There are three parts to this note about the AI-related memory supercycle. It starts with what happened/changed this week, moves on to what it means, and then analyzes the specific implications for NAND memory margins.
What Happened
Developments this week build on and add specifics to the argument from my Wednesday release: the AI-driven memory "supercycle" is looking less super and more like a typical cycle.
- China wants to own some memory markets by year-end 2027.
China’s YMTC is raising roughly $5 billion and says it intends to overtake Samsung and SK Hynix in NAND by the end of 2027. CXMT is pursuing the same opportunity in DRAM. Margins are too attractive to remain confined to three incumbents. Even if it is not successful, this promises immense pricing pressure on incumbents over the next 6-12 months.
- Scarcity is accelerating engineering efforts to use less memory.
Microsoft’s new Maia 200 is explicitly designed around controlling data movement and specialized memory more efficiently, rather than simply throwing more conventional compute at the problem.
A separate new paper puts the tradeoff even more starkly. When inference runs out of KV-cache memory, operators can add GPUs—or compress the cache. Across the configurations tested, compression was 1.2–2.0x cheaper than adding GPUs or memry, while increasing capacity per dollar by as much as 16.5x.
- OpenAI has launched its Jalapeño inference chip.
The chip reportedly uses Samsung HBM4, which will intensify the scramble for HBM. But custom silicon exists precisely because hyperscalers have enormous incentives to redesign around whatever component is expensive or scarce, making Jalapeño more extrapolative than innovative.
What It Means
- Scarcity creates supply. Extraordinary memory margins are pulling capital, capacity, and new competitors into the industry. Memory suppliers are scarred by prior boom-bust cycles, causing them to only add capacity slowly, which is creating appealing margins that big-spend would-be usurpers are now chasing.