Imagine getting exposed not by a leak, not by a whistleblower, but by your own random numbers. That’s the energy of the most nerdy-cool AI story of the week, and honestly it slaps.

The setup

Nvidia’s top chips, the Blackwell generation, need a US license to be exported to China, and per the report none has been reported for these parts. Will DePue, a former OpenAI engineer who now runs a research consultancy called Unemployment Labs, published a report on November 12 claiming that six models from DeepSeek, Moonshot AI, MiniMax and Ant Group carry the fingerprint of Blackwell GPUs. The Register picked it up the next day. Plot twist energy.

The list, per the report: DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ling-3.1-flash, all released since April 2026. Earlier models from the same labs (DeepSeek-V3, Kimi K2, Kimi K2.5…) match Nvidia’s older Hopper chips instead. So the glow-up is visible in the weights.

OK but how do random numbers snitch?

This is the genius part, so let’s break it down:

  1. Before training, every weight in a neural network needs a starting value. Those values come from a pseudorandom number generator. In PyTorch on Nvidia hardware, that’s an algorithm called Philox, run through Nvidia’s cuRAND library.
  2. “Pseudorandom” means it’s not actually random: same seed, same numbers. Every time.
  3. The catch: the work is split across many GPU threads, and how it’s split depends on the exact chip, for example how many processor cores (streaming multiprocessors) it has. So the same seed on a different GPU gives a different pattern.
  4. DePue says traces of that starting pattern survive training and are still readable in the published open weights.
  5. He searched more than six trillion combinations of chip layout and seed, and found matches. Per the report, B200/B300 chips have 148 cores and GB200/GB300 have 152, and models line up with one or the other.

For two models, DeepSeek-V4-Flash and Kimi K3, the report says the match was also confirmed by re-running it on physical Blackwell chips. The code is public and, per the report, the check runs on a laptop in a few seconds. That’s peer review speedrun any%.

The fine print (read this before you tweet)

DePue is careful here, and so are we. Per his own report, the fingerprint shows how the weights were initialized, which supports but doesn’t prove that Blackwell did the training. It also does not establish where the chips were, how access was obtained, or whether any export control was broken. One obvious possibility, flagged by The Register: renting Blackwell through cloud providers outside China, a loophole a proposed US bill (the Cloud Security Act) aims to close. The report says specially configured AMD hardware is the only alternative it found, and rules it out for DeepSeek-V4-Flash.

So the honest headline is “strong technical evidence of Blackwell access”, not “caught smuggling”. Nuance is not cringe.

Meanwhile, the GPUs are literally visible online

This isn’t the only receipt. In October, security firm Lava reported finding 32 Nvidia B300 GPUs (Blackwell Ultra) through publicly exposed monitoring endpoints with IP addresses geolocated to China, out of 4,724 GPUs it spotted on China-geolocated IPs. Lava itself warns that IP geolocation isn’t definitive (VPNs and proxies exist) and that this doesn’t prove a violation.

The bubble-watch bonus stat from the same research: across its global sample, nearly 60% of GPUs reported 0% utilization at the moment they were observed. Lava stresses it’s a snapshot, not an industry average. Still. A whole lot of very expensive chips were vibing at idle.

Why this matters

  • Export controls vs. reality: The Register notes that at least four US bills aiming to limit China’s access to cutting-edge AI hardware are waiting in Congress. Reports like this are fuel for that fight.
  • Open weights are a double-edged sword: publishing weights is great for research and for you running models locally, but it also means anyone can do forensics on them.
  • The meta-lesson: a “random” seed is deterministic. If you’ve ever generated an AI image with the same seed twice and gotten the same picture, congrats, you already understand the core of this investigation.

Respect to whoever looked at a pile of floating-point numbers and thought “I bet these remember their GPU”. Receipts below. 👇