DeepSeek Weights Show Traces of NVIDIA Blackwell GPUs: 445 of 469 Rows Match on B300
Former OpenAI researcher Will Depue found a trace of NVIDIA Blackwell GPUs, whose shipments to China are restricted by the US, in the weights of DeepSeek-V4-Flash. Before training, PyTorch fills the weights with random numbers, and their layout depends on the GPU model, much like the order in which a deck is dealt. On a B300 (a Blackwell chip), 445 of 469 weight rows matched the published values to the last bit, while the H100 produced no exact matches. This is evidence about how training started, not about all of the training.
AI-processed from TG: @ai_machinelearning_big_data (Machinelearning); edited by Hamidun News
Former OpenAI researcher Will Depue said that the open weights of Chinese models can reveal which GPUs they were initially trained on. In the weights of DeepSeek-V4-Flash, he found a signature of NVIDIA Blackwell, the latest generation of GPUs, whose shipments to China are restricted by the US. The idea is simple: before training, PyTorch fills the weights with starting random numbers directly on the GPU, and their layout depends on the GPU configuration. Depue took 469 rows from the published weights and reproduced them on a real NVIDIA B300 (a Blackwell chip). In 445 rows the numbers matched bit for bit, meaning down to the last bit, while the previous-generation H100 produced no exact matches.
How a GPU Leaves a Fingerprint in a Model
Export restrictions on NVIDIA's advanced chips have become one of the main levers of pressure on Chinese AI. But verifying which hardware was used to train a specific model is nearly impossible: companies either stay silent about their hardware or speak in general terms.
The fingerprint arises like this. Before training, the weights of a neural network are filled with "random" numbers, which are in fact pseudo-random. The layout of these numbers across a tensor depends on the GPU configuration. Imagine several people dealing the same deck, but each with their own dealing order. The deck is identical, but the cards in each hand come out differently.
In the published weights of DeepSeek-V4-Flash, some rows retained their original starting values: training did not erase them. Depue took 469 such rows and tried to reproduce them on real hardware. On the B300, 445 matched; on the H100, no matches were found for the same rows. From this he concludes that the initialization was done on a Blackwell-generation GPU.
What the Analysis of Six Models Showed
Depue checked the public weights of six models from DeepSeek, Moonshot AI, MiniMax and Ant Group. In all six he found traces of initialization on Blackwell. For earlier models from the same companies, the analysis pointed to Hopper, the previous generation.
The picture is logical: the fingerprint changes when the hardware changes. The method has a convenient property. Verification requires no access to data centers or internal documents. Public weights and access to the relevant GPUs for comparison are sufficient.
What This Result Does Not Prove
It is important not to read too much into it. The result applies only to the initialization stage, the very start of the model's path. Which GPUs were used in the remaining training stages, where they were located, and who owned them cannot be established by this analysis. The cards could have been rented, located outside China, or used one time only.
The match of 445 out of 469 rows is also not 100%, and the 469 rows were selected by the author himself. The original message contains no response from DeepSeek, Moonshot AI, MiniMax or Ant Group, and no independent verification is provided. Therefore, no legal conclusions about violating export rules can be drawn.
What This Means for Developers and Founders
First: public weights carry more information than it seems. They reveal not only architecture but also the hardware on which training began. If you publish your checkpoints, be aware that initialization can expose your infrastructure.
Second: chip export controls may gain a new audit tool. If other researchers confirm the method, regulators, analysts and competitors could apply it.
Third: for those who reproduce model training, this is a reminder that initialization results depend on the hardware. The same random seed on different GPUs can produce different starting weights. This matters for experiment reproducibility and for cross-checking results between clusters.
Bottom line: this is an interesting and testable clue, but not a verdict. It suggests that Blackwell likely participated in the start of training for six Chinese models. How the rest of the training proceeded remains unknown.
- Nov 21, 2026Chinese Tech Giants Double AI Spending to $139B, While Labs Recover Less Than 8%
- Nov 21, 2026Skein Agent Erases Finished Stages From Context: 97% Success Rate, But More Expensive Than OpenCode
- Nov 21, 2026SGLang with Humming Kernels Cut Time to First Token by 7–23% and Output Throughput by 17–34%
- Nov 21, 2026Open-Weight Inkling 2 Mini Matches Qwen3.8-27B on Independent Benchmark, Runs on 16 GB of Memory
Key Questions Answered
How a GPU Leaves a Fingerprint in a Model?
Export restrictions on NVIDIA's advanced chips have become one of the main levers of pressure on Chinese AI. But verifying which hardware was used to train a specific model is nearly impossible: companies either stay silent about their hardware or speak in general terms. The fingerprint arises like this. Before training, the weights of a neural network are filled with "random" numbers, which are in fact pseudo-random. The layout of these numbers across a tensor depends on the GPU configuration.
What the Analysis of Six Models Showed?
Depue checked the public weights of six models from DeepSeek, Moonshot AI, MiniMax and Ant Group. In all six he found traces of initialization on Blackwell. For earlier models from the same companies, the analysis pointed to Hopper, the previous generation. The picture is logical: the fingerprint changes when the hardware changes. The method has a convenient property. Verification requires no access to data centers or internal documents. Public weights and access to the relevant GPUs for comparison are sufficient.
What This Result Does Not Prove?
It is important not to read too much into it. The result applies only to the initialization stage, the very start of the model's path. Which GPUs were used in the remaining training stages, where they were located, and who owned them cannot be established by this analysis. The cards could have been rented, located outside China, or used one time only. The match of 445 out of 469 rows is also not 100%, and the 469 rows were selected by the author himself.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.