15吉瓦（GW）是什么概念？大约相当于1100万户家庭的日常用电，大约等于七座大型核电站的输出功率。这是SK电讯新成立的SK Hyper公司到2035年的AI数据中心建设目标。
仅仅一周前，英伟达和SK集团刚刚签署了5000亿美元的AI联盟，其中包括2GW的数据中心计划。现在SK电讯把目标直接拉到了15GW——翻了七倍多。与此同时，英伟达被曝拟为OpenAI的数据中心项目提供2500亿美元融资担保；7月28日当天，中国在雄安新区启动了"人工智能大模型IPv6能力提升专项行动"。
算力竞赛的规模和层级，正在以肉眼可见的速度升级。
从GW到国家战略：算力竞赛的三级跳
回顾2026年的算力基建新闻，你会看到一个清晰的数字膨胀轨迹。
几个月前，一个1GW的数据中心还是大新闻。Anthropic签了190亿美元20年的数据中心长约被认为是行业里程碑。英伟达和SK集团宣布2GW合作是封面级新闻。但到了这周，SK电讯的15GW计划直接把量级拉到了一个新台阶。
15GW不是一个公司能独立完成的事。它需要巨量的电力供应（相当于数个核电站的发电量）、大规模的土地资源、政府审批和支持、以及千亿甚至万亿美元级的资金投入。这已经不是企业级的基础设施投资，这是国家级别的能源和产业规划。
英伟达为OpenAI提供2500亿美元融资担保的消息也同样耐人寻味。这意味着什么？意味着芯片公司已经不仅仅是卖芯片了——它们在为客户提供买芯片、建数据中心的全链条融资支持。英伟达正在变成AI基础设施时代的"银行"，它赌的是整个AI生态的未来。
再看中国。7月28日在雄安新区启动的"人工智能大模型IPv6能力提升专项行动"，由中央网信办联合北京、上海、浙江、深圳四地网信办，会同5家头部大模型企业共同发起。这个行动的核心目标是推动生成式大模型应用全面支持IPv6，为智能时代筑牢网络根基。雄安新区同时宣布全面推进IPv6单栈部署，到2030年新建片区实现IPv6单栈规模化运行。这看起来是网络层的动作，但本质上是在为AI时代的网络基础设施做准备——AI Agent之间的大规模通信、海量设备的互联互通，都需要IPv6的地址空间和网络能力来支撑。
智能体走向终端：云端不是唯一答案
但算力竞赛不只是在云端一个方向上狂奔。量子位今天的文章指出了另一个同样重要的趋势：智能体正在走向终端。
全球首个Agentic扩散模型发布，实现了边行动边纠错，128K上下文长度追平自回归模型。端侧推理能力的突破意味着越来越多的AI计算可以在手机、PC、IoT设备上完成，不需要全部回传到云端。这对算力格局的影响是深远的：未来的算力架构不是"一切上云"，而是端云协同——简单任务在本地处理，复杂任务才调用云端算力。
这也解释了为什么高通、苹果、华为等芯片和终端厂商都在拼命做端侧AI——谁掌握了终端算力，谁就掌握了AI触达用户的最后一米。云端算力是"重工业"，终端算力是"轻工业"，两者缺一不可。
从企业买卡、到财团建数据中心、到国家规划算力网络——算力竞争的维度正在不断扩大。最终决定胜负的可能不是谁建的数据中心最大，而是谁能最有效率地在云端和终端之间分配算力、谁的电最便宜、谁的网络最通畅。
AI算力的"国家规划"时代已经开始。这不是某一家公司的战争，甚至不是某一个行业的战争——这是国家之间在基础设施层面的全面竞争。
明天见。
What does 15 gigawatts (GW) look like? Roughly the daily electricity needs of 11 million households. Roughly the output of seven large nuclear power plants. That's the AI data center target by 2035 for SK Telecom's newly launched SK Hyper subsidiary.
Just one week ago, Nvidia and SK Group signed a $500 billion AI alliance that included 2GW of data center plans. Now SK Telecom has pulled the target to 15GW — more than seven times larger. Meanwhile, Nvidia is reportedly providing a $250 billion financing guarantee for OpenAI's data center projects. And on July 28, China launched its "AI Large Model IPv6 Capability Enhancement Initiative" in Xiong'an New Area.
