A former OpenAI researcher says the newest open models from DeepSeek, Moonshot AI, MiniMax and Ant Group carry a hidden fingerprint of Nvidia Blackwell GPUs, the chips the US bars from export to China.
Will DePue published the analysis, “checkpoint forensics”, on 12 November 2026. TechCrunch and The Decoder covered it on 13 November, and it has been one of the most discussed AI stories of the week.
This guide explains how a GPU can leave a trace inside model weights, which six models were flagged, what the report does and does not prove, and what it means if you use DeepSeek V4, Kimi K3 or MiniMax M3 yourself.
- The claim: six Chinese open-weight models released since April 2026 show a Blackwell “fingerprint” in their starting weights: DeepSeek V4-Flash, V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax M3 and Ant Group’s Ling-3.1-flash.
- How: PyTorch splits random weight initialisation across GPU cores. Hopper has 132 cores (SMs), Blackwell has 148 or 152, and that split survives in rarely trained weights.
- Strength: the report calls DeepSeek V4-Flash “conclusive” (1,385 sigma) and the other five “strong”. It matched all 24 control models whose hardware is public.
- What it does not prove: where the chips were, who owned them, or that any law was broken. Offshore data centres could explain it.
- For users: nothing changes in how the models work, their licences or their downloads today. Companies with strict compliance rules may want to note it.
What Was Published and When
Will DePue worked at OpenAI on Sora, Sora 2, GPT-4o, o3 and GPT-5, according to The Decoder. He now runs a nonprofit research group called Unemployment Labs, which hosts the report.
On X, DePue wrote that he had found “conclusive forensic evidence” that the top four Chinese AI labs built their newest models on Blackwell chips. He also said senior researchers at OpenAI and Anthropic checked his findings independently (as reported by The Decoder).
| Date (2026) | Event | Source |
|---|---|---|
| 12 Nov | “Checkpoint forensics” report published; compute estimates revised the same day | Unemployment Labs |
| 12-13 Nov | DePue posts the result on X | X (@willdepue) |
| 13 Nov | TechCrunch, The Decoder, SFist and others report it | TechCrunch, The Decoder |
| 14 Nov | Follow-up coverage lists open questions (exact GPU, chip count, location) | Shattered |
| 15 Nov | Bloomberg reports Treasury Secretary Scott Bessent saying Nvidia needs a “dramatic shake-up” | Bloomberg (headline) |
How a GPU Leaves a Fingerprint in Model Weights
This is the clever part, and it needs no inside information. It only uses the weights that the labs published on Hugging Face.
Before training starts, every weight in a model is set to a random number. PyTorch makes those numbers with the Philox random number generator through Nvidia’s cuRAND library.
The GPU splits that job across its streaming multiprocessors (SMs). The number of SMs differs by chip, so the order of the random numbers carries a pattern that points to the chip family.
- Find weights that barely moved. Some weight slots, such as embeddings for tokens that almost never appear, change very little during training. A faint copy of the starting values survives there.
- Search for the seed and the core layout. The team tested about six trillion combinations of random seed and SM count to find one that reproduces those surviving values.
- Test on held-out weights. A match is only counted if the same seed and layout also predict weights that were not used in the search.
- Score the result in sigma. With so many combinations tried, the report treats about 8 sigma as what chance alone could produce. The flagged models score from 49 to 1,385 sigma on their strongest test.
- Reproduce on real hardware. Unmodified PyTorch reproduced DeepSeek V4-Flash’s preserved weights on a B300, DeepSeek V3’s on an H100, and Kimi K3’s on a rented GB200.
| SM count found | GPU family | Export status to China |
|---|---|---|
| 108 | A100 (Ampere) | Restricted |
| 132 | Hopper: H100, H200, H800 (cannot tell them apart) | H800 was the China version; H200 now allowed under a licence route |
| 148 | Blackwell: B200, B300 | Banned |
| 152 | Blackwell: GB200, GB300 | Banned |
Note: AMD’s MI300X can be set up to mimic 148 or 152 units. The report rules AMD out for V4-Flash, where 363 of 377 differing values match Nvidia and none match AMD, and calls AMD “very unlikely” for the other five.
Want to know what separates these chips in memory, FP4 speed and price? Our H100 vs H200 vs B200 vs B300 vs Rubin comparison covers every spec.
The Six Flagged Models (Full Table)
These are the report’s six main findings. Sigma values are the strongest single test for each model, as reported by Unemployment Labs.
| Lab | Model | Released | Layout found | Likely chips | Evidence |
|---|---|---|---|---|---|
| DeepSeek | V4-Flash | Apr 2026 | 148 SMs | B200 / B300 | 1,385σ, “conclusive” |
| DeepSeek | V4-Pro | Apr 2026 | 152 SMs | GB200 / GB300 | 49σ, “strong” |
| Moonshot AI | Kimi K3 | Jul 2026 | 152 SMs | GB200 / GB300 | 305σ, “strong” |
| Moonshot AI | Kimi K3.1-Pro | Nov 2026 | 152 SMs | GB200 / GB300 | 305σ, same seed as K3 |
| MiniMax | M3 | Jun 2026 | 148 SMs | B200 / B300 | 323σ, “strong” |
| Ant Group | Ling-3.1-flash | Oct 2026 | 152 SMs | GB200 / GB300 | 244σ, “strong” |
The weakest individual test across the six still scored 20 sigma, well above the 8 sigma chance line. Kimi K3.1-Pro sharing K3’s seed suggests it continued from the same starting checkpoint.
