Plan training runs with transparent, editable cost inputs. Model on demand, reserved, or spot pricing. Get totals, unit rates, and downloadable summaries instantly here.
This calculator separates costs into compute, storage, egress, and add-ons. It then applies support, contingency, and tax.
Illustrative values only. Replace with your quoted rates and contract terms.
| Example provider | GPU | GPU rate (USD/GPU-hr) | Typical use |
|---|---|---|---|
| Provider A | NVIDIA T4 (16GB) | 0.45 | Inference, prototypes, light finetuning |
| Provider B | NVIDIA A10 (24GB) | 1.20 | Balanced training and inference workloads |
| Provider C | NVIDIA A100 (40GB) | 3.50 | Large training runs, heavy compute |
| Provider D | NVIDIA H100 (80GB) | 6.75 | Premium training, fastest turnaround |
Compute spend is dominated by effective GPU-hours. The calculator adds a base instance rate to a per‑GPU rate, then multiplies by effective hours. For example, 4 GPUs at 3.50 USD/GPU‑hr plus 0.80 USD/hr base equals 14.80 USD/hr. If you run 8 hours/day for 14 days, planned hours are 112. With 70% utilization and 10% waste, effective hours become 86.24, and compute becomes about 1,276 USD before region and discounts.
Engineering teams rarely achieve 100% utilization because jobs queue, data pipelines stall, and checkpoints pause training. Utilization models sustained load, while waste captures retries, preemptions, and setup time. A spot workflow might show 60% discount, but if preemptions push waste to 25%, savings shrink. Track utilization from monitoring, and set waste from incident logs to keep estimates aligned with operations.
Datasets, checkpoints, and artifacts often outlive the compute window. This tool pro‑rates storage by days divided by 30.437, separating block and object tiers. Egress is modeled as GB times price per GB, which matters when exporting checkpoints, mirroring data across regions, or serving models externally. A 2 TB checkpoint export at 0.09 USD/GB adds 184.32 USD, even if compute is optimized.
Reserved and savings plans reduce the compute component through a discount multiplier, while the region multiplier scales compute and storage to reflect geographic uplifts. Compare at least three scenarios: on‑demand baseline, reserved for steady training, and spot for bursty experiments. Keep inputs consistent, then adjust only discount, waste, and utilization to see the true trade‑off between price and reliability.
Cost reviews move faster when assumptions are explicit. After calculating, export CSV for spreadsheets or PDF for approvals each quarter. Include provider label, GPU model, hours, utilization, storage sizes, and egress volume so stakeholders can reproduce results. Use cost per GPU‑hour and daily burn as KPIs to set budgets, define kill‑switch thresholds, and evaluate whether optimizations are worth the engineering time.
Utilization is the share of scheduled time where GPUs do productive work. It reduces planned hours into effective hours, reflecting queueing, data stalls, and under-filled batches.
Many GPU instances have a base charge for CPU, memory, and networking plus a GPU-specific charge. Splitting them helps you map quotes and test cheaper CPU footprints with the same GPU tier.
Use a higher discount for spot pricing, then increase waste to represent retries and restarts. If preemptions also lower steady throughput, reduce utilization to match observed training logs.
Providers bill storage monthly, but projects rarely run exactly one month. Using an average days-per-month value lets you estimate partial-month storage charges more smoothly across different project lengths.
In this tool, the region multiplier applies to compute and storage only. Model egress separately using the egress price per GB and volume, since it often follows different regional or tiered rules.
Enter the number of local currency units per 1 USD. Use the same rate your finance team uses for budgeting, or a conservative rate for stress testing when currency volatility is a concern.
Important Note: All the Calculators listed in this site are for educational purpose only and we do not guarentee the accuracy of results. Please do consult with other sources as well.