The 3–5 year fab lead time makes the 2027-vs-2028 distinction feel less important than the shape of demand. AI clusters can change the mix faster than fabs can change total capacity, so HBM and server DRAM may stay constrained even if consumer DDR4 cools off. For agent workloads I also wonder whether the answer is partly architectural: better batching, KV-cache reuse, and more deliberate context/state management can reduce pressure on both accelerator memory and the amount of replicated state each worker needs. That won’t erase the supply bottleneck, but it can change how much additional capacity turns into useful throughput.
1 Like