Memory Chips: Cycle or Supercycle?

Behind every AI model, smartphone, and data centre sits one essential component: the memory chip. The most common form is DRAM (Dynamic Random-Access Memory), found in everything from laptops to servers. Then there is HBM (High Bandwidth Memory), built by stacking layers of DRAM on top of each other so data can move in and out far faster. That matters in the age of AI because AI models constantly pull vast amounts of data from memory, much like a brain recalling a memory before acting on it, and ordinary DRAM can’t keep up. Every cutting-edge AI chip needs HBM attached. Nvidia’s latest Rubin GPU carries eight stacks of it.  HBM is also far harder to make than regular DRAM: each chip must be shaved to a fraction of its normal thickness, stacked a dozen high, and wired together with thousands of microscopic copper connections drilled vertically through silicon. A single defect anywhere in the stack can ruin the whole unit, and getting yields to a commercial level takes years of expertise and highly specialized equipment. That is why only three companies make this memory at scale: SK Hynix and Samsung of South Korea, and Micron of the United States.

The Boom

The numbers over the past year have been remarkable. With AI data centre spending on track to top $600 billion in 2026, demand has overwhelmed supply. DRAM prices nearly doubled in the first quarter, then jumped another 60% in the second. Part of the reason is that HBM and regular DRAM are made in the same factories from the same silicon, and HBM uses about four times as much silicon per gigabyte. HBM is also far more profitable. AI giants like Nvidia and the hyperscalers building AI data centres will pay a premium for a component their chips cannot run without. Regular DRAM, by contrast, sells into price-sensitive markets like phones and PCs. So, manufacturers shifted production toward HBM, supply of everything else shrank, and prices rose across the board. Micron’s Q2 2026 earnings confirmed this with revenue at $41.5 billion, up around 345% on the year.

Consumers are starting to feel the squeeze in price too, since laptops and phones rely on the ordinary memory chips that are now in short supply. Apple has confirmed that price rises are coming to its products, with CEO Tim Cook noting that memory makers are passing along huge cost increases. The storage in a base iPhone Pro reportedly cost Apple around $13 a year ago; that figure is now closer to $51.

Share prices followed. SK Hynix rose roughly 450% between November 2025 and its June 2026 peak, Samsung surged 290%, and the pair came to make up over 40% of South Korea’s entire stock index (KOSPI). A lot of that rally was bought with borrowed money. Korean margin loans hit a record ₩38.6 trillion, with over a third tied to these two stocks. Households were cashing in savings and insurance policies to join in, and when single-stock leveraged ETFs on the two names launched in late May, retail investors poured in some ₩13 trillion (around $9 billion) within weeks. At their peak these products accounted for as much as 70% of all trading in the two stocks.

When the Leverage Broke

In July, the structure gave way. As chip stocks worldwide pulled back, the borrowed money went into reverse. SK Hynix fell over 15% on July 13, and Samsung fell over 10% the same day. The KOSPI finished the month down over 22%. With so much leverage in the market this exacerbated the losses. Falling prices triggered margin calls, margin calls forced selling, and forced selling drove prices lower still. Over 1.2 million leveraged accounts, about 3.4% of Korea’s adult population, hit margin calls in July. Regulators have since banned new leveraged ETFs and tripled the cash required to trade existing ones.

Yet while the market was melting down, the underlying business kept booming. Korean chip exports rose 179% year over year in July to a near record $41 billion. The July crash came from leverage, not from any drop in demand for memory.

A Famously Cyclical Industry

This is not the industry’s first boom. Memory is a commodity business, and for decades it has traced the same pattern: demand outruns supply, prices spike, manufacturers race to add capacity, that capacity arrives all at once, and prices collapse. The 2017–18 boom, driven by cloud data centre build-outs, gave way to the 2019 downturn in which DRAM prices halved. The pandemic-era surge in PCs and devices was followed by the 2022–23 bust, in which Micron and SK Hynix lost billions of dollars after overestimating how long demand would last, and memory prices fell below cash cost for some producers.

This history is why the sector’s valuations deserve careful reading. After July’s selloff, Samsung and SK Hynix trade below 5x forward earnings, quite impressive for companies with operating margins in the mid-70s and order books sold out through 2026. While Micron sits around 6x, against roughly 20x for the S&P 500. On the surface, that looks remarkably cheap for companies growing revenue at triple-digit rates. But memory stocks have historically looked cheapest right at the top of the cycle, because analysts extrapolate boom prices into the future. When the cycle turns, the earnings collapse and apparent cheapness evaporates. The market is deliberately refusing to pay a full multiple for earnings it suspects may prove temporary. July’s events are a further reminder that in a leveraged market, stocks looking “cheap” does not stop forced selling.

The Case for a Supercycle

The case for a Supercycle rests on the idea that this time is structurally different, and that July was a financial accident rather than the top.

Start with supply. After their heavy losses in 2022–23, the big three have resisted the factory-building arms races of past cycles, and meaningful new supply won’t arrive before late 2027 at the earliest. Capacity is effectively sold out years ahead, increasingly under long-term contracts with price floors that never existed before. Micron expects the HBM market alone to triple to around $100 billion by 2028. The wild card is China, where state-backed CXMT is expanding fast and just listed in Shanghai’s biggest IPO of the year. For now, it holds only 4% of the market and can’t yet make HBM, but the oligopoly’s pricing power shouldn’t be taken as permanent.

The demand side is the stronger argument. Past cycles ran on a single engine: PCs in the 90s, smartphones in the 2010s, cloud servers in 2017. This one has several. AI data centres remain the anchor, but humanoid robots need onboard memory to process sensor data in real time, self-driving cars are effectively data centres on wheels, and early ventures are even exploring AI data centres in space. All of these are small markets today. What matters is the sequencing, with each new source of demand arriving just as the previous one would otherwise be maturing. That is what separates a Supercycle from an ordinary one.

None of this changes how the memory market ultimately works. New capacity will come eventually, and July showed how badly things can go for investors who arrive with borrowed money, supercycle or not. But sold-out capacity, record exports and rising AI spending suggest the upturn itself is intact. If even a fraction of what’s forecast for AI, robotics, and autonomous systems comes true, this one could run for years rather than quarters.

The information contained in this article is for information purposes only, and represent the views of the author. It is not intended as specific investment or financial advice, or a recommendation or solicitation to buy or sell any security. Any investments or strategies listed in this article may not be suitable for all investors. Past performance is not indicative of future performance, and as with any investment, prices may fluctuate. It is recommended that advice is sought from a qualified investment professional prior to implementing any financial plan. LOM has made every effort to ensure that the contents herein have been compiled from sources believed reliable, however LOM does not warrant the accuracy, adequacy, timeliness, or completeness of this information expressly disclaims liability for errors or omissions in this information.