Cloud providers sell the same GPU under several commercial models. AWS's documentation, for example, lists on-demand instances, Savings Plans, Reserved Instances, Spot Instances, dedicated hosts and instances, and capacity reservations, including Capacity Blocks for reserving clusters of GPU instances. Google Cloud's guidance for AI clusters similarly distinguishes on-demand, Spot, flex-start, calendar-mode future reservations and longer-term reserved blocks. [1][2]

These options vary along three dimensions: how much you pay per GPU-hour, how likely you are to obtain capacity when you want it, and how likely you are to lose it once running. No model wins on all three. The cheaper the rate, the more risk the buyer usually carries, either of interruption or of paying for capacity that sits idle. Kovara's separate article on spot pricing covers the interruptible end of the spectrum in more depth.

On-demand capacity is billed for the time it runs, with no commitment. AWS bills on-demand EC2 instances per second. Once an on-demand machine is running, the provider does not reclaim it; Google describes on-demand VMs as not subject to preemption and running until the user deletes them. [1][2]

The catch for GPUs is obtaining the machine in the first place. Google labels on-demand obtainability as best effort, meaning a request succeeds only if capacity happens to be free at that moment. For popular accelerators in popular regions, that can mean failed launches at busy times, which is why many teams pair on-demand usage with some form of reservation. [2]

Specialist providers follow the same pattern. Together AI bills on-demand GPU cluster capacity hourly with no commitment, and Vast.ai describes its on-demand rentals as fixed-price instances that run without interruption at a premium to its other options. [11][12]

Spot capacity is spare capacity sold at a discount on the condition that the provider can take it back. Google says its Spot VMs can be discounted by up to 91 percent, including for GPUs, but can be preempted at any time. AWS may interrupt Spot Instances when it needs the capacity, when the Spot price exceeds the buyer's maximum or when request constraints can no longer be met, and gives a two-minute warning. [3][4]

Warning periods differ. Azure provides a best-effort 30-second eviction notice for Spot VMs, offers no SLA for them, and lets users choose whether an evicted VM is deallocated or deleted. Marketplaces can work differently again: Vast.ai's interruptible instances are bid-based, and an instance is stopped if another user bids more for the same machine. [5][12]

Spot suits work that can be checkpointed and resumed: batch inference, hyperparameter sweeps, data processing and training jobs that save state frequently. It is a poor fit for multi-node training runs where losing one node stalls the whole job, or for latency-sensitive production inference without a fallback.

Reserved or committed pricing trades a term commitment for a lower rate. AWS Savings Plans commit the buyer to a dollar amount of hourly spend for one or three years, with all-upfront, partial-upfront or no-upfront payment. Azure reservations run for one or three years and can be paid upfront or monthly, with self-service refunds capped at 50,000 dollars in a rolling 12-month window. [6][8]

A key detail is whether the commitment reserves hardware or only price. AWS states that Savings Plans do not reserve capacity, and that regional Reserved Instances do not either, whereas zonal Reserved Instances do reserve capacity in a specific availability zone. A buyer who signs a discount commitment expecting guaranteed GPUs may find that the discount applies only to machines they can actually obtain. [6][7]

Specialist GPU clouds often offer shorter reservations. Together AI sells reserved cluster capacity for 1 to 90 days paid upfront and lets customers combine reserved and on-demand nodes in one cluster, while Vast.ai offers discounts of up to 50 percent for reserved rentals that keep on-demand priority. [11][12]

For training runs of days or weeks, hyperscalers now sell time-boxed GPU clusters. AWS Capacity Blocks for ML let customers reserve up to 64 GPU instances per block, and up to 256 across blocks, up to eight weeks in advance, placed in UltraClusters. Blocks cannot be cancelled and instances are terminated automatically at the end of the reservation. [9]

Capacity Block prices are set by supply and demand at the time of purchase, locked once reserved and charged upfront; Savings Plan and Reserved Instance discounts do not apply to them. Google offers comparable options: calendar-mode future reservations of up to 90 days for up to 80 VMs, flex-start VMs that run for up to seven days when capacity becomes available, and longer reserved blocks arranged through an account team up to a year ahead. [10][2]

At the largest scale, capacity is sold through negotiated contracts rather than published price lists. Nebius's 2025 agreement with Microsoft, for example, covers dedicated capacity from a specific new data center over multiple years. CoreWeave reports a revenue backlog made up of remaining performance obligations and other amounts expected under committed customer contracts, which stood at 66.8 billion dollars at the end of 2025. [13][14]

Such contracts give the buyer certainty over hardware, location and often networking design, and give the provider revenue it can borrow against. The trade-off is inflexibility: the buyer is committed for years to a hardware generation that newer chips will overtake, and the terms, including penalties, delivery milestones and what happens if capacity arrives late, are negotiated case by case and rarely public.

Most organisations end up with a mix. A common pattern is a committed baseline sized to steady workloads such as production inference, time-boxed reservations for known training runs, on-demand for bursts and spot for anything that can be interrupted. Google's own guidance frames this choice as a trade-off between obtainability, preemption risk and discount, and AWS documents its purchasing options as a menu of separate choices rather than a single price. [2][1]

The questions to ask any provider are simple but often left unasked: does this commitment reserve capacity or only price, how much warning will I get before an interruption, can I cancel or resell unused time, and what happens if the provider cannot deliver on the start date.

Kovara tracks published GPU prices across roughly 100 providers, so readers can see on the GPU prices page how the same accelerator is priced by different suppliers and, where providers publish them, under different commercial models. Kova can walk through the trade-offs for a particular workload, and buyers who need a time-boxed cluster or a longer commitment can request a quote to confirm real capacity rather than relying on a list price alone.