SK Hynix's $26.5bn Wall Street landing - memory is the new oil, priced accordingly

In plain terms

SK Hynix, the world's dominant supplier of HBM memory chips that live inside every AI accelerator, sold $26.5 billion in ADR shares on Nasdaq on 10 July 2026 - the biggest foreign IPO ever on Wall Street. Shares rose ~13% at the open. The company's chairman, Choi Tae-won, hinted at future US fab investment "if conditions are right." The market is pricing in a fresh capex supercycle in AI memory.

The move

Nikkei Asia and Korea Times both confirmed the raise: $26.5bn at an implied premium to the Kospi listing, with Wall Street underwriters collecting roughly $260m in fees (Tech in Asia). Trading opened +13%, which is not usually what happens on a raise of this size - supply-heavy IPOs typically trade flat or below. This traded like a demand-anchored deal: leveraged ETFs on SK Hynix ADRs launched almost immediately (ETNews), suggesting retail is already tuned in.

Separately, LNG has re-entered Korea's national energy playbook explicitly as strategic fuel for AI and chip capacity (Korea Times) - a tell that Seoul is planning for a compute-driven electricity load that renewables alone cannot yet backstop.

Under the hood

Why memory is where the money went:

• HBM3E/HBM4 is the bottleneck. Nvidia, AMD, and every hyperscaler compete for the same wafers.
• SK Hynix + Samsung + Micron are the only three suppliers at scale. SK Hynix currently holds ~50% share in HBM.
• A US listing gives SK Hynix cheaper access to dollar capital and political air cover for a future US fab - which is what Choi's remarks trailed.

The Nikkei parallel signal: Nanya Tech (Taiwan) announced it will quadruple 2026 capex to $6.2bn on the same day. Nanya is a DRAM commodity player, not HBM - but the message is that memory buildout is now sector-wide, not just at the leading edge.

So what

For decision-makers: the AI capex cycle is entering its memory phase after the accelerator phase (Nvidia) and ahead of the fab phase. Expect memory ASPs to stay firm through 2027 and margins to hold. For builders: assume HBM allocation constraints continue to shape which models get trained where - training costs are a memory-supply story now, not a GPU story. And Seoul betting LNG on AI is the tell of the week: compute demand is now dictating national energy policy.