The thesis is recursive, not sequential.
Each question produces a provisional answer. Later evidence is required to propagate backwards and amend earlier conclusions. We distinguish global AI trends from genuinely China-specific divergence, and architecture from economic proof.
Observe
Record primary evidence, company claims, standards, deployments and measurable operating data without forcing a narrative.
Classify
Separate observed fact, company claim, external analysis, inference and thesis-to-test.
Connect
Map the evidence across architecture layers and companies. Convergence is not automatically causality.
Back-propagate
Ask whether new evidence changes earlier questions. Preserve the revision rather than silently overwriting it.
Falsify economically
Engineering scale is not enough. Test utilisation, portability, reliability, watts, capex and ultimately RMB per useful token.
Evidence is the substrate, not a competing navigation layer.
Evidence now surfaces primarily inside company and relationship drill-downs. The full ledger remains available for audit, but the main experience starts from the system or an interaction and lets the reader inspect why the edge exists, what it means and what remains unknown.
What survived when architecture met the economics test?
The architecture survives Q10, but the claim becomes narrower. System optimisation can plausibly mitigate component constraints through utilisation, memory hierarchy, workload placement and abstraction. We still do not have independent evidence that the complete Chinese system generally beats a frontier Nvidia/CUDA system on total cost per useful task.
Not one giant computer. The more plausible architecture is heterogeneous homogeneous islands connected by an increasingly capable control plane.
Software + orchestration move toward the centre. Physical capacity matters only insofar as workloads can discover, access and productively use it.
Interconnect links almost every thesis layer: SuperPods, external context/storage, distributed inference, geographic scheduling and national resource pooling.
Policy, standards and engineering demonstrations increase viability. Only utilisation, revenue, reliability and economics establish value.
Q1–Q10 now form one thesis. Economics remains the governing test.
Abstraction is real and improving. Full mixed-vendor fungibility is neither established nor necessary; schedulable resource pools are the more plausible target.
Compute-energy coordination and workload geography are real directions, but dynamic hyperscaler arbitrage remains economically unproven.
Commercial inference, MaaS, agents and production workloads validate the need for the architecture faster than they validate a completed national fabric.
Architecture can improve utilisation and reduce dependence on scarce resources, but system-level cost superiority is not established. The key comparison is total system cost per useful completed task.