The Week DRAM Almost Caught Bitcoin — Memory Crosses Into Asset-Class Territory

On the last week of May 2026, the combined market capitalization of Samsung Electronics and SK hynix came close to matching Bitcoin's. Two companies that have been making memory since 1983 and 1992 are now being priced by capital markets at roughly the same weight as the asset that capital has been calling 'digital gold' for a decade.

This is not a market-cap parlor game. It is the terminal stage of memory being repriced from industrial commodity to asset class.

Why now

DDR5 16Gb spot closed May 31 at $41.9, up double-digits year-on-year. But the real story is not the spot tape — it is who cannot get supply. More than half the bill of materials of a modern AI accelerator is memory (HBM plus auxiliary DDR5), and Rubin-class racks reportedly consume more than 50% more HBM per rack than the prior generation.

On May 28, Micron began volume 1-alpha (1α) node DRAM production in Boise. The same day, Forbes argued Intel's foundry-leasing model is rewriting INTC's financial thesis. On May 29, Bank of America turned constructive on Korean semiconductors. The same day, news broke that Korean semiconductor inspection-equipment makers are facing their worst-ever component shortage on record.

These dots connect to a single line: structural supply tightness meeting structural demand pull.

The Bitcoin paradox

Bitcoin's supply is fixed by algorithm. Mining rewards halve every four years; total supply caps at 21 million. Memory has no such constraint — fabs can be expanded, capacity can be added. The fact that two memory makers can approach Bitcoin's market cap means the market has begun to believe the industry's capacity discipline as much as it believes Bitcoin's protocol.

Capital is pricing in the assumption that this cycle will not look like 2017 or 2022, when capacity gluts destroyed price. That assumption rests on three pillars.

First, AI cannot be turned off. Nvidia CFO Colette Kress used the May 31 earnings call to declare AI infrastructure 'no longer nice-to-have.' If AI infra is essential, memory capacity becomes essential by attachment.

Second, HBM and commodity DRAM lines have decoupled. Their capacity decisions no longer follow the same dynamic. HBM is effectively a derivative of AI accelerator unit volumes.

Third, the backend bottleneck — packaging, advanced inspection — is binding capacity expansion. The Korean inspection-equipment shortage is the canary.