What are neoclouds, and when should an enterprise use one?

Simon Harrsen - EVP North America Published October 14, 2026  ·  8 min read

A neocloud is a cloud provider that specializes in renting graphics processing units, or GPUs, for artificial intelligence work. A neocloud does not sell hundreds of services. A neocloud sells one thing: fast access to AI computing power. For companies weighing AI infrastructure financing, neoclouds are now a third option, next to the big public clouds and on-premise hardware.

This article explains what a neocloud is and how neoclouds compare with hyperscalers. The article also covers the risks an enterprise takes on as a neocloud customer, and when owning or leasing GPU servers makes more sense.

What is a neocloud?

A neocloud is a newer type of cloud company that rents GPU computing power for AI training and AI inference. AI training means teaching a model on large amounts of data. AI inference means running a finished model to answer requests. Neoclouds are built around GPUs from the ground up, instead of adding GPUs to a general-purpose cloud. The name is short for "new cloud," and some people write it as two words: neo cloud.

Many neoclouds rent bare metal servers. Bare metal means the customer gets a whole physical server, not a shared slice of one.

Most neoclouds offer the same basic choices. A customer can rent GPUs by the hour, on demand. A customer can reserve a dedicated cluster of GPUs for a fixed term. Some neoclouds also run AI models for customers and charge per request.

Renting GPU capacity this way is a form of GPU as a service. A neocloud is one type of GPU as a service provider. The big cloud platforms sell GPU as a service too, alongside everything else they offer.

How do neoclouds differ from hyperscalers?

A hyperscaler is one of the largest general-purpose cloud providers. Amazon Web Services, Microsoft Azure, and Google Cloud are the three biggest. Oracle is now often grouped with them. A hyperscaler sells thousands of services, from storage to databases to email. A neocloud focuses on GPUs.

The focus on GPUs shows up in three places: price, speed, and scope.

Price is the headline difference. In a February 2025 analysis, Uptime Institute found that an on-demand Nvidia DGX H100 instance averaged about $98 an hour at a hyperscaler. A roughly equivalent instance from a neocloud averaged about $34 an hour. The neocloud price was 66% lower. Two cautions apply. GPU rental prices have moved since early 2025. And the analysis did not include the discounts hyperscalers give for long-term commitments, which narrow the gap.

Speed matters too. Neoclouds often get new GPU generations into customers' hands quickly, because GPUs are their whole business.

Scope is where hyperscalers win. Neoclouds operate in fewer regions, offer fewer managed services, and expect customers to handle more of the setup themselves.

The table compares a neocloud, a hyperscaler, and leased on-premise hardware side by side.

Who are the major neocloud companies?

The major neocloud companies include CoreWeave, Lambda, Nebius, and Crusoe. Other names that come up often are Fluidstack and IREN.

CoreWeave is the best-known neocloud. CoreWeave started in cryptocurrency mining, moved into GPU cloud services, and listed on the Nasdaq stock exchange in March 2025. IREN also began as a bitcoin miner. The crypto roots are common in this market, because mining operations already had power contracts, data center space, and experience running large fleets of chips.

The list changes quickly. New providers launch, and existing ones merge or shift strategy. Any list of neocloud providers is a snapshot.

Why are enterprises looking at neoclouds?

Enterprises look at neoclouds for three reasons: access, price, and speed.

Access comes first. The newest GPUs can be hard to buy, and they need data center power and cooling that many companies do not have. A neocloud already has the hardware installed and running.

Price comes second. For on-demand GPU time, neoclouds have generally charged less per hour than hyperscalers.

Speed comes third. A team can start a project in days instead of waiting months to buy and install servers.

Consider a company that wants to fine-tune a language model on its own documents. Fine-tuning means further training an existing AI model on a company's own data. The work might take a few weeks of heavy GPU use, and then the GPUs sit idle. For a short, one-time project like that, renting from a neocloud is often the simplest choice.

What neocloud risk looks like from the customer's seat

Neocloud risk is mostly supplier risk. When an enterprise runs AI workloads on a neocloud, the enterprise depends on that provider staying healthy, staying full, and keeping its hardware current.

