D

calculator

Free AI Training Cost Estimator

Estimate the GPU-hours and dollar cost to pretrain or fine-tune a transformer model. Uses the standard 6N FLOP-per-token compute proxy, plus realistic 2026 GPU rates for H100, H200, B200, and A100 across the major cloud providers. Supports full pretraining, continued pretraining, LoRA, and full SFT scenarios. Includes electricity, networking, and utilization assumptions.

Inputs

$
%
$
%

Results

Total Training FLOPs (×10^21)

5

Total GPU-Hours Required

1,497

Wall-Clock Time at Cluster Size (Days)

0

Compute Cost (GPU Rental)

$5,240

Electricity Cost

$225

Infra Overhead Cost

$984

Total Estimated Cost

$6,448

Cost per Billion Parameters

$806

Cost per Billion Training Tokens

$64

By David Shadrake · Free, no signup required

AI & Machine Learning Resources

Strategic Playbooks

Roles That Use Tools Like This

Decisions to Compare

Explore Further

About the Author

David Shadrake

David Shadrake works on strategic business development and tech partnerships, with focus areas across AI, fintech, venture capital, growth, sales, SEO, blockchain, and broader tech innovation. Read more of his perspective on partnerships, market dynamics, and emerging technology at davidshadrake.com.