Integration is more visible; economic productivity is the harder proof.
Ulanqab makes the compute–grid dependency tangible, while Horinger connects a named domestic accelerator design with telecom infrastructure. The October AI export-control proposal also clarifies that internationally delivered AI services need not entail transferring the underlying technology. These developments strengthen the system thesis without proving national workload fungibility or attractive returns.
What would a connected AI infrastructure system actually look like?
The thesis is not that China is building one giant computer. It is that optimisation may increasingly happen across several nested layers — from the accelerator and cluster to the data-centre campus, regional resource pools and compute–energy coordination.
Important: the arrows describe the architecture being tested, not proof that every layer is already interoperable or economically fungible. See how we test the evidence →
What problem is China trying to solve?
AI demand can expand faster than the supply of frontier components, powered data-centre capacity and economically usable compute. The central question is therefore not simply how many chips or megawatts China can build, but how effectively the whole stack converts constrained inputs into completed AI work.
Why isn't building more chips and data centres enough?
A nominal accelerator, rack or megawatt only becomes valuable when software can use it, networks can feed it, memory can keep it productive, a data centre can power and cool it, and customers actually run workloads on it. This is why the thesis follows the chain from technology availability → powered infrastructure → useful compute → utilisation → economics.
What would genuinely connected infrastructure change?
At its strongest, system-level optimisation could allow workloads to be placed against the combination of compute, network capacity, latency and electricity that best fits them. That requires more than physical interconnection: resource visibility, compatible software, scheduling, settlement and workload portability all matter.
What suggests this is moving beyond policy ambition?
The evidence base has moved from broad national interconnection policy toward named provincial and cross-provincial mechanisms, operating resource pools and compute–energy experiments. The latest work strengthens the case for an emerging coordination layer, while stopping short of claiming that China already has a seamless national compute market.
Coordination is becoming more operational — but connection is not the same as fungibility.
Recent evidence adds operating cross-provincial aggregation and provisioning alongside provincial implementation and compute–electricity coordination. This advances the thesis from ambition toward execution; it does not establish frictionless workload movement or attractive economics.
Read the dated supporting evidence →A western park has cheap electricity and idle compute. Why can't eastern AI demand simply use it?
Because connected compute is not automatically usable compute. The July field evidence is valuable precisely because it shows the frictions the national architecture must overcome.
This is why stronger evidence of national coordination can coexist with weak utilisation or stranded capacity in individual projects. See the coordination evidence and constraints →
Why might China's route differ from the United States?
The comparison is useful as a mechanism, not a scorecard. US AI infrastructure is heavily assembled through large company-led estates and commercial ecosystems. China's architecture adds a stronger layer of national and provincial infrastructure coordination around company systems — including public compute networks, East Data–West Compute geography and energy policy.
Neither model is uniform, and the existence of coordination policy does not prove superior technical or economic outcomes.
Where could economic value emerge — if the architecture works?
System constraints can redirect value toward whichever layer controls a scarce, difficult-to-replicate capability: useful compute, orchestration, network bandwidth, powered AIDC capacity, grid access, cooling, storage or financing. But strategic relevance is not the same as shareholder economics.
That is why company pages are tested separately from the architecture thesis. A development can strengthen the national-system case without changing the economics of VNET, GDS, Kingsoft Cloud or any other company. See the company theses →
What would weaken or falsify the thesis?
The thesis weakens if coordination remains largely administrative rather than operational; heterogeneous compute cannot be made usefully schedulable; cross-region workloads remain too constrained by latency, data rules or settlement; powered capacity persistently outruns demand; or system integration fails to improve utilisation and economic returns.