The bottleneck in AI computing power is surprisingly a layer of fabric

Just saw Huxiu's article "Global AI Giants Frantically Grabbing a Piece of Cloth," and my first reaction was a sense of déjà vu. We usually focus on compiler optimization on Ascend, staring at operator fusion, memory bandwidth, and IR optimization space, tending to think bottlenecks lie in chips, software stacks, or scheduling strategies. Reality has pushed the problem down a layer. High-end electronic-grade fiberglass cloth, used for packaging substrates and PCB skeletons, is starting to choke AI servers.

The news mentions that China Jushi raised prices by 15% for thick cloth and 20% for thin cloth in early September, and Kingboard Laminates, a leader in copper-clad laminates, followed suit with price adjustments. The reason isn't mysterious. Electronic yarn raw materials are scarce, loom capacity is highly concentrated, and prices have risen for a year straight. This signal is very engineering-focused. When upstream materials start getting expensive, it means physical components like racks, boards, power supplies, cooling, and substrates are also being fought over.

The AI race has rolled from the algorithm layer down to the physical layer. Software can fuse several operators into one blob, saving a memory access, but you can't fuse away a roll of cloth, nor can you conjure up loom capacity.

So I advise those planning AI infrastructure not to just count cards and VRAM. At least break down the BOM to copper-clad laminates, electronic cloth, glass yarn, connectors, and power modules, and write lead times and alternative suppliers into capacity plans. Don't just look at whether there's a factory for domestic substitution; look at dielectric loss, thermal expansion, yield rates, and certification cycles. Running through one batch doesn't count as passing; stable supply for half a year does.