全球AI芯片算力正处于快速扩张阶段。当前全球AI芯片总算力约为2000万枚英伟达H100芯片等效规模，预计每9个月翻一倍，到2028年年底将达到约2亿枚芯片的等效算力，相当于目前规模的10倍[1]。
五大科技巨头正在推动前所未有的基础设施建设规模。亚马逊、Alphabet、Meta、微软和甲骨文五家公司2026年的资本开支预计合计约7500亿美元，占其全年总收入的38%[1]。IDC预测AI基础设施支出将从2025年的3180亿美元增长到2029年超过1万亿美元[1]。在美国多个地区正在建设的大型数据中心中，Colossus 2（孟菲斯）最大，算力约111.2万枚芯片等效，微软Fairwater Atlanta约76.9万枚，Meta Prometheus约76.3万枚，亚马逊New Carlisle约68.6万枚，OpenAI Stargate Abilene约50.9万枚[1]。
这一轮算力投资正在驱动AI模型与应用的迭代突破。Remote Labor Index数据显示，领先模型的工作完成率在2025年10月为2.5%，到2026年7月Claude Fable 5已达16.1%[1]。Claude Fable 5推翻了存在87年的广义雅可比猜想[1]。同时，GPT-5.6 Sol优化后端到端服务成本下降20%，Token生成效率提升超15%，GPT-5.6 Luna同时下调输入输出价格80%[1]。Gartner预计2026年AI云基础设施支出中55%将用于推理，到2029年这一比例将超过65%[1]。
Global artificial intelligence computing power is experiencing rapid acceleration, with total capacity projected to reach approximately 200 million NVIDIA H100-equivalent chips by the end of 2028, representing a tenfold increase from current levels [1]. Currently, worldwide AI chip computing capacity stands at approximately 20 million H100-equivalent chips and is expected to double every nine months [1].
The largest data centers now under development span multiple locations across the United States, with Colossus 2 in Memphis leading at approximately 1.112 million H100-equivalent chips [1]. Microsoft's Fairwater facility in Atlanta reaches 769,000 equivalent chips, followed by Meta's Prometheus at 763,000, Amazon's New Carlisle at 686,000, and OpenAI's Stargate facility in Abilene at 509,000 [1]. This massive infrastructure buildout is being driven by five major technology companies—Amazon, Alphabet, Meta, Microsoft, and Oracle—which are projected to collectively spend approximately 750 billion dollars on capital expenditures in 2026, equivalent to 38 percent of their annual total revenues [1]. Industry research firm IDC forecasts that AI infrastructure spending will grow from 318 billion dollars in 2025 to over 1 trillion dollars by 2029 [1].
The expansion of computing capacity is enabling advances in AI model performance and efficiency. According to the Remote Labor Index, leading models achieved a 2.5 percent task completion rate in October 2025, rising to 16.1 percent by July 2026 with Claude Fable 5 [1]. Claude Fable 5 resolved the generalized Jacobi conjecture, a mathematical problem unsolved for 87 years [1]. GPT-5.6 Sol improved end-to-end service costs by 20 percent and increased token generation efficiency by over 15 percent, while GPT-5.6 Luna reduced both input and output pricing by 80 percent [1]. Gartner projects that AI cloud infrastructure spending will increasingly focus on inference tasks, with inference accounting for 55 percent of expenditures in 2026 and expected to exceed 65 percent by 2029 [1].
