{"id":10015,"date":"2025-05-09T20:52:28","date_gmt":"2025-05-09T12:52:28","guid":{"rendered":"https:\/\/ai-stack.ai\/?p=10015"},"modified":"2025-05-09T20:57:49","modified_gmt":"2025-05-09T12:57:49","slug":"china-ai-close","status":"publish","type":"post","link":"https:\/\/ai-stack.ai\/en\/china-ai-close","title":{"rendered":"From Hardware to Ecosystem: Why\u202fJensen\u202fHuang Says China\u2019s\u202fAI Is \u201cRight\u202fBehind\u201d the\u202fU.S."},"content":{"rendered":"\n<h3 class=\"wp-block-heading\"><strong>1\u202f\u2502\u202fBackdrop \u2014 Huang\u2019s Warning on Capitol\u202fHill<\/strong><\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">During an impromptu interview in Washington,\u202fNVIDIA\u202fCEO Jensen\u202fHuang told reporters that <strong>\u201cChina is right behind us \u2014 the gap is very, very small.\u201d<\/strong> He added that sweeping export curbs on\u202fAI chips could hurt U.S. firms more than they slow Chinese progress.<a href=\"https:\/\/www.businessinsider.com\/jensen-huang-says-us-china-very-close-ai-chip-race-2025-4?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"> Business Insider<\/a><\/p>\n\n\n\n<figure class=\"wp-block-embed is-type-video is-provider-youtube wp-block-embed-youtube wp-embed-aspect-16-9 wp-has-aspect-ratio\"><div class=\"wp-block-embed__wrapper\">\n<span class=\"embed-youtube\" style=\"text-align:center; display: block;\"><iframe class=\"youtube-player\" width=\"1300\" height=\"732\" src=\"https:\/\/www.youtube.com\/embed\/EEMzZqLGLDk?version=3&#038;rel=1&#038;showsearch=0&#038;showinfo=1&#038;iv_load_policy=1&#038;fs=1&#038;hl=en-US&#038;autohide=2&#038;wmode=transparent\" allowfullscreen=\"true\" style=\"border:0;\" sandbox=\"allow-scripts allow-same-origin allow-popups allow-presentation allow-popups-to-escape-sandbox\"><\/iframe><\/span>\n<\/div><\/figure>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>2\u202f\u2502\u202fHardware Race \u2014 Domestic Chips + Rack\u2011Scale Clusters<\/strong><\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><td><strong>System<\/strong><\/td><td><strong>Scale \/ Peak FP16*<\/strong><\/td><td><strong>Key Take\u2011aways<\/strong><\/td><td><strong>Why It Matters<\/strong><\/td><\/tr><tr><td><strong>Huawei\u202fCloudMatrix\u202f384<\/strong><\/td><td>384\u202f\u00d7\u202fAscend\u202f910C, <strong>300\u202fPFLOPS<\/strong> (\u2248\u202f67\u202f% &gt;\u202fNVIDIA\u202fNVL72)<\/td><td>Optical \u201csuper\u2011node\u201d interconnect, 3\u00d7 NVL72 memory, but higher power draw and $8.2\u202fM price tag<\/td><td>First Chinese rack that can match a top U.S. GPU pod at system level<a href=\"https:\/\/www.ft.com\/content\/cac568a2-5fd1-455c-b985-f3a8ce31c097?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"> <\/a><a href=\"https:\/\/semianalysis.com\/2025\/04\/16\/huawei-ai-cloudmatrix-384-chinas-answer-to-nvidia-gb200-nvl72\/?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\">SemiAnalysis<\/a><\/td><\/tr><tr><td><strong>Baidu\u202fKunlun\u202fP800 Cluster<\/strong><\/td><td><strong>30\u202f000\u202fKunlun\u202fP800<\/strong> chips<\/td><td>Claimed to train 100\u202fB\u2011plus\u2011parameter models or fine\u2011tune 1\u202f000 tasks in parallel; already live in banks &amp; cloud<\/td><td>Shows Chinese vendors can scale bespoke silicon to hyperscale size<a href=\"https:\/\/www.reuters.com\/world\/china\/chinas-baidu-says-its-kunlun-chip-cluster-can-train-deepseek-like-models-2025-04-25\/?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"> Reuters<\/a><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">*FP16\/BF16 mixed precision commonly used for large\u2011model training.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>3\u202f\u2502\u202fModel Benchmarks \u2014 Open\u2011Source Titles Hitting the Leaderboards<\/strong><\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><td><strong>Model<\/strong><\/td><td><strong>Parameters<\/strong><\/td><td><strong>Latest Public Score<\/strong><\/td><\/tr><tr><td><strong>DeepSeek\u202fJanus\u2011Pro\u20117B<\/strong><\/td><td>7\u202fB<\/td><td>GenEval image\u2011generation <strong>0.80<\/strong>, beating DALL\u2011E\u202f3 (0.67) &amp; SDXL (0.74)<a href=\"https:\/\/huggingface.co\/deepseek-ai\/Janus-Pro-7B?