Compare GPU capacity and cost.
Define the fleet you are replacing, enter your quoted rate for each GPU, and compare total cost by term for the same theoretical capacity.
Define the capacity you are replacing
We match theoretical compute and aggregate memory, then take the larger requirement. Less memory per GPU is flagged as sharding.
Add quoted rates
Select a GPU category and model, then enter the hourly rate you were quoted. Reference rates are optional published figures for context.
Priced models
0 of 19 models entered · 19 in selected categoryNo rates yet. Enter a rate above to start the comparison.
Tip: you can add multiple GPUs. Each stays editable in the list below. Blank remains blank.Rates are used exactly as entered. Published references are optional context only and do not imply live inventory. NVL72 per GPU figures are normalized; rack sales and topology are quoted separately.
Compare total cost
Only priced models and the baseline appear. Ranked by selected term total. Theoretical equivalence only.
Enter a rate to see your comparison
Add at least one quoted rate above. We will show the fleet size required to match your baseline, the monthly run rate, and the total for the term you chose.
Baseline is H100 SXM5 × 32 · 1 year · bf16 tensor basisHow totals are calculated
Practical GPUs = max(ceil(baseline BF16 × 32 / alternative BF16), ceil(baseline GB × 32 / alternative GB)). Fleet per hour = practical GPUs × your quoted rate. Monthly = fleet per hour × 730. Total = fleet per hour × 8,760 hours for 1 year. Utilisation, kernels, batch shape, network and power are not modelled. NVL72 platform figures are per GPU normalized. Theoretical equivalence only.
Have ClusterBid review your quote and sourcing requirements
Send the baseline, your entered rates, and the term. We will review the requirements and help map them to suitable capacity. Theoretical equivalence is a starting point. Real performance needs a benchmark on your model, precision, and context length.