
Unemployment Labs报告称DeepSeek、月之暗面、MiniMax及蚂蚁集团六款模型权重显示Blackwell架构特征,但无法证明芯片来源或违规
前OpenAI研究工程师Will DePue声称从DeepSeek、月之暗面、MiniMax及蚂蚁集团六款模型权重中发现NVIDIA Blackwell GPU硬件指纹,引发对中国AI企业获取受美国出口管制高端芯片的关注,但研究无法证明芯片来源或违反出口管制。
AI-generated summary
美国对中国实施高端AI芯片出口管制,限制NVIDIA Blackwell等先进GPU销售。前OpenAI研究员Will DePue通过分析模型权重的GPU硬件指纹,声称发现多家中国AI企业模型使用Blackwell架构初始化权重。
曾參與 OpenAI Sora、GPT-4o 等模型開發的前研究工程師 Will DePue,11月12日在社群平台 X 公布一項調查,宣稱從中國 AI 業者公開的模型權重中,找到 NVIDIA 最新 Blackwell GPU 的硬體特徵。涉及的公司包括 DeepSeek、月之暗面(Moonshot AI)、MiniMax 與螞蟻集團,共有六款模型被判定使用 Blackwell 架構進行權重初始化。由於這些高階晶片受到美國對中國的出口管制,研究結果也引發外界對中國 AI 企業如何取得運算資源的關注。
DePue 創辦的研究組織 Unemployment Labs 在報告中指出,AI 模型開始訓練前,必須先為數十億甚至數千億個參數設定初始數值。這些看似隨機的數字,實際上會受到 GPU 運算架構與隨機種子(Seed)影響。不同晶片擁有不同數量的運算單元,即使使用相同種子,參數排列方式也可能不同。部分數值在後續訓練中沒有被修改,因此仍保留最初的排列特徵,形成研究人員所稱的「GPU 硬體指紋」,可用來反推當初建立模型權重的硬體類型。
研究團隊分析約6兆種晶片配置與隨機種子的組合後,發現 DeepSeek-V4-Flash 與 MiniMax-M3 的權重特徵符合 Blackwell B200、B300;DeepSeek-V4-Pro、Kimi K3、Kimi K3.1-Pro 及螞蟻集團 Ling-3.1-flash,則符合 Grace Blackwell GB200、GB300 系列。相較之下,DeepSeek-V3、Kimi K2.5 等較早版本,以及研究涵蓋的阿里巴巴 Qwen、智譜 GLM 等模型,仍呈現前代 Hopper 架構的特徵。這項方法主要辨識 GPU 的運算單元數量,例如 Hopper 的132個SM、Blackwell 的148個SM,以及 Grace Blackwell 的152個SM,無法進一步區分同組內的所有晶片型號。
為了確認結果,DePue 使用實際的 NVIDIA B300 與 GB200 晶片重新產生模型初始權重,成功重現 DeepSeek-V4-Flash 與 Kimi K3 中保留下來的數值;改用 H100 或更換隨機種子後,便無法得到相同結果。他也公布驗證程式碼,並表示已有 OpenAI、Anthropic 的資深研究人員獨立確認。另有研究者公開表示成功重現部分比對結果,但目前尚未看到兩家公司正式發表驗證聲明,這項方法的完整適用範圍仍有待更多研究檢驗。不過,硬體指紋能辨識模型權重的初始化特徵,並不能直接證明整個模型訓練過程都使用相同晶片,更無法判斷設備位於中國境內或海外資料中心。DePue 也在報告中承認,這些證據不足以確認晶片如何取得,以及是否有人違反美國出口管制。研究結果雖為中國 AI 業者可能接觸受管制高階 GPU 提供新的技術線索,但相關企業的實際硬體來源、訓練地點與採購方式,仍未獲得證實。
AI outlook — possibilities, not facts
美国政府或监管机构可能对相关中国AI企业的芯片采购展开调查
Possible · Within weeks
更多独立研究者将验证或质疑GPU硬件指纹方法的适用范围
Likely · Within weeks

On November 10, the Federal Aviation Administration announced the expansion of the deployment of the AI air traffic management system SMART to 22 control centers across the United States. Three days later, a nationwide flight suspension incident occurred. Officials have not yet confirmed a direct relationship between the two. The accident resulted in the cancellation of more than 1,000 flights and the delay of nearly 10,000 flights. The Secretary of Homeland Security stated that no signs of malicious cyber attacks were found.

American startup Thinking Machines Lab opened the Inkling 2 Small model weight download on November 12. It is licensed by Apache 2.0 and supports up to 1 million token context and multi-modal input. The NVFP4 version requires at least 500GB GPU memory. Taiwanese companies and institutions still need to evaluate data security and risks for use.

The US Federal Bureau of Investigation (FBI) issued a security alert on November 13, stating that the open source AI penetration testing tool ARTEX has been abused by hackers since August this year, attacking multiple industries such as finance, government, national defense, semiconductors, energy and technology. Some intrusions have obtained high-level permissions of enterprise cloud systems in less than 3 hours after locking the target. Taiwan's e-commerce SaaS platform provider 91APP has been attacked by ARTEX since October, which may affect up to 1,329 brands and about 4.8 million consumers, and some personal data may be leaked. The Cyberbiz e-commerce platform and a number of listed companies have also confirmed that they were hacked or had data leaked. The Ministry of Digital Development has issued an information security warning, reminding the public to beware of personal information fraud and calling on companies to check and fix vulnerabilities as soon as possible.

SpaceX launched the Starship giant rocket in Texas, challenging the "double capture" test of fully recovering the first and second stages for the first time. After liftoff, the booster malfunctioned, initiated the automatic abort procedure, crashed into the Gulf of Mexico and exploded, declaring a setback at the first hurdle; however, the spacecraft still successfully advanced towards orbit and strived to return on Saturday to complete mid-air clamping. The mission will deploy 26 Starlink satellites and orbit the earth 15 times at an altitude of approximately 275 kilometers.

Huida CEO Huang Renxun resigned on the 12th, shocking the global technology industry. Counterpoint Research pointed out that the short-term AI chip product route and supply chain cooperation are expected to remain stable, but the strategy of the new management team needs to be observed in the medium and long term. It is predicted that server AI ASIC shipments will surpass GPUs in 2027, and the AI computing market will develop in parallel towards GPUs and customized ASICs.

Lex Fridman, an American technology podcaster, posted on the