From individual chips to a national AI infrastructure system.
Use this only if you want the guided version. You can read the thesis, news flow and company research without completing any lessons. Here, the same ideas are unpacked step by step.
Build the system in seven steps
Read in order if this is new to you. If you already know a layer, jump directly to it.
How do thousands of AI processors work together?
Start with the basic scaling problem, then learn what Huawei means by a SuperPoD and why interconnect can determine useful performance.
Learn SuperPoD →Why can a fast AI chip still spend time waiting?
Learn how HBM, memory bandwidth and KV cache feed accelerators — and why compute performance is more than FLOPS.
Learn memory & HBM →What does AI change inside a data centre?
Connect the compute system to power density, cooling, fibre and usable MW. This is the physical AIDC layer.
Learn AIDC →Why is China building compute in different places?
Learn 东数西算 — East Data, West Computing: the geography of demand, land, energy and workloads that can tolerate distance.
Learn East Data–West Computing →How could separate compute resources become a more coordinated network?
Move from data-centre geography to 全国一体化算力网: visibility, interconnection, scheduling, settlement and the limits to true fungibility.
Learn the national compute network →Can computing workloads respond to where electricity is available?
Learn how portable workloads, schedulers, data centres, grids, storage and renewable generation can potentially interact.
Learn compute–energy coordination →What decides which computing resource does which job?
Finish with the control plane: the software problem of discovering, allocating and coordinating heterogeneous compute.
Learn orchestration →You now have the pieces. See the complete system.
The research thesis asks whether these layers can be coordinated well enough that China competes more effectively at the level of useful system output — despite important component constraints. The architecture page shows our current answer; the research page shows the evidence that changes it.
Learn first.
Interrogate second.
1. Understand the problem. We begin with the real-world constraint rather than jargon.
2. Learn the mechanism. Technical terminology is introduced once you have a mental model for it.
3. Examine the evidence. Policy documents, company disclosures and observed implementation are kept distinct from our inference.
4. Know the limits. Every lesson states what the evidence does not yet prove and points you toward the next concept.
Now interrogate a company announcement.
Practise separating reservations from delivered capacity, compatibility from orchestration, and policy connections from operating national coordination.