After HBM, the memory race moves down one floor: storage.
Reuters reported on September 18 that Solidigm (SK Hynix’s NAND arm) is studying a NAND fab in the United States, SK Hynix is negotiating production with Intel in Ohio, and China’s CXMT is preparing to enter the market. Three moves saying the same thing: global NAND supply is tightening, and the memory chain is being re-geopoliticized.
AI’s memory bill has three layers
The AI hype talks about GPUs and HBM, but the memory bill has three floors:
- HBM: the memory inside the accelerator, 2025’s crisis that reorganized the GPU queue.
- DRAM: the server’s memory feeding inference.
- NAND: where the data that trains and serves models rests: datasets, checkpoints, data lakes, backups.
Tightening NAND tightens the cost of everything storing AI, and AI’s data is the fastest-growing component: each model generation trains on more material, each training checkpoint is a full copy, each data lake grows with the product.
The geopolitical why of the American fab
Solidigm studying a US fab and SK Hynix negotiating with Intel in Ohio tell the same story HBM told: the Asia-concentrated memory chain became a strategic risk, and the blocs are building redundancy.
CXMT (China) entering NAND points to both vectors: more global supply in the medium term (price relief) and more fragmentation by bloc (supply risk). For the corporate buyer, the practical read is the dependency inventory applied to the byte: who supplies your storage, and what is the alternative if a bloc enters friction.
The sequential-scarcity script
AI hardware became a queue-based scarcity market: HBM first, DRAM in the middle, NAND now. The script repeats precisely: supply tightens, prices rise, governments enter with fab incentives, China announces entry, and the corporate buyer who did not plan pays the peak.
The same GPU script, sealed by AWS’s 15% on-demand hike: whoever understands the cycle contracts before the squeeze, whoever does not renegotiates after the adjustment. The difference between the two is a schedule, not luck.
What the ordinary company does with this
Three actions worth more than any price forecast:
- Inventory the three layers of memory the operation consumes (and each one’s cost per workload).
- Apply lifecycle to the byte: infinite snapshots, duplicated datasets, and old training checkpoints are wasted NAND, and scarce NAND punishes waste first.
- Renegotiate storage before the peak: a fixed contract in a calm market is what separates predictability from surprise.
Conclusion
The AI race is contested in layers: first the compute chip, now the memory holding what it learned. NAND entering the geopolitical dispute is the signal that data (AI’s input) also has a chain, a scarcity, and a sovereignty.
For anyone operating infrastructure in Brazil, the read is total cost: memory will oscillate, and the plan that survives is the one that knows its three layers, has a lifecycle for the byte, and renegotiates before the newspaper announces the squeeze.