Earlier research and evidence history (2)
China is connecting its AI infrastructure. Who will control the demand?
Shanghai's new model-and-compute marketplace, Alibaba's model-access decisions and Kingsoft Cloud's multi-model services raise a common question: who connects AI capacity to paying customers—and who captures the value?
Read the full analysis →PREVIOUS RESEARCH · 8 OCTOBER 2026
China is building more AI capacity. Who will use it — and who will earn a return?
This week's findings point to a growing divide between older data centres with spare space and newer facilities built for demanding AI workloads. They also raise fresh questions about how DeepSeek might finance expansion, whether VNET can offer more than space and electricity, and how far distributed computing can work in practice. The 5 October research into VNET's ownership, ByteDance links and Qingyang plans remains part of that story.
Read the consolidated research → · Evidence archive → · Architecture implications →
Earlier research · retained as the 8 October baseline.
The architecture is advancing; its economics remain the test.
New evidence links national computing–power policy to a real electricity dependency at VNET’s Ulanqab campus, while Huawei and China Telecom are developing an Ascend 950DT project at Horinger. China’s proposed AI technology-export rules distinguish transferring controlled technology from selling AI services abroad. None establishes commercial returns on its own.
Explore the main thesis · View the dashboard · See VNET evidence
China cannot solve its AI infrastructure constraints simply by building more of the same.
The bottleneck is not one thing. Leading accelerators and HBM are constrained; electricity and demand are concentrated in different places; installed compute is not automatically usable compute; and a more varied domestic hardware base creates additional software and interoperability work.
The question behind this site: can China compensate for weaknesses in individual components by coordinating the system around them?
A different architecture may be emerging from those constraints.
The proposition is not that China is building one giant national computer. It is that workloads, cloud software, compute, data centres, geography and electricity may be coordinated more deliberately — at several very different physical scales.
Key distinction: tight coupling inside a SuperPoD is not the same as moving workloads between distant data centres. Latency, software, data rules and economics change at each scale.
The important qualification
Connected does not mean fungible. Latency, chip/software compatibility, data rules, network constraints and cross-provincial settlement can all stop apparently spare compute from being useful to a distant customer.
See the real-world constraint case →The idea is starting to show up in real infrastructure.
These examples operate at different scales, but they address the same underlying problem: turning constrained resources into more usable AI capacity.
Xiong’an New Area · Hebei
Xiong’an, a state-backed development zone in Hebei province, has launched a platform connecting computing resources across provincial boundaries. It is practical evidence that coordination is moving beyond national policy documents.
See the evidence and limits →Shanghai → western / inland compute
Official Shanghai reporting documents a coordinated data-centre event and a three-day electricity-price-aware inference test across multiple regions. It is stronger evidence that some workloads can respond to power conditions — without proving national-scale economics.
See what the evidence does — and does not — establish →Huawei SuperPoD
At the opposite physical scale, Huawei is tightly connecting large numbers of domestic AI accelerators. The aim is to improve the performance of the complete compute system rather than relying only on the capability of an individual chip.
Understand SuperPoD →Different scales, same underlying problem: getting more useful AI output from the resources China can obtain and operate at scale.
Coordination is becoming visible. Economic advantage is not.
Evidence increasingly supports parts of the architecture. It does not yet show that compute is freely interchangeable across regions, that utilisation will consistently improve, or that the resulting system will produce AI more economically.
Those are the tests that matter. The detailed dashboard tracks the supporting evidence, contradictions and next questions without treating announced capacity as economic success.
Open the full running thesis →Why VNET’s links to CATL and ByteDance deserve a closer look
What changed: the buyer structure behind VNET’s ~38.1% stake is more closely aligned with CATL founder-family capital than the headline “CATL-affiliated” label suggests. Green Alchemist holds 71% of PJ Millennium’s economic interest and its disclosed UBO, Zeng Yuntian (曾云天), is identified in Chinese prospectus evidence as the son of CATL founder and chairman Zeng Yuqun (Robin Zeng). CATL Hong Kong holds the remaining 29%.
Why it matters: this sits alongside direct ByteDance–VNET fibre/transmission procurement footprints and VNET’s 9 September agreement with Qingyang to develop a GW-scale compute–power coordinated base. Together they strengthen the strategic alignment, customer-infrastructure and expansion layers of the VNET thesis — without proving CATL project funding or the identity of every unnamed hyperscale order.
Read the three VNET thesis movers →Follow the part of the story that matters to you.
Weekly research features
One consequential question at a time, explained from primary evidence with the live thesis underneath it.
Browse research features →UNDERSTAND THE SYSTEMArchitecture & policy
How compute, software, data centres, geography and energy may fit together.
Explore the architecture →FOLLOW THE COMPANIESCompanies & relationships
Who supplies, operates, funds and uses the infrastructure — and where economic value may accrue.
Explore companies →TEST THE EVIDENCESources & open questions
Inspect the evidence, contradictions, methodology and unresolved research tests.
Test the case →FOLLOW THE THESISRunning dashboard
See what has strengthened, weakened or remained unresolved as new evidence arrives.
Open the dashboard →New to the technical concepts? Use the plain-English explainer library →