前 OpenAI 研究员称六款中国 AI 模型暗藏 Blackwell 痕迹

Nano Banana Pro,提示词由 THE DECODER 提供
AI 研究员 Will DePue 称,一项取证分析发现了首个技术证据,表明中国领先的 AI 实验室尽管面对美国出口管制,仍能获取 NVIDIA 最新的 Blackwell 硬件。他表示,OpenAI 和 Anthropic 的资深研究员已经核查了该方法。
根据 Will DePue 的取证调查,中国 AI 公司 DeepSeek、Moonshot AI、MiniMax 和蚂蚁集团正在使用受美国出口管制的 NVIDIA Blackwell GPU。DePue 曾在 OpenAI 担任关键角色,参与打造 Sora 和 Sora 2 视频模型以及 GPT-4o、o3 和 GPT-5,之后创立了非营利研究组织 Unemployment Labs。
他说,自己的分析首次为这种获取行为提供了技术证据。美国官员和新闻媒体此前曾怀疑 DeepSeek 和 Moonshot AI 在使用 Blackwell 芯片,但从未公布证据。DePue 在 X 上写道,OpenAI 和 Anthropic 的资深研究员已独立核查并确认了他的发现。据 DePue 称,自 2026 年 4 月以来发布的六款中国模型带有 NVIDIA Blackwell 世代的独特特征,而 DeepSeek-V3 和 Kimi K2.5 等更早的模型则显示出 NVIDIA 上一代 Hopper 的痕迹。
证据直接来自已发布的语言模型。一个 AI 模型由数十亿个称为权重的数值组成。在训练开始前,标准软件框架 PyTorch 会用从一个称为 seed 的起始值生成的随机数填充这些空槽位。为了快速分配如此海量的数据,GPU 将工作拆分到各个计算核心上。报告称,由于不同芯片型号的计算核心数量不同,填充过程会产生一种特征性模式。较旧的 Hopper 芯片,如 H100 和 H800,使用 132 个计算核心,而较新的 Blackwell 芯片则根据型号(B200/B300 或 GB200/GB300)恰好使用 148 或 152 个核心。
随后,训练会数百万次调整这些数值,但一些用于未使用单词或符号的槽位仍保留其初始随机值。DePue 称,即使在模型经过大量训练的部分中,初始分布的微弱统计回声依然存在。他说,他反向推导了这种模式,并在真实的 Blackwell 芯片上精确复现了 DeepSeek-V4-Flash 和 Kimi K3 的结果。他已将验证代码发布在 GitHub 上。
训练运行成本超过 5 亿美元
该分析将 DeepSeek-V4-Flash 和 MiniMax-M3 对应到 B200 或 B300 芯片(148 个核心)。根据分析,DeepSeek-V4-Pro、Kimi K3、Kimi K3.1-Pro 以及蚂蚁集团的 Ling-3.1-flash 显示出 NVIDIA 液冷 GB200 或 GB300 旗舰机架系统的 152 核模式。DePue 表示,从统计上可以排除偶然性,并且该方法正确识别了所有 24 个硬件已知的对照模型,包括 Meta 的 Llama 3.1(H100)。
图片:Will DePue / Unemployment Labs
DePue 的报告估计,他考察的五次训练合计需要约 12,000 块 Blackwell GPU,算力大致相当于 30,000 块较旧的 H100 芯片。仅 Kimi K3 一项,就在约 70 天里需要估计 4,500 块 Blackwell 处理器,DeepSeek-V4-Pro 则需要约 3,000 块。按约 70,000 美元的单价计算,这些集群每一个的成本都在 2 亿至 3 亿美元之间,五个系统合计超过 5 亿美元。
这些发现也与 SemiAnalysis 首席分析师 Dylan Patel 的分析一致。在接受红杉资本采访时,Patel 表示 DeepSeek-V3 的模型架构仍是专门为 Nvidia 的 Hopper 芯片量身定制的,而 DeepSeek-V4 的架构则是为 Blackwell GPU 和华为的 Ascend 芯片打造的。根据 DePue 的报告,其他中国实验室的模型——包括阿里巴巴的 Qwen3.8 以及智谱、腾讯、小米和阶跃星辰的模型——仍只显示出 Hopper 的指纹特征,而美团的 LongCat-2.0 则指向华为硬件。
指纹并不能证明存在违法行为
严格来说,该方法只能显示是哪种硬件初始化了权重,不过训练本身很可能是在同一批系统上运行的。唯一的另一种技术解释是 AMD MI300X——它同样受到出口限制——被人为降频,从而产生相同的模式。DePue 表示,他利用芯片之间微小的舍入差异,为 DeepSeek-V4-Flash 排除了 AMD 的可能。
据这项调查,这些数据并未揭示这些 GPU 位于何处,也没有说明其所有者是谁。由于中国企业可以在中国境外的数据中心租用算力,报告称其技术证据显示这些公司获得了 Nvidia 的顶级硬件,但这本身并不能证明存在实体走私或违反美国出口法的行为。
没有炒作水的 AI 新闻——由人工精选
完整访问 THE DECODER 上的每一篇文章
参与评论和社区讨论
每周通过邮件获取 AI 新闻回顾
每年 6 次:“AI Radar”——深入探讨最重要的 AI 话题
每日 AI 新闻,始终紧跟最新动态
我们完整的十年归档
由拥有 10 年以上 AI 经验的团队报道
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Former OpenAI researcher says he found Nvidia Blackwell fingerprints hidden inside six Chinese AI models)
Nano Banana Pro prompted by THE DECODER
AI researcher Will DePue says a forensic analysis has turned up the first technical evidence that leading Chinese AI labs have access to Nvidia's latest Blackwell hardware despite US export controls. He says senior researchers at OpenAI and Anthropic have already checked the method.
According to a forensic investigation by Will DePue, Chinese AI companies DeepSeek, Moonshot AI, MiniMax, and Ant Group are using Nvidia's export-controlled Blackwell GPUs. DePue played a key role at OpenAI in building the Sora and Sora 2 video models as well as GPT-4o, o3, and GPT-5 before founding the nonprofit research group Unemployment Labs.
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.
AI News Without the Hype – Curated by Humans
Full access to every article on THE DECODER
Join the comments and community discussions
A weekly AI news recap via mail
6x/year: "AI Radar" — deep dives on the AI topics that matter most
Daily AI news, always up to date
Our full ten-year archive
Covered by a team with 10+ years in AI
原文为英文,正文由自动抓取并保留原文。
原文发布于 The Decoder(Former OpenAI researcher says he found Nvidia Blackwell fingerprints hidden inside six Chinese AI models),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。