Will DePue of Unemployment Labs reported on 12 November 2026 that six AI models from four Chinese developers contain numerical patterns consistent with Nvidia's Blackwell hardware.
In the clearest test, a Blackwell B300 reproduced 445 complete rows of values from DeepSeek-V4-Flash bit for bit.
The patterns come from the way the models' starting values were generated before training. They offer a clue to what hardware may have been used for training, while leaving open where the chips were, how the developers accessed them and whether any US export rules were breached.
How 445 Rows Pointed to Blackwell
An AI model's weights are numbers that shape how it processes information. Before training, software uses a number called a seed to generate starting weights and distributes the work across a chip's processing units.
SITUATION DETECTED: Former OpenAI researcher Will DePue says he has found strong forensic evidence that Chinese AI labs DeepSeek, Moonshot AI, MiniMax, and Ant Group all trained their newest frontier models on NVIDIA Blackwell GPUs, despite U.S. export controls.
— MTS (@MTSlive) November 12, 2026
The hardware layout can affect where the resulting values appear, and traces of those values can survive training. DePue's team worked backwards from the published weights to identify combinations of seeds and processing layouts that could reproduce the patterns.
Using the same seed, an older Nvidia H100 reproduced preserved starting values from DeepSeek-V3 but failed to reproduce the corresponding pattern in V4-Flash.
The B300 matched 445 of 469 selected V4-Flash rows exactly, while a rented Nvidia GB200 reproduced all 13 preserved rows tested from Moonshot AI's Kimi K3.
The team released code and selected weights so others could repeat specific checks. Its results were also consistent with hardware disclosures for 24 models, including related models and files saved before training began.
The Six Models
DeepSeek-V4-Flash and MiniMax-M3 matched a layout associated with Nvidia's B200 or B300. DeepSeek-V4-Pro, Moonshot AI's Kimi K3 and Kimi K3.1-Pro, and Ant Group's Ling-3.1-flash matched a layout associated with the GB200 or GB300.
Kimi K3.1-Pro shares K3's starting seed and may be a version of K3 that underwent further training. The shared seed leaves open whether the two releases inherited their starting weights from the same run and how much of their training they shared.
A specially configured AMD MI300X reproduced both layouts in the team's tests. A comparison of individual values allowed the researchers to rule out that AMD configuration for V4-Flash, while the same comparison left the other five results unresolved.
What US Export Rules Cover
In February, a senior US official alleged that DeepSeek had trained an upcoming model on Blackwell chips, which the official believed were in Inner Mongolia. In July, three people familiar with the matter were reported as saying Moonshot had used Blackwell chips to train Kimi K3.
Nvidia has disclosed US licensing requirements for exports to China of advanced products, including B200 chips and GB200 NVL72 systems.
Determining which rules applied in any of these cases would require details about the supplier, the customer and where the hardware was located.
A developer could generate starting weights on one system and move them to another for training, or rent access to hardware in a data centre abroad. The weights alone cannot show which company supplied the computing power or whether any chips crossed a border.
