HW insider says initial Atlas-950 deliveries will be 1024-card version. Demand for 8192-card SuperNode is not there. Each 1024-card is ~1 EFLOPS FP8.
It will be available on HW Cloud in Aug & external customers in Q4.
DS likely buying 1024-card nodes & connecting them together
Each yr, new Ascend-card 2x in compute, so it can plug in more powerful compute as time goes on.
HW is also deploying Kunpeng-950 Supernode w/ unified memory pool that DS may be interested in.
It is deploying 16k card systems that both Scale-Up (interconnect within a SuperNode) & Scale-Out (traditional network connection bw SuperNodes)
Comparable to Nvidia's NVLink InfiniBand
Again, each hyperscaler that buys Nvidia racks still need their own engineers to build out large Data centers.
Not a surprise that DS will also need to do that here if it's seeking GW-sized Data centers running out of Inner Mongolia (home to major CSPs & EDWC clusters).
I would expect it to use the standard that HW provided for Scaling-up & Out, since that's what US hyperscalers do when they buy NVL72 & put those racks in clusters.
For an AI lab that does much of the stuff on its own, is it any surprise that it wants to build its own AI cluster?
DeepSeek is going heavy-asset.
On June 9, the company posted an opening for IDC planning engineers, a role explicitly scoped to the design and delivery of MW-to-GW scale infrastructure. It follows April's hiring of data center O&M engineers in Ulanqab, Inner Mongolia. Taken together, this is the first time DeepSeek has fully shown its hand on owning compute infrastructure rather than just renting it.