Huawei's CloudMatrix 384 supernode in Inner Mongolia's Ulanqab Data Center ran the full-parameter post-training for DeepSeek V4 Pro, a 1.6-trillion-parameter MoE model released in April 2026 that supports 1M-token context. The June 2026 run on 384 Ascend 910C NPUs reportedly hit 30% utilization and finished without interruptions, though those specifics rest on X-sourced slides rather than broader independent confirmation.
it's strange that in the 950 generation Huawei is not advertising smaller clusters. SuperPOD is 8192 NPUs (though now we're hearing of an 1024-NPU slice). Also note: Inner Mongolia, Ulanqab DeepSeek is building there too, and no it's not Blackwells
Huawei Ascend CloudMatrix 384 Supernode: In-Depth Project Analysis
Link: https://chinaresearchcollective.substack.com/p/huawei-ascend-cloudmatrix-384-supernode?r=7m5sip
The setup pairs 384 Ascend 910C chips with 192 Kunpeng CPUs to deliver roughly 300 PFLOPS and 48 TB HBM, a footprint some analyses equate to more than 360 H100 servers. This scale invites closer looks at whether other labs will route later-stage training jobs onto comparable domestic clusters.
The 30% utilization and zero-interruption claims appear in recent commentary but lack cross-verified data so far. That gap keeps open how much further the same hardware might push on utilization or whether the run represents a one-off or repeatable pattern.
Users approve of Huawei's Ascend CloudMatrix 384 Supernode because the deployment makes sense as an upgrade path and the product is expected to sell strongly.
No Digg Deeper questions have been answered for this story yet.
it's strange that in the 950 generation Huawei is not advertising smaller clusters. SuperPOD is 8192 NPUs (though now we're hearing of an 1024-NPU slice). Also note: Inner Mongolia, Ulanqab DeepSeek is building there too, and no it's not Blackwells …or is this the same one?
Huawei Ascend CloudMatrix 384 Supernode: In-Depth Project Analysis
Link: https://chinaresearchcollective.substack.com/p/huawei-ascend-cloudmatrix-384-supernode?r=7m5sip
“In June 2026, the DeepSeek team successfully completed full-parameter post-training of the V4 Pro model on the CloudMatrix 384 supernode at Ulanqab” The first Mongolian Expert lives.
wait I'm wrong, they will upgrade CloudMatrix for 950 gen, this makes way more sense. This one will sell like pancakes.
wait I'm wrong, they will upgrade CloudMatrix for 950 gen, this makes way more sense. This one will sell like pancakes.
it's strange that in the 950 generation Huawei is not advertising smaller clusters. SuperPOD is 8192 NPUs (though now we're hearing of an 1024-NPU slice). Also note: Inner Mongolia, Ulanqab DeepSeek is building there too, and no it's not Blackwells …or is this the same one?
This is also better inference performance than what they were reporting for 950DTs at V4 launch.
on investment, so Capex is ≈150M RMB/0.4MW. That'd be 375B/GW, or $55B right now. Same as in the US, except in the US chips are much more performant per watt. And Capex is higher.
“In June 2026, the DeepSeek team successfully completed full-parameter post-training of the V4 Pro model on the CloudMatrix 384 supernode at Ulanqab” The first Mongolian Expert lives.
@teortaxesTex No, that's not true They are talking about the post training experiment done by Huawei and the university
“In June 2026, the DeepSeek team successfully completed full-parameter post-training of the V4 Pro model on the CloudMatrix 384 supernode at Ulanqab” The first Mongolian Expert lives.