Every AI model is trained and served on GPUs, and for a few years the only places to rent them at scale were the big three clouds. Then a new kind of provider appeared, one that does nothing but buy accelerators by the tens of thousands and rent them out: the neocloud. CoreWeave, Lambda, Crusoe, Nebius. This guide explains what a neocloud is, why they exploded onto the scene, how their economics really work, and when you would pick one over a hyperscaler.
It assumes no background beyond knowing that AI runs on GPUs. By the end you will be able to define a neocloud in a sentence, name the main players, read the unit economics that decide whether one survives, and choose between a neocloud and a hyperscaler for your own work.
What you'll learn
- What a neocloud is, and what the word means
- Why they appeared: the GPU supply gap
- How a neocloud differs from a hyperscaler, layer by layer
- The economics: capex, utilization, depreciation, and debt
- The landscape, from anchor tenants to GPU marketplaces
- How to choose, and the risks to keep an eye on
What is a neocloud?
A neocloud is a cloud provider built around a single product: GPUs for AI, rented by the hour. Where a traditional cloud sells hundreds of services, databases, queues, storage tiers, identity, a neocloud sells accelerators and the fast networking that ties them together, and not much else. The premise is focus: do the one thing the AI boom needs most, and do it cheaper and sooner than a generalist can.
The one-line version: a hyperscaler is a supermarket; a neocloud is a butcher. The supermarket carries everything. The butcher carries one thing, in more variety, often fresher and cheaper, and assumes you will get the rest elsewhere. "Neo" just means new, the new kind of cloud the AI era called for.
Why neoclouds appeared: the GPU gap
The story is supply and demand. When demand for AI compute went vertical, the hyperscalers could not pour GPUs into the market fast enough, they were also building for their own products, rationing capacity, and moving at the pace of a giant. That left a gap between what builders wanted and what the big clouds could hand them. Neoclouds are what rushed into that gap.
Two things made it possible to start one. First, Nvidia was willing to sell large allocations to new buyers, partly to widen the market beyond a handful of hyperscalers. Second, money was available: investors and lenders would fund GPU buildouts because the hardware itself could serve as collateral. Cheap-ish capital plus chip access equals a neocloud.
Neocloud vs. hyperscaler
The clearest way to see a neocloud is next to a hyperscaler, layer by layer. They are not competing to be the same thing; they are deliberately different shapes.
The practical differences fall out of those shapes. Neoclouds tend to offer the latest GPUs sooner, bare-metal access and high-speed InfiniBand for big training runs, and lower per-hour prices. Hyperscalers offer the surrounding services, global regions, compliance, and the convenience of one bill for everything. Most serious AI teams end up using both.
The economics: utilization vs. depreciation
A neocloud is, underneath, a financial machine for turning capital into GPU-hours. Understanding it comes down to one tension: a GPU costs a lot up front and loses value over time, so it has to be rented out enough, fast enough, to pay for itself before the next generation makes it cheap.
This is why you hear about neoclouds signing huge, multi-year contracts and taking on debt secured by their GPUs. A signed contract guarantees utilization, which guarantees the GPUs earn out, which is what makes the debt safe to take. It is also where the risk lives, which we get to below.
The neocloud landscape
The field sorts loosely into tiers. The names change as the market grows, but the shape is stable.
| Tier | Examples | Shape |
|---|---|---|
| Anchor | CoreWeave, Lambda | Large fleets, big-name customers, often public or late-stage |
| Challengers | Crusoe, Nebius, Together AI, Voltage Park | Fast-growing specialists, each with an angle (energy, inference, sovereignty) |
| Marketplaces | Vast.ai and similar | Aggregate spare capacity from many owners; cheapest, most variable |
| Hyperscalers | AWS, Google, Azure, Oracle | Not neoclouds, but rent GPUs too, with the full stack around them |
CoreWeave is the one to know. It began as a crypto-mining operation, pivoted to renting GPUs for AI, grew on a close Nvidia relationship and aggressive debt-financed buildouts, and went public in 2025. It is both the leader and the template the rest are measured against, and its results are a useful barometer for the whole category.
How to choose, and use, a neocloud
If you are renting GPUs rather than running a cloud, the decision is practical. A short checklist:
- Start from the workload. A big training run rewards a neocloud's newest GPUs and InfiniBand. A small inference service may be simpler and cheaper bundled into a cloud you already use.
- Price the whole job, not the GPU-hour. Add storage, egress, and the engineering time to wire up services a neocloud does not provide. The sticker rate is only the start of the bill.
- Check availability and contract terms. Can you get the chips you want, when you want them, and are you signing a long commitment or paying on demand? Scarcity and lock-in are the real constraints.
- Weigh reliability. Leaner stacks can mean less redundancy. For anything in production, ask what happens when a node fails, and test it.
