Single-stock record · MEMORY
Decision-support · not advice
Decision-support · not advice
Everyone watches Nvidia — the brain. Almost nobody watches the memory that feeds it. Yet AI is memory-bound, and one company makes ~60% of the HBM every AI accelerator needs — SK Hynix, which just pulled off the second-largest US listing in history. Here's the disclosed record, and the one question that decides everything: the cycle.
A modern AI server is only as fast as it can feed data to its processors — and that job falls to HBM, high-bandwidth memory: ultra-fast DRAM stacked in vertical layers and placed right beside Nvidia's GPUs. Without it, the GPU starves. As SK Hynix's own chairman put it, if the HBM is swapped out, the AI system may simply stop working — what used to be a peripheral part is now a core one.
Only three companies make HBM at scale — SK Hynix, Samsung and Micron — and SK Hynix leads with roughly 56–60% share, as Nvidia's lead supplier. Its output is reportedly sold out into 2027. For years, US investors couldn't buy it directly — it traded only in Seoul, and Micron became the proxy. The July 10 Nasdaq listing changed that.
Memory used to trade like a commodity — cheap, interchangeable, boom-and-bust. SK Hynix is now being priced like AI infrastructure, on the thesis that HBM demand runs structurally short of supply for years. The figures behind that re-rating are extreme:
Those are numbers a memory company almost never posts. The whole debate is whether they're the new normal — or a peak.
SK Hynix doesn't stand alone — it sits in the middle of the AI hardware chain. It buys ASML's EUV machines — the same sole-supplier tool from the last report — planning to spend nearly $7.8 billion on them by 2027 to etch its advanced memory. Then it sells HBM to Nvidia, supplying an estimated 70% of the memory for Nvidia's next-generation Vera Rubin platform. ASML makes the machine; SK Hynix makes the memory; Nvidia makes the brain; the world runs the AI. Three chokepoints, one chain — and SK Hynix just gave global investors a direct way in.
Memory is the most brutally cyclical business in technology. For decades it ran boom, then bust: shortage lifts prices and margins, everyone races to add capacity, capacity floods the market two to three years later, prices collapse. The entire bull case rests on one claim — that AI has broken that cycle for good. The bear case says it has only paused it.
SK Hynix looks cheap — roughly 8× earnings. But that's 8× peak earnings, at record margins, at the top of a cycle. If HBM supply catches up — and SK Hynix, Samsung and Micron are all expanding at once, together committing hundreds of billions — those margins compress, earnings fall, and a "cheap" 8× can balloon into an expensive multiple on shrinking profits. A low P/E at a cyclical peak is often a warning, not a bargain.
Memory is buried in acronyms. Here's the plain-English map — with the three that drive the AI supercycle marked.
| Acronym | Full form | What it is |
|---|---|---|
| DRAMDRIVER | Dynamic Random-Access Memory | Volatile working memory; holds active data, clears when power is off |
| NANDDRIVER | Not-AND (flash) | Non-volatile storage; keeps data without power — SSDs, phones |
| HBMDRIVER | High Bandwidth Memory | Stacked DRAM for AI accelerators and GPUs; the hottest segment |
| SRAM | Static Random-Access Memory | Faster volatile memory; needs 4–6 transistors per bit vs DRAM's one |
| NOR | Not-OR (flash) | Non-volatile flash for code storage; fast random read |
| DDR | Double Data Rate (SDRAM) | Mainstream DRAM standard; transfers data on both clock edges |
| SDRAM | Synchronous Dynamic RAM | DRAM synchronised to the system clock — the base for DDR |
| LPDDR | Low-Power DDR | Power-efficient DDR for mobile and laptop devices |
| GDDR | Graphics DDR | High-bandwidth DRAM for GPUs and AI inference |
| RDIMM | Registered Dual In-line Memory Module | Server DRAM module with a register buffer for stability |
| QLC | Quad-Level Cell | NAND storing 4 bits per cell; high density, lower endurance |
| SLC | Single-Level Cell | NAND storing 1 bit per cell; fastest, most durable |
| CXL | Compute Express Link | Interconnect standard for memory pooling and expansion |
| NVMe | Non-Volatile Memory Express | Protocol for fast SSD access over PCIe |
| SSD | Solid-State Drive | Storage device built on NAND flash |
| HDD | Hard Disk Drive | Mechanical magnetic storage; slower, cheaper per GB |
| HAMR | Heat-Assisted Magnetic Recording | HDD tech enabling ultra-high-capacity 30TB+ drives |
The three driving the AI supercycle are DRAM (working memory), NAND (storage), and HBM (stacked DRAM for AI chips). The rest — DDR, LPDDR, GDDR — are DRAM variants; SSD, QLC, SLC are NAND products. Master these three categories and the memory headlines suddenly read plainly.
This report does not tell you to buy, sell, or hold SK Hynix, and contains no price target. A record listing and spectacular numbers tell you a company is important — not that its shares are cheap, safe, or timed well for you.
Memory is deeply cyclical, this is a single, concentrated position with heavy customer and geopolitical exposure, and the stock has already run enormously — it can keep rising or fall hard. Figures are drawn from SK Hynix's filings, earnings and reputable news sources current to mid-July 2026 and will change; verify the latest before acting.
Educational, decision-support material only — not financial, investment, legal, or tax advice, and not a recommendation on any security. Do your own research and consider a licensed professional.
The chokepoint is real and the cycle risk is real — the edge is reading which way the institutional money actually leans, quarter by quarter. Join the free live masterclass and learn to read the disclosed record yourself, on any company. No hype. No price targets. Just the method.
Save my free seat →