Hopium just a few of these would go a long way to accelerate DeepSeek's project
3:38 PM · Jun 4, 2026 · 5.6K Views
AI commentator teortaxesTex argues DeepSeek could accelerate training by adopting Huawei Atlas-950 SuperPoDs over its 2,048 NVIDIA H800s
The Atlas-950 features 8,192 NPUs interconnected via 800GE links.
--0--
Sentiment
Sentiment building, check back later.
Cluster Engagement
-
Views
-
Comments
-
Reposts
-
Bookmarks
Expand data
Posts from X
Most Activity
Most Activity
VIEWS13.7KBOOKMARKS18LIKES64RETWEETS1
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
Atlas 950 SuperPODs would put DeepSeek in possession of roughly Vera Rubin class hardware, in some ways stronger. Just one SuperCluster would let them go toe to toe with the frontier… Except at current production volume the whole China will have like 2, tops. A man can dream…
8hViews 13.7KLikes 64Bookmarks 18
REPLIES6
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
But seriously. Does anyone believe this? Eric Xu at Huawei Connect 2025 (Sep 18) announced Ascend SuperCluster by Q4'2026. It's a system occupying 64K m^2, with 524K NPUs, tocal 75.5 *Peta*bytes of "HiZQ 2.0" HBM (4 Tb/s). The total BOM is >5x of all CloudMatrix 384's sold.
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
Atlas 950 SuperPODs would put DeepSeek in possession of roughly Vera Rubin class hardware, in some ways stronger. Just one SuperCluster would let them go toe to toe with the frontier… Except at current production volume the whole China will have like 2, tops. A man can dream…
2hViews 3KLikes 20Bookmarks 7