If you run any of these at home, our guides cover the hardware: how much VRAM Kimi K3 and K3.1 Pro need and how to run DeepSeek locally.
Which Chinese Models Showed Only Hopper
The report is not a blanket accusation. Many Chinese models read as Hopper (132 SMs), and some show no GPU signal at all.
| Result | Models |
|---|---|
| Hopper (132 SMs) | DeepSeek V3 (Dec 2024), Kimi K2 and K2.5, MiniMax-M2.x, Ling-3.0-flash, and models from Alibaba (Qwen), Zhipu (GLM), Tencent, Xiaomi and StepFun |
| Mixed | Kimi K2.7-Code: mostly Hopper with a 148-SM part; not counted among the six |
| No GPU fingerprint | Meituan LongCat-2.0 (Meituan says “AI ASIC superpods”), DeepSeek V4.1-Flash |
| No usable signal | CPU-initialised Qwen3 models, PaddlePaddle-based ERNIE-5.5 |
Tip: a Hopper reading cannot separate the legal H800 from the H100 or H200. It also does not rule out Blackwell being used later in training.
Interesting detail for local users: DeepSeek V4.1-Flash, the version many people run on a 128 GB PC, showed no fingerprint. Our V4.1 Flash local guide covers its GGUF sizes.
How Many Blackwell GPUs? The Compute Estimates
DePue also estimated cluster sizes. He labels these low confidence, because they depend on token counts, run lengths and utilisation that the labs never published.
| Training run | Blackwell GPUs (est.) | H100 equivalent | Run length |
|---|---|---|---|
| Kimi K3 | ~4,500 | ~11,500 | ~70 days |
| DeepSeek V4-Pro | ~3,000 | ~7,500 | ~70 days |
| Five runs combined | ~12,000 (range 7,000-22,000) | ~30,000 (per The Decoder) | – |
- Cost: about $200-300 million for the largest cluster, assuming $70,000 per GPU.
- All five together: more than half a billion dollars of hardware, by The Decoder’s reading of the numbers.
- Revision: the first version said 2,000-3,000 GPUs and $150-200 million. DePue replaced those figures on 12 November.
What the Report Proves and What It Does Not
This is where a lot of social media posts go too far. The report itself is careful, so here is the honest split.
| The report supports | The report does NOT show |
|---|---|
| The starting weights of six models were generated on hardware with Blackwell’s core layout | Where the chips were located (China or abroad) |
| The method works: 24 of 24 control models with public hardware were identified correctly | Who owned the chips, or how they were obtained |
| Nvidia, not AMD, is the likely maker for V4-Flash | That any export-control law was broken |
| Earlier models from the same labs read as Hopper | That all training ran on Blackwell (only initialisation is fingerprinted) |
Controls in the test set include Meta’s Llama 3.1 (H100), EleutherAI’s Pythia (A100) and Nvidia’s own Nemotron-3 Ultra (GB200). Closed-weight models such as GPT or Claude cannot be checked this way at all.
Warning: Chinese firms are known to rent compute in data centres outside China, notably in Malaysia. TechCrunch, The Decoder and the report itself all say offshore access could explain the result.
Export Rules and the Wider Nvidia Story
Blackwell chips are banned from export to China and to Chinese-owned firms. The less powerful H200 can be exported through a route that lets the US take a 25% cut, according to TechCrunch.
- Earlier claims: White House adviser Michael Kratsios said earlier in 2026 that Moonshot used export-banned Nvidia chips, citing US intelligence (TechCrunch).
- Chip flows: Epoch AI estimates about two million advanced chips entered China in 2024 and 2025 through routes that hid their final destination.
- Stockpile: Reuters reported about 600,000 H200 chips waiting in Hong Kong for shipment into China (5 November).
- Industry view: Omdia analyst Lian Jye Su said Chinese vendors are still using cutting-edge Nvidia chips to train models (via Shattered).
- Policy angle: Georgetown CSET’s Sam Bresnick has argued that controls should cover who uses chips, not just where they sit.
The timing is awkward for Nvidia. Jensen Huang left the company this week, Nvidia reports earnings on 17 November, and Commerce Secretary Howard Lutnick is due to testify about Nvidia’s China sales on 10 December.
As of 15 November, Nvidia had not confirmed or denied the research in anything reviewed by Shattered. DeepSeek, Moonshot, MiniMax and Ant Group had not responded publicly either.
What It Means If You Use These Models
For most readers, the practical answer is simple: nothing about the models has changed. The weights, licences, APIs and quality are exactly what they were last week.
If you run the weights locally
- Downloads: the models are still on Hugging Face under the same licences. The report does not ask anyone to remove them.
