Your AI workloads still need a service mesh - You always need a load-balancer/proxy/gateway, more generally, “middleware."