GPU is no longer the most critical bottleneck: The real constraints in the AI supply chain are shifting to ABF substrates, memory, and storage. TrendForce’s latest supply chain radar has revealed a crucial signal: The shortage in AI infrastructure is transitioning from “fighting for GPUs” to “scrambling for complete server components.” Currently, GPU lead times of 20–30 weeks are gradually returning to relative balance; however, other key components remain significantly constrained: - ABF substrates: 48–56 weeks (normal: ~12 weeks) - HDD: ~50 weeks (normal: ~16 weeks) - DRAM: ~20 weeks (normal: ~8 weeks) - NAND and MLCC: Supply remains tight This highlights a very practical reality: Owning GPUs does not guarantee smooth server delivery. An AI server is an integrated system. While GPUs are the most expensive and visible component, if any single part—ABF substrate, DRAM, NAND, HDD, or MLCC—is in short supply, the entire unit’s delivery will be held up. Therefore, the next phase of AI infrastructure investment should not focus solely on $NVDA itself, but on the broader opportunities in supply chain expansion and catch-up across components. **ABF Substrates** As advanced AI chip packaging becomes increasingly complex, demand for high-end substrates continues to rise. **DRAM / HBM** AI training and inference are driving higher memory capacity and bandwidth requirements per server, further increasing the value of memory components. **NAND / HDD** AI-generated data is growing exponentially. Storage demand extends beyond high-performance computing—cold data, training datasets, logs, and model files all require significantly larger storage capacities. **MLCC** Higher server power consumption and more complex circuit boards increase both the quantity and specifications required for passive components. What’s truly significant is that the AI investment thesis has evolved from: “GPU shortage” to: “Simultaneous expansion across GPU + HBM + ABF + storage + passive components + power.” Going forward, the market is likely to pay increasing attention to components previously overshadowed by the spotlight on GPUs. If judged purely by supply tightness, I would prioritize monitoring: ABF substrates, DRAM / HBM, and HDD— because these segments currently exhibit the most significant deviations from normal lead times and have the most direct impact on overall server assembly schedules. The real challenge in the next phase of the AI supply chain may no longer be: “Do we have GPUs?” but rather: “Can we assemble a complete AI server on time?”