我的征尘是星辰大海。。。
The dirt and dust from my pilgrimage forms oceans of stars...
-------当记忆的篇章变得零碎,当追忆的图片变得模糊,我们只能求助于数字存储的永恒的回忆
作者:黄教授
手机视频列表
道路是曲折的前途是光明的
视频
音频
原始脚本
道路是曲折的,工信部算力互联行动背后的现实困境与破局起点。 当大模型训练需要千卡集群的算力支撑,当东部企业为一张 H800显卡竞价,当西部数据中心的服务器利用率不足60%,这组矛盾的画面正是当下中国算力产业的真实缩影。 工信部推动的算力互联互通行动,看似是算力一张网的宏大蓝图。 但落地之初,必须直面三重残酷现实,被 NVIDIA 主导的存量生态、供需错配的算力乱象以及外部技术封锁的双重挤压。 一、70%的市场被卡脖子,存量 GPU 生态的封闭困局。 打开国内任意一家头部互联网公司的算力机房。 大概率会看到同一景象,机架上密集排列的 NVIDIA GPU。 据2025年行业数据, NVIDIA 在国内 AI 算力芯片市场的份额高达70%,华为昇腾等国产芯片仅占23%,其余厂商瓜分剩余7%。 这些 GPU 从 A800到 H20,是当前大模型训练、工业 AI 推理的主力军,但也是算力互联绕不开的第一道坎。 问题的核心在于 NVIDIA 的生态封闭性。 这些 GPU 的算力释放完全依赖 CUDA 软件生态,从底层的 CUDA Toolkit 到 中间层的 CUDNN 加速库,再到上层的 PyTorch、TensorFlow 适配,形成了一套硬件、软件、工具的闭环。 更关键的是 NVIDIA 通过技术限制,如 CUDA EULA 禁止跨硬件模拟和硬件阉割,如特供版 GPU 的 NVLink 带宽砍半。 从根源上阻断了非 NVIDIA 架构整合其算力的可能。 这意味着什么?国内现存的数百万张 NVIDIA GPU,看似是算力资产,实则是封闭的孤岛。 阿里的 GPU 集群无法直接对接腾讯的算力池。 中小企业的单卡算力更难参与到大模型训练中,因为没有统一的协议能让这些 GPU 突破 cuda 的限制,实现跨主体协同。 工信部要做的算力互联,第一步就要面对70%的存量算力被一家外企绑定的现实,而解套的难度堪比在别人的操作系统上重构新生态。 二,小而散的算力乱象。 需求急缺与闲置浪费的双重矛盾。 我们有10台单卡服务器。 但训练一个中等模型需要八卡集群,这些单卡根本用不上。 这是很多中小 AI 企业的共同困扰。 当前国内算力资源的分布呈现出严重的结构性失衡。 一方面是需求端的迫切缺口,大模型训练需要高密度、低延迟的集群算力,如1024卡、2048卡的统一调度。 但国内能提供此类算力的,只有阿里、腾讯、字节等头部企业的自建集群,且大多优先满足内部需求。 另一方面是供给端的大量闲置。 据行业统计,国内中小数据中心的算力上架率平均不足60%。 西部部分地区的智算中心甚至长期处于半闲置状态,原因是本地需求少,跨地域调度难。 更棘手的是算力碎片化,不同数据中心的硬件架构,GPU、NPU、TPU 混杂,接口协议,各自的调度系统、服务标准、计费、时延 承诺完全不同意。 比如一家广州的 AI 公司想租用成都某数据中心的 NPU 算力,不仅要解决数据跨地域传输的延迟问题,还要面对模型适配不同芯片的重构成本,最后往往因麻烦且不划算放弃。 这种有算力用不上、需算力找不到的乱象,正是工信部要破解的核心痛点。 但要把这些小而散的算力拧成一股绳,远比新建一个超算中心更复杂。 三、外部封锁与内部短板,国产算力的双重突围压力。 如果说存量 GPU 的封闭性是内部困境,那么外部技术封锁则让算力互联的道路更添阻力。 美国对高端 AI 芯片的出口管制持续升级,从最初限制 H100等高端型号,到后来对 H800、A800等特供版也加码限制。 甚至试图阻断先进制成芯片的对华供应。 这意味着国内想通过买更多 NVIDIA GPU 来缓解算力缺口的路正在被逐步堵死。 而国产算力的替代之路目前仍处于补短板阶段。 华为昇腾虽推出384卡的 Cloud Matrix 超节点,通过 Matrix Link 实现了高带宽互联,但产能和生态成熟度仍需时间。 昇腾芯片的量产能力尚未满足大规模需求。 C A N N 开源生态的开发者数量约40万,仅为 CUDA 超400万的1/10。 谷歌 TPU 虽在专用 AI 计算上有优势,但生态绑定 TensorFlow,且受限于谷歌云服务,难以融入国内算力网络。 这就形成了双重压力,短期无法依赖进口芯片填补缺口,长期国产芯片的生态建设又需时间。 工信部推动算力互联,本质上是在外部卡脖子、内部待成熟的夹缝中寻找一条不依赖单一厂商、不被技术封锁捆住的路径。 但这条路径的每一步都要面对技术突破和生态建设的双重考验。 四,2026年的破局起点,从算力标识到骨干网的务实第一步。 面对如此多的困境,工信部并没有急于追求全国算力一张网的理想形态。 而是在算力互联互通行动计划中给出了2026年的务实目标。 先解决能互联能调度的基础问题,再谈高效协同。 第一步是统一语言,建立全国统一的算力标识体系。 简单说就是给每一份算力发身份证,标注清楚硬件类型,是 NVIDIA A800,还是华为昇腾910。 算力规格,FP16算力多少?Teraops,内存带宽,HBM 容量多少?服务能力,支持训练还是推理,时延多少?有了这个身份证,不同数据中心的算力才能被认得出、比得了。 就像电商平台上的商品标注参数,企业可以清晰判断这份算力是否符合我的需求。 第二步是打通通路,推动三大运营商升级骨干网。 算力要跨地域调度,首先得有高速公路。 工信部明确要求推广800G1400G光纤传输技术,强化国家级互联网骨干直联点,建设算力专网。 目前中国移动的400 8百 G 光模块也实现了国产化量产,如瑞海光电的产品已通过阿里云认证。 这些基础设施的升级正是为了让东部的需求能高效对接西部的算力,减少跨地域传输的延迟和损耗。 第三步是试点破冰,建国家区域两级算力互联平台。 目前国家超算互联联网平台已打通。 20多家超算智算中心,累计提供近百亿计算时。 比如某科研团队在上海,可通过平台调用兰州超算中心的算力,无需自己搭建集群。 这种先试点再推广的模式,正是为了在2026年前验证跨地域、跨架构算力调度的可行性,为后续全国互联积累经验。 这些举措看似是基础工作,实则是破解当前困境的关键。 通过标识解决算力不互通的问题,通过骨干网解决调度不高效的问题,通过平台解决资源不匹配的问题。 虽然离全国算力一张网的理想还有距离,但至少为存量算力利用和国产算力突围找到了一个现实的起点。 道路的曲折在于每一步都要直面存量生态的封闭、供需矛盾的尖锐、外部封锁的压力。 但破局的希望,也正藏在这些务实的第一步中,当算力有了统 统一的身份证,当跨地域的高速路建成,当分散的算力能被高效调度,中国算力产业才能真正从被动依赖走向主动可控。 前途是光明的,中国算力产业破局之路。 尽管当前中国算力产业在前行路上荆棘丛生,但拨开层层迷雾,曙光已然显现。 从 NVIDIA 的发展瓶颈,到中国自身优势的逐步凸显,再到关键技术突破带来的新机遇,中国算力产业的未来充满希望,前途一片光明。 Nvidia 的发展困境,Nvidia 在 GPU 领域长期占据主导地位。 然而其发展正面临着物理层面的瓶颈。 在追求单卡高算力以实现高密度算力的进程中,物理上限问题愈发显著。 单台服务器受限于能耗以及主板支持的 PCI 插槽数量,能够容纳的 GPU 显卡数量有限,无法无限制的增加 GPU 以提升算力。 即便通过 NVLink 技术将多台服务器组成集群,构建万卡数据中心,又会遭遇电力难题。 