Affiliate disclosure: Some links on this page may be affiliate or sponsored links. We may earn a commission at no extra cost to you. This does not change our editorial criteria.
Rent GPUs vs build a data center
This page is a beginner guide to a money-and-time choice. You can rent GPU cloud by the hour. You can buy (or lease) the chips and put them in someone else's building. Or you can look for a site with real power because the machines never turn off.
A GPU is a graphics processing unit: a chip that is very good at the math in AI, simulation, and rendering. GPU cloud means you pay a company for those chips, then shut them off. Building, in this guide, means taking on hardware, a hall, and the power that hall needs. Powered land is a development site with a credible path to large-load electricity. It is not a finished data center.
Who it is for: the person asked "should we just buy the GPUs?" or "do we need our own data center?" You do not need to be a utility engineer to follow this page.
Rent when you need cards this week, how often you will run is still a guess, and you would rather not own a building. Stop treating rental as the architecture when the same GPUs stay busy for years. The land-and-power half of that fork is when GPU clusters need powered land. The definition is what powered land is.
Four options, in plain English
People mash four products into one sentence. Separate them.
1. Rent GPU cloud. You log into someone else's GPUs. You pay while they run (on-demand) or for a reserved window. The meter can hit zero. You do not control the utility. Time-to-GPU can be hours or days if the chips exist.
2. Reserved cloud, still rental. You commit to months so the rate looks friendlier. You still do not own the building. You still cannot raise the power density. A three-year reservation is a financing choice, not a data center.
3. Colocation. Colocation means you (or your vendor) put hardware in someone else's hall and pay for space, power, and cooling. You are closer to owning the chips. You still depend on a building that already has headroom. When that headroom is gone, you are back to a site problem.
4. A site you control, starting from powered land. You are not buying a slogan about acreage near a line. You are buying a credible electricity path, then designing the hall. Buildings, cooling, and the GPU floor still have to be built. This is slow compared with a cloud login. It is the option that matches a cluster that is the product.
CapEx is money you spend to own something (chips, a hall, a substation). OpEx is money you spend to run something (hourly GPUs, colo rent, power bills). Rental is mostly OpEx. Build is CapEx plus a long OpEx tail of power, people, and parts. Neither is "cheaper" until you count hours.
When renting wins
Rental wins on time-to-GPU. A reservation can exist this afternoon. A substation cannot. If the research question might die in a month, buying land is the wrong move.
Rental wins on unknown utilization. Utilization means how many hours the machines actually run. Cloud GPU lets the meter fall to zero. Owned hardware does not. Until you can show a calendar of busy hours, you do not have a cluster plan.
Rental wins when the model size is still moving, and when the software is still portable. CUDA is NVIDIA's software layer for NVIDIA chips. LLM training is the classic large job whose shape is still a guess. Life-sciences bursts live on GPU cloud for life sciences. Rental also wins as overflow after you site a cluster. The mistake is using overflow logic for the nodes that never turn off.
When building (or siting) wins
Build or site when the cluster is no longer temporary. Utilization is high on a planning horizon measured in years. The GPU mix is stable enough to design power and cooling around. You want NVLink (NVIDIA's fast in-server link), a cooling loop, a power density, or a security boundary a shared cloud will not give you. You are hiring people to keep the cluster up.
The scarce input is then electricity on a date. Powered land is a development site where that path is already credible. Do not shop by invented campus megawatts. A megawatt is a large unit of electrical power. "We will just drop these H100 trays in any vacant cage" fails when the cage cannot cool or feed them. Colocation is a valid middle and still not powered land. When colo headroom is gone, read when GPU clusters need powered land and browse the U.S. data center directory.
How to count the decision without a fake hourly rate
We will not invent a dollar-per-GPU-hour, and we will not invent a break-even month. See the GPU cloud pricing index for how quotes work. Bands there will be dated when they exist.
Count GPUs, memory class, and hours in a real year: not a peak week times 52. A three-week train is a rental. A chatbot that answers all year is a utilization story. If next year's model wants a different memory tank, owning last year's cards is a sunk cost. Rent includes a provider. Build includes people, spare parts, and a night when a node dies. If you can get GPUs this quarter, rent. If the queue you cannot buy your way out of is power, you are in the land conversation.
Common mistakes
Buying GPUs with nowhere honest to plug them. Treating a long cloud reservation as "we built it." Using one expensive month as proof you should own a cluster. Building a hall for a model shape that will change twice before the concrete cures. Ignoring hybrid. The usual adult answer is both: rent the bursts, site the baseload.
GPU cloud partners (placeholder)
Placeholder partner table for the rental side of this fork. No URLs. No hourly prices. When those links go live, they will use rel="sponsored". Map and directory links are not affiliate placements.
| Best for | GPU types | Notes |
|---|---|---|
| Partner TBD - experiments and unknown utilization | Rented NVIDIA GPUs | Placeholder. Time-to-GPU and a meter that can hit zero. rel="sponsored" when live. |
| Partner TBD - overflow on top of a sited cluster | Burst NVIDIA capacity | Placeholder. Keep rental for peaks after baseload is sited. rel="sponsored" when live. |
| Partner TBD - not a substitute for powered land | Dedicated cloud nodes | Placeholder. A reserved instance is still not a utility interconnection. rel="sponsored" when live. |
This table is not a ranking. It will reflect partner relationships once partners exist.
GPU cloud pricing varies. Verify live rates.
There is no honest single dollar-per-hour number on this page. List prices move with GPU type, region, and whether you pay only when the machine is on. Quoted "from" rates often skip storage, data transfer, and idle reservations. See the GPU cloud pricing index for how we talk about rates. We will not invent hourly prices here.
Count GPUs times hours you actually run. Do not use a peak week as the year. If the honest description is "we would rent the same shape all year," rental is a financing choice, not a bargain by default.
Next step
If the honest description is "we might need a hundred hours," stay on rental. If it is "these nodes are the product, and we need a place with power," open the map, read powered land, and use when GPU clusters need powered land. That guide is the land-and-megawatt half. This page is the rent-versus-build choice that gets you there.
Find powered land when you need a site. Rent GPUs when you need capacity now.
Related
- When GPU clusters need powered land
- GPU cloud for life sciences and drug discovery
- Rent GPUs for LLM training
FAQ
Should we rent GPUs or build a data center?
Rent if utilization is unknown, the experiment might end, or you need cards this week. Look at building or siting when the cluster is steady, multi-year, and limited by electricity and a hall. Many teams do both: rent the bursts, site the work that never turns off.
Is a reserved GPU cloud the same as owning?
No. A reservation is capacity on someone else's floor. You still do not control power, cooling, or the next density jump. Owning starts when you control the hardware and, eventually, the site path.
Can we colocate instead of controlling land?
Yes, and many clusters live that way. Colocation still depends on a building that already has power and cooling headroom. When that headroom is gone, you are back to the site problem the directory and map exist to show.
What utilization means rental has stopped making sense?
When the same GPU shape would be busy most hours for a planning horizon in years, and you keep extending the reservation because turning it off would break the product. Count real hours. Do not use a peak week as the year.
Do you publish a break-even dollar-per-hour?
No. We do not invent rates. Use the pricing index to learn how quotes work, then run your own hours against a live provider check.
Where should I start if I still need GPUs this quarter?
Stay on rental. Come back to this guide, and to powered land for GPU clusters, when the machines stop turning off.
Sources
- What is powered land
- U.S. data center directory
- Powered Lands data center map
- NVIDIA H100 GPU
- NVIDIA NVLink
- NVIDIA CUDA Toolkit
Need land or power instead? See the map or contact.