A report published on 12 November by Unemployment Labs, Will DePue’s research group, says six models from DeepSeek, Moonshot AI, MiniMax and Ant Group carry the numerical fingerprint of Nvidia’s Blackwell GPUs, which need a US licence for export to China. It estimates that five of the final training runs used about 12,000 Blackwell GPUs, clusters it puts at “more than half a billion dollars” together.
At the report’s own price of roughly $70,000 per GPU, its best estimate of 12,000 chips comes to about $840 million, by our arithmetic. The report says its findings “do not establish where the chips were located, how access was obtained, or whether export controls were violated”.
Business Pill · Neoclouds Explained: Renting the Chips by the Hour
An evergreen Business Engineer explainer on neoclouds, the providers that rent GPUs by the hour. It teaches the general idea only and does not discuss the Unemployment Labs report or any Chinese lab.
The key insight: As we read it, the report moves the argument from officials’ claims to evidence others can rerun: it reads the chip class from weights the labs published themselves and posts its code, while saying nothing about where the chips sat or who owned them.
The report reads six models as initialised on Blackwell: DeepSeek-V4-Flash and DeepSeek-V4-Pro (April 2026), Moonshot’s Kimi K3 (July 2026) and Kimi K3.1-Pro (November 2026), MiniMax-M3 (June 2026) and Ant Group’s Ling-3.1-flash (October 2026). DeepSeek-V4-Flash and MiniMax-M3 match the B200 or B300; the other four match the GB200 or GB300 in Nvidia’s rack-scale systems.
Earlier models from the same four labs read as Nvidia Hopper, including DeepSeek-V3 (December 2024), Kimi K2.5 (January 2026), MiniMax-M2.7 (April 2026) and Ling-3.0-flash (August 2026). The report says a Hopper reading cannot tell an H800 from an H100 or H200.
The newest models it could read from Alibaba (Qwen), Zhipu (GLM), Tencent, Xiaomi and StepFun all read as Hopper. It found no Nvidia fingerprint in Meituan’s LongCat-2.0, which the company says was trained on “AI ASIC superpods”.
Best estimate of Nvidia Blackwell GPUs per final training run, with the report’s ranges and estimated run lengths, from the Unemployment Labs report of 12 November 2026 (revised). The report calls these estimates low-confidence; the $840 million is our arithmetic at the report’s ~$70,000 per GPU.
What Was Posted on X
I have a major result to share.
I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.
It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.
These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.
Will DePue, who wrote the Unemployment Labs report, announced it in this post on 12 November. It gives the method in plain words and the older estimate of 2,000 to 3,000 Blackwell GPUs for the largest runs; the report itself was revised the same day to 4,500 for Kimi K3 and 3,000 for DeepSeek-V4-Pro.
If you’d like to see more work like this, Unemployment Labs is raising $5M to fund our research for the next 6 months. As a subsidiary of https://macrogrants.com, we’re a 501c3 pending non-profit. Please reach out to sponsor@unemploymentlabs.com.
In a follow-up post the same evening, DePue said Unemployment Labs is raising $5M to fund six months of research and is a subsidiary of Macrogrants, with non-profit status pending. We include it as context on who funds the work.
How the Fingerprint Works
The method rests on how PyTorch fills a model’s starting weights. A seeded random-number generator spreads the work across GPU threads, and the thread count depends on the chip: 132 processor cores (SMs) on Hopper, 148 on the B200/B300 and 152 on the GB200/GB300. The same seed on a different chip puts the numbers in different places.
Some weights are never updated in training and keep their starting values, so the report searched about six trillion seed-and-layout combinations to find the ones that recreate them. On a physical B300 it says unmodified PyTorch reproduced 445 of 469 selected DeepSeek-V4-Flash rows bit for bit, and a rented GB200 reproduced all 13 of Kimi K3’s preserved rows. Changing the seed by one reproduced none.
The report says the strongest chance match across the search would reach about 8 sigma, while the strongest evidence for each of the six models ranges from 49 to 1,385 sigma. It also says the method agreed with all 24 cases where a developer has publicly named its hardware, including Nvidia’s own Nemotron-3 Ultra, which reads as 152 SMs.
The Numbers, by Our Arithmetic
Cluster value. 12,000 Blackwell GPUs at roughly $70,000 each is about $840 million; the report’s range of 7,000 to 22,000 GPUs gives $490 million to $1.54 billion. The $70,000 is the report’s figure, derived from a Wall Street Journal report on Blackwell chips in Malaysia.
