JD Cloud Plans 100,000-Card GPU Cluster, First to Adopt Domestic Moore Threads Chips
nashnova research
JD Cloud announced plans to build a 100,000-card compute cluster using Moore Threads' domestic GPUs — the first time a homegrown GPU enters a leading cloud provider's core training infrastructure at this scale.
What is JD Cloud building?
JD Cloud said on September 9 it will use Moore Threads full-function GPUs to build a compute cluster scaled to 100,000 cards.
Full-function GPU — a chip that handles both training and inference — will power core workloads including model training.
This means → JD Cloud is not running a small pilot; it is placing a domestic GPU at the heart of its most performance-intensive work.
Why does this count as a milestone?
Prior deployments of domestic GPUs in ultra-large clusters have been very limited; none has reached the 100,000-card tier publicly.
JD Cloud describes this as the first time a domestic GPU enters a leading cloud provider's 100,000-card core compute cluster.
In plain terms = domestic GPUs have mostly appeared in small-scale trials; this is the first time a top-tier cloud provider plans to use one as a mainline resource at mega-cluster scale.
What is the key test ahead?
A plan is not a deployment — whether Moore Threads can sustain stable training loads at the 100,000-card level remains the central question.
This reflects the biggest uncertainty around domestic GPUs today: not "can they be made," but "can they run reliably at scale."
If the cluster goes live as planned, it will be Moore Threads' first publicly verified large-scale deployment inside a leading cloud provider.
市场有风险，内容仅供研究参考，不构成投资建议。