Neocloud

Neocloud (short for "new cloud", also called a GPU cloud or AI cloud) is a category of cloud service provider whose business is built almost entirely around renting out accelerated computing — clusters of GPUs and other AI accelerators — rather than the broad mix of storage, databases and virtual machines offered by traditional hyperscalers.^[1] Neoclouds sell GPU-as-a-Service (GPUaaS) on demand or under reserved contracts, and their customers are typically AI laboratories, model developers, enterprises fine-tuning large models and research institutions.

History

The term gained usage in 2023 and 2024, when demand for NVIDIA GPUs for large language model training far outstripped the capacity that Amazon, Microsoft and Google could supply, and long waiting lists pushed buyers toward specialists. CoreWeave, which began as a cryptocurrency mining operator, converted its data centres to AI compute and listed on Nasdaq in March 2025; Nebius, Lambda, Crusoe and Nscale expanded rapidly alongside it.^[2]

Synergy Research Group recorded neocloud revenues of roughly US$9 billion in the fourth quarter of 2025 — up 223% year over year — and more than US$25 billion for the full year, forecasting the segment would approach US$400 billion by 2031 on a sustained 58% compound annual growth rate.^[1] ABI Research separately estimated GPUaaS revenue from neoclouds at US$42 billion in 2025, rising above US$250 billion by 2030.^[2]

Technology and Business Model

Neoclouds differ from hyperscalers in three practical ways. First, they operate dense GPU clusters — racks of accelerators interconnected with high-bandwidth fabrics such as InfiniBand or RoCE and usually liquid-cooled — rather than general-purpose server fleets. Second, they offer elastic AI capacity through on-demand instances, reserved commitments and managed Kubernetes or Slurm schedulers tuned for distributed training. Third, their economics rest on long-term capacity contracts and heavy capital expenditure: CoreWeave reported second-quarter 2026 revenue of US$2.575 billion, up 112% year over year, with a backlog of roughly US$104 billion and full-year capital expenditure guidance of US$35–39 billion.^[3] Crusoe reported 4.9 gigawatts of contracted AI infrastructure across five United States campuses in June 2026, with a pipeline above 40 GW.^[4]

Published pricing is typically lower than list prices for equivalent hyperscaler instances; Lambda, for example, publishes on-demand rates for single GPUs and clusters directly rather than through sales-quoted enterprise agreements.^[5] The trade-offs are concentration risk — most neoclouds depend on a single chip supplier, NVIDIA, and on a handful of very large customers — and geographic footprint, since the majority of capacity sits in North America and Europe.^[2] Groq illustrates the category's volatility: after licensing its inference technology to NVIDIA in December 2025, the company refocused as an inference neocloud and raised US$350 million in August 2026.^[6]

Applications

Neocloud capacity is used for pre-training and fine-tuning foundation models, batch inference for AI products, and high-performance computing workloads such as drug discovery and climate modelling. Enterprises also use it as burst capacity alongside their primary cloud, and AI laboratories rent superclusters for single large training runs without owning a data centre.

>See Also

• GPU Cluster
• AI Data Centres in Malaysia
• Sovereign AI
• NVIDIA
• National AI Compute Exchange
• Liquid Cooling