HBM memory is the next AI bottleneck, and SK hynix just showed why

Discussion of AI hardware focuses on GPUs, but every accelerator depends on high-bandwidth memory (HBM) stacked beside the processor. This week's results from the largest HBM maker show that supply of that memory is committed well ahead.

The numbers

SK hynix, reporting on October 29, posted quarterly sales of KRW 24.45 trillion, up 39% year over year, and operating profit of KRW 11.38 trillion, up 62% with a 47% margin. It was the first quarter in which operating profit exceeded KRW 10 trillion ^[1]. Net profit was KRW 12.6 trillion ^[2].

Blocks & Files attributes the results to sales of 12-high HBM3E and DDR5 server memory ^[2]. TrendForce reports that all 2026 HBM supply negotiations with key customers are complete ^[1], and Blocks & Files characterizes SK hynix's HBM, standard DRAM and NAND output as "effectively sold-out through 2026" ^[2]. That phrase is the publication's wording, not a company statement. HBM4 finished development in September ^[1], with shipments starting in the fourth quarter and full-scale sales planned for 2026 ^[1][2]. The company also plans to raise 2026 capital spending significantly, with a new Cheongju fab line (M15X) slated to begin production next year ^[1].

Demand signals upstream

In the first days of October, Samsung and SK hynix each signed letters of intent with OpenAI tied to its Stargate project, according to The Register's October 2 report. The article says OpenAI expects the Korean partners to be tasked with around 900,000 DRAM wafer starts a month, a scale only "the biggest fabs on the planet" can support ^[3]. Letters of intent are not purchase orders, so treat this as a statement of ambition rather than contracted volume.

Why it matters

When memory is committed a year ahead, new accelerator capacity is gated by more than GPU availability. For buyers, that means constrained supply and firm pricing in memory-heavy configurations are plausible planning assumptions, though none of the sources here forecast prices.

Practical implication

If you plan to buy or lease GPU servers for 2026, lock in memory-inclusive delivery dates with suppliers early, and confirm which HBM generation the system uses. For API-based AI, expect the effect to show up as capacity limits more than as list-price changes.