SK Hynix and Micron Both Hit $1 Trillion. Memory Just Joined the AI Trade.
Two memory chip makers crossed the trillion-dollar threshold on the same day, sending the S&P to a new record. HBM is suddenly the most valuable commodity in semiconductors. Two years ago, this industry was in a glut.
What Happened
SK Hynix (KRX: 000660) and Micron Technology (NASDAQ: MU) both crossed $1 trillion in market capitalization on May 26, joining a club that until recently included only Apple, Microsoft, Nvidia, Alphabet, Amazon, and Saudi Aramco. The dual milestone helped push U.S. stocks to fresh record highs, with the S&P 500 closing at 7,022.95 on April 15 and continuing higher in May. The memory chip industry, which spent most of 2022 and 2023 mired in a brutal oversupply cycle, has now produced two trillion-dollar companies inside three years.
The driver is HBM, or high-bandwidth memory: the stacked DRAM that sits next to Nvidia’s GPUs and feeds them data fast enough to keep them from idling. Without HBM, even the most powerful AI chip is bottlenecked. SK Hynix and Micron are two of the three companies in the world (Samsung is the third) capable of producing it at the volume and speed Nvidia and AMD require.
Why It Matters
The memory cycle is dead, long live HBM. Memory has been the most cyclical business in semiconductors for forty years. Boom, bust, boom, bust, driven by capacity additions that always overshoot demand. HBM has broken that pattern. It is a custom, advanced-packaging product that cannot be manufactured at the same scale as commodity DRAM. The customers (Nvidia, AMD, hyperscalers) are signing multi-year supply agreements, which gives Hynix and Micron the kind of revenue visibility they never had selling DDR5 into PCs. The same companies that were marked-to-market at $30 per share two years ago are now trillion-dollar entities because their product profile changed fundamentally.
Nvidia is the customer that made this possible. Nvidia bought $25.8 billion in inventory and committed to $95.2 billion in supply contracts in Q1 alone. A huge share of that is HBM. Every Blackwell GPU consumes multiples of HBM compared to prior generations, and Nvidia is shipping more Blackwell units than any previous architecture in its history. The math is straightforward: as long as AI infrastructure spending continues, HBM consumption continues, and memory makers compound.
The risk is the same one Nvidia faces. If AI capex peaks, HBM demand peaks with it. There is no diversified end market underneath. Unlike DRAM (used in phones, PCs, servers, automotive), HBM is overwhelmingly an AI training and inference product. The trillion-dollar valuations bake in continued growth. A pause in hyperscaler capex would expose memory makers to the same multiple compression that Nvidia would face.
- AI capex pause: HBM is overwhelmingly an AI product. A slowdown in hyperscaler spending cuts both volume and pricing power.
- Samsung catching up: Samsung was slower to qualify HBM3E with Nvidia. If they close the gap, the duopoly becomes a three-way market with more competitive pricing.
- Multi-year supply agreements: HBM customers are locked in through 2027. Visibility is unprecedented for memory.
- Capacity constrained: HBM manufacturing requires advanced packaging that takes years to build. Supply additions lag demand.