CHINA AI EVOLUTIONSearch
Choose how much time you have:2 minGive me the picture5–10 minWhat changed & whyCompany viewWho matters & economicsEvidenceSources & uncertaintyLearnTeach me the system
Home / Explore the six-layer value chain
CHINA AI INFRASTRUCTURE · LIVING SYSTEMS RESEARCH

Where is value created across China’s AI infrastructure stack?

AI infrastructure is not one market. Demand passes through software, compute, interconnect, data centres and energy before it becomes usable AI output. This guide shows what each layer contributes, which companies operate there and what has to happen for investment in that layer to create economic value.

22 SEPTEMBER 2026 · THESIS UPDATE

The link between compute, geography and energy is becoming operational

The new evidence is easier to interpret once the underlying layers are clear. The next section therefore reconstructs the system from demand down to power.

The Fujian–Ningxia pilot adds an execution example spanning the control plane, network, data centre and power layers. National monitoring and Xiong’an platform reach are separate evidence streams, with separate tests for commercial value.

Read the evidence and implications →
LAYER 01

AI demand & models

Demand originators and model ecosystems.

Explore →LAYER 02

Control plane

Cloud, runtime and workload scheduling.

Explore →LAYER 03

Compute / memory

Accelerators, DRAM, HBM and the storage hierarchy.

Explore →LAYER 04

Interconnect

Scale-up links, optical systems, wide-area networks and private networking.

Explore →LAYER 05

AIDC

Operators, facility technology suppliers and captive developers.

Explore →LAYER 06

Power & energy

Storage, power electronics, cooling and grid access.

Explore →
04 • VALUE-CHAIN SPINE

The site is about the system. Companies are evidence of how its interfaces are being built.

These interfaces are where the thesis becomes testable: a company matters when it provides evidence that two layers are actually being connected.

Every relationship is now classified first by the infrastructure problem it helps us observe. This keeps the project from becoming a collection of company profiles: architecture defines the question; corporate relationships provide evidence; economics determines whether the relationship matters.

AI DEMAND & MODELSDeepSeek • Qwen • ByteDance/Doubao • Xiaomi
→
CONTROL PLANECloud • MaaS • runtime • scheduler
→
COMPUTE / MEMORYAccelerators • HBM/DDR • storage hierarchy
→
INTERCONNECTScale-up • scale-out • WAN • optics
→
AIDCAI-ready MW • cooling • physical deployment
→
POWER & ENERGYHVDC • BESS • grid • green power
ByteDance is now inside the architecture, not outside it: ByteDance/Doubao supplies hyperscale AI demand; Volcano Engine is a control-plane/MaaS layer; Huawei, Cambricon and Iluvatar represent heterogeneous compute inputs; veRoCE/EthLink and ByteDance storage engineering sit in interconnect/data; and its Wuhu/Wuwei campuses connect the stack to physical AIDC and power. Open ByteDance ecosystem →
A

DEMAND × CONTROL PLANE

Why: convert application demand into schedulable, monetizable compute. Cases: Xiaomi × KC; ByteDance × Volcano Engine.

B

DEMAND × HETEROGENEOUS COMPUTE

Why: keep AI output scaling despite fragmented accelerator supply. Case: ByteDance × Huawei/Cambricon/Iluvatar.

C

COMPUTE × NETWORK

Why: make separated resources behave more like usable pools. Case: VNET × Huawei/DYXnet.

D

AIDC × ENERGY

Why: turn electricity into reliable AI-ready MW. Case: VNET × CATL.

E

ENERGY × POWER ELECTRONICS

Why: deliver high-density electrical architecture to AI racks. Case: CATL × Zhongheng.

F

ENERGY × CONTROL PLANE

Why: optimize the physical energy/computing system itself. Case: CATL × DeepCtrls.

Reading rule: company names tell us where to look. The value-chain interface tells us why we care. Evidence status tells us how much we are entitled to conclude.