The scale and level of the compute race is escalating at a visible pace.
From GW to National Strategy: The Compute Race Triple Jump
Looking back at compute infrastructure news in 2026, you can see a clear trajectory of ballooning numbers.
A few months ago, a 1GW data center was still big news. Anthropic signing a $19 billion, 20-year data center deal was considered an industry milestone. Nvidia and SK Group announcing 2GW cooperation was cover-story-level news. But this week, SK Telecom's 15GW plan pulls the order of magnitude to a completely new level.
15GW isn't something a single company can pull off independently. It requires enormous power supply (equivalent to several nuclear plants), massive land resources, government approval and support, and hundreds of billions if not trillions in capital. This is no longer enterprise-level infrastructure investment — this is national-level energy and industrial planning.
Nvidia's reported $250 billion financing guarantee for OpenAI is equally revealing. What does it mean? It means chip companies aren't just selling chips anymore — they're providing full-chain financial support for customers to buy chips and build data centers. Nvidia is becoming the "bank" of the AI infrastructure era, betting on the future of the entire AI ecosystem.
Then look at China. The "AI Large Model IPv6 Capability Enhancement Initiative" launched in Xiong'an on July 28 is co-sponsored by the Central Cyberspace Administration together with Beijing, Shanghai, Zhejiang, and Shenzhen, plus five top LLM companies. The core goal is to push generative AI applications to fully support IPv6, building network foundations for the intelligent era. Xiong'an also announced full IPv6 single-stack deployment, with new districts achieving full-scale IPv6 single-stack operation by 2030. This looks like a network-layer move, but it's essentially preparing the network infrastructure for the AI era — large-scale communication between AI agents and interconnection of massive devices all require IPv6's address space and network capabilities.
Agents Move to Edge Devices: Cloud Isn't the Only Answer
But the compute race isn't just sprinting in the cloud direction. QbitAI's article today points to another equally important trend: agents are moving to edge devices.
The world's first agentic diffusion model was released, capable of acting and self-correcting with 128K context length matching autoregressive models. Breakthroughs in on-device inference mean more AI computation can happen on phones, PCs, and IoT devices without round-tripping to the cloud. The impact on compute architecture is profound: the future isn't "everything to the cloud" but edge-cloud coordination — simple tasks processed locally, complex tasks invoking cloud compute.
This explains why Qualcomm, Apple, Huawei and other chip and device makers are all pushing hard for on-device AI — whoever controls edge compute controls the last meter between AI and users. Cloud compute is "heavy industry," edge compute is "light industry," and both are essential.
From companies buying chips, to consortiums building data centers, to nations planning compute networks — the dimensions of compute competition keep expanding. Ultimately, victory may not go to whoever builds the largest data center, but whoever can most efficiently distribute compute between cloud and edge, whoever has the cheapest electricity, whoever has the smoothest networks.
The "national planning" era of AI compute has begun. This isn't a war between companies, or even between industries — it's full-scale competition between nations at the infrastructure level.
See you tomorrow.
算力竞赛正在从"谁的GPU多"升级到"谁的电便宜、谁的地多、谁的网络通"。AI基建的终局是国家规划，不是企业采购。
—— Dawn Vision编辑部
The compute race is escalating from 'who has more GPUs' to 'who has cheaper power, more land, better networks.' The endgame of AI infrastructure is national planning, not corporate procurement.
—— The Dawn Vision Editorial Desk
SK电讯 · SK Hyper · 15GW · 英伟达2500亿担保 · IPv6专项行动 · 算力基建 · 国家规划 · 端云协同 · Agent终端化 · 雄安新区
SK Telecom · SK Hyper · 15GW · Nvidia $250B guarantee · IPv6 initiative · compute infrastructure · national planning · edge-cloud coordination · agents to edge · Xiong'an
Sources · 信源 Sources
本文基于 Dawn Vision 认知引擎处理的 10 个源信号生成，经编辑部人工审核。素材来源：爱范儿、36氪、量子位。
This article was generated by the Dawn Vision cognitive engine processing 10 source signals, with human editorial review. Sources: ifanr, 36Kr, QbitAI.
