Key points
- Six models released since April 2026 contain numerical patterns matching NVIDIA Blackwell GPUs, according to the published report
- The analysis excludes specially configured AMD hardware as the source for these patterns in five of the six cases.
- The findings do not establish where the chips were located or whether export controls were violated, the report states.
Unemployment Labs published a forensic analysis claiming that six AI models from Chinese developers carry signatures of NVIDIA Blackwell GPUs, which are restricted for export to China. The report examines publicly released weights from DeepSeek, Moonshot AI, MiniMax and Ant Group to identify hardware fingerprints embedded during creation.Researchers found that six AI models from Chinese companies contain data patterns linked to restricted NVIDIA chips.Forensic analysis of model weights reveals numerical signatures consistent with NVIDIA Blackwell hardware in six recent Chinese AI releases.
The claim
The analysis identifies numerical patterns in model weights that match specific Blackwell hardware configurations. DeepSeek-V4-Flash and MiniMax-M3 align with B200 or B300 chips featuring 148 processor cores. DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash match GB200 or GB300 chips with 152 cores. Researchers confirmed matches on physical Blackwell chips for DeepSeek-V4-Flash and Kimi K3.
The response
U.S. officials and press outlets had previously alleged that DeepSeek and Moonshot AI used Blackwell chips, but they did not publish technical evidence. Unemployment Labs states their findings provide technical evidence of access to advanced computing hardware, filling a gap in prior reporting that lacked such data.
What the evidence means
The fingerprint identifies how weights were created and supports, without proving, that Blackwell performed the training. The researchers excluded specially configured AMD hardware as the source for DeepSeek-V4-Flash and consider it very unlikely for the other five models.reported The report explicitly states it does not establish chip location, access methods or export control violations.
Evidence: reported
Chinese frontier AI models carry the fingerprint of export-controlled GPUsaifleet-hn-ai
In practice
Enterprise IT buyers cannot infer supply chain compliance or hardware provenance from these models based on this report. The analysis relies on numerical patterns in published weights, not leaks or insider accounts. Users should note that earlier models from these labs matched older Hopper generation hardware, indicating a shift in training infrastructure since April 2026.reported
Evidence: reported
Chinese frontier AI models carry the fingerprint of export-controlled GPUsaifleet-hn-ai
Why it matters
- Buyers assessing Chinese AI vendors now have technical indicators of Blackwell hardware use, rather than relying solely on unverified press reports.
- The exclusion of AMD hardware as an alternative source strengthens the case for NVIDIA chip usage in these specific model releases.
- The inability to determine chip location or export violation status means compliance risks remain unquantified for procurement teams.
- The distinction between Hopper and Blackwell fingerprints allows buyers to track hardware generation shifts in competitor model releases.
Topics: AI models, NVIDIA, hardware forensics, export controls, DeepSeek
Updated
Then / now / next
Earlier models from these Chinese labs matched NVIDIA’s older Hopper generation hardware.
Six models released since April 2026 contain numerical patterns matching Blackwell-class chips.
Watch for further technical evidence regarding chip location or export license status.
Story timeline
- Then Earlier models from these Chinese labs matched NVIDIA’s older Hopper generation hardware.
- 12 Nov 2026 Chinese frontier AI models carry the fingerprint of export-controlled GPUs
- Now Six models released since April 2026 contain numerical patterns matching Blackwell-class chips.
- Next Watch for further technical evidence regarding chip location or export license status.
For readers
Business
- Technical evidence now exists for Blackwell hardware use in Chinese AI models.
- Compliance status remains unknown as chip location and access methods are unestablished.
Researchers
- Weights from six models match Blackwell hardware patterns since April 2026.
- AMD hardware was excluded as the source for these patterns in five cases.
Policy
- Findings do not establish whether export controls were violated.
- U.S. officials had previously alleged Blackwell use without technical evidence.
Ask the article
Local keyword search of this article; no generated answers.Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
Which specific models show Blackwell signatures?
DeepSeek-V4-Flash, MiniMax-M3, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash show patterns matching Blackwell hardware.
Does this prove export controls were violated?
No, the report states the findings do not establish whether export controls were violated or how access was obtained.
Can AMD hardware produce these same patterns?
Researchers excluded AMD hardware for DeepSeek-V4-Flash and consider it very unlikely for the other five models.
Connected stories
Reading receipt
- Six models released since April 2026 contain numerical patterns matching NVIDIA Blackwell GPUs, according to the published report
- The analysis excludes specially configured AMD hardware as the source for these patterns in five of the six cases.
- The findings do not establish where the chips were located or whether export controls were violated, the report states.
Still open: Where were the Blackwell chips located during model training?
Still open: How did these companies obtain access to the restricted hardware?