Reserved and on-demand compute, explained.

The difference is the commitment.

On-demand compute generally gives you access to available capacity without a long fixed reservation. Reserved compute allocates a defined amount of capacity for an agreed period, with commercial obligations tied to that period.

Neither label fully describes the contract. Start and end dates, interruption rights, billing increments, delivery terms, and cancellation provisions still need to be checked.

On-demand suits changing requirements.

Short experiments and variable workloads may benefit from a flexible rental. You can often adapt the resource to the work as your needs become clearer.

The tradeoff is availability: the hardware and quantity you want may not be available when you return. Also distinguish ordinary on-demand access from interruptible spot capacity, which can be reclaimed according to the provider’s rules.

Reservations suit a known capacity window.

A planned training run or an ongoing inference requirement can justify securing capacity for a defined period. The reservation should identify the exact resources, location, delivery window, and service commitments.

A commitment to capacity does not necessarily mean full upfront payment. The agreement may specify a deposit, periodic billing, or another payment schedule. The amount owed and the date it is due are separate questions.

Compare offers on the same basis.

Normalize the quantity, hardware, and duration before comparing the price. Then account for the terms that change what you can actually do with the resource.

• Is the price per GPU-hour, per node-hour, or for the entire term?
• Do idle reserved hours remain payable?
• What is included for storage, networking, support, and setup?
• When is the resource delivered, and what happens if delivery is late?
• Can the reservation be extended, transferred, or sublet?

Start with the work you know you need.

Estimate the duration and scale of your workload, then identify which constraints are fixed and which can move. Region flexibility, hardware alternatives, and a wider start window may create more sourcing options.

Use a capacity request to make those tradeoffs explicit. A quote is an offer to evaluate; it is not necessarily an inventory hold or a running machine.