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"> Hugging Face<\/a><a href=\"https:\/\/janusai.pro\/?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\">janusai.pro<\/a><\/td><\/tr><tr><td><strong>Alibaba\u202fQwen\u202f2.5\u202fOmni<\/strong><\/td><td>3\u202fB\u202f\u2192\u202f72\u202fB series<\/td><td>Jumped to #1 on Hugging\u202fFace open\u2011source board in early April\u202f2025<a href=\"https:\/\/huggingface.co\/Qwen\/Qwen2.5-Omni-7B?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"> Hugging Face<\/a><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">Algorithm tweaks and massive instruction data let smaller Chinese models equal or surpass U.S. peers of the same size.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>4\u202f\u2502\u202fMarket Flywheel \u2014 Hundred\u2011Million\u2011User Apps Feed More Data<\/strong><\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">In the January\u202f2025 <strong>Global AI\u2011App MAU<\/strong> chart, China took two of the top five slots:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Doubao (ByteDance)\u202f\u2014\u202f78\u202fM MAU<br><\/li>\n\n\n\n<li>DeepSeek\u202f\u2014\u202f34\u202fM MAU<br><\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">That scale supplies fresh conversation data and third\u2011party plug\u2011in traffic, accelerating iterative improvement.<a href=\"https:\/\/backlinko.com\/most-popular-ai-apps?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"> Backlinko<\/a><a href=\"https:\/\/backlinko.com\/deepseek-stats?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\">Backlinko<\/a><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>5\u202f\u2502\u202fPatent &amp; R&amp;D Density \u2014 \u201cQuantity Has a Quality of Its Own\u201d<\/strong><\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">WIPO\u2019s 2024 landscape report shows <strong>38\u202f000 GenAI patent families from China (2014\u20112023), six\u202ftimes the U.S. count.<\/strong><a href=\"https:\/\/www.wipo.int\/pressroom\/en\/articles\/2024\/article_0009.html?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"> WIPO<br><\/a> Generous local subsidies mean more experiments move from paper to product.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>6\u202f\u2502\u202fCompute Infrastructure \u2014 The 300\u202fEFLOPS National Target<\/strong><\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">China\u2019s six\u2011ministry action plan aims for <strong>\u2265\u202f300\u202fEFLOPS<\/strong> of aggregate compute by 2025, up from 197\u202fEFLOPS in\u202f2023.<a href=\"https:\/\/www.globaltimes.cn\/page\/202310\/1299520.shtml?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"> <\/a><a href=\"https:\/\/english.www.gov.cn\/news\/202407\/07\/content_WS668a8075c6d0868f4e8e8f76.html?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\">english.www.gov.cn<\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>What\u2019s an\u202fEFLOPS?<\/strong>\u202fFLOPS are <em>floating\u2011point operations per second<\/em>.<br><em>Exa\u2011<\/em> (10\u00b9\u2078) means <strong>1\u202fEFLOPS\u202f=\u202f1\u202fquintillion (a million\u2011trillion) calculations every second<\/strong> \u2014 the \u201cexascale\u201d class occupied today by U.S. supercomputer Frontier (\u2248\u202f1.1\u202fEFLOPS). Hitting 300\u202fEFLOPS would be like running almost 300 Frontiers in parallel.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Generous land, power\u2011price and financing incentives are driving a wave of cool\u2011climate AI data\u2011center construction to reach that figure.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>7\u202f\u2502\u202fCost Efficiency \u2014 DeepSeek\u2011V3 Breaks the \u201cBillion\u2011Dollar Club\u201d<\/strong><\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">DeepSeek reports training a <strong>671\u202fB\u2011parameter MoE<\/strong> with <strong>2.78\u202fM H800\u202fGPU\u2011hours (\u2248\u202f57\u202fdays, US\u202f$5.6\u202fM)<\/strong> \u2014 orders of magnitude below the rumored GPT\u20114 bill.<a href=\"https:\/\/www.unite.ai\/how-deepseek-cracked-the-cost-barrier-with-5-6m\/?