- Safety: how a model was trained says nothing about backdoors or data collection. Running weights offline keeps your prompts on your PC either way.
- Alternatives: if you prefer models with disclosed hardware, Llama 3.1 and Nemotron-3 Ultra were among the 24 confirmed controls.
If you use the official APIs or apps
- Service: DeepSeek, Kimi and MiniMax APIs work as before. Prices and limits have not changed because of this report.
- Privacy: the bigger question for API users is still where your data is stored. Read is DeepSeek safe? for the data-in-China details.
If you choose models for a company
- Check your policy. Some firms, especially US government contractors, already restrict Chinese-made models. This report may add weight to those rules.
- Ask your legal or compliance team. The report makes no claim against users, but only your lawyers can say what your contracts require.
- Keep a model register. Note the model name, version, source and licence for each one you deploy, so you can swap quickly if rules change.
- Have a fallback. Compare options in our DeepSeek vs Claude guide before you need one.
How to Check a Model’s Training Hardware Yourself
You do not need DePue’s six-trillion search to get a rough answer for most models. Start with what the lab discloses.
- Read the model card on Hugging Face. Look for a “Hardware” or “Training infrastructure” section. Meta and Nvidia usually list GPU type and hours.
- Read the technical report. Labs often state the cluster in the paper (for example DeepSeek’s V3 report named H800s).
- Check the report’s Appendix 2. It lists about 40 models, US and European ones included, with the hardware class each one reads as.
- Treat export flags carefully. The appendix flags describe the chip class under US rules. The report says they are not findings of any violation.
- For the deep check: DePue says he published verification code on GitHub, linked from the report. It needs PyTorch and access to the matching GPU type to reproduce a fingerprint.
What Is Still Unknown
- Exact chips: 148 SMs could be B200 or B300; 152 could be GB200 or GB300.
- Location and owner: the weights cannot show this, and no one has published it.
- Independent replication: no unaffiliated team had published a replication as of 14 November (per Shattered).
- Lab responses: none of the four labs had commented. Moonshot previously said Kimi K2.5 was trained on H800 chips, which matches its Hopper reading.
- Government action: no official US response to this specific report has been published yet.
My Take
The method is the real story for me. Using leftover randomness in published weights to identify the hardware is elegant, it validated on every known control, and anyone with the right GPU can test it.
But I would not read it as proof of smuggling. Renting Blackwell servers abroad is a known route, and the report is upfront that it cannot tell the difference. Expect labs to start initialising weights on CPUs or scrubbing unused slots, which would end this trick quickly.
For people who just run Kimi K3 or DeepSeek V4 on their own hardware, this changes nothing practical today. I will update this post if a lab responds, a replication lands or regulators act.
FAQ
Which AI models were found to use Nvidia Blackwell chips?
DeepSeek V4-Flash and V4-Pro, Moonshot’s Kimi K3 and K3.1-Pro, MiniMax M3 and Ant Group’s Ling-3.1-flash, according to Will DePue’s 12 November 2026 report.
Does this prove China smuggled Nvidia chips?
No. The report shows Blackwell-type hardware made the starting weights, but it cannot show where the chips were or who owned them. Offshore data centres are a possible explanation.
How can model weights reveal which GPU was used?
PyTorch splits random initialisation across the GPU’s cores, and Hopper and Blackwell have different core counts. That pattern survives in weights that training barely changes.
Is it still safe and legal to use DeepSeek V4 or Kimi K3?
The report makes no claim against users, and the models, licences and downloads are unchanged. Companies with strict rules on Chinese models should check with their own compliance team.
Did Qwen or GLM models use Blackwell?
No. Alibaba’s Qwen and Zhipu’s GLM models read as Hopper in the report. Some Qwen3 models were initialised on CPU and gave no signal.
Could the chips have been AMD instead of Nvidia?
The report says AMD’s MI300X can mimic the layout, but rules it out for V4-Flash and calls it very unlikely for the other five models.
Has Nvidia or DeepSeek responded?
Not as of 15 November 2026. Nvidia, DeepSeek, Moonshot, MiniMax and Ant Group had not publicly commented on the research.
Sources
- Unemployment Labs, Will DePue, “Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs” (12 Nov 2026), unemploymentlabs.com/checkpoint-forensics
- Will DePue on X, announcement post (Nov 2026)
- TechCrunch, “Export-banned Nvidia chips are behind China’s most powerful AI models” (13 Nov 2026)
- The Decoder, “Former OpenAI researcher says he found Nvidia Blackwell fingerprints hidden inside six Chinese AI models” (13 Nov 2026)
- Shattered, “Banned Blackwell Chips Tied to Kimi K3, DeepSeek V4: 4 Gaps” (14 Nov 2026)
- Bloomberg, “Bessent says Nvidia needs dramatic shakeup after Huang ouster” (15 Nov 2026, headline)
- Reuters, H200 stockpile in Hong Kong report (5 Nov 2026), via TechCrunch
- Epoch AI, chip import estimates for 2024-2025, via TechCrunch



