9月11日，2026中国算力大会在河北廊坊开幕。工信部发布的《全国一体化算力网建设规划》给出了一句关键定调："有序部署万卡、十万卡及以上智算集群"。
这不是口号，是已经在发生的事。2025年，中国已建成万卡智算集群42个，智能算力规模超过1590EFLOPS。2026年的目标是落地50+万卡集群，实现10万卡集群从0到1的突破。中科曙光的曙光8000（登峰）已经落成，成为首个全国产十万卡AI超集群。
国产芯片"爆单"：需求超供给10倍
算力扩张的背后，是国产AI芯片的供不应求。
根据央视财经数据，2026年国产AI芯片需求约400万颗，交付约300万颗，百万级缺口。高端芯片更是"一卡难求"——单排产周期长达3年。寒武纪、摩尔线程、壁仞等国产芯片厂商的营收都在高速增长。
华为的昇腾950PR芯片已经成为国产算力的"顶流"。据英国《金融时报》报道，华为预计2026年AI芯片收入将高达120亿美元，较2025年的75亿美元大幅增长60%以上。
"枢纽—区域—边缘"：算力网的三层架构
工信部的规划明确提出构建"枢纽—区域—边缘"多层次算力设施体系。这意味着算力不再是"建一个大机房"的问题，而是一个分布式网络的问题。
枢纽层负责大规模训练，区域层负责中等规模推理，边缘层负责实时响应。这种分层架构可以让算力资源更高效地匹配不同场景的需求——训练任务去枢纽，推理任务去区域，实时任务去边缘。
芯模协同：国产生态的闭环
规划特别强调了"加大力度适配国产算力芯片"。这意味着国家层面正在推动芯模协同——让国产芯片和国产大模型深度适配，形成闭环生态。
渤海证券指出，国内头部国产大模型企业正在深化芯模协同开发，将国产算力芯片作为支撑模型推理的核心算力硬件底座。上海证券也认为，在海外芯片受限的背景下，国产算力芯片将成为AI基础设施的核心。
"十万卡集群不是终点，是起点。当算力成为国家基础设施，整个AI产业的格局都会被重塑。"—— 一位算力行业分析师
对AI产业的影响
算力扩张对AI产业的影响是多方面的。
对于大模型公司来说，更多的算力意味着可以训练更大的模型、处理更多的推理请求。对于AI应用公司来说，算力成本的下降意味着更多AI应用可以落地。
但最大的影响可能是国产替代的加速。当国产芯片能够满足十万卡集群的需求时，中国AI产业对海外芯片的依赖将大幅降低。这不仅是一个技术问题，更是一个战略问题。
明天见。
On September 11, the 2026 China Compute Conference opened in Langfang, Hebei. The MIIT's 'National Integrated Computing Power Network Construction Plan' delivered a key statement: 'orderly deployment of 10K, 100K+ GPU intelligent computing clusters.'
This isn't a slogan — it's already happening. In 2025, China built 42 clusters with 10K+ GPUs, with intelligent computing capacity exceeding 1,590 EFLOPS. The 2026 target is deploying 50+ 10K-GPU clusters and achieving breakthroughs from 0 to 1 in 100K-GPU clusters. Sugon's Aurora 8000 (Dengfeng) is already operational as the first fully domestic 100K-GPU AI supercluster.
Domestic Chips 'Order Explosion': Demand Exceeds Supply 10x
Behind computing expansion is the supply shortage of domestic AI chips.
According to CCTV Finance, 2026 domestic AI chip demand is approximately 4 million units, with deliveries around 3 million — a million-unit gap. High-end chips are even harder to get, with single-order production cycles reaching 3 years. Domestic chip makers like Cambricon, Moore Threads, and Biren are all seeing rapid revenue growth.
Huawei's Ascend 950PR chip has become the 'top star' of domestic computing. According to the Financial Times, Huawei expects 2026 AI chip revenue to reach $12 billion, up over 60% from $7.5 billion in 2025.
'Hub-Region-Edge': Three-Layer Computing Architecture
The MIIT plan explicitly proposes building a 'hub-region-edge' multi-level computing facility system. This means computing is no longer about 'building one big data center,' but about a distributed network.
The hub layer handles large-scale training, the region layer handles medium-scale inference, and the edge layer handles real-time responses. This layered architecture lets computing resources more efficiently match different scenario demands.
Chip-Model Synergy: Domestic Ecosystem Loop
The plan emphasizes 'increasing efforts to adapt domestic computing chips.' This means the national level is pushing chip-model synergy — deep adaptation between domestic chips and domestic large models, forming a closed-loop ecosystem.
Bohai Securities notes that leading domestic large model companies are deepening chip-model collaborative development, using domestic computing chips as core hardware for model inference. Shanghai Securities also believes that with overseas chip restrictions, domestic computing chips will become AI infrastructure's core.
"100K-GPU clusters aren't the end — they're the beginning. When computing becomes national infrastructure, the entire AI industry landscape will be reshaped."— A computing industry analyst
Impact on AI Industry
Computing expansion impacts the AI industry in multiple ways.
For large model companies, more computing means training larger models and handling more inference requests. For AI application companies, falling computing costs mean more AI applications can be deployed.
But the biggest impact may be accelerating domestic substitution. When domestic chips can meet 100K-GPU cluster demands, China's AI industry dependence on overseas chips will significantly decrease. This isn't just a technical issue — it's a strategic one.
See you tomorrow.
十万卡集群不是终点，是起点。当算力成为国家基础设施，整个AI产业的格局都会被重塑。
—— 一位算力行业分析师
100K-GPU clusters aren't the end — they're the beginning. When computing becomes national infrastructure, the entire AI industry landscape will be reshaped.
— A computing industry analyst
China Compute Conference · 100K GPU clusters · domestic chips · MIIT · computing network · Sugon 8000 · Huawei Ascend · chip-model synergy
Sources · 信源 Sources
本文基于 Dawn Vision 认知引擎处理的 15 个源信号生成，经编辑部人工审核。素材来源：中新网、光明网、新华网、中国经济网、央视财经。
This article was generated by the Dawn Vision cognitive engine processing 15 source signals, with human editorial review. Sources: CNS, Guangming, Xinhua, CE.cn, CCTV Finance.
