Provincial procurement and operator-run model marketplaces are beginning to reveal who might stand between AI customers and the underlying infrastructure. Zhejiang has selected computing-service suppliers, while Wuhan reports actual marketplace transactions. The open question is whether cloud providers, telecom operators or provincial platforms will control billing and scheduling—and therefore capture the most durable share of the economics.
Who controls model access and workload routing?
Shanghai’s Haina launch connects a common model interface to several cloud and regional computing providers. Alibaba’s selective endpoint retirements and Kingsoft Cloud’s multi-model offering show different ways to control developer access. A launch or product catalogue does not demonstrate profitable nationwide orchestration.
1 OCTOBER 2026 · SOFTWARE PORTABILITY UPDATE
Portability is beginning below the scheduler: common model infrastructure is being mapped directly onto Ascend.
DeepSeek’s Ascend-native DeepEP and DeepGEMM implementations preserve familiar public interfaces while replacing lower-level communication and matrix kernels with Ascend-specific paths. That does not make heterogeneous accelerators interchangeable, but it can reduce the amount of application-layer rewriting required before a scheduler or cloud platform can place a workload on different hardware pools.
Boundary: this is workload-specific portability evidence, not proof that China has a universal cross-vendor control plane.
22 SEPTEMBER 2026 · THESIS UPDATE
From provider scheduling to provincial and cross-province coordination
Hebei illustrates aggregation of existing local schedulers and resource pools. Xiong’an illustrates access across regions. Fujian–Ningxia supplies a specific pilot linking a workload scheduler with electricity-load systems.
The updated framework separates monitoring, marketplace matching, workload placement and token/model routing. Each has a different commercial test.
Read the evidence and implications →Control plane now extends above individual providers.
Registered national-network standards add monitoring, scheduling, resource identification, billing and matching/trading to the architecture. This supports an emerging national coordination layer; operational maturity remains unproven.
The layer that decides whether installed infrastructure becomes useful compute.
This is now a first-class value-chain layer. The control plane is not one product: it is a stack of software and management systems that observe, schedule, route, recover and optimise AI workloads from model/API demand down through compute, network, facilities and — increasingly — energy. The key research question is which actors can move from optimising their own assets to orchestrating resources across providers, architectures and geographies.
Who can make multiple sites, owners and architectures behave like a more valuable resource?
The economic opportunity may move beyond owning capacity. A higher-order control plane needs enough visibility and authority to decide what workload should run where, when, over which network, on which accelerator pool, with which storage/memory path and — eventually — under which power conditions. China's national compute-interconnection policy makes this a system-level question rather than merely an internal cloud optimisation problem.
Current company lens: KC approaches the opportunity from workload/compute orchestration; VNET from AIDC, network and operational visibility; telecoms from connectivity and regional/national fabric; Huawei, Alibaba and ByteDance from deeply integrated private control planes. None is yet proven to be the neutral national orchestrator.
Why the policy matters: MIIT's programme explicitly calls for unified compute identifiers, resource-query interfaces, state awareness, resource selection, supply-demand matching and scheduling across different subjects, architectures and geographies. The 2026 node programme turns that into a 1+M+N operating structure. Guangdong's regional node description goes as far as compute registration, transaction matching and cross-provincial scheduling. These are implementation signals, not proof of seamless commercial fungibility.
Who currently controls what?
Kingsoft Cloud
StarFlow explicitly spans heterogeneous-resource scheduling, training/inference orchestration, topology-aware RDMA scheduling, observability and GPU fault self-healing. This is a genuine compute/workload control plane. The unresolved question is breadth across current domestic accelerator vendors and whether KC can orchestrate capacity beyond resources it directly manages.
VNET
Smart Navigation is described by VNET as its next-generation “O&M brain”, with visual management, intelligent scheduling and cloud collaboration, deployed across more than 90% of self-built data centres. VNET therefore has a real lower control-plane position across facility operations, energy and capacity — not yet evidence of a cross-provider compute scheduler.
Huawei
Agentic Infra joins token production, tiered context memory and unified general/AI scheduling. CCE Volcano Next is Huawei's explicit scheduling layer above heterogeneous infrastructure; Huawei claims shared training/inference pooling and fragmentation consolidation improve utilisation by over 30%.
ByteDance / Volcano Engine
ByteDance Seed works on distributed training, high-performance inference and heterogeneous-hardware compilation. Volcano Engine exposes GPU/mGPU scheduling, topology-aware placement and shared-resource mechanisms. It is a strong demand-to-control-plane case.
Alibaba Cloud
Cloud-native AI tooling combines heterogeneous compute, storage and network resource management with workload scheduling, GPU sharing and model/inference services. Alibaba is one of the clearest vertically integrated control-plane benchmarks.
National + telecom fabric
MIIT's 1+M+N compute-interconnection architecture explicitly targets resource aggregation, selection, monitoring, trading and cross-subject/cross-architecture/cross-region scheduling. This creates a possible layer above individual clouds — but standards and nodes do not yet prove seamless commercial fungibility.
Primary research: Kingsoft Cloud StarFlow • VNET Innovation / Smart Navigation • Huawei Agentic Infra • ByteDance Seed Infrastructure • Volcano Engine GPU scheduling • MIIT 1+M+N nodes. Company performance claims remain company-reported unless independently corroborated.