AI Cloud Providers: The Neoclouds Powering the AI Boom

The AI boom created a new kind of cloud company. Alongside Amazon Web Services, Microsoft Azure and Google Cloud, a group of specialised providers, often called neoclouds, now rent graphics processing units (GPUs) and AI computing to model developers and businesses. Some build huge data centres for training frontier models; others offer software platforms that run AI models efficiently for applications. CoreWeave's public listing in 2025 showed how large this market could be, and private neoclouds have since raised billions at soaring valuations. This guide ranks the private AI cloud providers in our database, explains the different business models and covers the risks, from debt to customer concentration.

Every company in the table links to a full profile with its sourced valuation, funding history and investors. We include private companies that provide AI computing as a service, from GPU capacity to inference platforms. Public companies such as CoreWeave and Nebius are discussed for context. Data infrastructure companies such as VAST Data are covered in our guide to AI infrastructure companies.

The ranking

Two kinds of AI cloud

AI cloud providers fall broadly into two groups. Capacity providers, such as Crusoe, Nscale, Fluidstack and Lambda, build or lease data centres, buy GPUs and rent large clusters, often to AI labs training frontier models or to companies running large workloads. Their businesses resemble infrastructure investment: capital intensive, financed with debt, and dependent on long-term contracts. Platform providers, such as Baseten, Together AI, Modal, Fal.ai and Fireworks AI, focus on software that makes AI models run quickly and cheaply, especially for inference, the process of using trained models to answer requests. They may rent GPUs from capacity providers or hyperscalers and charge customers per use. The line between the two is blurring as companies expand.

Capacity providers: Crusoe, Nscale, Fluidstack and Lambda

Crusoe, valued at US$30.9 billion, began by using stranded energy to power computing and became a major developer of AI data centres, including the Abilene, Texas campus associated with OpenAI's Stargate project. Nscale, based in London and valued at about US$14.6 billion, is building AI data centres in Europe, including in Norway. Fluidstack, also in London, raised an US$830 million Series A at US$7.5 billion and was reported to be in talks at a much higher valuation in 2026. Lambda, one of the earliest GPU cloud companies, was reported at about US$6 billion and signed a large deal with Microsoft. These companies compete on securing power, chips and sites quickly.

Inference platforms: Baseten, Together AI, Fireworks AI and Modal

Baseten, valued at US$13 billion, helps companies deploy and scale AI models in production with high reliability. Together AI, at US$8.3 billion, offers a cloud for open-source models, from inference to fine-tuning and training. Fireworks AI, founded by PyTorch veterans, focuses on fast, efficient inference. Modal Labs provides serverless computing that lets developers run AI code without managing servers, and has grown with demand for sandboxes for AI agents. As AI applications scale, inference spending is expected to exceed training spending, which benefits these platforms.

Generative media: Fal.ai

Fal.ai specialises in running image, video and audio models, and raised at US$4.5 billion in December 2025 with reported annualised revenue above US$200 million. Its focus on one workload shows how specialisation can win in a crowded market.

Why neoclouds emerged

When demand for AI computing exploded after ChatGPT's launch, Nvidia's GPUs were scarce and hyperscalers could not build fast enough. Neoclouds moved quickly, buying GPUs, signing data centre leases and offering capacity to AI labs. Nvidia supported this, allocating chips to neoclouds and sometimes investing in them, which diversified its customer base. Neoclouds also offered configurations optimised for AI, such as high-speed networking between GPUs, and flexible contracts. Some hyperscalers became customers themselves, renting neocloud capacity to meet their own demand.

The economics of GPU clouds

A GPU cloud's economics depend on the cost of chips, power and data centre space, the price customers pay per GPU hour, and utilisation. GPUs are expensive and lose value as newer generations arrive, typically every one to two years. Providers try to sign long-term contracts that cover the cost of hardware with a margin, ideally with creditworthy customers. When capacity is scarce, prices are high and margins strong; when supply catches up, prices fall. Inference platforms add software value on top of raw capacity, which can support higher margins if their optimisations deliver real savings to customers.

Financing with debt

Building AI data centres requires enormous capital, often billions of dollars per project. Neoclouds fund this with a mix of equity, debt secured on GPUs and customer contracts, and partnerships with infrastructure investors. Debt allows rapid growth but increases risk: if customers fail to renew or GPU values fall faster than expected, lenders and shareholders could face losses. Analysts watch leverage, contract terms and customer concentration closely. The financing structures used in this market are a key topic for investors.

Power is the new bottleneck

As chip supply improved, electricity became the main constraint on building AI data centres. Large campuses need hundreds of megawatts or even gigawatts of power, and grid connections can take years. Providers that secure power quickly, including through on-site generation or partnerships with energy companies, gain an advantage. This has linked AI cloud providers with energy developers, including natural gas, renewables and, in the future, nuclear. Our guide to nuclear energy startups covers one part of this story.

How to evaluate an AI cloud provider

Useful questions include: who are its largest customers, and what share of revenue do they represent? How long are its contracts, and what are the terms? How much debt does it carry, and on what terms? How quickly can it deliver new capacity, and where is its power coming from? For platforms: what performance and cost advantages does its software deliver, and how easily could customers switch?

Risks

Key risks include customer concentration, with a few AI labs accounting for much demand; leverage; hardware obsolescence; falling GPU prices if supply exceeds demand; power constraints; and competition from hyperscalers that are expanding capacity and building their own chips. If AI spending slows, neoclouds with heavy debt could face pressure. Circular arrangements, where chipmakers or AI labs invest in their own suppliers or customers, have also drawn scrutiny.

What to watch

Watch for IPOs from Crusoe, Lambda and Nscale, new multi-billion-dollar data centre projects, contract announcements with AI labs, debt financing terms and inference pricing trends. Our guides to AI infrastructure companies and AI chip startups provide related context.

Key takeaways

AI cloud providers have become some of the fastest-growing private companies in the world. Crusoe and Nscale lead capacity providers, while Baseten, Together AI, Fireworks AI, Modal and Fal.ai lead inference platforms. The opportunity is enormous, but so are the capital needs, and debt, customer concentration and power constraints make this a high-risk segment.

Sovereign AI and regional clouds

Governments increasingly want AI computing capacity within their own borders, for reasons of national security, data protection and economic development. This trend, often called sovereign AI, has created opportunities for regional providers. Nscale and Fluidstack, both based in London, have been involved in European projects, and Middle Eastern and Asian governments have funded large AI data centres. Sovereign projects can bring long-term contracts and public support, but they also involve political considerations and may favour local partners.

How customers choose a provider

AI labs training frontier models prioritise the availability of the newest GPUs in very large clusters, reliable networking and speed of delivery. Application companies running inference care more about latency, reliability, cost per request and ease of use. Many companies use multiple providers to secure capacity and avoid dependence on one supplier.