Cloud GPU prices and availability move quickly. Compare providers using a specific workload and a quote or current checkout price, rather than a table of rates captured months ago.
Start with the actual GPU
Record the GPU model, memory, GPU count, interconnect, CPU, RAM, local storage, region, and billing unit. A single GPU VM and an eight-GPU NVLink node are different products even when both list "B200."
For NVIDIA HGX B200, NVIDIA lists 180 GB HBM3e per GPU. A GB200 NVL72 offering is a different rack-scale system.
Compare the complete offer
| Question | Why it matters |
|---|---|
| Is the exact GPU configuration available in your region? | A listed instance type may have no capacity for your target date. |
| Is the price on demand, reserved, or spot? | Interruptions and commitments change the economics. |
| What are storage and data-transfer charges? | These can matter for large datasets and checkpoints. |
| How are GPUs connected within and across nodes? | Training performance depends on interconnect and topology. |
| What is the minimum billing period? | Idle and setup time can dominate short jobs. |
| What software and support come with the instance? | Driver, container, and support differences cost engineering time. |
Providers including Lambda, CoreWeave, RunPod, and the hyperscalers publish product and pricing pages. Open the current page or request a capacity quote for the region and configuration you need. Use the provider's current rate and capacity for your comparison.
Test cost per result
Run the same model and precision on each candidate. Record throughput, time to first result, total GPU-hours, storage use, and data transfer. Multiply by the actual current rate and add non-GPU charges. A faster GPU can cost more per hour and less per finished job, or the reverse.
If you expect steady high utilization, compare the rental result with a written quote for a complete owned server, including facility and support costs.
