研究机构 Unemployment Labs 于2026年11月12日发布的一项法医分析报告指出,包括 DeepSeek、月之暗面(Moonshot AI)、MiniMax 和蚂蚁集团在内的多家中国前沿人工智能实验室,其最新发布的六个大模型中存在受美国出口管制的英伟达 Blackwell GPU 硬件指纹 1。该研究通过分析模型权重初始化时的随机数生成模式,证实这些模型在2026年4月至10月期间开始在 Blackwell 架构芯片上进行初始化或训练,而此前基于 Hopper 架构的模型则未呈现此类特征 1。
具体而言,DeepSeek-V4-Flash、DeepSeek-V4-Pro、Kimi K3、Kimi K3.1-Pro、MiniMax-M3 和 Ling-3.1-flash 等模型被检测到包含与英伟达 B200/B300 或 GB200/GB300 系列匹配的数值模式 1。技术原理在于 PyTorch 使用 Philox 伪随机数生成器时,不同 GPU 流多处理器(SM)数量导致的线程分布差异:Blackwell B200/B300 拥有148个 SM,GB200/GB300 拥有152个 SM,而旧款 Hopper H100/H800 仅有132个 SM 1。时间线数据显示,DeepSeek 从2024年12月基于 Hopper 训练的 V3 版本转向2026年4月基于 Blackwell 训练的 V4 版本;月之暗面也从2026年1月的 K2.5(Hopper)转向同年7月的 K3(Blackwell) 1。
报告强调统计显著性极高,在约6万亿种组合搜索中,意外匹配的预期上限约为8 sigma,而这六个模型的匹配强度介于49至1,385 sigma之间 1。研究还排除了 AMD MI300X 模拟 Blackwell 布局的可能性,通过检查微小的数值差异(如 DeepSeek-V4-Flash 在377个关键值上与英伟达而非 AMD 匹配)予以否定 1。估算显示,最大的两个运行任务 Kimi K3 和 DeepSeek-V4-Pro 各需数千个 Blackwell GPU 运行2至3个月,总集群成本可能超过5亿美元 1。
尽管证据强烈指向对受控硬件的访问,但报告明确指出这仅证明了硬件使用权的存在,并未直接证实走私或违反出口管制的具体路径 1。此外,并非所有中国科技公司都采用了 Blackwell 架构,阿里巴巴(Qwen)、智谱(GLM)、腾讯、小米和阶跃星辰的最新模型仍显示为 Hopper 训练特征 1。美团 LongCat-2.0 声称使用 ASIC 超节点,且未发现英伟达指纹 1。
A forensic study published on November 12, 2026, by Unemployment Labs has identified hardware signatures consistent with export-controlled NVIDIA Blackwell GPUs within six recent frontier AI models developed by leading Chinese laboratories 1. The analysis focused on DeepSeek, Moonshot AI (Kimi), MiniMax, and Ant Group, revealing that their latest releases—specifically DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3, and Ling-3.1-flash—contain numerical patterns matching the B200/B300 or GB200/GB300 architectures 1. While the evidence strongly suggests access to restricted hardware, the report clarifies that it confirms only the utilization of such chips during initialization or training, without directly proving smuggling operations or specific violations of export control pathways 1.
The methodology relied on detecting variations in thread distribution caused by different Stream Multiprocessor (SM) counts when PyTorch employs the Philox pseudo-random number generator 1. Blackwell chips feature either 148 SMs (B200/B300) or 152 SMs (GB200/GB300), distinct from the 132 SMs found in older Hopper H100/H800 units 1. This architectural shift marks a clear timeline for adoption: DeepSeek transitioned from Hopper-based V3 models in December 2024 to Blackwell-based V4 models starting in April 2026, while Moonshot AI moved from Hopper-based K2.5 in January 2026 to Blackwell-based K3 in July 2026 1. To rule out emulation via alternative hardware, researchers examined minute numerical discrepancies; for instance, DeepSeek-V4-Flash matched NVIDIA-specific values across 377 key points, effectively excluding AMD MI300X configurations that could theoretically mimic Blackwell’s SM layout 1.
The statistical confidence of these findings is exceptionally high, with match strengths ranging from 49 to 1,385 sigma against an expected accidental match ceiling of approximately 8 sigma within a search space of roughly 6 trillion combinations 1. The scale of this infrastructure investment is substantial; the largest training runs for Kimi K3 and DeepSeek-V4-Pro are estimated to have required thousands of Blackwell GPUs operating for two to three months, implying total cluster costs exceeding $500 million 1. Despite these discoveries, the use of Blackwell hardware is not uniform among Chinese tech giants 1. Recent models from Alibaba (Qwen), Zhipu AI (GLM), Tencent, Xiaomi, and StepFun continue to display training characteristics associated with Hopper architecture 1. Additionally, Meituan’s LongCat-2.0 claims to utilize ASIC supernodes, with no NVIDIA fingerprints detected in its weights 1.
评论
还没有评论,欢迎留下第一条。