Sector Intelligence

The AI Memory War: Why SK Hynix, Samsung and Micron Matter

Jul 27, 2026

By Ieva Bružaitė | Founder, Orion Capital | July 2026

Executive Summary

SK Hynix sits at the center of the AI memory revolution. Every technological revolution has a visible layer and an invisible one. In the AI era, the visible layer belongs to Nvidia and AMD — the companies whose names move markets and dominate headlines. The invisible layer belongs to memory. Without High Bandwidth Memory, the most advanced AI accelerator in the world is an expensive piece of silicon waiting for data that cannot arrive fast enough.

July 2026 made this distinction impossible to ignore. Semiconductor stocks entered their sharpest correction since early 2025, with the Philadelphia Semiconductor Index falling roughly a quarter from its late-June peak. Yet beneath the volatility, the structural story strengthened rather than weakened: the HBM market has grown from roughly $16 billion in 2024 to an estimated $58 billion this year, SK Hynix completed one of the largest US listings in history, and 2026 memory capacity is effectively sold out. This report examines the memory war between SK Hynix, Samsung and Micron — and why we believe it has become the most consequential rivalry in the AI supply chain.

Why Semiconductor Stocks Recently Sold Off

The July correction was severe by any measure. After the semiconductor complex posted a record second quarter, the sector gave back nearly 24% from its June high in the space of three weeks — technically a bear market, and the largest weekly decline since March 2025. Micron fell 13% in a single session. Intel dropped 21% over seven trading days after reports that its most advanced foundry process would not reach profitable yields on schedule.

What matters is not the magnitude of the decline but its anatomy. Three forces converged. First, profit-taking: after a 71% quarterly gain in the leading semiconductor ETF, some selling was arithmetic, not analysis. Second, a hawkish turn in Federal Reserve minutes reminded investors that high-multiple growth stocks remain hostage to the discount rate. Third — and most telling — the market began punishing good news. Samsung reported a preliminary quarterly operating profit nineteen times higher than a year earlier, a staggering result, and its shares fell roughly 7% anyway. When record earnings trigger selling, the market is no longer asking whether the numbers are good. It is asking whether they can stay this good.

Our reading: this is a repricing of expectations, not a repudiation of demand. Hyperscaler capital expenditure continues to rise — Alphabet lifted its 2026 capex guidance to as much as $205 billion — and memory suppliers report capacity committed through year-end. Corrections of this kind are what a supercycle looks like from the inside.

The AI Memory Race

Six companies define the AI hardware ecosystem. Only three of them make the memory that binds it together.

Company	Core Strength	AI Position	Key Risk
SK Hynix	HBM leadership	~58–62% of the HBM market; lead supplier for Nvidia’s Rubin generation	Memory cycle; expectations now priced for perfection
Samsung	Scale & manufacturing	Broad semiconductor ecosystem; HBM4 recovery push	HBM execution and customer qualification
Micron	US memory leader	Rose from ~2% HBM share in 2023 to ~20%	Competitive pressure from two larger rivals
TSMC	Advanced foundry	Manufactures the logic behind nearly every AI chip	Capacity and capex intensity
Nvidia	AI GPUs	AI compute leader; sets the pace for HBM demand	Valuation; concentration of expectations
AMD	AI accelerators	Fast-growing challenger platform	Execution against an entrenched leader

Why HBM Matters

High Bandwidth Memory is a three-dimensional structure: DRAM dies stacked vertically and connected by microscopic channels, delivering five to ten times the bandwidth of conventional memory. Modern AI models are not limited by how fast a chip can compute, but by how fast data can reach it. Memory bandwidth — not raw compute — has become the binding constraint of artificial intelligence.

The economics reflect this. The HBM market was worth roughly $16 billion in 2024. It reached about $38 billion in 2025 and is projected at $58 billion in 2026, with credible estimates pointing toward $100 billion by 2028. Operating margins at the leading supplier have reached levels — around 75% — that most software companies would envy. When a component manufacturer earns software margins, it has stopped being a commodity business. It has become strategic infrastructure.

The barriers protecting this position are physical, not legal: three-dimensional stacking, advanced packaging, and capital cycles measured in years. This is why the industry remains a three-player oligopoly, and why no fourth player is arriving soon.

Why Nvidia and AMD Depend on SK Hynix

Nvidia and AMD design processors. They do not make the memory those processors require. Every accelerator that leaves TSMC’s fabs carries stacks of HBM sourced from SK Hynix, Samsung or Micron — and the allocation of those stacks quietly determines which AI platforms can scale.

SK Hynix holds roughly 58–62% of this market, and its position appears to be carrying over into the next generation: analysts expect it to supply the majority of HBM4 for Nvidia’s Rubin platform. Samsung, after slipping to third place, is mounting an aggressive recovery through HBM4 and pricing. Micron has executed one of the quiet success stories of the cycle, growing from 2% share in 2023 to roughly 20% today.

The strategic conclusion is simple: the AI supply chain must be analysed as an ecosystem, not as isolated tickers. Alphabet spends, Nvidia sells systems, SK Hynix ships the memory. Each link inherits the demand — and the fragility — of the others.

What to Watch This Week

SK Hynix reports second-quarter results on 29 July — its first earnings as a Nasdaq-listed company following a landmark US debut earlier this month. Consensus points to revenue near ₩84 trillion and a record operating profit around ₩64 trillion, implying an operating margin of roughly 75–77%. After a month in which the market sold record profits, the numbers themselves may matter less than the guidance: any signal on HBM4 timing, 2027 capacity commitments and pricing will set the tone for the entire AI complex into autumn.

Orion Capital Perspective

There are two clocks in every market. The reactive clock ticks with headlines — a selloff, a Fed minute, a delayed process node. The anticipatory clock ticks with structure — capacity, margins, dependence. In July 2026 the reactive clock rang loudly, and it was worth hearing: valuations had run ahead of patience, and some air deserved to come out.

But the anticipatory clock tells a different story. Foundation models keep growing. Data centres keep expanding. Memory capacity for this year is already spoken for, and the technology required to produce it belongs to three companies on Earth. Investors who look only at GPU designers are watching the visible layer of the revolution and ignoring the layer that makes it possible.

Orion Capital believes the defining question of AI investing has shifted. It is no longer “who designs the fastest chip?” It is “who controls the bandwidth?” For now, the answer sits in Icheon, South Korea — with two determined rivals closing in.

Related Research:

About the Author

Ieva Bružaitė is the Founder of Orion Capital and an independent market researcher focused on artificial intelligence, semiconductor infrastructure and long-term investing. Her research focuses on AI infrastructure, semiconductors, data centres and the companies shaping the next generation of technology. Her work combines independent market research with long-term investment analysis for technology-focused investors.

Read more about Ieva Bružaitė and Orion Capital→

References

Disclaimer: This publication is independent research and reflects the views of the author. It is provided for informational purposes only and does not constitute investment advice or a recommendation to buy or sell any security. Investors should conduct their own analysis before making decisions.

← Back to Orion Capital