Customer concentration is the first risk to understand. Before listing on the stock market, CoreWeave filed public financial documents with regulators. Those documents showed that two customers made up 77% of CoreWeave's 2024 revenue. Microsoft alone made up 62%. A provider that depends on one or two large customers can change its plans quickly if one of those customers leaves.

Debt is the second risk. Neoclouds borrow heavily to buy GPUs. CoreWeave reported total debt of $35 billion as of June 30, 2026.

Thin margins are the third risk. McKinsey has written that if GPU rental prices dip slightly, or if utilization falls below about 80%, neocloud returns flatten out. Utilization means the share of time the GPUs are rented and earning money. A provider in that position may raise prices, change terms, or cut back.

Hardware age is the fourth risk. GPUs lose value fast as new generations arrive. A neocloud must keep buying new hardware to stay competitive. If the neocloud cannot keep up, its customers are left on older chips.

Contract terms are the fifth risk. Cisco notes that reserved neocloud clusters typically run for one to three years. A one-to-three-year reservation is about as long as an equipment lease. The difference is what happens at the end. A neocloud customer keeps nothing.

None of these risks means a neocloud will fail. The risks mean an enterprise should ask the same questions it would ask any critical supplier. Who are your largest customers? How is your hardware financed? What happens to our workloads and data if you are acquired or go out of business? Computer Weekly's reporting on neocloud risk makes the same point: plan for the possibility that a provider fails.

Neocloud, hyperscaler, or on-premise: how to decide

The right choice depends mostly on one number: how much of the time your GPUs will be busy.

Uptime Institute's February 2025 analysis put numbers on this. Running dedicated hardware cost less than renting from a hyperscaler on demand once utilization passed about 22%. Running dedicated hardware cost less than renting from a neocloud once utilization passed about 66%. In plain terms, if your GPUs would be busy about two-thirds of the time or more, owning the capacity beat the neocloud's hourly rate in that model.

Here is what two-thirds looks like in practice. Picture a team running one eight-GPU server. A month has about 730 hours. Two-thirds of a month is about 480 hours, or roughly 16 hours a day. In Uptime's model, if the team kept the server busy about 16 hours a day or more, owning the server cost less than renting the same capacity from a neocloud. If the server would sit idle most nights and weekends, renting cost less.

These figures are a snapshot. Cloud prices have moved since, and the comparison did not include hyperscaler commitment discounts.

The two-thirds mark gives enterprises a simple way to decide.

A neocloud fits when the work is short, bursty, or experimental. Pilots, one-time training runs, and projects that need the newest GPUs right away belong here.

A hyperscaler fits when the AI workload is tied to other services the company already runs in that cloud, such as its data platform or applications.

On-premise hardware fits when the workload is steady and predictable, when data must stay under the company's control, or when the budget needs to be fixed. On-premise means servers in the company's own data center or in a colocation facility. A colocation facility is a data center where a company rents space, power, and cooling for its own servers.

Many enterprises will use more than one option. A common pattern is to run steady inference workloads on owned or leased hardware, and to rent extra capacity for short training bursts.

Leasing changes the on-premise math. A purchase needs a large up-front payment. A lease spreads the cost into fixed monthly payments over the term.

A fair market value lease also changes who carries the risk of falling hardware value. Here is an example. A company needs a GPU server that costs $400,000 new. Over three years, newer GPU generations arrive, and the server loses much of its value. If the company buys the server, the company absorbs the full loss in value. In a fair market value lease, the monthly payments are based mainly on the value the leasing company expects the server to lose during the term. If the server is worth less than expected at the end, the leasing company takes that loss, not the customer.

End-of-term value matters for GPUs, because GPUs lose value quickly. At the end of the term, the company can return the equipment and move to the next GPU generation.

CHG-MERIDIAN does not operate a neocloud or sell GPU as a service. CHG-MERIDIAN offers server and data center equipment leasing for enterprises that want to run AI on their own terms.

About the author Simon Harrsen EVP North America, CHG-MERIDIAN Simon Harrsen is EVP North America at CHG-MERIDIAN, with nearly 15 years of experience in the equipment leasing and technology lifecycle industry. He leads the company's growth and technology lifecycle strategy across the North American market, and has been quoted in and contributed bylines to publications including Monitor Daily, Channel Insider, Channel Dive, and Procurement Magazine, among many others. Connect on LinkedIn › | Email Simon ›