SuperPoD
Huawei's term for a tightly connected group of AI processors and supporting components designed to operate more like one large compute system. It is a local/system-scale concept, not the same thing as a nationwide compute network.
AIDC
AI data centre: infrastructure designed for dense AI computing, where power delivery, liquid cooling, networking and rack design can matter as much as floor space.
East Data–West Compute
China's programme to develop major computing hubs and move suitable workloads toward regions with land, energy and renewable-resource advantages, while keeping latency-sensitive work closer to users where needed.
Control plane / orchestration
The software layer that discovers resources and decides where and how workloads run. In this thesis it can range from scheduling inside a cluster to coordinating capacity across sites or providers.
KV cache
Stored key/value information from a model's attention calculations. Reusing it can avoid recomputing earlier context, but large or long-running workloads can make memory capacity and movement increasingly important.
HBM
High Bandwidth Memory: very fast memory placed close to AI accelerators. Supply, packaging and bandwidth can constrain how much useful performance an accelerator system actually delivers.
Heterogeneous compute
A computing environment containing different processor types or accelerator families. Software compatibility and scheduling become important if workloads are to move efficiently between them.
Useful system throughput
Our analytical focus on completed, usable AI work rather than headline chip specifications or installed FLOPS alone. It asks what the whole system delivers after memory, networking, software, reliability, power and utilisation are considered.
Compute–energy coordination
Coordinating when or where suitable computing workloads run with electricity availability, grid conditions or storage. Evidence of pilots does not automatically prove large-scale commercial optimisation.
Token
A unit used by language models to process and generate text or other encoded information. Token volume can be a useful demand indicator, but tokens alone do not measure profitability or total infrastructure efficiency.