Google Launches Gemini 4 Hydrogen for Agentic AI
Study Links DeepSeek and Kimi to Banned Nvidia Chips
A forensic weight study by former OpenAI engineer Will DePue alleges that six Chinese open-weight models from DeepSeek, Moonshot AI, MiniMax, and Ant Group carry mathematical initialization fingerprints matching export-restricted Nvidia Blackwell processors.
Oladipupo Ajayi | · 6 min read

Publicly released artificial intelligence models carry hidden mathematical signatures inside their opening numbers that reveal the exact computer chips used to build them. For two years, Washington export restrictions barred semiconductor makers from shipping Nvidia top-tier processors to mainland China, aiming to slow down foreign frontier model development. Yet Chinese laboratories continued releasing competitive open-weight models throughout 2026, claiming they achieved parity through clever software algorithms and domestic silicon. A technical investigation published this week challenges that official narrative. Former OpenAI researcher Will DePue released a forensic weight analysis alleging that six prominent Chinese models—including DeepSeek-V4-Pro, Moonshot AI Kimi K3, MiniMax-M3, and Ant Group Ling-3.1-flash—were initialized on Nvidia Blackwell graphics processors that are strictly banned from export to China. Relying purely on mathematical analysis of public model files rather than anonymous leaks, the findings add fresh fuel to congressional investigations into semiconductor smuggling and offshore cloud rentals. The technical allegations arrive as Chinese model makers seek massive public market valuations, continuing coverage we tracked when DeepSeek targeted a $74B valuation ahead of a public listing.
Published through research group Unemployment Labs, the study examined open-weight checkpoints downloaded from public developer repositories. As detailed in Dataconomy coverage of the forensic chip analysis, DePue checked the starting parameter distributions of dozens of commercial models against six trillion combinations of processor layouts and random number seeds. Four of the six flagged Chinese models matched Nvidia GB200 or GB300 Grace Blackwell superchips, while two matched standalone B200 or B300 Blackwell accelerators. Estimating that training two large frontier models requires roughly 2,000 to 3,000 Blackwell processors worth more than $500M, the report lands at a sensitive moment in Washington. Days earlier, Senators Elizabeth Warren and Chris Coons urged federal officials to preserve Nvidia export records following the sudden resignation of Nvidia chief executive Jensen Huang over ties to Southeast Asian data center buyers.
How Weight Forensics Expose Underlying Silicon
To understand how an independent researcher can identify a physical graphics processor simply by inspecting a software file, one must look at how neural networks begin training. Before a model reads a single word of internet text, engineers initialize billions of weight parameters with pseudorandom numbers generated on the GPU.
While random number generators follow deterministic mathematical formulas, the exact sequence of floating-point numbers written into memory depends on how threads are scheduled across physical processor cores, memory banks, and tensor units. Because Nvidia altered the internal streaming multiprocessor layout and thread-block scheduling between its older Hopper (H100/H200) architecture and its newer Blackwell (B200/GB200) architecture, the two chip families produce distinct numerical patterns when running identical initialization seeds. DePue validated his detection method across twenty-four reference models with verified hardware histories before auditing Chinese releases. In a striking demonstration, he took the public weights of DeepSeek-V4-Flash and reproduced its exact starting weight matrices bit-for-bit on a physical Nvidia B300 processor. Further technical reporting by Unite.AI on the Unemployment Labs findings noted that the methodology detects hardware layouts without needing insider access to private training clusters.
Comparing Flagged Chinese Models and Hardware Fingerprints
| Model Name | Developer Organization | Release Window | Detected Chip Fingerprint | Prior Model Fingerprint |
|---|---|---|---|---|
| DeepSeek-V4-Pro | DeepSeek | April 2026 | Nvidia GB200 / GB300 Grace Blackwell | Hopper (DeepSeek-V3, Dec 2024) |
| DeepSeek-V4-Flash | DeepSeek | April 2026 | Nvidia B200 / B300 Blackwell | Hopper (DeepSeek-V3, Dec 2024) |
| Kimi K3 & K3.1-Pro | Moonshot AI | July 2026 | Nvidia GB200 / GB300 Grace Blackwell | Hopper (Kimi K2.5, Jan 2026) |
| MiniMax-M3 | MiniMax | June 2026 | Nvidia B200 / B300 Blackwell | Hopper (MiniMax-M2.7, Apr 2026) |
| Ling-3.1-flash | Ant Group | October 2026 | Nvidia GB200 / GB300 Grace Blackwell | Hopper (Ling-3.0-flash, Aug 2026) |
The chronological progression shown in the table reveals a clear hardware transition across four major Chinese laboratories. Earlier checkpoints released by DeepSeek, Moonshot AI, MiniMax, and Ant Group between late 2024 and early 2026 carried Hopper signatures. Between April and October 2026, all four organizations transitioned to Blackwell hardware for subsequent model generations. Notably, the pattern did not appear across every Chinese tech giant: recent readable releases from Alibaba, Zhipu, Tencent, Xiaomi, and StepFun still registered Hopper signatures rather than Blackwell layouts.