Signal vs chance. The weakest of the six headline results, 49 sigma, is about 6.1 times the 8-sigma chance benchmark; DeepSeek-V4-Flash’s 1,385 sigma is about 173 times it.
The revision. The first version, and DePue’s X post, put the largest runs at 2,000 to 3,000 Blackwell GPUs each. The revised best estimate for Kimi K3 is 4,500, 1.5 times the top of that range, after the report assumed lower chip efficiency.
The AMD check. In the 377 values where Nvidia and AMD MI300X output differ, DeepSeek-V4-Flash’s weights match Nvidia in 363, about 96%, and AMD in none, per the report.
Responses and Earlier Claims
We found no public response to the report from DeepSeek, Moonshot AI, MiniMax, Ant Group, Nvidia or the Commerce Department’s Bureau of Industry and Security as of 15 November. The Register, which called DePue an ex-OpenAI engineer, and The Decoder reported the findings on 13 November.
The report cites earlier claims without technical evidence: a Reuters report of 23 February 2026 in which a US official said DeepSeek trained on Nvidia’s best chip, and White House OSTP director Michael Kratsios on Moonshot AI on 22 July 2026. It also cites a June 2025 GitHub comment by a DeepSeek engineer: “we don’t have SM100 GPUs”, SM100 being Blackwell.
As we read the timeline in the report, the switch is recent and lab by lab: each of the four labs has a Hopper-era model followed by a Blackwell-era one between April and October 2026, while the newest Qwen, GLM, Tencent, Xiaomi and StepFun models it could read still read as Hopper.
Unemployment Labs report by Will DePue, 12 November 2026
“Notably, our findings provide technical evidence of access to advanced computing hardware, but do not establish where the chips were located, how access was obtained, or whether export controls were violated.”
Three Implications
FOR EXPORT-CONTROL POLICY The report says Blackwell systems can be rented from data centres outside China; a fingerprint shows access to a chip class, not its location or owner.
FOR MODEL BUYERS The method works only on open weights from PyTorch on GPUs, so closed models and some frameworks, such as Baidu’s PaddlePaddle, cannot be checked this way, per the report.
FOR THE COST OF COMPUTE At the report’s ~$70,000 per GPU, its best estimate of 12,000 Blackwell GPUs for five runs is about $840 million, by our arithmetic; the report says total spending is likely in the billions.
The framework places export controls in its top layer, governance: “Capability gating inside labs, procurement frameworks inside enterprises, export controls between nation-states, litigation over contested deployments. No longer a passive layer; an active priced variable.”
As we read it, the report tests that governance layer against the silicon layer beneath it: a chip class that needs a US licence for China shows up, per its forensics, in models from four Chinese labs. The framework predates the report and does not discuss it.
Strongest evidence per model, in sigma, from Appendix 1 of the Unemployment Labs report (12 November 2026); the report puts the best chance match across its search at about 8 sigma. Not independently verified; the multiples of 8 sigma are our arithmetic.
What Is Not Established
The report says the fingerprint shows the hardware that created the starting weights; it “supports, without proving, the inference that Blackwell also performed the training”. A lab could create weights on one system and train on another.
The weights cannot show who owned the chips or where they were. The report notes that Blackwell systems can be rented from data centres outside China, and says the evidence does not by itself establish that an export-control violation occurred.
The GPU counts are low-confidence: the report says “We are not confident in these estimates”, because chip efficiency, token counts and run lengths are its assumptions. Specially configured AMD MI300X chips can produce the same layouts; the report excludes them only for DeepSeek-V4-Flash. DePue says senior researchers at OpenAI and Anthropic checked the result; they are not named.
This Week In The AI Supercycle
AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source
This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.
Unemployment Labs says six models from four Chinese labs, starting with DeepSeek-V4 in April 2026, carry a Blackwell fingerprint at 49 to 1,385 sigma against an 8-sigma chance benchmark. Its best estimate of 12,000 GPUs for five runs is about $840 million at its $70,000 price, by our arithmetic; the report does not establish where the chips were or whether any rule was broken.
A note on sourcing. Every finding, chip reading, sigma score and GPU estimate here comes from the Unemployment Labs report by Will DePue (12 November 2026, revised the same day) and his X post; we have not reproduced the analysis. We found no public response from the companies named or from the US Commerce Department as of 15 November 2026 (about 13:00 UTC). The $840 million cluster value, the range in dollars and the multiples of the 8-sigma benchmark are our arithmetic on the report’s figures.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
