AI Infrastructure Investment Report 2026

Key Takeaways

  • AI infra spending: 340B by 2027
  • NVIDIA H100/H200 cluster demand exceeds supply through 2026
  • Inference infrastructure overtaking training infrastructure in investment share
  • New category emerging: “Agent Infrastructure” distinct from model infrastructure
  • China sovereign AI infra accelerating post-export controls; Huawei Ascend gaining share
- AI基础设施支出:2025年全球达1800亿美元,预计2027年将达3400亿美元 - 英伟达H100/H200集群需求在2026年前持续供不应求 - 推理基础设施在投资占比上正超越训练基础设施 - 新兴类别:“智能体基础设施”(Agent Infrastructure),区别于模型基础设施 - 中国主权AI基础设施在出口管制后加速发展;华为昇腾份额持续增长

Summary

This investment landscape report (March 2026) maps the AI infrastructure stack and capital flows. The headline finding: inference infrastructure — previously an afterthought — now commands 45% of new AI spending, up from 18% in 2024. This reflects the shift from model research to production deployment.

本投资格局报告(2026年3月)梳理了AI基础设施堆栈与资本流向。核心发现显示:推理基础设施——此前仅被视为次要环节——如今已占据新增AI支出的45%,较2024年的18%显著上升。这反映了重心从模型研究向生产部署的转移。

The “Agent Infrastructure” category is newly distinguished from general model infrastructure. Agent infra includes: orchestration layers (LangChain, LlamaIndex), memory/state stores (Letta, Weaviate), sandbox environments (E2B, Firecracker), and skill/tool registries (AgentSkills.io, Smithery). Investors are betting that agent infra will be stickier than model API access, which faces commoditization pressure.

“智能体基础设施”类别近期从通用模型基础设施中独立划分出来。智能体基础设施包括:编排层(LangChain、LlamaIndex)、记忆/状态存储(Letta、Weaviate)、沙箱环境(E2B、Firecracker)以及技能/工具注册表(AgentSkills.io、Smithery)。投资者押注智能体基础设施将比面临同质化压力的模型 API 访问具有更高的用户粘性。

China sovereign AI infra commentary: Huawei’s Ascend 910B clusters are now used in production at Baidu and ByteDance, reducing H100 dependency. Performance gap versus NVIDIA narrows for inference workloads (vs. training), making Ascend viable for production agent deployments.

中国主权AI基础设施评论:华为昇腾910B集群现已在百度和字节跳动投入生产使用,降低了对H100的依赖。其在推理任务(相较于训练)上与英伟达的性能差距正在缩小,使昇腾成为生产级智能体(Agent)部署的可行方案。

Relevant Concepts

Relevant Entities