The artificial intelligence boom has cycled through headline-grabbing winners such as GPU designers, cloud hyperscalers, and data centre REITs and now gathers around “memory”: beneath the H100s and Blackwell chips is specialized memory hardware that can feed data to AI accelerators fast enough to justify their extraordinary cost.

This has become a bottleneck that has recently put two South Korean companies into the spotlight: Samsung Electronics and SK Hynix.

Comebacks and Pure Plays

For most of its history, Samsung Electronics has been the world’s largest memory chipmaker by an unchallenged margin. Samsung’s product mix spans conventional DRAM, NAND flash, and the increasingly critical High Bandwidth Memory (HBM) segment — the stacked-chip architecture that sits directly alongside AI accelerator dies to deliver the enormous data throughput that modern AI models demand.

Samsung’s HBM market share plummeted from 41% in Q2 2024 to just 17% in Q2 2025 as the company struggled to pass Nvidia’s qualification tests, leaving it largely dependent on older-generation HBM3 chips while competitors shipped HBM3E. Nvidia — which designs the GPUs that consume the lion’s share of all HBM produced — is the gatekeeper of the ecosystem, and being locked out of its supply chain was an existential commercial problem for the company.

Samsung eventually passed Nvidia’s qualification tests for its fifth-generation 12-layer HBM3E product in mid-2025, roughly 18 months after completing development of the chip, following a redesign of the DRAM core to address thermal performance issues. As a result, recovery has been dramatic: Samsung reported Q1 2026 revenue of ₩133.9 trillion — up approximately 70% year on year — with operating profit climbing more than 750% from a year earlier to a fresh record of ₩57.2 trillion.

Where Samsung stumbled in the transition to HBM3E, SK Hynix sprinted – which was Nvidia’s primary HBM supplier years before the AI boom hit its current fever pitch. The financial rewards correlate with this relationship: the company posted full-year 2024 revenues of ₩66.19 trillion ($45.97 billion), up 49.5%, with a remarkable turnaround from a ₩9.14 trillion net loss the prior year to a ₩19.8 trillion net profit. HBM accounted for more than 40% of total DRAM revenue by Q4 2024, and SK Hynix’s HBM revenue had grown by more than 4.5 times over 2023.

In March 2025, SK Hynix became the first in the world to deliver 12-layer HBM4 samples to major customers, and the company is now racing to establish leadership in the HBM4 generation that will underpin Nvidia’s forthcoming Vera Rubin architecture. By late 2025, SK Hynix had already sold out its planned 2026 memory supply, underscoring the severity of the supply-demand imbalance. Q1 2026 was a record-shattering quarter, with revenue hitting ₩52.6 trillion — up 60% quarter on quarter and 198% year on year — alongside a 72% operating margin.

By the end of 2025, the two South Korean companies are estimated to have cornered around 80% of the global HBM market.

Websim is the retail division of Intermonte, the primary intermediary of the Italian stock exchange for institutional investors. Leverage Shares often features in its speculative analysis based on macros/fundamentals. However, the information is published in Italian. To provide better information for our non-Italian investors, we bring to you a quick translation of the analysis they present to Italian retail investors. To ensure rapid delivery, text in the charts will not be translated. The views expressed here are of Websim. Leverage Shares in no way endorses these views. If you are unsure about the suitability of an investment, please seek financial advice. View the original at

Source: TrendForce, Introl, Counterpoint Research, Leverage Shares analysis

SK Hynix is actually the more focused of the two, where revenue breaks down as:

DRAM (around 60–70% of revenue), which includes conventional server DRAM, PC/mobile DRAM, and HBM
NAND Flash (around 30–35%), which includes enterprise solid-state drives (SSDs) sold under the Solidigm brand and consumer/mobile NAND
CMOS Image Sensors (CIS), a small and declining segment

Customer concentration, however, has become strikingly acute. In 2023, no single client accounted for more than 10% of SK Hynix’s revenue; by 2024, Nvidia alone contributed approximately 16%. In the first half of 2025, revenue from a single “major customer” — widely understood to be Nvidia — reached ₩10.89 trillion, representing roughly 27% of SK Hynix’s total consolidated revenue for that period.

