Podcast Atlas
concept Updated 2026-08-18 Topics: Technology

Neo Cloud

The Future of Everything: What CEOs of Circle, CrowdStrike & More See Coming in 2026 adds Crusoe as an energy-first neocloud case. The source places Crusoe among early neoclouds with CoreWeave, but makes abundant power, data-center construction, gas turbines, batteries, and customer leases the differentiating layer.

贾扬清：我所经历的「人工智能已死」到「AI 颠覆世界」的数年巨变丨串台「声东击西」S10E24 adds Lepton AI as a founder-operator view of the neocloud category. Jia Yangqing distinguishes AI cloud from traditional cloud by the need for tightly connected GPU capacity, scheduling, model-serving layers, and hardware-software fit.

E246｜何谓蒸馏？聊聊硅谷如何看中国开放模型逼近前沿 adds the open-model competition version. 王铁镇 argues that neoclouds can use strong open-weight models such as Kimi K3 to compete against closed API providers, because the serving layer can focus on hardware efficiency, scheduling, model hosting, and lower token prices rather than owning a closed frontier model.

Neo cloud is the AI-native GPU-cloud model discussed in E230｜1万亿收入预期背后：英伟达的巅峰与软肋. Alex contrasts neoclouds with hyperscalers: hyperscalers grew from CPU and storage cloud and often expose VM-oriented abstractions, while neoclouds are more likely to use k8s clusters and bare-metal access to preserve GPU efficiency.

The concept belongs under MaaS Infrastructure because customers are not only renting machines. In the source, a stronger neocloud offers earlier Nvidia GPU access, cluster management, model services, and kernel optimization so AI teams can convert accelerators into useful training and inference capacity.

Key Claims

  • Neoclouds compete on AI workload fit rather than generic cloud breadth.
  • Bare-metal efficiency can matter when virtualization overhead reduces expensive GPU utilization.
  • K8s cluster management, model services, and kernel optimization can turn raw hardware into a more defensible product.
  • Neoclouds still face land and power, supply-chain, and SLA constraints.
  • Strong open weights can let neoclouds sell model serving and optimization without first building a proprietary frontier model.
  • Crusoe adds that the category can compete on energy development and project execution, not only GPU scheduling or model hosting.

Connections