据相关数据显示,一个万卡算力中心的耗电量相当于一座中型城市。 如到2027年,美国预计将有40%的数据中心因能源问题受限。 如此庞大的电力需求,让数据中心的可持续发展面临挑战。 难以依靠集中式的超大型数据中心来满足不断增长的算力需求。 NVIDIA GPU 为了兼顾游戏、显示挖矿以及 多种科学运算等多方面功能,设计上存在较大冗余。 这不仅使得成本居高不下,耗能也远超专为人工智能设计的芯片。 在构建大规模算力中心时,进一步加剧了成本与能耗负担。 并且 NVIDIA 构建的 CUDA 闭源系统,虽成为其商业护城河,但也限制了用户的使用自由,阻碍了大集群计算的发展。 例如对虚拟设备、虚拟总线等技术应用设置重重障碍,限制了算力的灵活调配与扩展。 中国算力产业的优势与机遇,政策与基础设施优势。 中国政府在算力产业发展上展现出强大的决心与行动力。 工信部积极推动算力互联互通行动。 国家有足够的财力支持建设下一代高速互联网骨干网,采用800G1400G光纤来联通数据中心,实现多数据中心的互联互通。 这一举措不仅能将将算力资源分散到电力供应更平衡的地区,解决能源瓶颈问题,还能整合全国算力资源,提高资源利用率。 相比之下,即使是美国富可敌国的顶级公司,也难以凭借一己之力完成如此庞大的基础设施建设。 通过国家的统筹规划与推动,中国在算力基础设施建设上具备了领先优势,为算力产业的长远发展奠定了坚实基础。 能源与资源调配优势,中国在能源领域积极布局,大力推广清洁能源,为数据中心提供可持续的能源供应保障。 在算力资源分布上,充分发挥西部算力资源丰富、东部算力需求旺盛的特点。 通过构建高速互联网络,实现中东部需求与西部算力 中心的有效连接。 优化资源配置,降低算力成本,提升整体算力服务效率。 这种能源与资源的合理调配模式,是中国算力产业发展的独特优势,能够更好地支撑算力产业的持续扩张与升级。 技术突破带来的新机遇,华为的高速互联总线技术成为中国算力产业破局的关 关键力量。 该技术突破了传统冯诺依曼架构的限制,能够将远端的 TPU、NPU 等计算芯片通过高速互联总线连接起来,使其在运算时如同本地计算卡一般,构建成一个超级计算机。 TPU、NPU 专为人工智能设计,在成本和优化效率上优势明显。 虽然当前存在集成度和调度灵活性不足等问题。 但随着技术的不断发展,以及高速互联总线技术的应用,有望实现 质的飞跃。 通过将大量的 TPU、NPU 集成,形成天文数量的计算核心。 利用800G1400G光纤的高速连接,在算力上超越 NVIDIA 的 GPU 集群并非遥不可及。 这种技术创新为中国在人工智能算力领域实现弯道超车提供了可能,开辟了一条全新的发展道路。 综上所述,虽然当前中国算力产业面临着 NVIDIA 主导市场以及外部技术封锁等诸多挑战。 但从长远来看,NVIDIA 自身发展的瓶颈为中国提供了破局的契机。 中国凭借政策支持、基础设施建设、能源资源调配以及关键技术突破等多方面优势,在算力产业发展上已经找准方向,迈出了坚实的步伐。 只要持续加大研发投入,完善产业生态,中国算力产业必将在全球竞争中脱颖而出,迎来光明的未来,为人工智能等新兴产业的发展提供源源不断的强大动力。
修正脚本
道路是曲折的,工信部算力互联行动背后的现实困境与破局起点。 当大模型训练需要千卡集群的算力支撑,当东部企业为一张 H800显卡竞价,当西部数据中心的服务器利用率不足60%,这组矛盾的画面正是当下中国算力产业的真实缩影。 工信部推动的算力互联互通行动,看似是算力一张网的宏大蓝图。 但落地之初,必须直面三重残酷现实,被 NVIDIA 主导的存量生态、供需错配的算力乱象以及外部技术封锁的双重挤压。 一、70%的市场被卡脖子,存量 GPU 生态的封闭困局。 打开国内任意一家头部互联网公司的算力机房。 大概率会看到同一景象,机架上密集排列的 NVIDIA GPU。 据2025年行业数据, NVIDIA 在国内 AI 算力芯片市场的份额高达70%,华为昇腾等国产芯片仅占23%,其余厂商瓜分剩余7%。 这些 GPU 从 A800到 H20,是当前大模型训练、工业 AI 推理的主力军,但也是算力互联绕不开的第一道坎。 问题的核心在于 NVIDIA 的生态封闭性。 这些 GPU 的算力释放完全依赖 CUDA 软件生态,从底层的 CUDA Toolkit 到中间层的 CUDNN 加速库,再到上层的 PyTorch、TensorFlow 适配,形成了一套硬件、软件、工具的闭环。 更关键的是 NVIDIA 通过技术限制,如 CUDA EULA 禁止跨硬件模拟和硬件阉割,如特供版 GPU 的 NVLink 带宽砍半。 从根源上阻断了非 NVIDIA 架构整合其算力的可能。 这意味着什么?国内现存的数百万张 NVIDIA GPU,看似是算力资产,实则是封闭的孤岛。 阿里的 GPU 集群无法直接对接腾讯的算力池。 中小企业的单卡算力更难参与到大模型训练中,因为没有统一的协议能让这些 GPU 突破 CUDA 的限制,实现跨主体协同。 工信部要做的算力互联,第一步就要面对70%的存量算力被一家外企绑定的现实,而解套的难度堪比在别人的操作系统上重构新生态。 二、小而散的算力乱象。 需求急缺与闲置浪费的双重矛盾。 我们有10台单卡服务器。 但训练一个中等模型需要八卡集群,这些单卡根本用不上。 这是很多中小 AI 企业的共同困扰。 当前国内算力资源的分布呈现出严重的结构性失衡。 一方面是需求端的迫切缺口,大模型训练需要高密度、低延迟的集群算力,如1024卡、2048卡的统一调度。 但国内能提供此类算力的,只有阿里、腾讯、字节等头部企业的自建集群,且大多优先满足内部需求。 另一方面是供给端的大量闲置。 据行业统计,国内中小数据中心的算力上架率平均不足60%。 西部部分地区的智算中心甚至长期处于半闲置状态,原因是本地需求少,跨地域调度难。 更棘手的是算力碎片化,不同数据中心的硬件架构,GPU、NPU、TPU 混杂,接口协议、各自的调度系统、服务标准、计费、时延承诺完全不统一。 比如一家广州的 AI 公司想租用成都某数据中心的 NPU 算力,不仅要解决数据跨地域传输的延迟问题,还要面对模型适配不同芯片的重构成本,最后往往因麻烦且不划算放弃。 这种有算力用不上、需算力找不到的乱象,正是工信部要破解的核心痛点。 但要把这些小而散的算力拧成一股绳,远比新建一个超算中心更复杂。 三、外部封锁与内部短板,国产算力的双重突围压力。 如果说存量 GPU 的封闭性是内部困境,那么外部技术封锁则让算力互联的道路更添阻力。 美国对高端 AI 芯片的出口管制持续升级,从最初限制 H100等高端型号,到后来对 H800、A800等特供版也加码限制。 甚至试图阻断先进制程芯片的对华供应。 这意味着国内想通过买更多 NVIDIA GPU 来缓解算力缺口的路正在被逐步堵死。 而国产算力的替代之路目前仍处于补短板阶段。 华为昇腾虽推出384卡的 Cloud Matrix 超节点,通过 Matrix Link 实现了高带宽互联,但产能和生态成熟度仍需时间。 昇腾芯片的量产能力尚未满足大规模需求。 CANN 开源生态的开发者数量约40万,仅为 CUDA 超400万的1/10。 谷歌 TPU 虽在专用 AI 计算上有优势,但生态绑定 TensorFlow,且受限于谷歌云服务,难以融入国内算力网络。 这就形成了双重压力,短期无法依赖进口芯片填补缺口,长期国产芯片的生态建设又需时间。 工信部推动算力互联,本质上是在外部卡脖子、内部待成熟的夹缝中寻找一条不依赖单一厂商、不被技术封锁捆住的路径。 但这条路径的每一步都要面对技术突破和生态建设的双重考验。 四、2026年的破局起点,从算力标识到骨干网的务实第一步。 