CONTROL PLANE

StarFlow provides full-process training/inference management, heterogeneous-resource scheduling, topology-aware RDMA placement, observability and GPU fault self-healing. KC has therefore moved beyond raw compute rental into workload and compute management.

Smart Navigation is a facility-level “O&M brain” deployed across >90% of self-built sites, combining intelligent scheduling, cloud collaboration and autonomous operations. This creates a lower control-plane asset spanning physical capacity, cooling and energy.

Agentic Infra combines token production, context memory, runtime and unified general/AI scheduling. Its architecture increasingly joins physical and software planes.

Seed develops distributed training, high-performance inference and heterogeneous-hardware compilation. Volcano Engine exposes GPU and mGPU scheduling, topology and sharing controls.

A vertically integrated benchmark across models, cloud, workload scheduling, compute, storage and networking. It demonstrates what a full-stack control plane can look like when demand and infrastructure sit inside one platform.

MIIT's 1+M+N architecture explicitly calls for resource aggregation, selection, monitoring and interconnection scheduling across different subjects, architectures and regions. This is policy-backed market infrastructure, not yet proof of frictionless compute fungibility.

Higher utilisation turns sunk accelerator capex into more billable output. Scheduling, sharing and failure recovery can therefore create value without adding silicon.

If software hides enough hardware differences, buyers can choose among more compute pools. That can reduce dependence on one accelerator source while increasing the value of the abstraction layer.

The largest white space is a trusted layer above individual providers that can discover, compare and route work across operators, regions and architectures. Policy is building some prerequisites; commercial ownership is unresolved.

VNET shows why the control plane extends below Kubernetes. AI-ready capacity depends on cooling, power, maintenance, reliability and site-level decisions.

CATL/DeepCtrls and VNET's energy-management work suggest a future boundary where compute scheduling and physical energy optimisation interact. Hyperscale dynamic coordination remains unproven.

If every provider remains a closed island, accelerator portability stays poor and national scheduling is mostly directory/trading infrastructure, the control-plane value pool will remain fragmented rather than becoming a dominant neutral layer.