Cover art for Top Chinese AI Models Use NVIDIA Blackwell GPUs

Top Chinese AI Models Use NVIDIA Blackwell GPUs

Chinese AI labs use NVIDIA Blackwell GPUs despite export restrictions, according to forensic analysis.

Nov 12, 2026

10 cards · 1m 20s · tap to begin

From · · 1m 20s

Top Chinese AI Models Use NVIDIA Blackwell GPUs

Chinese AI labs use NVIDIA Blackwell GPUs despite export restrictions, according to forensic analysis.

In brief

Chinese AI labs use NVIDIA Blackwell GPUs despite export restrictions, according to forensic analysis. Six Chinese AI models were initialized on export-controlled NVIDIA Blackwell GPUs, suggesting access to restricted hardware. Originally reported by Unemployment Labs.

Forensic Evidence

Table of ten models from DeepSeek, Moonshot AI, MiniMax and Ant Group. Each lab’s earlier model was created on NVIDIA Hopper and its newest models on NVIDIA Bla

Analysis of six Chinese AI models shows numerical patterns matching NVIDIA Blackwell GPUs despite export restrictions.

“Our forensic analysis of publicly released AI models finds evidence that DeepSeek, Moonshot AI, MiniMax and Ant Group have gained access to NVIDIA’s Blackwell GPUs”

Hardware Fingerprint

Figure 1

Model weights reveal 148 or 152 processor cores matching Blackwell B200/B300 or GB200/GB300 chips.

“DeepSeek-V4-Flash and MiniMax-M3 match NVIDIA’s Blackwell-class B200 or B300 chips, which have 148 processor cores”

Timeline of Models

Figure 2

DeepSeek models switched from Hopper to Blackwell in April 2026; Moonshot AI followed in July 2026.

“DeepSeek’s models read as Hopper-trained through DeepSeek-V3 (released in December 2024), and as Blackwell from DeepSeek V4 (released in April 2026)”

Six Models Affected

Figure 3

DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ling-3.1-flash use Blackwell GPUs.

“Six models from four Chinese labs read as Blackwell-trained”

Training Scale

Figure 4

Each model required 500 to 3,000 Blackwell GPUs for two to three months of training.

“From each model’s size and the amount of data it was trained on, we estimate that each run needed roughly 500 to 3,000 Blackwell GPUs”

AMD Alternative

Figure 5

AMD MI300X could theoretically match Blackwell patterns but evidence shows NVIDIA usage.

“On a physical MI300X, DeepSeek-V4-Flash’s preserved weights point to NVIDIA: in the 377 values where the two vendors’ results differ, the published weight matches NVIDIA’s in 363 and AMD’s in none”

Statistical Evidence

Evidence ranges from 49 to 1,385 sigma, far stronger than random chance.

“The strongest evidence for each of the six models ranges from 49 to 1,385 sigma”

U.S. Export Controls

Blackwell GPUs require U.S. export licenses for China; none have been reported.

“• Requires a U.S. license for export to China; none have been reported”

Verification

Figure 6

Reproduced weights match Blackwell patterns on NVIDIA hardware without modification.

“On the B300, unmodified PyTorch reproduced 461 of the 469 selected rows in DeepSeek-V4-Flash with correlations above 0.99”

Conclusion

Strong evidence suggests six Chinese AI models were initialized on export-controlled NVIDIA hardware.

“Based on our forensic investigation, we believe that we have strong evidence that six Chinese AI models were initialized on export-controlled NVIDIA hardware”

The one thing

Six Chinese AI models were initialized on export-controlled NVIDIA Blackwell GPUs, suggesting access to restricted hardware.

Read the original on Unemployment Labs

How we make summaries

React

Sign in to react and comment.

Comments (0)

Life is short. Keep it sweet. Respect others' opinions and be kind!

    Recommended next

    More decks on nvidia blackwell and related topics.