Training and fine-tuning economics
GPU training cost calculator
Calculate time and cloud cost from the tokens you will process and throughput measured on your exact cluster. Then inspect two published MLPerf time-to-quality references—kept separate from your estimate.
- No default training throughput
- Transparent epochs and overhead
- 20 hashed source runs across two quality targets
Your measured scenario
Estimate one complete training job
Cluster throughput must come from your own model, precision, sequence length, software and hardware test. Overhead adds checkpointing, evaluation and restart allowance.
Complete every field with measured or planned inputs.
Exact published system
Two reproducible training reference runs
Every result below uses the same Nebius eight-GPU B300 system and is re-priced with the current official on demand $7.85 per-GPU-hour catalog rate, equal to $62.80 per hour for the exact eight-GPU reference system. This is a transparent reference calculation, not a promise for another workload.
trainingTime to train to quality target
GPT-OSS 20B pretraining
$87.3772 / reference run
- Published result
- 5,008.89 seconds
- Reference time
- 83.48 min to published quality target
- System
- 8× NVIDIA B300
- Reference cluster rate
- $62.80 / h
- Reference yield @ $100 compute
- 1.14 reference runs
- Purchase model
- on demand
- Price observed
- 2026-08-13 02:40 UTC
- Capacity
- 8 GPUs available together: not verified
- Evidence
- 10 verified runs
- Source revision
eabf23a07b2a
C4 · 3.34 log perplexity
Inspect source evidence ↗trainingTime to train to quality target
Llama 3.1 8B pretraining
$75.3729 / reference run
- Published result
- 4,320.74 seconds
- Reference time
- 72.01 min to published quality target
- System
- 8× NVIDIA B300
- Reference cluster rate
- $62.80 / h
- Reference yield @ $100 compute
- 1.33 reference runs
- Purchase model
- on demand
- Price observed
- 2026-08-13 02:40 UTC
- Capacity
- 8 GPUs available together: not verified
- Evidence
- 10 verified runs
- Source revision
eabf23a07b2a
C4 · 3.3 log perplexity
Inspect source evidence ↗Calculation
Formula you can audit
Processed tokens = dataset tokens × epochs. Seconds = processed tokens ÷ measured cluster tokens/second × (1 + overhead). Cost = seconds ÷ 3,600 × GPUs × price per GPU-hour.