Skip to main content

Report links six AI models to Blackwell initialization patterns

14 NOVEMBER 2026·2 MIN READ·4 SOURCES

A report from Unemployment Labs says six models released since April 2026 carry numerical patterns matching Blackwell hardware. The report says the fingerprint identifies the hardware used to initialize model weights, not necessarily every chip used for training.

Key takeaways · 4

  • 01

    Treat the fingerprint as evidence about weight initialization, not a complete record of training hardware.

  • 02

    The report names six models, released since April 2026, with patterns matching Blackwell hardware.

  • 03

    The reported GPU counts and costs are estimates that depend on assumptions the weights cannot confirm.

  • 04

    The report says its findings do not establish chip location, access method or an export-control violation.

What the report claims

Unemployment Labs says six models released since April 2026 contain numerical patterns matching Blackwell hardware[2]. The models are DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ling-3.1-flash[2]. The report says the fingerprint reveals how model weights were created by identifying traces of their initial values[2]. It does not claim that the fingerprint identifies every chip involved in subsequent training[3].

How the match was assessed

DePue says the method recovers traces of initial values, then searches for hardware and initial configurations that could reproduce them[2]. He says he searched roughly six trillion combinations of seed and chip layout[2]. The strongest evidence reported for the six models ranges from 49 to 1,385 sigma[2]. DePue says the method agreed with public hardware disclosures for all 24 control models he tested, and that he confirmed the Blackwell match for DeepSeek-V4-Flash and Kimi K3 on physical Blackwell chips[2].

Estimates and limitations

DePue estimates about 12,000 Blackwell GPUs across five runs, with best estimates of about 4,500 and 3,000 GPUs[2]. The report estimates the largest cluster would cost roughly $200 million to $300 million, and the five runs together more than half a billion dollars[2]. DePue says these GPU-count and cost estimates are uncertain and depend on assumptions that the weights cannot confirm[2]. The report says specially configured AMD hardware was the only alternative it identified, and that it ruled out that explanation for DeepSeek-V4-Flash[2].

Context and open questions

The report says the latest models it could read from Alibaba, Zhipu, Tencent, Xiaomi and StepFun still showed Hopper rather than Blackwell fingerprints[2]. Startup Fortune said the report names no reviewer and does not mention OpenAI or Anthropic[3]. It also said it found no comment from DeepSeek, Moonshot, MiniMax, Ant Group or Nvidia[3]. Earlier reporting had alleged Blackwell use by DeepSeek, which denied it, and Moonshot, which U.S. officials accused of buying Blackwell systems and renting time on foreign clouds, Startup Fortune said[3].

Independent comparison

Bojie Li says he scanned 265 open models and obtained a hardware reading for 185[4]. He says his analysis and DePue’s report identified the same GPU class for 65 models, while neither found a fingerprint in nine others[4]. These figures offer a comparison between two analyses, but they do not resolve the report’s stated limits on determining chip location, access or export-control compliance[2][4].

For teams evaluating model provenance, the key distinction is between evidence about initialization and a complete account of training. Treat estimates of hardware scale and cost as provisional, and avoid turning a technical fingerprint into a conclusion about chip access or compliance.

Why it matters
For teams

AI training for your team

kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.

How this developed

  1. 14 November 2026

    Report links six AI models to Blackwell initialization patterns

Sources

AI fluency, one session a day, built for your work.