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"> Unite.AI<br><\/a> Even if that excludes R&amp;D trial runs, it proves that careful data\u2011mixing and sparse routing can slash capital needs.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>8\u202f\u2502\u202fRemaining Gaps<\/strong><\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><td><strong>Area<\/strong><\/td><td><strong>Current Bottleneck<\/strong><\/td><\/tr><tr><td><strong>Advanced Process<\/strong><\/td><td>Ascend\u202f910C still 7\u202fnm; energy\u2011efficiency lags 4\u202fnm\u202fBlackwell\u2011class GPUs<\/td><\/tr><tr><td><strong>Software Ecosystem<\/strong><\/td><td>MindSpore \/ CANN maturity trails CUDA stack; porting costs remain high<\/td><\/tr><tr><td><strong>Talent Gravity<\/strong><\/td><td>Top researchers still cluster in U.S. big\u2011tech labs; long\u2011term flow uncertain<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>9\u202f\u2502\u202fOverall Assessment<\/strong><\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>System\u2011level stacking &amp; optical links<\/strong> let China field usable top\u2011tier compute despite chip sanctions.<br><\/li>\n\n\n\n<li><strong>Open\u2011source + community<\/strong> tactics propel benchmark parity and rapid app diffusion.<br><\/li>\n\n\n\n<li><strong>State\u2011backed infrastructure<\/strong> (300\u202fEFLOPS) supplies the \u201cpower station\u201d for continued scaling.<br><\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Bottom line:<\/strong> the contest has shifted from a \u201cgeneration gap\u201d to a <strong>neck\u2011and\u2011neck sprint<\/strong>. Future separation will hinge on access to sub\u20114\u202fnm manufacturing, energy efficiency breakthroughs, and how export\u2011control dynamics evolve. For global firms, risk\u2011balanced strategies that tap <em>both<\/em> ecosystems are fast becoming table stakes.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>1\u202f\u2502\u202fBackdrop \u2014 Huang\u2019s Warning on Capitol\u202fHill During an impromptu interview in Washington,\u202fNVIDIA\u202fCEO Jensen\u202fHuang told reporters that \u201cChina is right behind us \u2014 the gap is very, very small.\u201d He added that sweeping export curbs on\u202fAI chips could hurt U.S. firms more than they slow Chinese progress. Business Insider 2\u202f\u2502\u202fHardware Race \u2014 Domestic Chips + Rack\u2011Scale Clusters System Scale \/ Peak FP16* Key Take\u2011aways Why It Matters Huawei\u202fCloudMatrix\u202f384 384\u202f\u00d7\u202fAscend\u202f910C, 300\u202fPFLOPS (\u2248\u202f67\u202f% &gt;\u202fNVIDIA\u202fNVL72) Optical \u201csuper\u2011node\u201d interconnect, 3\u00d7 NVL72 memory, but higher power draw and $8.2\u202fM price tag First Chinese rack that can match a top U.S. GPU pod at system level SemiAnalysis Baidu\u202fKunlun\u202fP800&#8230;<\/p>\n","protected":false},"author":253372376,"featured_media":10016,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"_crdt_document":"","jetpack_post_was_ever_published":false,"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[96987604,96987592],"tags":[96987654,96988068],"class_list":["post-10015","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-news","category-featured-articles","tag-ai-en","tag-china-2"],"blocksy_meta":[],"acf":[],"jetpack_featured_media_url":"https:\/\/i0.wp.com\/ai-stack.ai\/wp-content\/uploads\/2025\/05\/%E6%A8%A1%E5%9E%8BA-9.jpg?fit=1920%2C1080&quality=100&ct=202603031250&ssl=1","jetpack_shortlink":"https:\/\/wp.me\/ph344V-2Bx","jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/posts\/10015","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/users\/253372376"}],"replies":[{"embeddable":true,"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/comments?post=10015"}],"version-history":[{"count":0,"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/posts\/10015\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/media\/10016"}],"wp:attachment":[{"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/media?parent=10015"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/categories?post=10015"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/tags?post=10015"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}