GPU Cost Calculator
Price the silicon — GPUs × hours × the published GPU-hour, with the idle-billing trap priced beside it.
GPU Cost Calculator
Results recalculate instantly on every keystroke. Nothing you type is transmitted.
What this result does not account for
- Card-hours only — storage and dataset egress bill separately
- On-demand rates; spot lines named in the note
In short: 8 GPUs for 12 hours at $2.69 a GPU-hour bill $258.24 — one training night on the mid-market cloud. Left running the weekend (60 hours) the same fleet bills $1,291.20, because idle GPUs bill exactly like busy ones. One card is $64.56 a day at this rate; the same night on a hyperscaler's per-GPU line reads $660.48. The rate spread is the decision.
Formula
bill = GPUs × hours × rate; idle = GPUs × 60 × rate
GPU clouds bill the card by the hour (per-second on some), and the 2026 H100 spread is wide: marketplace hosts from $1.49, RunPod $1.99–$2.69, Lambda $3.29–$3.99, CoreWeave about $4.25–$6.16, hyperscalers to $6.98 a GPU-hour. Spot and interruptible tiers read 60–91% under on-demand. The weekend column is the honest one: a forgotten pod bills the same rate as a training run.
Worked Example
- Enter the fleet, the wall hours and the published rate.
- Read the run, the idle weekend, and the per-card day.
- Compare the ladder in the note before committing the job.
Defaults: $258.24 for the night, $1,291.20 for the weekend it was forgotten. Drive the rate to the hyperscaler's $6.88 and the night reads $660.48 — same job, 2.56× the bill, zero extra accuracy.
Strengths & Limits Of This Model
Where this engine is strong
- The idle weekend priced — the real GPU surprise
- Published rate presets across the provider spread
Where it stops
- Availability differs wildly from price
Practical Use Cases
Training budgets
the run priced before the launch
Provider choice
the spread, not the average
Finops audits
the idle lines in last month's bill
Methodology & Editorial Standards
Computation runs in IEEE-754 double precision at full internal precision; rounding to two decimal places occurs strictly at the display layer, so no cumulative drift enters the result. All monetary outputs use accounting presentation — grouped thousands, two decimals, negatives in parentheses — so figures can be transcribed directly into a model or working paper. Division-by-zero and out-of-domain inputs return an em-dash rather than a misleading number.
This engine was reconciled against an independent reference implementation and hand-verified for the worked example above before release. Our full five-stage review process is published on the About Us page.
Disclaimer. This calculator is provided for informational and modelling purposes only and does not constitute financial, tax, legal, medical, or engineering advice. Verify all figures with a qualified professional before acting on them.
GPU Cost Calculator — 8 Expert FAQs
8 analyst-written answers to the questions practitioners actually ask — optimised for voice and answer-engine retrieval.
What does an H100 actually cost per hour?
Published 2026 on-demand lines: marketplace hosts from $1.49, RunPod $1.99 (PCIe) to $2.69 (SXM), Lambda $3.29–$3.99, CoreWeave about $4.25–$6.16, AWS's 8-GPU node works out near $6.88 a card, Azure lists up to $6.98. Same chip; the spread is the business model.
Why is the spread so wide?
Hyperscalers bundle enterprise SLAs, on-demand availability and egress; marketplaces auction idle consumer-grade capacity. The chip is identical — what you buy is the_ops around it. For batch jobs that survive interruption, the cheap end wins outright.
What does spot/interruptible save?
Published spot lines run 60–91% under on-demand — the default shape's $2.69 class reads $0.02 or less on CPU spot, and H100 spot from about $1.66. The price is eviction: checkpoint often, or don't bother.
Why price the idle weekend?
Because it is the most common GPU surprise on every finops audit: a debug pod left running over a weekend bills 60 hours at full rate. $1,291.20 of nothing is the default fleet's weekend.
How does this compare to renting tokens?
The token page prices other people's GPUs by the output token — no ops, no idle risk, margin on top. Self-hosting wins at high steady utilization and loses anywhere it idles; the crossover is a utilization question, not a rate question.
Does storage or egress bill extra?
Often yes, though the GPU-native clouds increasingly include egress. Datasets, checkpoints and snapshots bill on the storage side — price the shelf on the S3 page and the wire on the transfer page.
Why 12 hours as the default run?
It is one honest overnight — the unit most teams actually budget in. A fine-tune fits in one; a foundational run is ten thousand of them, which is why the per-card day line matters more at scale.
What about A100s and L40S?
Cheaper: A100 80 GB reads $1.19–$2.79 across the same providers and an L40S about $0.79. The presets pin the H100 class because it is the pricing benchmark; the field takes any published rate.