Weight Initialization Forensics Links Six Chinese AI Models to Nvidia Blackwell GPUs

The Three-Question Framework

What Happened?

On November 12, 2026, researcher Will DePue of Unemployment Labs published technical evidence showing six open-weight Chinese models released since April 2026 contain parameter initialization layouts matching Nvidia B200/B300 (148 SMs) and GB200/GB300 (152 SMs) GPUs.

Why It Matters

Before training adjusts model weights, PyTorch's cuRAND Philox generator stripes initial random values across 2,048 threads per Streaming Multiprocessor, leaving an indelible hardware signature in unused or weakly modified weights without requiring leaks or insiders.

What Could Happen Next

With DeepSeek, Moonshot AI, MiniMax, and Ant Group shifting newer models onto Blackwell signatures while Alibaba, Zhipu, Tencent, Xiaomi, and StepFun still read as Hopper-era chips, export-control analysts must trace rented cloud clusters versus domestic hardware.

Executive Key Takeaways

DeepSeek-V4-Flash and MiniMax-M3 match the 148-SM (303,104-thread) B200/B300 layout, while DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, and Ant Group's Ling-3.1-flash match the 152-SM (311,296-thread) GB200/GB300 layout.

Statistical significance across the six models spans 49 to 1,385 sigma, far above the 8-sigma threshold expected by chance across roughly 6 trillion tested seed-and-layout combinations.

Unmodified PyTorch on a rented B300 reproduced 445 of 469 preserved rows in DeepSeek-V4-Flash bit for bit from DeepSeek's seed (953049583), while a GB200 reproduced all 13 preserved rows in Kimi K3 from seed 20256586.

To rule out partitioned AMD MI300X GPUs (which can mimic 148 or 152 compute units), a vendor-divergence test on DeepSeek-V4-Flash matched Nvidia on 363 of 377 differing entries and AMD on zero.

Source: Technical report by Will DePue at Unemployment Labs, published November 12, 2026.

The forensic pipeline exploits deterministic hardware parallelism during step-zero neural network creation. Large PyTorch initializations via Nvidia's cuRAND library assign up to 2,048 threads per Streaming Multiprocessor (SM) to populate contiguous tensor elements with Philox pseudorandom draws. Earlier Hopper H100 SXM, H200, and compliant H800 GPUs pack 132 SMs (270,336 total threads), whereas Blackwell B200/B300 chips carry 148 SMs and GB200/GB300 superchips expose 152 SMs. Embedding rows tied to rare tokens that never update during training preserve initial values almost identically, while even modified attention layers retain a faint rank-order correlation—averaging r = 0.066 (57 sigma) across 42 layers of DeepSeek-V4-Flash.

Validation against 24 models whose developers disclosed hardware confirmed 100% agreement, including Meta's Llama 3.1 (132 SMs), EleutherAI's Pythia suite (108 SMs for A100), and Nvidia's Nemotron-3 Ultra (152 SMs for GB200). Tracking each developer chronologically revealed a clear hardware inflection in mid-2026: DeepSeek shifted from Hopper on V3 (December 2024) to Blackwell on V4 (April 2026); Moonshot moved from Hopper on Kimi K2.5 (January 2026) through a merged K2.7-Code checkpoint to GB200/GB300 on K3 (July 2026); MiniMax transitioned from M2.7 (April 2026) to M3 (June 2026); and Ant Group moved from Ling-3.0-flash (August 2026) to Ling-3.1-flash (October 2026). Even DeepSeek-V4-Flash's weakest single layer test hit 20 sigma.

Unemployment Labs roughly estimated the final training runs for five of these models (treating K3.1-Pro as a continued checkpoint of K3) required around 12,000 Blackwell GPUs overall—between 1,400 and 3,400 chips per model over 45 to 100 days. At an implied $70,000 per Blackwell GPU cited from Wall Street Journal reporting on ByteDance infrastructure in Malaysia, the final training compute alone represents more than $500 million in cluster value, excluding earlier experiments or inference fleets.

Limits: The report emphasizes that weight initialization patterns prove only what chip created starting parameters and cannot establish where chips were physically located, whether systems were legally rented offshore versus smuggled, or whether export laws were violated. Closed-weight APIs, PaddlePaddle frameworks (Baidu's ERNIE-5.5), CPU-initialized models (several Qwen3 variants), Meituan's LongCat-2.0, and DeepSeek-V4.1-Flash yielded no recoverable GPU signature.

The Three-Question Framework

What Happened?

Why It Matters

What Could Happen Next

Executive Key Takeaways

Primary Sources & Citations