面对如此多的困境,工信部并没有急于追求全国算力一张网的理想形态。 而是在算力互联互通行动计划中给出了2026年的务实目标。 先解决能互联能调度的基础问题,再谈高效协同。 第一步是统一语言,建立全国统一的算力标识体系。 简单说就是给每一份算力发身份证,标注清楚硬件类型,是 NVIDIA A800,还是华为昇腾910。 算力规格,FP16算力多少Teraops,内存带宽,HBM 容量多少?服务能力,支持训练还是推理,时延多少?有了这个身份证,不同数据中心的算力才能被认得出、比得了。 就像电商平台上的商品标注参数,企业可以清晰判断这份算力是否符合我的需求。 第二步是打通通路,推动三大运营商升级骨干网。 算力要跨地域调度,首先得有高速公路。 工信部明确要求推广800G、1400G光纤传输技术,强化国家级互联网骨干直联点,建设算力专网。 目前中国移动的400G、800G 光模块也实现了国产化量产,如瑞海光电的产品已通过阿里云认证。 这些基础设施的升级正是为了让东部的需求能高效对接西部的算力,减少跨地域传输的延迟和损耗。 第三步是试点破冰,建国家区域两级算力互联平台。 目前国家超算互联网平台已打通20多家超算智算中心,累计提供近百亿计算时。 比如某科研团队在上海,可通过平台调用兰州超算中心的算力,无需自己搭建集群。 这种先试点再推广的模式,正是为了在2026年前验证跨地域、跨架构算力调度的可行性,为后续全国互联积累经验。 这些举措看似是基础工作,实则是破解当前困境的关键。 通过标识解决算力不互通的问题,通过骨干网解决调度不高效的问题,通过平台解决资源不匹配的问题。 虽然离全国算力一张网的理想还有距离,但至少为存量算力利用和国产算力突围找到了一个现实的起点。 道路的曲折在于每一步都要直面存量生态的封闭、供需矛盾的尖锐、外部封锁的压力。 但破局的希望,也正藏在这些务实的第一步中,当算力有了统一的身份证,当跨地域的高速路建成,当分散的算力能被高效调度,中国算力产业才能真正从被动依赖走向主动可控。 前途是光明的,中国算力产业破局之路。 尽管当前中国算力产业在前行路上荆棘丛生,但拨开层层迷雾,曙光已然显现。 从 NVIDIA 的发展瓶颈,到中国自身优势的逐步凸显,再到关键技术突破带来的新机遇,中国算力产业的未来充满希望,前途一片光明。 Nvidia 的发展困境,Nvidia 在 GPU 领域长期占据主导地位。 然而其发展正面临着物理层面的瓶颈。 在追求单卡高算力以实现高密度算力的进程中,物理上限问题愈发显著。 单台服务器受限于能耗以及主板支持的 PCI 插槽数量,能够容纳的 GPU 显卡数量有限,无法无限制地增加 GPU 以提升算力。 即便通过 NVLink 技术将多台服务器组成集群,构建万卡数据中心,又会遭遇电力难题。 据相关数据显示,一个万卡算力中心的耗电量相当于一座中型城市。 如到2027年,美国预计将有40%的数据中心因能源问题受限。 如此庞大的电力需求,让数据中心的可持续发展面临挑战。 难以依靠集中式的超大型数据中心来满足不断增长的算力需求。 NVIDIA GPU 为了兼顾游戏、显示、挖矿以及多种科学运算等多方面功能,设计上存在较大冗余。 这不仅使得成本居高不下,耗能也远超专为人工智能设计的芯片。 在构建大规模算力中心时,进一步加剧了成本与能耗负担。 并且 NVIDIA 构建的 CUDA 闭源系统,虽成为其商业护城河,但也限制了用户的使用自由,阻碍了大集群计算的发展。 例如对虚拟设备、虚拟总线等技术应用设置重重障碍,限制了算力的灵活调配与扩展。 中国算力产业的优势与机遇,政策与基础设施优势。 中国政府在算力产业发展上展现出强大的决心与行动力。 工信部积极推动算力互联互通行动。 国家有足够的财力支持建设下一代高速互联网骨干网,采用800G、1400G光纤来联通数据中心,实现多数据中心的互联互通。 这一举措不仅能将算力资源分散到电力供应更平衡的地区,解决能源瓶颈问题,还能整合全国算力资源,提高资源利用率。 相比之下,即使是美国富可敌国的顶级公司,也难以凭借一己之力完成如此庞大的基础设施建设。 通过国家的统筹规划与推动,中国在算力基础设施建设上具备了领先优势,为算力产业的长远发展奠定了坚实基础。 能源与资源调配优势,中国在能源领域积极布局,大力推广清洁能源,为数据中心提供可持续的能源供应保障。 在算力资源分布上,充分发挥西部算力资源丰富、东部算力需求旺盛的特点。 通过构建高速互联网络,实现中东部需求与西部算力中心的有效连接。 优化资源配置,降低算力成本,提升整体算力服务效率。 这种能源与资源的合理调配模式,是中国算力产业发展的独特优势,能够更好地支撑算力产业持续扩张与升级。 技术突破带来的新机遇,华为的高速互联总线技术成为中国算力产业破局的关键力量。 该技术突破了传统冯诺依曼架构的限制,能够将远端的 TPU、NPU 等计算芯片通过高速互联总线连接起来,使其在运算时如同本地计算卡一般,构建成一个超级计算机。 TPU、NPU 专为人工智能设计,在成本和优化效率上优势明显。 虽然当前存在集成度和调度灵活性不足等问题。 但随着技术的不断发展,以及高速互联总线技术的应用,有望实现质的飞跃。 通过将大量的 TPU、NPU 集成,形成天文数量的计算核心。 利用800G、1400G光纤的高速连接,在算力上超越 NVIDIA 的 GPU 集群并非遥不可及。 这种技术创新为中国在人工智能算力领域实现弯道超车提供了可能,开辟了一条全新的发展道路。 综上所述,虽然当前中国算力产业面临着 NVIDIA 主导市场以及外部技术封锁等诸多挑战。 但从长远来看,NVIDIA 自身发展的瓶颈为中国提供了破局的契机。 中国凭借政策支持、基础设施建设、能源资源调配以及关键技术突破等多方面优势,在算力产业发展上已经找准方向,迈出了坚实的步伐。 只要持续加大研发投入,完善产业生态,中国算力产业必将在全球竞争中脱颖而出,迎来光明的未来,为人工智能等新兴产业的发展提供源源不断的强大动力。
英文翻译
The road is winding. The practical dilemmas and starting points for breakthroughs behind the Ministry of Industry and Information Technology's computing power interconnection initiative. When large-scale model training requires the computing power support of thousand-card clusters, when enterprises in the east compete for a single H800 GPU, and when server utilization in western data centers is less than 60%, this set of contradictory scenes is a true microcosm of China's current computing power industry. The computing power interconnection initiative promoted by the Ministry of Industry and Information Technology appears to be a grand blueprint for a unified computing power network. But at the outset of implementation, three harsh realities must be faced head-on: the existing ecosystem dominated by NVIDIA, the chaotic supply-demand mismatch in computing power, and the dual pressures of external technological封锁. 