GPU pricing across the four clouds.

Here is how the four mechanisms compare across the providers, and how to run the comparison that actually decides where AI workloads belong. Specific rates move constantly, so treat any figure here as indicative and verify against each provider's current pricing page.

What drives the real cost of a GPU, not the sticker rate?

Three things sit between the hourly rate and your bill. First, throughput: two GPUs at different hourly rates can deliver very different work per hour on the same model, so the rate per useful token or per training step matters more than the rate per hour. Newer accelerators often cost more per hour but finish the job faster, which can make them cheaper per result. Second, utilization: a GPU billed continuously but kept busy only part of the time has an effective cost far above its rate, and GPUs are expensive enough that idle time is the single largest source of AI waste. Third, the surrounding instance: GPU instances bundle CPU, memory, fast local storage, and high speed networking, and for distributed training the interconnect can decide whether the cluster scales efficiently or stalls. Compare configured instances on your workload, not GPUs on a spec sheet.

How do the purchase models compare across the providers?

The structure is consistent even though the names differ. For steady, always on inference, reserved or committed pricing wins on every provider, because the committed rate applies to hours you would consume anyway: AWS through Savings Plans and Reserved Instances, Azure through Reservations and the Azure Savings Plan, GCP through Committed Use Discounts, and OCI through Universal Credits. For interruptible and fault tolerant work such as batch training and experimentation, spot and preemptible capacity offer the deepest discounts, with the tradeoff that the provider can reclaim the capacity. For scarce top end GPUs that are hard to get on demand at all, capacity blocks and reservations guarantee access for a window. OCI is frequently competitive on raw GPU rates and notably cheaper on egress, which matters for data heavy training; the hyperscalers offer broader instance variety and deeper managed tooling. The right provider is workload specific, so the comparison has to be run per workload.

When does spot or preemptible GPU capacity pay off?

Spot and preemptible GPUs carry the deepest discount because the provider can reclaim them with little notice, so they pay off precisely when your work can absorb interruption: batch training with checkpointing, hyperparameter sweeps, offline scoring, and experimentation that can restart from a saved state. They are the wrong tool for latency sensitive production inference, where an interruption is a customer facing outage. The discipline is to architect training to checkpoint frequently so a reclaim costs minutes of recompute rather than a whole run, and to keep production inference on committed or on demand capacity. Done well, the split sends the largest and most flexible part of the GPU bill to the cheapest capacity while protecting the part that must not be interrupted.

How does egress and data gravity change the GPU decision?

GPU rate is not the whole bill for data heavy AI. Training pulls large datasets, and moving that data between providers, regions, or back out to your own systems incurs egress, which differs markedly across the four clouds. OCI prices egress materially lower than the hyperscalers, which can tip a data heavy training workload in its favour even when the GPU rates are close. The buyer side discipline is to model the full cost, GPU hours under the right purchase model plus storage plus egress, and to keep data and compute co located so you are not paying to shuttle terabytes across a metered boundary. A cheap GPU rate undone by expensive egress is not a saving. We set this in the wider context in the cross cloud cost optimization guide.

Where GPU pricing fits the wider AI cost program

GPU rate is one input into AI unit economics, alongside the choice between on demand and provisioned inference, model selection, and capacity strategy. Read the full approach in the cross cloud cost optimization guide, understand why GenAI broke the cloud budget, and size commitments with AI capacity reservations and commitments. The cheapest GPU is the right purchase model on the right provider, kept busy, measured per unit of work.

Compare GPUs on cost per unit of work, not the sticker

We run the GPU comparison that matters, cost per unit of useful work under the right purchase model on the right provider, across AWS, Azure, GCP, and OCI, with egress and utilization in the model, as an independent advisory that takes zero provider commissions. Our guarantee: we reduce your cloud spend or we reimburse our service fee, on a Fixed Fee scoped up front or a no risk Gainshare basis. Download the cross cloud guide, or read training versus inference cost profiles.