跳到正文
原文
The Decoder:AI News· Maximilian Schreiner·· 3 天前AI 评分63

前OpenAI研究员称在6个中国AI模型权重中发现Nvidia Blackwell芯片指纹

Former OpenAI researcher says he found Nvidia Blackwell fingerprints hidden inside six Chinese AI models

文章要点
  • 1前OpenAI研究员DePue称在6个2026年4月后发布的中国模型权重中检出Nvidia Blackwell芯片特征
  • 2DePue称OpenAI与Anthropic资深研究员已独立核验其结论,该方法对24个已知硬件对照模型全部识别正确
  • 3DePue估算5次训练合计约12,000块Blackwell芯片,相当于约30,000块H100,五个系统总成本超过5亿美元
  • 4报告称证据不能证明走私或违反美国出口法,也不显示芯片位置与归属,算力可能来自中国境外数据中心租用
决策要点

权重初始化痕迹把芯片来源变成第三方可检验的技术问题,出口管制的执行依据从传闻转向公开权重的统计证据。

  • 影响:采用被点名开源模型的企业在采购与合规评估中多了一层出口管制暴露的变量;对训练方而言,权重初始化会留下可被外部复现检验的算力指纹;报告不认定违法,境外租用等渠道未被排除,定性应保持谨慎。
  • 参考:可考虑在引入DeepSeek-V4、Kimi K3等被点名模型时,把该分析与GitHub验证代码列入尽调材料并跟踪第三方复现;自研团队值得先盘点算力采购与租用合同,评估来源声明能否经受同类权重检验。
  • 依据:判断基于DePue的报告与GitHub验证代码,24个对照模型识别结果与OpenAI、Anthropic研究员核验均为本人口径,暂无第三方公开复现;SemiAnalysis的Dylan Patel称DeepSeek-V4架构面向Blackwell构建,方向一致。
正文

His analysis, he says, provides the first technical evidence of this access. US officials and news outlets had suspected that DeepSeek and Moonshot AI were using Blackwell chips but never published proof. DePue wrote on X that senior researchers at OpenAI and Anthropic have already independently checked and confirmed his findings. According to DePue, six Chinese models released since April 2026 carry the distinct signature of Nvidia's Blackwell generation, while earlier models such as DeepSeek-V3 and Kimi K2.5 show traces of Nvidia's older Hopper generation.

The evidence comes directly from the published language models. An AI model consists of billions of numerical values called weights. Before training starts, PyTorch, the standard software framework, fills these empty slots with random numbers generated from a starting value called a seed. To distribute this massive amount of data quickly, the GPU splits the work across its individual compute cores. Because different chip models have different core counts, the filling process produces a characteristic pattern, according to the report. Older Hopper chips such as the H100 and H800 use 132 compute cores, while the newer Blackwell chips use exactly 148 or 152 cores depending on the model (B200/B300 or GB200/GB300).

Training then adjusts these values millions of times, but some slots for unused words or symbols keep their initial random values. Even in heavily trained parts of the model, a faint statistical echo of the initial distribution survives, according to DePue. He says he worked the pattern backward and reproduced the results for DeepSeek-V4-Flash and Kimi K3 exactly on real Blackwell chips. He has published the verification code on GitHub.

Training runs costing more than half a billion dollars

The analysis maps DeepSeek-V4-Flash and MiniMax-M3 to B200 or B300 chips (148 cores). DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, and Ant Group's Ling-3.1-flash show, according to the analysis, the 152-core pattern of Nvidia's liquid-cooled GB200 or GB300 flagship rack systems. DePue says chance can be ruled out statistically and that the method correctly identified all 24 control models with known hardware, including Meta's Llama 3.1 (H100).

Image: Will DePue / Unemployment Labs

DePue's report estimates that the five training runs he examined required about 12,000 Blackwell GPUs combined, as much computing power as roughly 30,000 older H100 chips. Kimi K3 alone required an estimated 4,500 Blackwell processors over about 70 days, and DeepSeek-V4-Pro about 3,000. At a unit price of about $70,000, each of those clusters would cost $200 million to $300 million, and all five systems together more than half a billion dollars.

The findings also align with analysis from Dylan Patel, chief analyst at SemiAnalysis. In an interview with Sequoia Capital, Patel said DeepSeek-V3's model architecture was still specifically tailored for Nvidia's Hopper chips, while DeepSeek-V4's architecture was built for Blackwell GPUs and Huawei's Ascend chips. According to DePue's report, models from other Chinese labs, including Alibaba's Qwen3.8 and those from Zhipu, Tencent, Xiaomi, and StepFun, still show only Hopper signatures, while Meituan's LongCat-2.0 points to Huawei hardware.

The fingerprint doesn't prove a legal violation

Strictly speaking, the method shows only which hardware initialized the weights, though the training itself very likely ran on the same systems. The only other technical explanation would be an AMD MI300X, which is also restricted, artificially throttled so it produces the same pattern. DePue says he ruled out AMD for DeepSeek-V4-Flash using tiny rounding differences between the chips.

The data doesn't reveal where the GPUs are or who owns them, according to the investigation. Because Chinese companies can rent computing capacity at data centers outside China, the report says its technical evidence shows access to Nvidia's top hardware but does not by itself prove physical smuggling or a violation of US export law.

来源:The Decoder:AI News · the-decoder.com