The Memory Chip You Have Never Heard of Is the Reason AI Models Can Think at All. Mirae Asset Just Bet Bigger on SK Hynix.
Every conversation about which companies win the AI era eventually arrives at Nvidia. But behind every Nvidia GPU doing billion‑token computations for OpenAI, Anthropic, and Google sits a memory chip without which none of those computations would be fast enough to matter commercially. That chip is High Bandwidth Memory. And the company that makes more of it than anyone else on earth is SK Hynix.
Mirae Asset Securities raised its target price for SK Hynix to 360,000 Korean won per share, equivalent to approximately $265 at current exchange rates, citing sustained and growing demand for High Bandwidth Memory from AI chip manufacturers. The upgrade was reported by Tech in Asia and reflects a broader analyst consensus that SK Hynix's position as the world's leading HBM supplier is more durable and more commercially valuable than its current valuation implies.
High Bandwidth Memory is a specific class of memory chip designed for exactly the problem that makes AI training and inference so computationally demanding: moving enormous amounts of data between the processor and the memory that stores the model's parameters. Standard DDR memory chips, which power the memory in conventional computers, are connected to the processor through relatively narrow data channels and have bandwidth measured in tens of gigabytes per second. HBM stacks multiple memory chips vertically using through‑silicon vias, a packaging technique that creates thousands of simultaneous data connections between the memory stack and the processor, enabling bandwidths measured in terabytes per second.
Without HBM, the GPUs running frontier AI models would spend most of their time waiting for data to arrive from memory rather than performing computations. The memory bandwidth bottleneck is specifically the constraint that Fractile's in‑memory compute architecture is trying to eliminate from the chip level. HBM is the current industry's answer to the same constraint at the packaging level.
SK Hynix is understood to supply more than 60 percent of global HBM3E, the current‑generation high bandwidth memory used in Nvidia's Blackwell architecture GPUs. Its relationship with Nvidia is strategic and exclusive at the bleeding edge: the HBM3E 16‑stack products that will power Nvidia's upcoming Rubin AI chip architecture carry pricing premiums of approximately 10 to 20 percent above the 12‑stack products they supersede, and SK Hynix is the primary qualified supplier.
The revenue implications of that pricing premium, multiplied across the volume of AI chips that Nvidia, AMD, and Google are deploying globally, are the foundation of Mirae Asset's upgraded price target. Each incremental generation of AI chip requires more HBM per chip, at higher bandwidth specifications, from a supplier pool that Samsung and Micron are still qualifying into. SK Hynix's technological lead in HBM3E qualification is not a permanent moat, but it is a two to four quarter lead that generates pricing power and margin expansion at exactly the moment AI chip demand is accelerating fastest.
The Rubin AI chip architecture that Nvidia is expected to ramp in late 2026 and 2027, and which Jensen Huang has described as delivering 35x lower cost per token than the H100 generation, will require more HBM capacity per chip than any previous Nvidia architecture. For SK Hynix, Rubin's production ramp represents the next leg of the AI memory demand story that Mirae Asset's price target upgrade is reflecting.