1. 70% of the market is choked. The closed predicament of the existing GPU ecosystem. Open the computing power room of any major domestic internet company. You will likely see the same scene: racks densely packed with NVIDIA GPUs. According to 2025 industry data, NVIDIA holds a 70% market share in China's AI computing power chip market, while domestic chips like Huawei's Ascend account for only 23%, with the remaining 7% divided among other manufacturers. These GPUs, from A800 to H20, are the main force for current large-scale model training and industrial AI inference, but they are also the first hurdle that computing power interconnection cannot avoid. The core issue lies in NVIDIA's closed ecosystem. The computing power release of these GPUs relies entirely on the CUDA software ecosystem, from the underlying CUDA Toolkit to the intermediate CUDNN acceleration library, and to the upper-layer PyTorch and TensorFlow adaptations, forming a closed loop of hardware, software, and tools. More critically, NVIDIA, through technical restrictions such as the CUDA EULA prohibiting cross-hardware emulation and hardware阉割 (e.g., halving the NVLink bandwidth of special edition GPUs), fundamentally blocks the possibility of integrating its computing power with non-NVIDIA architectures. What does this mean? The millions of existing NVIDIA GPUs in China, seemingly assets of computing power, are in reality closed islands. Alibaba's GPU clusters cannot directly interface with Tencent's computing power pool. It is even harder for the single-card computing power of small and medium-sized enterprises to participate in large-scale model training because there is no unified protocol to allow these GPUs to break through CUDA's limitations and achieve cross-entity collaboration. The first step for the Ministry of Industry and Information Technology to achieve computing power interconnection is to face the reality that 70% of the existing computing power is tied to a foreign company. The difficulty of extrication is comparable to rebuilding a new ecosystem on someone else's operating system. 2. The chaotic computing power landscape of small and scattered resources. The dual contradiction between urgent demand and idle waste. "We have 10 single-card servers, but training a medium-sized model requires an eight-card cluster. These single cards are simply useless." This is a common困扰 for many small and medium-sized AI enterprises. The current distribution of domestic computing power resources shows severe structural imbalance. On the demand side, there is an urgent gap: large-scale model training requires high-density, low-latency cluster computing power, such as unified scheduling of 1024 or 2048 cards. However, only the self-built clusters of leading enterprises like Alibaba, Tencent, and ByteDance can provide such computing power domestically, and most of these are prioritized for internal needs. On the supply side, there is a large amount of idle resources. According to industry statistics, the