delivered to them fast enough. With larger models, longer context windows and greater use of inference, demand for fast memory is rising, and High Bandwidth Memory (HBM) has become a core component of AI infrastructure. For investors, the issue lies less in the technical specifications than in their economic implications: memory is no longer a largely interchangeable component. Without HBM, even expensive computing capacity cannot be used efficiently, which strengthens the bargaining power of memory makers.
One industry analysis gives a sense of the potential economic scale, estimating the HBM market at between $50 billion and $100 billion over the 2026–2028 period. But starting positions within the sector differ markedly. SK Hynix is described as the HBM market leader, with an estimated 50% to 70% share of HBM4, giving it the strongest exposure to AI and the highest margins of the three major memory makers. According to the same analysis, Micron is delivering the strongest relative growth in HBM, while Samsung, the largest and most diversified memory producer, continues to close the gap in HBM4.
Micron has signed 16 so-called “Strategic Customer Agreements”. Many of them run until 2030 and include firm take-or-pay commitments, obliging customers to pay for the agreed volumes whether or not they take delivery. Over time, these agreements are expected to cover around half of the company’s revenue. Micron has already sold its entire output for 2026 and says it can currently meet only about 50% of actual demand. For investors, this is unusual: the shortage is extreme, while a growing share of future demand is already locked in by contract.
One of the defining features of the traditional memory business is therefore changing. Historically, manufacturers had to expand capacity without knowing precisely what demand would look like a few years later, which exposed the sector to the risk of overinvestment followed by falling prices. Longer contracts, by contrast, provide greater visibility for investment planning. At its investor day, Sandisk announced new multi-year agreements with eight customers, expected to cover around half of its bit output capacity in fiscal 2027 and nearly two-thirds in fiscal 2028. For a sector whose earnings have traditionally been highly volatile, that kind of demand visibility would mark a fundamental shift.
The counterarguments are no less important. Manufacturers are investing heavily: Micron intends to raise capital expenditure to more than $25 billion, while Samsung is planning $73 billion. In previous cycles, such figures would have been a fairly reliable warning signal: higher prices lead to heavy investment, which creates excess capacity, which in turn pushes prices down. At the same time, the chairman of SK Group is warning of a wafer supply deficit of more than 20% by 2030 (wafers being the thin silicon discs from which chips are made). It is precisely this contrast between rapid capacity expansion and persistent shortages that makes the current cycle so difficult to interpret, and that lies at the heart of the dilemma facing investors.
In the bearish scenario, the market is already close to its peak, and the increase in capacity will trigger the next downturn by 2027 or 2028 at the latest. The more optimistic view, conversely, is that AI is creating a structurally new demand base, while multi-year contracts could prevent capacity expansion from becoming as uncoordinated as in previous cycles. We therefore contrast the classic boom-and-bust curve with an S-curve of technological adoption. Such a development would be decisive for valuations: not because cyclicality would disappear, but because demand could settle at a structurally higher level after each downturn.
From a technological standpoint, too, several factors suggest that memory needs will continue to broaden. HBM4 is expected to increase available bandwidth significantly once again. At the same time, technologies such as High Bandwidth Flash, CXL, and in-memory and near-memory computing are designed to address different bottlenecks in capacity, speed and power consumption. Sandisk offers a concrete example with High Bandwidth Flash (HBF): the company says the technology could deliver bandwidth comparable to HBM, with around eight to sixteen times the capacity at a similar cost. Japan’s Mizuho is even considering a 16-layer stacked solution offering HBM-like bandwidth at about one-tenth of the cost, with first revenue contributions envisaged from 2028. For investors, this would matter because it would allow NAND memory to capture a larger share of the value created by AI memory, without fully replacing HBM.
Another effect could prove decisive for longer-term demand. More efficient and less expensive AI does not necessarily translate into lower memory consumption. Under the Jevons paradox, lower costs can instead create new uses and additional consumption. More queries, longer contexts, the development of AI agents, and applications on smartphones and robots could increase overall demand by more than efficiency gains reduce it. For investors, the memory boom is therefore not simply a bet on continued chip price increases. The crucial issue is whether structural demand is growing faster than new production capacity.
Investing in this theme does not require betting on a single memory maker. Lam Research, which supplies equipment to all three major producers, is one example: it could benefit from this dynamic whichever player gains market share in the HBM race, while offering a less cyclical profile than a pure memory stock. The investment case therefore extends from chipmakers to the equipment suppliers supporting capacity expansion. If AI infrastructure remains insufficient and manufacturers manage to secure a larger share of their output under multi-year contracts, the sector’s profitability profile could stabilise. Conversely, if new capacity comes on stream faster than expected, the familiar pattern of oversupply and falling prices could return. The central valuation debate is therefore no longer only about when the memory cycle will peak, but about a more fundamental question: does memory still have the same cyclical profile it had before the AI boom?
Retrouvez l’ensemble de nos articles Strategic Case