Approximately 90% of Nvidia’s HBM currently comes from SK Hynix. That single statistic captures the depth of the interdependency: Nvidia’s GPUs, which power most of the world’s AI model training and inference infrastructure, are essentially bottlenecked by SK Hynix’s production capacity. The linkages extend beyond Nvidia: Google, Meta, Amazon, and Microsoft all depend on SK Hynix’s output, directly or through Nvidia system purchases.

Samsung, on the other hand, is a highly diversified conglomerate, which is implicitly a smartphone company, an OLED panel supplier, a car audio brand, and a home appliance maker. In Q4 2025, the Device eXperience (DX) division – which includes smartphones, tablets, PCs, wearables, and home appliances like TVs, fridges, and washing machines – contributed roughly the same top-line revenue as Device Solutions (DS) — the semiconductor division, which itself splits into Memory (DRAM, NAND, HBM), System LSI, and Samsung Foundry (contract chip manufacturing for third-party fabless customers) — but with far lower margins. Memory is where Samsung makes the bulk of its profit; the System LSI and Foundry units within DS have also been loss-making or marginally profitable in recent years, as Samsung Foundry lags TSMC in advanced node yield and customer confidence.

This is partly why it has been argued that Samsung trades at a “conglomerate discount” relative to the pure-play memory upside SK Hynix offers.

A “AI Memory” Thesis Play

SK Hynix, Micron, and Samsung account for roughly 73% of DRAM’s assets, making concentration risk a genuine concern — if memory pricing weakens or AI spending slows, the fund will feel it sharply. The memory industry is also historically cyclical, prone to boom-bust dynamics when capacity additions eventually overshoot demand. Currently, however, training and running AI models at the scale now being deployed by hyperscalers requires memory bandwidth that only HBM can supply, and the capital intensity of building new HBM capacity ensures that supply will remain constrained for years.

Samsung and SK Hynix are, at this moment, considered to be highly indispensable infrastructure providers. Google’s seventh-generation TPU is said to integrate eight HBM3E stacks per chip, while Amazon’s Trainium3 reportedly uses four stacks, with both companies relying on Samsung and SK Hynix for supply. Meanwhile, the average selling price (ASP) for DRAM rose by mid-60% quarter on quarter in Q1 2026 while that of NAND surged mid-70%.

The financial vehicle for capturing this thesis around memory arrived on April 2, 2026 after Roundhill Investments launched the Roundhill Memory ETF under the ticker DRAM. The first exchange-traded fund (ETF) built exclusively around memory chip companies – with primary holdings including Samsung, SK Hynix, and Micron — DRAM amassed $6.5 billion in assets under management (AUM) in its first 27 trading days, making it the fastest-growing ETF launch in history and surpassing the previous record held by the iShares Bitcoin Trust (IBIT). RAM then crossed $20 billion in AUM within roughly 80 days of launch.

Source: Morningstar, StockAnalysis, Leverage Shares analysis

Investing During the Supercycle

While current conditions might be described as a “Hyper-Bull” phase which has now eclipsed the historic peak memory companies witnessed in 2018, it simultaneously becomes the most compelling argument for the trade and a reminder of the eventual mean reversion that followed that prior supercycle. Professional investors in Europe can now consider the 3X Long Memory ETP (DRM3) and the 3X Short Memory ETP (DRMS) – both with the DRAM ETF as the underlying – during bullish and bearish periods of volatility within the supercycle.

Also available are the 3X Long SK Hynix ETP (HNX3) and the 3X Samsung Electronics ETP (SMG3) for focused bets during bullish phases of the trajectory on the respective chipmakers’ stocks.

Your capital is at risk if you invest. You could lose all your investment. Please see the full risk warning here.