average computing power utilization rate of small and medium-sized domestic data centers is less than 60%. Some smart computing centers in western regions are even semi-idle for long periods due to low local demand and difficulty in cross-regional scheduling. More troublesome is the fragmentation of computing power: different data centers have mixed hardware architectures (GPU, NPU, TPU), interface protocols,各自的 scheduling systems, service standards, billing, and latency commitments are completely inconsistent. For example, an AI company in Guangzhou wants to rent NPU computing power from a data center in Chengdu. It must not only solve the latency issue of cross-regional data transmission but also face the reconstruction cost of adapting the model to different chips. Often, they give up because it is troublesome and uneconomical. This chaos—where computing power is available but unusable, or needed but unfindable—is the core pain point the Ministry of Industry and Information Technology aims to solve. But pulling together these scattered and small computing power resources is far more complex than building a new supercomputing center. 3. External封锁 and internal shortcomings: the dual pressure of domestic computing power breaking through. If the closure of existing GPUs is an internal dilemma, then external technological封锁 adds more resistance to the path of computing power interconnection. The United States' export controls on high-end AI chips have continuously escalated, from initially restricting high-end models like H100 to later adding restrictions on special editions like H800 and A800, and even attempting to block the supply of advanced process chips to China. This means that the domestic approach of alleviating the computing power gap by buying more NVIDIA GPUs is gradually being blocked. The alternative path of domestic computing power is still in the stage of making up for shortcomings. Although Huawei's Ascend has launched the 384-card Cloud Matrix super node, achieving high-bandwidth interconnection through Matrix Link, production capacity and ecosystem maturity still need time. The mass production capability of Ascend chips has not yet met large-scale demand. The number of developers in the CANN open-source ecosystem is about 400,000, only one-tenth of CUDA's over 4 million. Although Google's TPU has advantages in dedicated AI computing, its ecosystem is tied to TensorFlow and limited to Google Cloud services, making it difficult to integrate into domestic computing power networks. This creates dual pressure: in the short term, reliance on imported chips cannot fill the gap; in the long term, the construction of a domestic chip ecosystem takes time. The Ministry of Industry and Information Technology's promotion of computing power interconnection is essentially seeking a path in the gap between external chokeholds and internal immaturity—a path that does not depend on a single vendor and is not bound by technological封锁. But every step of this path faces the dual tests of technological breakthrough and ecological construction. 