The Factor Substitution Theorem tells us that if AI demand is strong enough, eventually all semiconductor production capacity—regardless of whether it's advanced or mature—will be filled. Simply put: When a resource becomes scarce or expensive, the system will switch to using other, more abundant resources. AI computing systems are essentially a set of substitutable production factors: GPU computing power, storage, CPU scheduling, network bandwidth, algorithm efficiency, power, and even software complexity. What do engineers do when GPUs are strained? They don't wait for production capacity; instead, they change their technological approach. For example: Reducing redundant computation with a larger KV cache Trading storage for computing power Offloading some tasks with the CPU Improving GPU utilization with a scheduling system Reducing GPU memory requirements with quantization models. The problem of insufficient computing power is transformed into a storage problem, a scheduling problem, or an algorithm problem. And when storage or bandwidth becomes a bottleneck, the direction reverses: Increasing recompute Increasing computational density Designing dedicated ASICs Optimizing communication topology. The bottleneck never disappears; it just moves. The history of AI infrastructure development is essentially a history of constantly shifting bottlenecks. However, there's a point easily misunderstood here. Factor substitution doesn't mean demand will expand indefinitely, nor does it mean all resources will eventually be fully utilized. Because substitution is predicated on efficiency. Companies replace GPUs with more CPUs because it's still cost-effective; Storage is increased because overall costs decrease. Once the benefits of substitution fall below the costs, expansion will cease. AI computing power is ultimately subject to a simple constraint: The value created by AI must exceed the resources consumed by AI. Otherwise, no amount of production capacity will be utilized. Factor substitution expands demand, but it still depends on the growth of demand points. From a longer-term perspective, the law of factor substitution also reveals a deeper trend: AI development will not be limited by a single bottleneck. After GPUs comes storage, After storage comes interconnect, After interconnect comes electricity, After electricity may come scheduling and system efficiency. Each bottleneck shift redefines the direction of the next round of industrial investment. Technological progress is never about eliminating limitations. Instead, it moves the restrictions to new places. This is precisely how the AI era truly operates.