What does each company tell us about China’s AI infrastructure build-out?
This is not a directory of company profiles. Each company is used to answer a different infrastructure question: who supplies physical capacity, who coordinates heterogeneous compute, who creates demand, who integrates energy, and who connects the technology stack. Start with the question you want answered.
START WITH THE MAPSee the system before opening a company thesis.
CURRENT RESEARCH · OCTOBER 2026
Company exposure differs across the infrastructure stack.
The latest evidence concerns national compute–energy policy, Huawei–China Telecom’s Horinger project and proposed Chinese AI technology-export controls. None should be translated automatically into new VNET, GDS or Kingsoft Cloud revenue. Financing risk also differs depending on who owns the chips, buildings and electrical assets.
VNETGrid energisation, wholesale delivery and CATL-linked optionality.
GDSWholesale utilisation, project financing and offshore asset economics.
Kingsoft CloudAI cloud demand, model serving, hardware costs and margins.
HuaweiAscend deployment, software interoperability and useful throughput.
One system. Four lenses. Start with the relationships before the individual companies.
This is the orientation layer for the company research: see the system first, then move into the companies and evidence behind each connection.
The ecosystem map is a visual evidence layer beneath the value-chain spine. Use the four lenses to move between the whole system, the control plane, individual companies and specific relationships. Solid edges are evidenced relationships; dashed edges are research gaps or architectural convergence and are never promoted into partnerships.
CONTROL-PLANE LENS
Who can see, schedule and optimise which resources?
This lens separates internal resource management from the more valuable but harder problem of cross-provider orchestration.
White space: a neutral control layer capable of discovering and routing workloads across different operators, accelerator architectures, regions and eventually energy conditions. MIIT policy explicitly points toward parts of this market structure; commercial capture remains open.
This is the relationship graph.It is generated from company interactions rather than a generic supply chain. Select a node to isolate its first-order ecosystem. Select an edge to inspect the relationship directly. The Relationships lens is the complete interaction register and gateway to drill-down pages.
How the machine fits together
China AI Infrastructure · chinaaievolution.com · See accompanying evidence and review date.
Ecosystem shapes emerging from the evidence
New central archetype — ByteDance: demand-to-infrastructure hyperscaler. Douyin, Doubao, Seed, Coze and Volcano Engine create token demand that pulls ByteDance downward into schedulers, heterogeneous accelerators, proprietary network protocols, storage, custom CPUs and hyperscale power infrastructure. OPEN BYTE DANCE DEEP DIVE →
China AI Infrastructure · chinaaievolution.com · See accompanying evidence and review date.
KV/context storage architectural convergence; no commercial edge established
Memory / storage
Convergence
Evidence
Alibaba
VNET
Historical data-centre customer relationship
AIDC
Historical
Evidence
Alibaba Cloud
GDS
Physical Express Connect access-point presence
Network / AIDC
Confirmed scope
Evidence
Moore Threads
JD Cloud
Domestic accelerator cluster cooperation
Compute / cloud
Announced
Evidence
Register rule: this is the master interaction register, not a highlights matrix. Relationships can carry several dimensions simultaneously. Co-participation, adjacency and architectural similarity remain distinct from commercial relationships.
WHY THESE COMPANIES?
Companies are selected for the architectural question they help us test
The names below are not a general China-tech watchlist. Each is included because it controls or connects a layer that can test the system thesis.
This is not a directory of the largest Chinese technology companies. We track companies when their assets, technology, customers or partnerships illuminate a consequential part of the system thesis. Conventional labels such as “data centre”, “cloud” or “battery” are not enough: the research asks what each company controls, operates, connects to, or may be able to extend into.
Core thesis companiesExplicit, evolving company theses where architecture, value capture and economics are tested separately.
System architects & enablersCompanies whose technology or infrastructure can determine how multiple layers work together.
Demand anchorsModel, cloud and application companies whose workloads pull investment through the infrastructure stack.
Get to know the companies
A short orientation before the deeper research. Roles can overlap and can change as new evidence arrives.
Selection rule: inclusion does not imply an investment view or equal importance. A company is added, promoted or removed from this index when evidence changes its relevance to the architecture. Relationship labels remain evidence-specific: ownership/strategic, commercial/customer, ecosystem overlap, technical interoperability, or analytical hypothesis.
FOLLOW THE RELATIONSHIPS
Move from the company view into the evidence between companies
Individual company theses explain what each business controls, operates and may capture economically. The deep dives below examine the specific relationships, infrastructure links and ecosystem evidence that connect those companies to the wider system.
Follow the evidence behind selected company-to-company relationships, infrastructure links and control-plane questions. These are narrower studies than the company theses above.