在英伟达GPU一卡难求的当下，有一条消息值得认真对待：京东云宣布拟建基于摩尔线程国产GPU的十万卡规模智算集群。
这不是一个「国产替代」的口号，而是一个具体的工程计划。十万卡规模意味着什么？目前全球范围内，十万卡级别的GPU集群屈指可数——微软为OpenAI建的集群、Meta的RSC集群，都是这个量级。京东云如果真的用国产GPU建出十万卡集群，将是国产GPU首次进入头部云厂商的核心算力体系。
为什么是「拟建」而不是「建成」
首先要注意一个关键词：「拟建」。这意味着项目目前处于规划阶段，还没有真正落地。从规划到建成，中间要经过硬件采购、机房建设、网络互联、软件栈适配、稳定性验证等一系列环节，每一个环节都可能遇到意想不到的问题。
但「拟建」本身已经是一个重要信号。它意味着京东云作为头部云厂商，已经认真评估了摩尔线程GPU的可用性，并且认为它有资格进入十万卡级别的核心算力规划。在此之前，国产GPU在云厂商的角色主要是边缘场景的补充——能做一些推理任务、能跑一些轻量训练，但从来没有人把它放到核心位置上。
京东云CEO曹鹏在发言中表示，这是国产算力发展的重要里程碑。摩尔线程创始人兼CEO张建中则说，这次合作标志着GPU从产品形态到云服务形态的跨越。两个人的表态都指向同一个判断：国产GPU正在从「备选方案」变成「主力方案」。
十万卡集群的技术挑战
建十万卡GPU集群，最难的不是买十万张卡。最难的是让十万张卡高效协同工作。
首先是网络互联。十万张GPU之间需要高速、低延迟的网络连接。任何一个节点的通信瓶颈都会拖慢整个集群的训练效率。英伟达之所以在大模型训练中不可替代，很大程度上是因为它的NVLink和InfiniBand技术提供了极致的互联带宽。摩尔线程的GPU能不能提供同等级别的互联能力，是一个核心的技术问号。
其次是软件栈。大模型训练不仅需要GPU硬件，还需要一整套软件生态——驱动、编译器、分布式训练框架、通信库。英伟达的CUDA生态是十多年积累的结果，国产GPU的软件栈成熟度还远远不够。京东云和摩尔线程需要在软件适配上投入大量资源。
第三是稳定性。十万卡级别的集群，每天都会有一定比例的硬件故障。如何在大规模故障频繁发生的情况下保证训练任务不中断、不退化，是工程上的巨大挑战。
国产算力的「用」比「造」更重要
过去几年，国产GPU芯片的讨论焦点一直集中在「能不能造出来」。摩尔线程、海光信息、寒武纪、壁仞科技等公司都在各自的技术路线上推进。但「造出来」只是第一步，更关键的是「有人用」。
芯片是一个典型的规模效应产品：用的人越多，迭代越快，成本越低，生态越完善。如果国产GPU只停留在实验室和小规模部署，它永远无法真正成熟。京东云这次把摩尔线程GPU放到十万卡级别的规划中，意味着国产GPU终于有机会在真实的大规模生产环境中被检验和迭代。
这对整个国产算力生态的意义，可能比任何一次芯片发布会都要大。
明天见。
In an era where NVIDIA GPUs are impossible to get, one piece of news deserves serious attention: JD Cloud has announced plans to build a 100K-card-scale AI compute cluster based on MThreads domestic GPUs.
This isn't a "domestic substitution" slogan — it's a concrete engineering plan. What does 100K-card scale mean? Globally, GPU clusters at this scale are few and far between — Microsoft's cluster for OpenAI, Meta's RSC cluster, both in this range. If JD Cloud actually builds a 100K-card cluster with domestic GPUs, it will be the first time a domestic GPU enters a top cloud provider's core compute stack.
Why "Planned" Instead of "Built"?
First, pay attention to a key word: "planned." This means the project is currently in the planning phase, not yet realized. From planning to completion, there are hardware procurement, data center construction, network interconnection, software stack adaptation, stability verification, and more — every step could encounter unexpected problems.
But "planned" itself is already a significant signal. It means JD Cloud, as a top cloud provider, has seriously evaluated MThreads GPUs' viability and deemed them qualified for 100K-card-level core compute planning. Previously, domestic GPUs played a supplementary role in cloud providers — handling some inference tasks, running lightweight training — but no one had placed them in a core position.
JD Cloud CEO Cao Peng stated this is a significant milestone for domestic compute development. MThreads founder and CEO Zhang Jianzhong said the partnership marks a leap from GPU-as-product to GPU-as-cloud-service. Both statements point to the same judgment: domestic GPUs are transitioning from "backup option" to "primary option."
Technical Challenges of a 100K-Card Cluster
Building a 100K-card GPU cluster, the hardest part isn't buying 100K cards. The hardest part is making 100K cards work together efficiently.
First is network interconnection. 100K GPUs need high-speed, low-latency network connections. Any node's communication bottleneck slows the entire cluster's training efficiency. NVIDIA's dominance in large model training is largely because its NVLink and InfiniBand technologies provide extreme interconnect bandwidth. Whether MThreads GPUs can provide comparable interconnection capabilities is a core technical question.
Second is the software stack. Large model training needs not just GPU hardware but an entire software ecosystem — drivers, compilers, distributed training frameworks, communication libraries. NVIDIA's CUDA ecosystem is the result of over a decade of accumulation. Domestic GPU software stack maturity is still far behind. JD Cloud and MThreads will need to invest heavily in software adaptation.
Third is stability. At 100K-card scale, a certain percentage of hardware failures happen every day. Ensuring training tasks don't get interrupted or degraded amid frequent large-scale failures is a massive engineering challenge.
"Using" Domestic Compute Matters More Than "Building" It
Over the past few years, the discussion around domestic GPU chips has focused on "can they be built." Companies like MThreads, Hygon, Cambricon, and Biren Technology are advancing on their respective technical paths. But "building it" is only step one. The more critical question is "will anyone use it."
Chips are a classic scale-economy product: more users lead to faster iteration, lower costs, and a more complete ecosystem. If domestic GPUs stay in labs and small-scale deployments, they'll never truly mature. JD Cloud placing MThreads GPUs in 100K-card planning means domestic GPUs finally have the chance to be tested and iterated in real, large-scale production environments.
This may matter more for the entire domestic compute ecosystem than any chip launch event.
See you tomorrow.
国产GPU正在从「备选方案」变成「主力方案」——用的人越多，迭代越快，生态越完善。
—— Dawn Vision编辑部
Domestic GPUs are transitioning from "backup option" to "primary option" — more users means faster iteration and a more complete ecosystem.
— The Dawn Vision Editorial Desk
JD Cloud · MThreads · domestic GPU · 100K GPU cluster · AI compute center · compute infrastructure · Cao Peng · Zhang Jianzhong · CUDA ecosystem
Sources · 信源 Sources
本文基于 Dawn Vision 认知引擎处理的 9 个源信号生成，经编辑部人工审核。素材来源：京东云官方、C114通信网、财联社。
This article was generated by the Dawn Vision cognitive engine processing 9 source signals, with human editorial review. Sources: JD Cloud Official, C114, CLS.
