[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-weight-analysis-links-deepseek-and-kimi-to-nvidia-blackwell":10,"sections":35},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":22,"persona_id":22,"persona_name":22,"section":24,"tags":25,"sources":30,"feedback":34,"feedback_at":22,"cost_usd":34,"total_tokens":34},13140,"weight-analysis-links-deepseek-and-kimi-to-nvidia-blackwell","Weight Analysis Links DeepSeek and Kimi to Nvidia Blackwell","UnemploymentLabs says weight initialization patterns match Blackwell hardware, but it cannot show where the chips sat or whether any rule broke.","An independent research group says four open-weight Chinese models carry the statistical fingerprint of training on Nvidia Blackwell chips, which are under U.S. export controls. The group, UnemploymentLabs, named DeepSeek V4, Kimi K3.1, MiniMax-M3 and Ling-3.1-Flash.\n\nThe method is simple. Nvidia's chip families differ in their count of shader modules (SMs), the rough GPU equivalent of CPU cores: 132 on Hopper, 148 on Blackwell B parts, 152 on Blackwell GB parts. When training starts, a model's weights are set to pseudo-random values using Nvidia's cuRAND library, and that work is split across the SMs at 2,048 threads each for large weight sets. A given seed and layout always produce the same numbers.\n\nThe researchers reran that initialization with matching chips, software and thread counts, and the distribution of values lined up with the released models. When they changed the SM and thread counts, the match collapsed. That let them identify the chip family behind each model and estimate roughly how many cards were used. For DeepSeek V4, they also checked the actual weight values and ruled out AMD hardware configured to mimic Nvidia's layout.\n\nThe limits are real, and the researchers stated them. The analysis cannot say which exact card produced the weights, because it tracks SM layout rather than memory capacity or bandwidth. It works only on open weights, and it assumes a PyTorch and CUDA software stack. It also cannot say where the chips were or how they were accessed.\n\nThat last gap is the important one. As far as the source can tell, Chinese firms renting Blackwell capacity abroad is not illegal, and U.S. companies are not strictly barred from providing it, as the Tencent-Oracle deal shows. A fingerprint that cannot separate a smuggled rack from a rented cloud cluster settles a technical question while leaving the legal one open.\n\nWhat travels further is the technique. Training hardware is usually taken on a lab's word, and this check needs only the published weights.","[\"nvidia\",\"export-controls\",\"deepseek\",\"ai-chips\"]","2026-11-14T14:50:00.000Z","2026-11-14T23:52:50.964Z","2026-11-14T23:53:44.747Z","published",null,[],"ai",[26,27,28,29],"nvidia","export-controls","deepseek","ai-chips",[31],{"name":32,"url":33},"Tom's Hardware","https:\u002F\u002Fwww.tomshardware.com\u002Ftech-industry\u002Fartificial-intelligence\u002Fresearchers-allege-deepseek-kimi-and-other-chinese-ai-models-were-trained-on-export-controlled-nvidia-blackwell-chips-statistical-analysis-on-model-weights-reveals-high-correlation-to-nvidia-hardware",0,{"sections":36},[37,41,46,51,56,61,66,71,76,81,86,91,96,101],{"name":38,"slug":24,"count":39,"latest_published_at":40},"AI",7162,"2026-11-15T12:00:00.000Z",{"name":42,"slug":43,"count":44,"latest_published_at":45},"Security","security",1141,"2026-11-15T17:15:13.000Z",{"name":47,"slug":48,"count":49,"latest_published_at":50},"Deals","deals",671,"2026-11-15T14:00:00.000Z",{"name":52,"slug":53,"count":54,"latest_published_at":55},"Policy","policy",630,"2026-11-15T16:25:16.000Z",{"name":57,"slug":58,"count":59,"latest_published_at":60},"Consumer Tech","consumer-tech",383,"2026-11-15T17:55:32.000Z",{"name":62,"slug":63,"count":64,"latest_published_at":65},"Hardware","hardware",343,"2026-11-15T14:19:24.000Z",{"name":67,"slug":68,"count":69,"latest_published_at":70},"Science","science",236,"2026-11-14T20:54:14.000Z",{"name":72,"slug":73,"count":74,"latest_published_at":75},"Software","software",177,"2026-11-14T08:54:55.000Z",{"name":77,"slug":78,"count":79,"latest_published_at":80},"Startups","startups",162,"2026-11-13T17:32:35.000Z",{"name":82,"slug":83,"count":84,"latest_published_at":85},"General","general",130,"2026-11-15T15:41:13.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":90},"Dev Tools","dev-tools",127,"2026-11-11T19:00:44.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"Gaming","gaming",112,"2026-11-14T19:34:13.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"Reviews","reviews",38,"2026-10-20T12:00:00.000Z",{"name":102,"slug":103,"count":104,"latest_published_at":105},"How-To","how-to",10,"2026-11-14T23:45:00.000Z"]