Study claims six model weights, including DeepSeek and Moonshot, carry NVIDIA Blackwell chip fingerprints
Will DePue, a researcher at Unemployment Labs, released a technical investigation report on November 12, claiming that digital features matching NVIDIA's Blackwell architecture appear in six model weights released in recent years by DeepSeek, Moonshot, MiniMax and Ant Group, while earlier models mainly exhibited Hopper architecture features. The study exploits differences in the initial weights generated by PyTorch's Philox pseudorandom number algorithm on GPUs with different numbers of SMs to reverse-identify the random seed and hardware layout, and reproduced the corresponding weights of DeepSeek V3 and V4-Flash on H100 and B300. Based on this, the report estimates that some training clusters require hundreds to thousands of Blackwell GPUs, but the authors point out that the fingerprint only reflects the initialization hardware and cannot prove the chip's location or any violation of export controls.
Why read this
AI score 78
The study uses pseudorandom initialization remnants in model weights to infer the GPU architecture used for training, and provides a reproducible verification process and applicable boundaries.
Topics
Body · AI translation
On November 12, Will DePue, a researcher at Unemployment Labs, published a technical investigation report analyzing the model weights publicly released in recent years by DeepSeek, Moonshot AI, MiniMax, and Ant Group. The study found that six new models from these four companies all exhibit numerical signatures matching NVIDIA's Blackwell architecture, whereas their previously released models mainly showed signatures of the previous-generation Hopper architecture.
AI models are composed of billions or even hundreds of billions of adjustable numerical parameters, known as model weights. Before formal training begins, all weights must first be assigned a set of initial values. The researchers found that these initial values are not entirely random, but are generated by specific pseudorandom number generation algorithms, and the way they are generated and assigned is closely tied to GPU hardware architecture.
The PyTorch framework used in the study initializes model weights using the Philox pseudorandom number generation algorithm, and distributes computational tasks across a large number of parallel threads inside the GPU via NVIDIA's cuRAND library. Different GPU models have different numbers of streaming multiprocessors (SMs) and use corresponding thread layouts, so when the same random seed is run on different GPU architectures, the generated numbers are assigned to different weight positions.
In other words, even if two GPUs use the same random seed, the final arrangement of initial weights will differ because the number of internal compute units differs. This difference is like a hardware fingerprint hidden within the model.
Under normal circumstances, after a model undergoes extensive training, the initial weights are continuously adjusted, and the original arrangement signatures may gradually disappear. However, the researchers found that some weights barely change during training. For example, certain parameters in the model used to represent rare words or symbols may retain their initial values when the training data does not cover such content. In addition, even if some weights have changed, a weak statistical correlation may still remain between the initial values and the final values.
Using these residual signatures, the research team worked backward to find the random seed and GPU thread layout originally used, and then compared the computed results with the public model weights. To avoid accidental matches, the researchers also held out some weights as an independent validation sample; after determining the random seed and hardware configuration, they performed cross-validation using data that had not participated in the search.
The researchers noted that NVIDIA's Hopper architecture H100, H200, and H800 have 132 SMs, while Blackwell B200 and B300 have 148 SMs, and GB200 and GB300, which use the Grace-Blackwell design, have 152 SMs. By analyzing the numerical arrangement in model weights, it is possible to determine to some extent which type of GPU was used during initialization.
The investigation found that DeepSeek's V3 model, released in December 2024, still shows Hopper architecture signatures, but the V4 series released in April 2026 already shows Blackwell signatures. Moonshot AI's earlier Kimi K2.5 still matches the Hopper architecture, while Kimi K3, released in July 2026, matches GB200 or GB300. MiniMax showed a similar change from M2.7 to M3, and Ant Group's Ling series shifted from Hopper signatures in Ling-3.0-flash to Blackwell signatures in Ling-3.1-flash.
Among them, DeepSeek V4-Flash and MiniMax M3 match the 148-SM Blackwell layout, while the other four models match the 152-SM Grace-Blackwell layout.
To further verify the findings, DePue also reproduced the results on real NVIDIA GPUs. He used one Hopper architecture H100 and one Blackwell architecture B300, ran unmodified PyTorch programs, and tested with the DeepSeek random seed 953049583 identified in the study.
