ThunderCompute GPU Cloud Pricing: Complete Guide ($/hr for Every GPU)

Contents

• Introduction
• ThunderCompute Overview
• Complete Pricing Breakdown
• Pricing Comparison
• Choosing the Right GPU
• Cost Optimization
• FAQ
• Related Resources
• Sources

Introduction

ThunderCompute offers GPU cloud infrastructure targeting machine learning practitioners and researchers. The platform emphasizes transparent per-hour pricing across diverse hardware options. This guide provides complete pricing information for every GPU type available on ThunderCompute as of March 2026, with detailed comparisons to competing providers.

ThunderCompute Overview

ThunderCompute operates global GPU infrastructure with focus on cost efficiency and simplicity. Unlike marketplace providers like Vast AI, ThunderCompute manages hardware directly, ensuring consistent quality and availability. Unlike premium providers like Lambda, ThunderCompute prioritizes competitive pricing.

Platform strengths:

• Transparent hourly pricing (no hidden fees)
• Multiple GPU options from budget to premium
• Global data center presence
• Simple web interface and CLI provisioning

Key limitations:

• Smaller provider with fewer GPUs than AWS/GCP
• Limited multi-region redundancy
• No production SLAs or dedicated support

Complete Pricing Breakdown

ThunderCompute pricing varies by GPU type and data center region. US regions typically offer lowest rates; international regions command 10-20% premiums.

Budget GPUs

RTX 3090

• Price: $0.28/hour (US-Central)
• VRAM: 24GB
• Bandwidth: 576 GB/s
• Use case: Entry-level ML, prototyping

RTX 4080

• Price: $0.45/hour
• VRAM: 12GB
• Bandwidth: 576 GB/s
• Use case: Smaller models, inference

L4

• Price: $0.52/hour
• VRAM: 24GB
• Bandwidth: 300 GB/s
• Use case: Inference focus (lower power)

Mid-Tier GPUs

RTX 4090

• Price: $0.48/hour
• VRAM: 24GB
• Bandwidth: 576 GB/s
• Use case: Most popular training GPU

A10

• Price: $0.67/hour
• VRAM: 24GB
• Bandwidth: 600 GB/s
• Use case: Multi-workload (training + inference)

RTX A6000

• Price: $0.74/hour
• VRAM: 48GB
• Bandwidth: 576 GB/s
• Use case: Larger models requiring 48GB

High-End GPUs

A100 PCIe (40GB)

• Price: $1.35/hour
• VRAM: 40GB
• Bandwidth: 1.6TB/s (HBM2e)
• Use case: Production 13-34B model training

A100 SXM (80GB)

• Price: $1.68/hour
• VRAM: 80GB
• Bandwidth: 2TB/s
• Use case: Large model training, multi-GPU clusters

H100 PCIe (80GB)

• Price: $1.38/hour
• VRAM: 80GB
• Bandwidth: 2TB/s (HBM3)
• Use case: Fast training, largest models

H100 SXM (80GB)

• Price: $2.49/hour
• VRAM: 80GB
• Bandwidth: 3.35TB/s
• Use case: Maximum performance, multi-GPU setups

Premium GPUs

GH200 Grace Hopper

• Price: $4.50/hour
• VRAM: 120GB
• Bandwidth: 4.5TB/s
• Use case: Largest models, GPU+CPU acceleration

B200

• Price: $7.20/hour
• VRAM: 192GB
• Bandwidth: 8TB/s
• Use case: Next-generation large models (70B+)

Pricing Comparison

ThunderCompute vs. Vast AI

Vast AI targets budget buyers via peer-to-peer marketplace:

Vast AI undercuts ThunderCompute 40-50% but with reliability trade-offs. Unvetted providers mean inconsistent performance and frequent terminations.

ThunderCompute vs. Lambda Labs

Lambda emphasizes reliability and support:

ThunderCompute undercuts Lambda on H100 SXM as well as on PCIe and A100. Lambda includes managed support; ThunderCompute emphasizes self-service.

ThunderCompute vs. JarvisLabs

JarvisLabs bundles integrated development environments with GPU access:

ThunderCompute offers lower raw pricing. JarvisLabs includes pre-installed tools (Jupyter, Git) reducing setup overhead.