The risks
The same traits that let neoclouds grow fast make them fragile, and it is worth seeing clearly.
Frequently asked questions
What is a neocloud?
A neocloud is a cloud provider that specializes in renting out GPUs for AI work, training and inference, instead of the general-purpose computing the big clouds offer. Examples include CoreWeave, Lambda, Crusoe, and Nebius. The name marks them as a new kind of cloud, built GPU-first for the AI era.
What does neocloud mean?
Neo means new, so a neocloud is a new-style cloud focused on AI accelerators rather than general compute. The term spread through 2023 and 2024 as GPU-specialist providers grew large enough to rival parts of the hyperscalers, and it has stuck as the label for the category.
What is the difference between a neocloud and a hyperscaler?
A hyperscaler, such as AWS, Google Cloud, or Azure, offers a huge menu of managed services across general computing. A neocloud does one thing: rent GPUs, usually as bare metal with fast networking, often cheaper per GPU-hour and with quicker access to the newest chips, but with fewer managed services and a smaller global footprint.
What are some neocloud companies?
CoreWeave is the largest and has been public since 2025. Others include Lambda, Crusoe, Nebius, Together AI, and Voltage Park, plus marketplaces like Vast.ai that resell capacity from many providers. New ones appear regularly; the common thread is GPUs as the core product, not a side option.
Are neoclouds cheaper than the big clouds?
Often yes for raw GPU time, because they specialize and run leaner. But cheaper per GPU-hour is not the whole bill: the big clouds bundle storage, networking, security, and managed services a neocloud may not. Compare total cost for your actual workload, plus availability and contract length, not just the sticker rate.
How does a neocloud make money?
It buys GPUs, a large up-front cost, then rents them by the hour. The math is utilization versus depreciation: a GPU has to earn back its price before it loses value as newer chips arrive. Many neoclouds finance the GPUs with debt secured against the hardware and lock in big customers on long contracts to de-risk the buildout.
Is CoreWeave a neocloud?
Yes, CoreWeave is the best-known neocloud. It began as a crypto-mining operation, pivoted to renting GPUs for AI, grew on a close Nvidia relationship and debt-financed buildouts, and went public in 2025. It is the template most other neoclouds are measured against.
What are the risks of neoclouds?
Three stand out: concentration, since a few giant customers can be most of the revenue; depreciation, since GPUs lose value as new ones ship; and debt, since buildouts are often heavily financed. If GPU demand or prices soften, the leverage that fueled growth cuts the other way. For builders who are just renting, the practical risks are availability and reliability.
Should I use a neocloud or a hyperscaler?
Use a neocloud when GPUs are the point and you want the newest chips, fast networking, and a better rate, and you can handle a thinner managed stack. Use a hyperscaler when you need the surrounding services, global reach, and deep integration. Many teams use both: a neocloud for training runs, a hyperscaler for the product around them.
Glossary
- Neocloud
- A cloud provider specialized in renting GPUs for AI, rather than offering general-purpose computing. Also called a GPU cloud or AI cloud.
- Hyperscaler
- A massive general-purpose cloud (AWS, Google Cloud, Azure) offering hundreds of services. Rents GPUs too, with the full stack around them.
- GPU-hour
- The unit a neocloud sells: one GPU rented for one hour. The price times utilization is the revenue side of the business.
- Utilization
- The share of time a fleet's GPUs are actually rented and earning. The single biggest driver of whether a neocloud is profitable.
- Depreciation
- The loss of a GPU's value over time as newer, faster chips ship and rental rates fall. The clock a neocloud races against.
- Bare metal
- Direct access to a physical machine with no virtualization layer in between, common on neoclouds for maximum performance.
- InfiniBand
- A high-speed, low-latency networking standard used to link many GPUs into one training cluster. A neocloud selling point.
- Capex
- Capital expenditure: the large up-front cost of buying GPUs and building data centers, often financed with debt.
- CoreWeave
- The largest and best-known neocloud; a former crypto miner that pivoted to GPU rental and went public in 2025.
Where to go next
You now have the whole picture: what a neocloud is, the GPU gap that created the category, how it differs from a hyperscaler, the utilization-versus-depreciation math that decides who survives, and how to choose one. Neoclouds are the rental layer of the AI economy, and how it is priced ripples up to what every builder pays.
For the parts inside the box, the companion guide on the AI hardware stack explains the GPUs, HBM, and interconnects a neocloud rents out. For the money side, our GPU and inference economics briefing goes deeper on the costs. And for the live moves, deals, buildouts, and the next CoreWeave, our AI infrastructure news hub tracks the compute layer as it shifts.
For the daily moves in chips, clouds, and the economics that tie them together, the daily briefing reads the wire every morning and closes each edition with one falsifiable call we settle in public. This guide is part of The Primer, our growing library of ground-up explainers, re-checked against the live landscape each month so the details stay current.