4. The starting point for breakthroughs in 2026: from computing power identification to the pragmatic first step of the backbone network. Faced with so many difficulties, the Ministry of Industry and Information Technology did not rush to pursue the ideal form of a national unified computing power network. Instead, in the computing power interconnection action plan, it set a pragmatic target for 2026: First, solve the basic problem of enabling interconnection and scheduling, then discuss efficient collaboration. The first step is to unify the language: establish a nationwide unified computing power identification system. Simply put, issue an "ID card" for each piece of computing power, clearly标注 the hardware type (e.g., NVIDIA A800 or Huawei Ascend 910), computing power specifications (e.g., FP16 computing power in Teraops, memory bandwidth, HBM capacity), and service capabilities (e.g., supporting training or inference, latency). With this ID card, computing power from different data centers can be recognized and compared. Just like product parameters on e-commerce platforms, enterprises can clearly judge whether this computing power meets their needs. The second step is to open up channels: promote the upgrade of backbone networks by the three major telecom operators. For cross-regional scheduling of computing power, a highway is first needed. The Ministry of Industry and Information Technology explicitly requires the promotion of 800G and 1400G optical fiber transmission technologies, strengthening national-level internet backbone direct connection points, and building dedicated computing power networks. Currently, China Mobile's 400G and 800G optical modules have achieved domestic mass production, such as products from Ruihai Optoelectronics that have passed Alibaba Cloud certification. These infrastructure upgrades aim to efficiently connect eastern demand with western computing power, reducing cross-regional transmission latency and loss. The third step is to pilot breakthroughs: build national and regional two-level computing power interconnection platforms. Currently, the National Supercomputing Internet Platform has connected over 20 supercomputing and smart computing centers, providing hundreds of billions of computing hours cumulatively. For example, a research team in Shanghai can call computing power from the Lanzhou Supercomputing Center through the platform without building their own cluster. This pilot-first, then scale-up model is to verify the feasibility of cross-regional and cross-architecture computing power scheduling before 2026, accumulating experience for subsequent nationwide interconnection. These measures may seem like basic work, but they are crucial to solving current difficulties. By solving the problem of non-interoperable computing power through identification, inefficient scheduling through backbone networks, and resource mismatches through platforms. Although there is still a gap from the ideal of a national unified computing power network, it at least provides a realistic starting point for utilizing existing computing power and breaking through with domestic computing power. The winding road lies in facing the closed existing ecosystem, the sharp supply-demand contradiction, and the pressure of external封锁 at every step. But the hope for breakthroughs is hidden in these pragmatic