Choosing the Right GPU

By Workload Type

Development & Prototyping

• RTX 4090 ($0.48/hr) for speed
• Or RTX 3090 ($0.28/hr) for budget
• Saves $1-2 per experiment vs. A100

7B Model Fine-Tuning

• RTX 4090 recommended ($0.48/hr, 8 hours) = $3.84
• A100 acceptable ($1.35/hr, 6 hours) = $8.10
• RTX 4090 faster, cheaper overall

13-34B Model Fine-Tuning

• A100 SXM ($1.68/hr, 15 hours) = $25.20
• H100 ($1.38/hr, 10 hours) = $13.80
• H100 faster despite higher hourly rate

70B+ Model Training

• H100 SXM ($2.49/hr) for single GPU
• 2x H100 SXM ($4.98/hr total, 40 hours) = $199.20
• B200 ($7.20/hr) if available

By Budget

< $5

• RTX 3090 or RTX 4090 spot pricing
• Fits 7B model validation

$5-25

• A100 PCIe for 13B exploration
• Or RTX 4090 for multiple experiments

$25-100

• A100 SXM for multi-epoch 13B fine-tuning
• Or multiple A100 experiments

> $100

• H100 for 70B models
• Or multi-GPU setups

Cost Optimization

1. Spot Instance Discounts

ThunderCompute offers spot pricing 40-60% below on-demand:

Spot rates:

• RTX 4090: $0.19/hr (60% discount)
• A100: $0.55/hr (59% discount)
• H100: $0.57/hr (59% discount)

Strategy: Use spot for fault-tolerant workloads with checkpointing. Implement checkpoint every 30 minutes to recover from interruptions.

Cost reduction: 60% for interruption-tolerant training.

2. Regional Pricing

US-Central typically lowest. International regions 10-20% higher.

Optimization: Use US region for cost-sensitive work if geographic latency acceptable.

Cost reduction: 10-20% by regional selection.

3. Committed Usage Discounts

ThunderCompute provides discounts for 1, 3, or 12 month commitments:

• 1 month commitment: 15% discount
• 3 month commitment: 25% discount
• 12 month commitment: 35% discount

Strategy: Commit if baseline GPU volume known (e.g., ongoing research with predictable needs).

Cost reduction: 15-35% for committed users.

4. LoRA + Quantization Stacking

Combine LoRA (4-10x speedup) with quantization (50% memory reduction) and use cheaper GPUs:

Example: 13B model on RTX 4090 with LoRA + quantization

• Standard: A100 SXM ($1.68/hr, 15 hours) = $25.20
• Optimized: RTX 4090 ($0.48/hr, 2 hours) = $0.96
• Savings: $24.24 (96% reduction)

Trade-off: LoRA adapters cannot fine-tune all parameters. Quantization reduces accuracy marginally. Validate output quality.

5. Batch Experiment Runs

Run 5-10 related experiments on single instance. Amortize setup overhead across experiments.

Cost: Single GPU rental covers multiple training runs without repeating initialization.

Cost reduction: 20-40% by avoiding repeated setup.

FAQ

How does ThunderCompute pricing compare to AWS?

AWS A100 on-demand: $4.08/hour for single GPU ThunderCompute A100 SXM: $1.68/hour

ThunderCompute costs 59% less. AWS justifies premium through SLAs, compliance certifications, and integrated ecosystem.

Does ThunderCompute offer free tier or credits?

ThunderCompute does not offer free tier. Deposit required to activate account. New users may contact sales for startup credits.

Can I launch multiple instances?

Yes. Provision as many instances as account quota allows. Quotas increase with account age and payment history.

What about data transfer costs?

ThunderCompute charges outbound data transfer: $0.10-0.30 per GB depending on destination. Inbound transfer free. Minimize by caching data locally on instance.

How reliable is ThunderCompute compared to Lambda?

ThunderCompute maintains 99.5% monthly uptime SLA. Lambda provides 99.9% SLA with dedicated support. Difference: ThunderCompute fails 3.6 hours/month vs. Lambda 0.7 hours/month. For most ML workloads, this difference is acceptable.

Does ThunderCompute support distributed training?

Yes. Provision multiple instances and configure with PyTorch Distributed or Horovod. ThunderCompute's network connectivity enables inter-instance communication at acceptable latency.

Related Resources

• Vast AI Pricing Guide
• Lambda Cloud GPU Pricing
• JarvisLabs GPU Pricing
• GPU Pricing Guide

Sources

• ThunderCompute Official Pricing: https://www.thundercompute.io/pricing
• ThunderCompute Documentation: https://docs.thundercompute.io