Xunzhuo
Liu

In the old network, routing mostly moved packets. In the AI stack, the thing being moved is semantic work: intent, uncertainty, privacy risk, reasoning demand, memory, tool use, and action. Once traffic starts carrying meaning, the routing layer cannot stay a thin forwarding layer.

This is why tokens feel like the wrong primitive to optimize alone. Tokens are easy to count, but they are not equal. A token produced by a small open model, a specialized model, a frontier closed model, or a local edge model has a very different cost, latency, energy footprint, and risk profile. The deeper question is whether the system spent the right kind of intelligence in the right place.

• Tokens are not equal. The cost gap between model paths can be orders of magnitude, so token economics has to ask whether each token was spent in the right intelligence tier.
• Energy is the hidden unit of intelligence. A model is not only capability; it is hardware, power, latency, supply, and operating cost.
• The durable problem is coordination. The future is not one frontier model serving everything, but a heterogeneous fabric of closed models, open models, tools, verifiers, memory, edge devices, and different generations of hardware.

That is the lens behind Semantic Routing as Energy Infrastructure. To me, semantic routing is the control layer that decides where semantic work should live, when to stay cheap and local, when to escalate, when to retrieve, when to verify, and when to spend the expensive intelligence. It is not just model selection. It is resource scheduling for intelligence.

My research direction is to make this layer real and measurable: workload signals, routing memory, policy languages, evaluation, cost-quality frontiers, privacy boundaries, and cross-layer scheduling. vLLM Semantic Router is one concrete step toward an open semantic control plane for AI systems: inspectable, composable, and shared by design.