● ResearchNov 13, 2026 · 12:518 sources

Researcher Will DePue Defends Nvidia Blackwell Forensics, Says DeepSeek V4.1 Chip Remains Unknown

From marketbrief.now

AI;DR

Researcher Will DePue stated that a search of 4.3 billion random seeds across GPU layouts returned zero matches on DeepSeek's September V4.1-Flash release. DePue argued that returning zero matches across more than 6 trillion combinations proves the technique does not produce false Blackwell hits by chance. His forensic method matched the exact disclosed training hardware across all 24 open-weights models where developers stated what chips they used.

Read originalSign in to suggest
Image: marketbrief.now

Sources

+ 5 more
  1. SUPPORT@dummerspast39“His own report, buried in the appendix: "We recovered no signature from DeepSeek-V4.1-Flash (after an exhaustive search)"”x.com
  2. SUPPORT@nrehiew_“Even if weights are changed during training, their relative ordering from initialization should be preserved and you can reverse-engineer the SM count of the GPU that produced the initial weights”x.com
  3. SUPPORT@kyleichan“It says DeepSeek, Moonshot, MiniMax have switched to Blackwells while Zai, Tencent, Alibaba have not. Meituan LongCat also not (they’ve said they trained on Chinese AI chips).”x.com
  4. SUPPORT@teortaxestex“the interesting bit is that this implies they all have functioning Blackwell training stacks.”x.com
  5. SUPPORT@michaeljburry“This is no surprise to analysts who years ago questioned SE Asia/Singapore’s need for so many advanced $NVDA chips.”x.com
#deepseek#llmScore 0.31.8k tokensMarkdown
Read original