GPU cloud pricing chart: compare real hourly costs

Direct answer: A useful GPU cloud pricing chart compares more than the advertised hourly rate. It should normalize the GPU model and count, VRAM, region, access model, billing method, included storage and network resources, interruption risk, availability, and support. If those fields are missing, two prices that look comparable may describe different services.

This guide provides a repeatable way to compare GPU cloud offers without relying on a permanent “cheapest provider” ranking. Prices, capacity, and product terms change, so every comparison should carry a visible date and link to the provider’s current pricing or console.

What a GPU cloud pricing chart must normalize

Start by recording the following fields for every option:

• GPU model and count: One RTX 5090 is not comparable to a two-GPU instance, even when both pages lead with the same model name.
• VRAM: Record memory per GPU and do not assume that multiple GPUs behave like one larger shared memory pool.
• Region: Price and availability can vary by location.
• Access model: Separate on-demand capacity from Spot or preemptible capacity.
• Dedicated or shared access: Confirm whether the workload receives the listed GPU or a virtualized share.
• Billing method: Record the displayed hourly rate and the actual billing increment or minimum charge.
• Included resources: Note the vCPU, system RAM, disk, and network allowance attached to the preset.
• Additional charges: Check persistent storage, outbound data transfer, public IP addresses, support, and other optional resources.
• Lifecycle behavior: Record what happens when an instance is stopped, interrupted, or terminated.
• Availability and support: A listed price does not prove that capacity is currently launchable or reserved.

This structure prevents a common mistake: comparing a low Spot rate with a standard on-demand rate and treating the difference as a pure discount.

Calculate the real workload cost

Use a formula based on your workload rather than a generic market average:

Real workload cost = active compute + persistent resources + data transfer + interruption and restart cost + operational overhead.

Active compute is the displayed rate multiplied by the time the instance is billable. Persistent resources can continue after compute stops, depending on the provider and configuration. Interruption cost includes lost work, recovery time, and duplicate compute when a Spot instance ends before a checkpoint. Operational overhead covers the engineering time needed to provision, monitor, and recover the workload.

Keep every input visible. If a provider does not publish a charge or limit, mark it as unknown instead of inserting an industry average.

For practical checks on capacity, interruptions, data movement, and support, see our GPU rental pitfalls guide.

On-demand and Spot pricing answer different needs

On-demand pricing usually describes capacity that you start and stop without a long-term reservation. It does not automatically guarantee that a particular GPU is available in every region. Providers may offer separate reservation or capacity-assurance products.

Spot or preemptible capacity can be less expensive because the provider may reclaim it. The interruption rules and notice periods are provider-specific. Review the current documentation before using Spot capacity for training, rendering, inference, or another long-running job:

• AWS Spot interruption notices
• Google Cloud Spot VMs
• Azure Spot Virtual Machines

Use Spot capacity for work that can checkpoint and resume. Use a standard or reserved path when interruption would invalidate the run, break a service commitment, or cost more than the nominal saving.

Compare capacity assurance separately

A billing model and a capacity commitment are separate decisions. For example, AWS documents On-Demand Instances and Capacity Reservations as distinct products. Apply the same distinction to every provider: verify whether a price is a usage rate, a reservation, an interruptible offer, or a marketplace listing.

Use dated provider examples

A provider example should state its source date, region, GPU or instance type, access model, billing unit, included resources, and additional charges. Avoid copying a single number from an aggregator when the provider’s official page or console is available.

For Hivenet, the public pricing page lists RTX 5090 rental from €0.75 per hour. It also says that standard instance pricing includes the storage and network volume shown for the preset, without separate add-on charges for those included amounts. Optional or persistent resources can have their own lifecycle and charges, so check the selected preset and billing view before launch.

Hivenet Docs currently state that the RTX 4090 fleet has been retired and is no longer available for new Compute workloads. The current fleet reference lists RTX 5090 instances with 32 GB of GDDR7 VRAM per GPU. Exact CPU, RAM, disk, bandwidth, location, availability, and active price can vary. The GPU types reference directs users to the Compute console as the source of truth for the preset they can launch at that moment.

A practical comparison workflow

1. Define the workload, required VRAM, expected runtime, region, interruption tolerance, and data movement.
2. Open each provider’s current pricing page or console on the same day.
3. Record the comparison fields from this guide without converting unknowns into assumptions.
4. Calculate the real workload cost using your expected runtime and resource lifecycle.
5. Check whether the required capacity is launchable in the chosen region.
6. Run a short representative benchmark before committing a long workload.

For a model-specific discussion, see the updated guide to RTX 4090 and RTX 5090 cloud pricing, including the retirement of the RTX 4090 fleet on Hivenet.

Common GPU pricing chart mistakes

Mixing retail purchase prices with cloud rental rates

A graphics card’s retail price and a cloud instance’s hourly rate answer different questions. Retail comparisons include hardware ownership, power, cooling, maintenance, and utilization. Cloud comparisons include access terms, attached resources, location, availability, and service operations. Keep them in separate charts.

Using an old price without an effective date

A once-valid launch price can remain indexed after the offer changes or the hardware is retired. Every price row should show when it was checked and link to the current source.

Treating “listed” as “available”

A marketing page, historical article, or third-party chart may list hardware that is not currently launchable. Confirm the live preset before planning a workload.

Assuming every attached resource is free forever

Included disk or network allowances apply to the stated preset and lifecycle. Persistent storage may continue after compute stops. Read the provider’s stop, terminate, and billing documentation before leaving resources attached.

Frequently asked questions

What is the best way to compare GPU cloud prices?

Normalize the GPU, region, access model, billing increment, attached resources, interruption rules, availability, and support. Then calculate the cost of your workload rather than comparing headline rates alone.

Is Spot GPU pricing always cheaper?

The listed rate may be lower, but interruption and recovery can increase the real cost. Spot works best for fault-tolerant jobs that checkpoint frequently.

Does on-demand pricing guarantee GPU availability?

No. On-demand normally describes how usage is purchased. Capacity assurance or reservation may be a separate product, and regional inventory can still be limited.

Which GPUs are currently available on Hivenet?

Hivenet Docs list RTX 5090 as the current GPU family and state that the RTX 4090 fleet has been retired. Check the Compute console for the presets, locations, capacity, and active prices available when you launch.

Are storage and network resources included in Hivenet Compute pricing?

The Hivenet pricing page says standard instance pricing includes the storage and network volume listed for that preset. Optional or persistent resources may have separate charges, and storage costs may continue after compute is stopped. Verify the selected configuration in the console and billing view.

Next step

Build a dated comparison sheet with the fields above, then test the smallest configuration that fits your workload. If you are considering Hivenet, review the current pricing page and confirm the live preset in the Compute console before launch.