The Geolocation Caveat: Smuggled Chips or Overseas Cloud Rentals?
While the mathematical evidence linking the model weights to Blackwell silicon is precise, DePue included an important technical caveat: initialization fingerprints identify the type of processor that generated the starting numbers, not the physical address where the server rack sits.
There are two distinct pathways through which a Chinese laboratory can initialize weights on Blackwell hardware without smuggling physical server racks across customs borders:
- Overseas Cloud Leasing: Chinese research teams can rent computing clusters hosted inside commercial data centers in Singapore, Malaysia, or the Middle East, where local operators legally purchased Blackwell racks from American server vendors.
- Split Initialization and Training: A lab could theoretically initialize model weights on a small rented Blackwell node overseas and transfer the checkpoint file to domestic clusters running Huawei Ascend or older Hopper chips for the main training run, though researchers consider that workflow unlikely given the massive bandwidth required.
- Illicit Hardware Diversion: Third-party shell companies and gray-market brokers may have physically routed restricted B200 and GB200 server blades into mainland China despite Department of Commerce export bans.
Separating physical smuggling from remote cloud access sits at the center of ongoing federal inquiries in Washington. Earlier this week, congressional investigators summoned former Nvidia chief Jensen Huang to explain multi-billion-dollar hardware shipments to Singapore cloud provider Megaspeed. Federal trade compliance and distributor investigations were detailed in our report on Congress summoning Jensen Huang over distributor scrutiny.
Why Export Controls Struggle Against Borderless Compute
The forensic findings highlight a structural weakness in semiconductor export policy. Rules written by the Bureau of Industry and Security were designed to stop physical cargo ships and air freight pallets from delivering restricted chips to Chinese ports.
In modern cloud computing, physical geography matters far less than fiber optic bandwidth. Training weights and gradient updates travel across encrypted international internet cables in seconds. Unless Washington pairs hardware export bans with strict know-your-customer verification rules for foreign data center operators, restricted processors installed in neutral third countries will continue powering foreign frontier models. Lawmakers on Capitol Hill are already pushing legislation such as the AI Overwatch Act to require physical geolocation telemetry inside advanced server processors, ensuring chipmakers can verify where their hardware operates throughout its lifespan. Infrastructure constraints and data center expansion battles were explored in our report on the $2.1T data center power bottleneck.
What the Forensics Mean for Global AI Competition
Will DePue weight analysis proves that open-weight model releases contain far more forensic metadata than developers realized. By turning pseudorandom number initialization into an auditing tool, independent researchers have given regulators and security analysts a mathematical window into private server rooms.
Neither Nvidia, DeepSeek, Moonshot AI, MiniMax, nor Ant Group issued immediate public responses to the study. Yet the political consequences in Washington are already building. As senators demand export records and committees prepare public hearings, the assumption that export controls alone can freeze foreign model progress looks increasingly fragile. In the race for computational leadership, controlling physical borders is no longer enough when intelligence moves at the speed of light over global cloud networks.
Read More on TechRobust:

Oladipupo Ajayi
Oladipupo Ajayi
Expertise:Artificial Intelligence, Machine Learning Trends, Data Infrastructure, Enterprise AI Strategy, Frontier Tech Commentary
Award:TechRobust AI & Data Voice of the Year 2025
Ola is an Editor-at-Large at TechRobust, delivering authoritative commentary, high-level analysis, and investigative features across the frontiers of machine intelligence and big data. He tracks frontier model developments, enterprise AI adoption, data governance, and the societal shifts driven by computational breakthroughs.
Related Stories
Google Launches Gemini 4 Hydrogen for Agentic AI
The $2.1T AI Data Center Power Bottleneck Explained
Inside the 2026 AI Agent Crisis: Escapes and Zero-Days