first steps. When computing power has a unified ID card, when cross-regional highways are built, and when scattered computing power can be efficiently scheduled, China's computing power industry can truly move from passive dependence to active control. The future is bright: The path of breakthroughs for China's computing power industry. Although China's computing power industry is currently full of thorns on its journey, once the layers of fog are cleared, the dawn is already visible. From NVIDIA's development bottlenecks to the gradual prominence of China's own advantages, and to new opportunities brought by key technological breakthroughs, the future of China's computing power industry is full of hope and bright prospects. NVIDIA's development difficulties: NVIDIA has long dominated the GPU field. However, its development is facing physical bottlenecks. In the pursuit of high single-card computing power to achieve high-density computing, the issue of physical limits is becoming increasingly significant. A single server, limited by energy consumption and the number of PCI slots supported by the motherboard, can accommodate a limited number of GPUs, and cannot endlessly increase GPUs to boost computing power. Even by connecting multiple servers into a cluster through NVLink technology to build a 10,000-card data center, power issues arise. According to relevant data, a 10,000-card computing center consumes electricity equivalent to a medium-sized city. For example, by 2027, 40% of data centers in the United States are expected to be constrained by energy issues. Such massive power demand challenges the sustainable development of data centers. It is difficult to rely on centralized mega-data centers to meet growing computing power needs. NVIDIA GPUs, designed to兼顾 games, displays, mining, and various scientific computations, have significant redundancy. This not only keeps costs high but also makes energy consumption far exceed chips专门 designed for artificial intelligence. When building large-scale computing power centers, this further exacerbates cost and energy burdens. Moreover, the closed-source CUDA system built by NVIDIA, while forming its commercial moat, also limits user freedom and hinders the development of large-cluster computing. For example, it sets numerous obstacles to the application of technologies like virtual devices and virtual buses, restricting the flexible allocation and expansion of computing power. China's computing power industry advantages and opportunities: Policy and infrastructure advantages. The Chinese government has shown strong determination and action in developing the computing power industry. The Ministry of Industry and Information Technology actively promotes computing power interconnection initiatives. The state has sufficient financial resources to support the construction of the next-generation high-speed internet backbone network, using 800G and 1400G optical fibers to connect data centers and achieve interoperability. This measure can not only disperse computing power resources to areas with more balanced electricity supply, solving energy bottlenecks, but also integrate national computing power resources and improve utilization