Skip to main content

Checkpoint Forensics: Tracing the Chips Behind AI Models

A new report uses checkpoint forensics to claim six Chinese models carry the fingerprint of export-controlled Blackwell chips. Here is what it shows and what it can't prove.

By Mohammad Muneer Ahmed
Published: Nov 16, 2026
4 min read
👁️ 75 Unique Views
Checkpoint Forensics: Tracing the Chips Behind AI Models
Premium Insight

Why It Matters

India relies on global AI models and on imported chips, so rules about who can buy advanced hardware affect Indian startups and researchers too. If tools like checkpoint forensics make export controls easier to check, chip rules could get tighter or more closely watched worldwide.

What the Report Says

On November 12, 2026, Unemployment Labs published a report about a method it calls checkpoint forensics. The lab is a nonprofit led by Will DePue, who used to work at OpenAI. The report says six Chinese AI models have patterns that match Nvidia’s Blackwell chips. Those chips need a US license to be exported to China. The report doesn’t prove that anyone broke the law.

What Happened

The six models are DeepSeek-V4-Flash and V4-Pro, Kimi K3 and K3.1-Pro from Moonshot AI, MiniMax-M3, and Ant Group’s Ling-3.1-flash. They all came out since April 2026. Older models from the same labs, like DeepSeek-V3, match the older Hopper chips.

How Checkpoint Forensics Works

Every AI model starts out with random numbers, and those are made on a chip. DePue’s idea is that a model’s published weights can keep a trace of the chip layout that made those starting numbers. It’s kind of like a fingerprint. He says he searched more than six trillion combinations of hardware layouts and random seeds to find matches. Blackwell chips have layouts with 148 or 152 streaming multiprocessors, and that helps their signature stand out.

What It Can and Can’t Show

This only shows where a training run started. It doesn’t prove the whole model was trained on Blackwell. The study also hasn’t been checked by independent experts. Some models can start on regular processors, which leave no fingerprint, so a missing signal doesn’t prove a lab used older chips. Chips can also reach China in other ways, like cloud access in other countries or stock bought before the ban, and the report doesn’t say which happened. Trending Topics says the labs haven’t commented. It also says Nvidia earlier said claims about chips being sent to China were investigated and proven wrong.

State of the Art: A New Kind of Evidence

In past export-control cases, people mostly used purchase records, leaks, or tips. This method reads the model itself. DePue also put a verification repository on GitHub with weight rows and scripts for five models, so other people can try to repeat his results.

Future Potential and Applications

If other researchers confirm the method works, regulators and auditors could use it to check claims about what hardware built a model. It might also help figure out where a model came from. But labs could try to hide the signal later, for example by changing how they start their models. That’s just my own guess.

Future Impact and Business Opportunities

A tool like this could change the export-control debate, because it gives lawmakers something they can measure. Chipmakers and cloud companies might feel pressure to check their customers more carefully. It could also create a market for companies that audit AI models. All of this depends on whether the method holds up.

Found this analysis insightful?

Share with colleagues, engineers, and your network.

Link copied to clipboard!