The results showed that the H100 could accurately reproduce the initial weights retained in DeepSeek V3, but could not reproduce V4-Flash; the B300 was exactly the opposite, able to reproduce the relevant weights in V4-Flash. In the B300 test, 461 of the 469 selected weight rows achieved a correlation above 0.99, of which 445 achieved bit-for-bit exact matches.
The researchers also verified Moonshot AI's Kimi K3 on a rented GB200 server, using the identified random seed 20256586, and successfully reproduced the 13 sets of initial weights retained in the model. After changing the random seed by just one value, the match results disappeared entirely.
From a statistical perspective, the researchers said the strongest matching evidence for the six models reached 49 to 1385 standard deviations (Sigma), far exceeding the threshold of about 8 standard deviations they set to rule out accidental matches arising from large-scale searches. Even the weakest single test reached 20 standard deviations.
Training clusters may scale to thousands of GPUs
In addition to determining the chip architecture, the researchers also estimated the relevant training clusters based on model size, number of activated parameters, and number of training tokens.
The report concludes that the larger-scale training tasks among the six models may require hundreds to thousands of Blackwell GPUs, with a single training run lasting about two to three months. Among them, DeepSeek V4-Pro and Moonshot AI's Kimi K3 have the largest training scales, with best estimates requiring 2,000 to 2,500 and 2,000 to 3,000 Blackwell GPUs, respectively.
The best estimate for DeepSeek V4-Flash is 1,000 to 1,500 GPUs, MiniMax M3 is about 500 to 600, and Ant Group's Ling-3.1-flash is about 1,100 to 1,650.
The research report gives wider ranges for individual training scales. For example, DeepSeek V4-Flash may require 250 to 4,500 GPUs, and the estimated range for Kimi K3 is 1,000 to 8,000. The 3,000 to 4,500 figure highlighted in the Wccftech headline falls within the larger-scale scenarios of these estimate ranges, and is not a uniform cluster size that the researchers derived for all models.
The researchers also attempted to estimate the cost of this hardware. At a price of about $70,000 per Blackwell GPU, the largest training clusters could be worth $150 million to $200 million, and the total hardware cost of the training tasks associated with the six models exceeds $500 million. These estimates cover only the final training phase and do not include preliminary experiments, model development, or other computing resource investments.
Chip fingerprints do not prove that illegal exports occurred
Although this study provides new technical evidence, its conclusions still have important limits.
First, the fingerprints in the model weights directly reflect the hardware configuration used when the initial weights were generated, and cannot prove with 100% certainty that subsequent training was also completed on the same batch of GPUs. In theory, a development team could initialize a model on one type of hardware and then transfer the weights to other devices for training.
Second, the study cannot determine where these Blackwell chips are actually located, nor can it determine whether Chinese companies purchased the chips, rented computing resources through third-party data centers, or used other means to gain access.
In particular, overseas data center leasing may be an important explanation. Chinese companies could remotely use Blackwell GPUs in data centers outside China without having to import the chips directly into China. Recent reports that Chinese AI companies are booking overseas computing capacity have also drawn attention to this possibility.
The researchers also noted that specially configured AMD hardware could in theory also produce some similar features. However, the research team said it has ruled out this alternative explanation for DeepSeek V4-Flash and considers it very unlikely that the other five models used this approach.
In addition, this study is currently a technical investigation released by independent researchers and still needs to be independently reproduced by more external teams. The researchers said they have provided reproduction scripts to some reviewers, but the full public implementation details have yet to be released.
Therefore, this finding cannot directly prove that Chinese AI companies violated U.S. export control regulations. What it can support is the conclusion that the initial weights of the relevant models closely match the specific computing layout of Nvidia Blackwell GPUs, which indicates that the model development teams very likely had access to this type of hardware.
If subsequent research can independently reproduce these results, its significance may go far beyond this investigation of Chinese AI companies. Because model weights themselves are public, researchers may in the future be able to analyze public models to reverse-identify the hardware used for their initialization, and even track the computing platforms used by different AI labs at different stages.
This also means that AI models may not only be products of algorithms and training data, but may also leave in their digital parameters some kind of technical trace that can be used to trace their computing environment. For the United States, which is implementing advanced chip export controls, this method could become an entirely new technical forensics tool.