Rent, reserve or own: ways to get AI hardware

Think about how people get somewhere to live. Some take a short let. Some sign a longer lease. Some buy a home. Each route suits different plans and budgets.

Businesses getting AI hardware face a similar choice. They can pay as they go, reserve capacity for a set period, or buy and run their own. This piece explains each route in terms of cost, control and commitment. It is general information, not legal or financial advice.

Route one: pay as you go

The simplest route is to use capacity on demand. You send work to a provider and pay for what you use, often per request or per hour.

Cost. There is no upfront spend. But the price per unit of work can be higher than on a longer commitment, and it can change.

Control. You have little say over where the hardware is or exactly what it is. The provider decides.

Commitment. Very low. You can stop at any time. The catch is that capacity may not be there when demand is high. We discuss busy hours in our piece on how AI services answer many people at once.

This route suits early experiments, uneven demand, and work that can wait.

Route two: reserve capacity

The middle route is to reserve a set amount of capacity for an agreed period. You pay for it whether or not you use all of it.

Cost. The rate can be lower than pay as you go, because the provider has certainty. But unused capacity is still paid for.

Control. You often get more say over the type of hardware and, sometimes, where it runs. That can matter for data residency. See Data residency explained.

Commitment. Medium. You sign up for a term, and leaving early can be costly.

This route suits steady, predictable work, such as a product already in use.

Route three: own the hardware

The third route is to buy hardware and run it yourself, in your own building or in space you rent from a data centre.

Cost. There is a large upfront spend, followed by ongoing costs for power, cooling, space, staff and maintenance. We list these in Hidden costs of AI capacity that aren't on the quote. Over a long period of heavy use, the cost per unit of work can be lower.

Control. The most control. You choose the hardware, the location and how it is run.

Commitment. High. Hardware ages, and AI hardware moves on quickly. You also take on the work of keeping it running, and the risk if it sits idle.

This route suits large, steady workloads and organisations with the skills to operate hardware. See Operators, not developers.

Mixing routes

Many businesses do not pick just one. A common pattern is to cover steady demand with reserved or owned capacity, and use pay as you go for peaks.

This gives a lower average cost for the base load, and flexibility for surprises. It does need a way to send work to the right place at the right time. We describe what that might look like in Control tower for AI work.

The comparison is not only about price. The same task can cost very different amounts depending on route, timing and location. See AI inference cost: why one task gets different prices.

Questions to ask before choosing

These questions help match the route to your situation.

1. How steady is our demand, and how well can we predict it?
2. Do we have rules on where our data can be processed?
3. Do we have the people and space to run hardware ourselves?
4. What happens to our plans if hardware is delayed? See our piece on why AI chips are hard to get.
5. If we reserve, what happens if our needs change part way through?
6. Can we see dated prices, so we can compare routes fairly, for the same task under the same conditions?

For per-use model prices, see Starfield's Model Watch.

What we are building

We think buyers should be able to commit to capacity with clear terms, and see where and how their work was delivered.

We're building standardized compute close to power, connectivity and demand, connected through one programmable network for AI inference. The aim is to place workloads by policy, on latency, data residency, model and load, with an auditable record of commitment, placement and delivery for each workload.

That record is designed to show what was committed, where the work ran and what was delivered. We explain it in A receipt for AI work.

If you are deciding how to get AI capacity, use Request capacity below.