Neocloud

A neocloud is a cloud provider that rents out GPU computing power specialized for AI training and inference, rather than the general-purpose services of hyperscalers like AWS or Azure. Neoclouds such as CoreWeave, Nebius and IREN run GPU-dense data centers wired with high-speed interconnects, offering faster access to the newest chips, flexible terms, and often lower prices.

What a neocloud is, in plain terms

A **neocloud** is a cloud company that does essentially one thing: rent out GPU computing power for artificial intelligence. Traditional hyperscalers (Amazon AWS, Microsoft Azure, Google Cloud) offer hundreds of general-purpose services. A neocloud strips that down to a single focus, GPU compute for training and running AI models, and optimizes everything around it. The name signals a *new kind* of cloud, purpose-built for the AI era rather than the web-app era. Because they specialize, neoclouds often provide faster access to the latest GPU generations, more flexible contracts, and pricing that can undercut hyperscalers, sometimes substantially, for comparable raw compute.

How a neocloud actually works

Under the hood, a neocloud packs data centers with thousands of GPUs (today mostly NVIDIA's Hopper H100/H200 and newer Blackwell GB200 and GB300 systems) and wires them together with very fast networking, NVIDIA NVLink inside a rack and InfiniBand or high-speed Ethernet between racks, so many GPUs can act as one large machine for training a single model. On top of the hardware sits an orchestration layer: customers spin up GPU clusters on demand via API, get pre-configured machine-learning environments, and can scale from a handful to thousands of GPUs quickly. The business model is capacity rental, usually a mix of on-demand usage and large multi-year reserved contracts that underwrite the enormous upfront cost of the chips and buildings.

Why neoclouds matter for AI

Modern AI is bottlenecked by compute. Training frontier models and serving billions of inference requests requires more GPUs than the hyperscalers can build out on their own, and AI labs want guaranteed, dedicated capacity. Neoclouds emerged to fill that gap. The category has grown explosively: industry estimates put neocloud revenue above \$25 billion in 2025, and by some counts roughly one in three AI workloads now runs on a neocloud rather than a traditional hyperscaler. The biggest AI buyers, OpenAI, Microsoft and Meta, have signed multibillion-dollar, multi-year deals with neoclouds to secure the compute they cannot build fast enough themselves.

Where neoclouds sit in the supply chain

A neocloud is the layer between the chipmaker and the AI lab. NVIDIA (with rivals like AMD) designs the GPUs; a deep supply chain, TSMC for fabrication, memory makers for HBM, plus server, power and cooling vendors, builds the systems; and neoclouds assemble these into operational data centers that AI companies rent. Networking and **photonics** are an increasingly critical part of that stack. As clusters scale to hundreds of thousands of GPUs, NVIDIA is pushing **co-packaged optics (CPO)**, moving from copper to light for chip-to-chip communication, with its Quantum-X InfiniBand and Spectrum-X Photonics switches rolling out through 2026. NVIDIA frames optical interconnects not as an upgrade but as a structural requirement for next-generation AI data centers.

Who the key players are

**CoreWeave (CRWV)** is the benchmark pure-play. Founded in 2017 as a crypto-mining startup, it pivoted to GPU cloud, IPO'd in March 2025, and reached \$5.1 billion in 2025 revenue (up ~170%), with backlog near \$100 billion from customers including OpenAI, Microsoft and Meta. **Nebius (NBIS)** is the leading full-stack challenger, spun out of Russia's Yandex and relisted on Nasdaq in late 2024. It builds a vertically integrated AI cloud and reported Q1 2026 revenue of \$399 million, up 684% year over year. **IREN (IREN)** transitioned from Bitcoin mining to AI infrastructure, validated by a direct NVIDIA cloud-services contract and a multibillion-dollar GPU-cloud deal. Other names in the space include Lambda, Nscale, Crusoe and Applied Digital (APLD).

What's changing now

Two shifts define the current phase. First, the competition is moving from a GPU race to a **power war**: securing gigawatts of electricity and the land, grid connections and cooling to run them has become the binding constraint. CoreWeave and Nebius have each contracted multiple gigawatts of capacity. Second, the sector runs on **circular financing**. NVIDIA invests directly in neoclouds (a multibillion-dollar stake in CoreWeave, \$2 billion in Nebius, backing for others) while also being their main chip supplier and, in some cases, a customer renting back capacity. This tight loop has fueled extraordinary growth but also draws scrutiny over how durable the demand and the financing really are.

Frequently asked

How is a neocloud different from AWS or Azure?

Hyperscalers like AWS and Azure are general-purpose, offering thousands of services for almost any computing need. A neocloud focuses narrowly on GPU compute for AI, which lets it optimize for that single workload, often delivering newer hardware faster, more flexible terms, and lower prices for raw compute, while leaving the broad service catalog to the hyperscalers.

What are the main neocloud stocks?

The most-discussed publicly traded neoclouds are CoreWeave (CRWV), Nebius Group (NBIS), and IREN (IREN). Applied Digital (APLD) is often grouped with them, and several large players such as Lambda, Crusoe and Nscale remain private. These are example companies for context, not investment advice.

Why is NVIDIA investing in its own customers?

NVIDIA takes equity stakes in neoclouds like CoreWeave and Nebius to help finance the buildout of capacity that runs on its GPUs, ensuring demand for its chips and a healthy ecosystem of buyers. Critics call this 'circular financing,' since NVIDIA is simultaneously supplier, investor, and sometimes customer.

What GPUs do neoclouds use?

Mostly NVIDIA data-center GPUs: the Hopper generation (H100, H200) and the newer Blackwell platform, including GB200 NVL72 racks and the GB300 (Blackwell Ultra) systems ramping through 2026. Some also deploy AMD Instinct accelerators. The newest chips are scarce, so early access is a key competitive advantage.

Is the neocloud boom sustainable?

It depends on whether AI compute demand keeps growing fast enough to justify the spending. Bulls point to massive multi-year contracts from OpenAI, Microsoft and Meta and persistent GPU scarcity. Skeptics worry about the circular financing, heavy debt loads tied to GPUs, and the cost of securing enough power, the sector's emerging bottleneck.

Related companies