Atlas of AI Buildout

Chokepoint · Memory & Storage

The bottleneck isn't the GPU. It's the memory beside it

SK hynix and the HBM memory bottleneck

63.2%of 2025 HBM revenue, one supplier · grade A, Reported

An AI accelerator is useless without high-bandwidth memory stacked beside it, and SK hynix took 63.2% of HBM revenue in 2025. Samsung and Micron split most of the rest. Korea ranks #1 on the Memory control basis as a direct result.

HBM carries a substitutability weight of 0.95 here, near the top of the scale, because qualifying a memory vendor with an accelerator designer takes years. A customer cannot switch HBM supplier the way it reorders commodity DRAM. Supply gates GPU output directly, which is why HBM allocation, not wafer starts, has been the binding constraint on accelerator shipments.

The financial evidence is unambiguous. SK hynix closed 2025 with ₩97.15tn of revenue and ₩47.21tn of operating profit — a 49% margin — and overtook Samsung on annual profit for the first time in its history. When a bottleneck is real, it shows up in the margin.

Source — TrendForce · Counterpoint · IDC, DRAM / HBM / NAND revenue rankings (2026)
DRAM top three >90% of revenue. Q1 2026 Samsung 38%, SK hynix 29%. HBM 2025: SK hynix 63.2%. NAND 2025: SK hynix 20.9% (second).

Explore this in the interactive atlas →

The thinking behind the atlas is written up in Cheap thought has a physical address on Substack.

More findings