efficiency. In contrast, even the most resource-rich top companies in the United States would find it difficult to complete such large-scale infrastructure construction on their own. Through national overall planning and promotion, China has a leading advantage in computing power infrastructure construction, laying a solid foundation for the long-term development of the computing power industry. Energy and resource allocation advantages: China is actively deploying in the energy sector, vigorously promoting clean energy to provide sustainable energy supply for data centers. In terms of computing power resource distribution, it fully leverages the characteristics of abundant computing power resources in the west and strong demand in the east. By building high-speed interconnection networks, effective connections between central and eastern demands and western computing power centers are achieved. Optimize resource allocation, reduce computing power costs, and improve overall computing power service efficiency. This efficient model of energy and resource allocation is a unique advantage for China's computing power industry development, better supporting the continuous expansion and upgrade of the computing power industry. New opportunities from technological breakthroughs: Huawei's high-speed interconnection bus technology has become a key force in China's computing power industry breakthrough. This technology breaks through the limitations of the traditional von Neumann architecture, connecting remote TPUs, NPUs, and other computing chips through a high-speed interconnection bus, making them operate like local computing cards, thus building a supercomputer. TPUs and NPUs are专门 designed for artificial intelligence, with obvious advantages in cost and optimization efficiency. Although there are currently issues such as insufficient integration and scheduling flexibility. With the continuous development of technology and the application of high-speed interconnection bus technology, a qualitative leap is expected. By integrating a large number of TPUs and NPUs,形成 astronomical numbers of computing cores. Using the high-speed connection of 800G and 1400G optical fibers, surpassing NVIDIA's GPU clusters in computing power is not out of reach. This technological innovation provides a possibility for China to overtake competition in the field of AI computing power, opening up a new development path. In summary, although China's computing power industry currently faces challenges such as NVIDIA's market dominance and external technological封锁. In the long run, NVIDIA's own development bottlenecks provide China with opportunities for breakthroughs. With advantages in policy support, infrastructure construction, energy and resource allocation, and key technological breakthroughs, China's computing power industry has already found the right direction and taken solid steps. As long as it continues to increase R&D investment and improve the industrial ecosystem, China's computing power industry will surely stand out in global competition, ushering in a bright future and providing a steady stream of powerful impetus for the development of emerging industries such as artificial intelligence.
back to top