GPU Training → Cost Estimator
The paper says 72 hours on 8×A100. Budget for 90. Spot preemption, checkpoint I/O, and that one GPU throwing Xid errors — they all add 20–50%. Model real training economics across 6 providers. Not the idealized numbers in the arXiv methodology section.
🟢📐 Training Run Parameters
Configure your estimated training workload. GPU pricing fluctuates; this calculator provides order-of-magnitude estimates for budgeting.
🟢📉 Savings Scenarios
Compare how pricing model choice changes your total bill. Spot instances offer the deepest discount but can be preempted mid-training — only suitable for checkpoint-resumable workloads.
| Scenario | Discount | Est. Total Cost | Risk |
|---|
Cloud GPU Pricing Landscape (2026)
GPU pricing for ML training is undergoing a structural shift. Three forces are reshaping the market: H100/B200 supply ramp is pushing down A100 pricing on tier-2 clouds; Lambda Labs and RunPod have built reservation-free GPU clouds that undercut hyperscaler on-demand rates by 40–70%; and AMD MI300X and Intel Gaudi 3 are creating credible alternatives to NVIDIA for specific workloads.
This calculator models the six most relevant providers for independent ML teams. AWS/GCP/Azure are included for enterprise procurement comparisons, but most independent teams will find the best pricing on Lambda Labs, RunPod, or Vast.ai.
| GPU Model | VRAM | FP16 TFLOPS | Typical $/hr (On-Demand) | Best-For |
|---|---|---|---|---|
| H100 (80 GB) | 80 GB HBM3 | 990 | $2.50–5.00 | Large-scale LLM training (70B+ params) |
| A100 (80 GB) | 80 GB HBM2e | 312 | $1.10–3.00 | Foundation model training, diffusion models |
| A100 (40 GB) | 40 GB HBM2e | 312 | $0.80–2.50 | 7B–13B model fine-tuning |
| L40S (48 GB) | 48 GB GDDR6 | 362 | $0.60–1.50 | Inference + light fine-tuning |
| RTX A6000 (48 GB) | 48 GB GDDR6 | 192 | $0.40–1.00 | LoRA fine-tuning, inference serving |
| RTX 4090 (24 GB) | 24 GB GDDR6X | 330 | $0.30–0.70 | Small model training, experiments |
| RTX 6000 Ada (48 GB) | 48 GB GDDR6 | 364 | $0.50–1.20 | Fine-tuning 7B–13B, quantized inference |
How to Use This Estimator for Training Budget Planning
1. Estimate GPU-Hours Honestly
Most teams underestimate training time by 2–3×. Factor in: hyperparameter sweeps (multiply by N), failed runs (10–20% overhead), checkpointing I/O time (5–15% of wall clock), and debugging iterations (2–3 reruns per model version). Apply the Utilization Factor to model these overheads.
2. Spot vs. On-Demand Trade-off
Spot/preemptible instances are 50–70% cheaper but can be terminated at any time. For training runs longer than 24 hours, ensure your training loop supports resumable checkpoints to object storage (S3/GCS). Without checkpointing, spot savings are a false economy — a preemption 60 hours into a 72-hour run wastes everything.
3. Don't Forget Egress and Storage
Training a 70B-parameter model generates ~140 GB of checkpoints per save. Over 72 hours with hourly checkpointing, that's ~10 TB of writes. This calculator includes a rough storage + network allowance per provider, but for production budgeting, use the Cloud Storage Cost Comparison tool to model storage costs separately.
Pricing Basis, Sources & Assumptions
Every rate on this page is a published vendor list price — no negotiated discounts, private pricing, or credit offsets. Totals are in USD and exclude tax. Rates were last checked against the sources below on .
| Pricing input | Basis used on this page |
|---|---|
| Region | No single region. Rates are a range across the GPU rental market, from hyperscaler on-demand at the top to specialist GPU clouds at the bottom; the model anchors on the low end and lets the utilisation factor carry the rest. |
| Currency | USD — on-demand list price, tax excluded |
| Last checked | — GPU specifications come from the vendor data-centre pages; per-hour rate ranges are carried from the linked rental marketplaces as of this date. |
Modelling assumptions
- Rates are USD per GPU-hour for on-demand rental, quoted as ranges: H100 80 GB $2.50–5.00, A100 80 GB $1.10–3.00, A100 40 GB $0.80–2.50, L40S $0.60–1.50, RTX A6000 $0.40–1.00, RTX 4090 $0.30–0.70.
- Training time is an input, not a prediction — it comes from your own measured run at the batch size and sequence length you actually use.
- The utilisation factor scales effective hours: at 80% the same run costs 25% more, because the GPU is billed for wall-clock time, not for useful FLOPs.
- HBM capacity, memory bandwidth, and FP16 throughput in the table are vendor-published specifications, not measured by jslet.
What this model excludes
- Storage, checkpointing, and dataset egress, which are priced separately and often reach 10–20% of a training bill.
- Orchestration overhead, failed runs that must be resubmitted, and hyperparameter sweeps.
- Preemption risk on spot or interruptible capacity, and reserved/committed discounts that can cut 30–60% off the hourly rate.
Sources
- NVIDIA (2026). "H100 Tensor Core GPU." HBM3 capacity, memory bandwidth, and dense FP16 throughput. nvidia.com
- NVIDIA (2026). "A100 Tensor Core GPU." HBM2e capacity and bandwidth for both the 80 GB and 40 GB parts. nvidia.com
- RunPod (2026). "GPU Cloud Pricing." On-demand per-GPU-hour rates used for the mid range. runpod.io
- Lambda (2026). "Lambda GPU Cloud." On-demand and reserved GPU instance rates. lambdalabs.com
- Vast.ai (2026). "GPU Rental Marketplace." Marketplace floor rates — the low end of every range on this page. vast.ai
Vendor list prices change without notice — re-check the linked pages before committing spend. jslet takes no vendor sponsorship and carries no affiliate links; see about.
🟢📎 Embed This Calculator
Add the GPU training cost estimator to your team docs, internal wiki, or blog. No API key. No tracking. 100% client-side — inputs never leave the browser.
<iframe src="https://www.jslet.com/embed/gpu-training-cost" width="100%" height="700" style="border:none;border-radius:10px;" title="GPU Training Cost Estimator" loading="lazy"></iframe>