What Is a Neocloud? – GPU Cloud Providers and Their Power Needs
Not long ago, anyone who wanted serious computing power in the cloud went to one of the big three cloud companies and took what they had. Then demand for AI chips exploded, the big clouds ran short, and a new group of companies stepped in with racks full of GPUs and a simple pitch: rent them by the hour, now. They have become some of the most aggressive buyers of data center power in the country. What is a neocloud, and why does it matter so much for power?
A neocloud is a cloud provider built mainly to rent out GPU computing for AI training and inference, rather than offering the full menu of services a hyperscale cloud provides. Well-known examples include CoreWeave, Lambda, Crusoe and Nebius. Neoclouds typically sell access to dense GPU clusters on contracts of a few years, move quickly and often price below the largest clouds. Because their product is almost pure GPU capacity, power is their biggest constraint: every new cluster needs tens to hundreds of megawatts, delivered fast.
What Makes a Neocloud Different From a Hyperscaler?
Focus. A hyperscale cloud provider sells hundreds of services: storage, databases, networking, analytics, and also GPUs. A neocloud sells mainly one thing, which is large amounts of accelerated computing, often as bare-metal clusters with fast networking and the software tools AI developers want.
That narrow focus changes the business. Customers tend to be AI labs, model developers, enterprises training their own models and, interestingly, the big clouds themselves when they need overflow capacity. Contracts are often for reserved clusters over a set period instead of pay-as-you-go use.
It also changes the price. Bisnow, citing Uptime Institute estimates, reported in 2025 that neocloud GPU capacity can cost about 66 percent less than hyperscalers, and that neocloud revenue had grown at an 82 percent compound annual rate since 2021, according to JLL.
Where Did Neoclouds Come From?
Several of them grew out of cryptocurrency mining. Mining taught these companies how to buy GPUs in bulk, find cheap power, cool dense hardware and stand up sites quickly. When mining economics shifted, the same skills and some of the same hardware pointed toward AI and other GPU computing.
CoreWeave began in crypto mining before pivoting to GPU cloud services. Crusoe started by putting computing next to oil field flares: Data Center Dynamics reported in 2021 that Crusoe ran about 40 flare-powered modular data centers across North Dakota, Montana, Wyoming and Colorado, before launching a cloud service on that energy. Today Crusoe is better known as a developer of very large AI campuses.
That energy-first heritage still shapes how neoclouds think. Many of them treat power as the product and computing as the way to sell it.
Why Do Neoclouds Build So Fast?
Because their customers will not wait, and the hardware loses value every month it sits idle. A new GPU generation arrives roughly every year or two. A cluster that comes online six months late earns less over its useful life, so speed is worth paying for.
To move quickly, neoclouds tend to:
- Lease space from wholesale data center developers rather than build every site themselves.
- Take over existing powered sites, including former crypto mines with grid connections already in place.
- Use prefabricated modules and standard liquid-cooling designs to shorten construction.
- Sign power and capacity deals early, sometimes before a tenant contract is final.
Speed has costs. Faster sites often come with less redundancy, higher lease rates or temporary power arrangements, and every shortcut has to be paid back later.
How Much Power Does a Neocloud Need?
A lot, and in big blocks. A useful rule of thumb from chip-level math is that each modern GPU needs roughly 1.5 to 2 kilowatts at the utility meter once the server, networking, cooling and electrical losses are counted. Here is the arithmetic for illustrative cluster sizes, using 1.7 kW per GPU:
| Cluster size | Power at the meter | Yearly energy at 80% load |
|---|---|---|
| 10,000 GPUs | about 17 MW | about 119,000 MWh |
| 50,000 GPUs | about 85 MW | about 596,000 MWh |
| 100,000 GPUs | about 170 MW | about 1.19 million MWh |
The math for the top row: 10,000 × 1.7 kW = 17,000 kW, or 17 MW. Over a year at 80 percent average load, 17 MW × 0.8 × 8,760 hours ≈ 119,000 megawatt-hours (MWh). Newer chips push the per-GPU figure higher, so a cluster of the same count built a generation later needs more power, not less.
For a neocloud with several sites, total demand easily reaches hundreds of megawatts, which puts these companies in the same conversations with utilities as the biggest cloud providers.
Where Do Neoclouds Get Their Power?
Wherever they can get it fastest. In practice, neocloud power strategies fall into a few patterns:
- Leasing from colocation developers who have already secured grid capacity, often in established data center markets.
- Hosting deals with former bitcoin miners whose sites already hold large grid connections and substations.
- Energy-first campuses in places with available generation, sometimes pairing a grid connection with on-site gas generation or batteries.
- Direct utility arrangements for the largest builds, which bring the same large load studies and tariffs everyone else faces.
The bitcoin-mining route is especially interesting for the power industry. Miners built sites in places like Texas with big interconnections and cheap power, then learned that renting space to GPU tenants could pay more than mining. Converting those sites usually means adding liquid cooling, better redundancy and stronger fiber, but the hardest part, the power connection, is already there.
What Are the Risks of the Neocloud Model?
The biggest risk is a timing mismatch. Bisnow reported that neocloud GPU contracts often run two to five years, while the data center leases behind them typically run 10 to 15 years. If demand cools or customers leave when their contracts end, the neocloud still owes rent and power commitments.
Other risks are familiar to anyone in the power business. Many neoclouds depend on a few large customers. GPUs lose value quickly as newer models arrive. Much of the buildout is financed with debt secured by the hardware and the contracts. For utilities and landlords, that means checking credit and contract terms carefully before committing big infrastructure to a single neocloud tenant.
Power efficiency is one lever neoclouds can pull to stay competitive. Better chips, liquid cooling and higher utilization all lower cost per job, which is why AI energy efficiency gets so much attention. Our guide to AI energy efficiency looks at how far those gains go.
What Should a Site Owner Ask a Neocloud Tenant?
Ask the questions you would ask any large industrial customer, plus a few that are specific to GPU computing. From the operator's side of the table, these are the ones that matter most:
- What rack density and cooling type do you need now, and what will the next hardware generation need?
- How much of the load needs full backup power, and how much can ride through an outage or be shed?
- How fast will the load ramp, and how much does it swing during training runs?
- Who is the end customer, and how long is that contract?
- Who pays for grid upgrades, on-site generation and cooling plant if the tenant leaves early?
The density question deserves extra time. A site built for 40 kW racks may need major rework for 100 kW liquid-cooled racks, and a neocloud chasing the newest chips will want the denser design within a few years. Planning the cooling loops, floor loading and electrical capacity for that next step up front is far cheaper than retrofitting a live hall.
The load swing question matters for anyone running on-site generation. Training clusters can move tens of megawatts up and down quickly, and engines, turbines and batteries all need to be specified for that duty, not just for steady output.
Neoclouds Turn Power Into GPU Hours
Strip away the branding and a neocloud is a machine for turning megawatts into rentable GPU hours. That is why the companies that grew out of crypto mining still think energy-first, and why their growth is limited less by chips than by how fast they can find power in large blocks.
For anyone holding a powered site, a gas supply or a grid position, that makes neoclouds likely buyers. Ask about contract length, customer concentration and who carries the risk if the GPUs go quiet. Those answers matter as much as the megawatts.