Task Definitions & Baseline Models
This report calculates platform costs based on three typical AI workloads:
1. LLM Fine-tuning (7B/13B/70B)
LoRA/QLoRA fine-tuning of Llama 3/Mistral series, batch_size=4, seq_len=4096.
2. Diffusion Model Training (Stable Diffusion 3)
Training SD3 from scratch, resolution 1024×1024, batch_size=8.
3. Inference Deployment (7B/13B/70B)
vLLM inference service, 50 concurrent requests/s.
Performance Score Methodology
Value = Relative Performance Score / Hourly Cost ($)
Relative score uses H100 SXM5 as baseline 100, other GPUs scaled by tokens/s or iterations/s.
LLM Fine-tuning (70B LoRA) Cost Ranking
| Rank | Provider | GPU | $/h | Score | Value |
|---|---|---|---|---|---|
| 1 | Vast.ai | H100 | $1.5 | 95 | 63.3 |
| 2 | RunPod | H100 | $2.5 | 98 | 39.2 |
| 3 | GCP (Spot) | A100-80GB | $0.7 | 80 | 114.3 |
| 4 | Voltage Park | A100-80GB | $1.5 | 82 | 54.7 |
| 5 | AWS (Spot) | H100 | $1.0 | 95 | 95.0 |
GCP Spot A100-80GB has the highest value (114.3), but Spot instances carry interruption risk. Among on-demand, Vast.ai H100 leads (63.3).
Diffusion Training Cost Ranking
| Rank | Provider | GPU | $/h | Score | Value |
|---|---|---|---|---|---|
| 1 | Vast.ai | RTX 4090 | $0.15 | 45 | 300.0 |
| 2 | RunPod | RTX 4090 | $0.25 | 48 | 192.0 |
| 3 | GCP (Spot) | L4 | $0.25 | 30 | 120.0 |
| 4 | AWS (Spot) | A10G | $0.30 | 32 | 106.7 |
| 5 | Vast.ai | A100-40GB | $1.0 | 75 | 75.0 |
RTX 4090 dominates diffusion training value because SD3's VRAM needs (24GB) match perfectly.
Inference Deployment Ranking (70B vLLM)
| Rank | Provider | GPU | $/h | tokens/s | $/M tokens |
|---|---|---|---|---|---|
| 1 | RunPod | H100 (Serverless) | $2.5 | 3500 | $0.71 |
| 2 | Vast.ai | H100 | $1.5 | 3200 | $0.47 |
| 3 | AWS (Spot) | H100 | $1.0 | 3400 | $0.29 |
| 4 | GCP (Spot) | A100-80GB | $0.7 | 2800 | $0.25 |
| 5 | Lambda Cloud | H100 | $2.8 | 3500 | $0.80 |
For inference, use Serverless endpoints (e.g., RunPod Serverless) — pay-per-request is cheaper than always-on instances.
Recommended Combo Plans
Budget Plan (Monthly < $500)
- Training: Vast.ai RTX 4090 × 4 ($0.60/h total)
- Inference: RunPod Serverless H100 ($0.71/M tokens)
- Monthly 100h training: $60
- Monthly 100M tokens: $71
- Total: $131/month
Balanced Plan ($500-2000/month)
- Training: GCP Spot A100-80GB × 8 ($5.6/h total)
- Inference: Vast.ai H100 always-on ($1.5/h)
- Monthly 100h training: $560
- Monthly 200M tokens: $94
- Total: $654/month
Enterprise Plan ($2000+/month)
- Training: Lambda Cloud H100 cluster × 8 ($22.4/h)
- Inference: Lambda Cloud H100 × 2 always-on ($5.6/h)
- Monthly 100h training: $2240
- Monthly always-on: $4032
- Total: $6272/month (99.9% SLA)
Hidden Costs
| Hidden Cost | Affected | Est. Extra |
|---|---|---|
| Data Egress | AWS/GCP/Azure | $0.05-0.12/GB |
| Storage (SSD) | All | $0.05-0.15/GB/mo |
| IP Address | AWS | $3.6/mo/EIP |
| Snapshots | AWS/GCP | $0.05/GB/mo |
Dedicated GPU clouds like Vast.ai and RunPod typically don't charge data egress — an advantage for large data training.
Methodology
Data sources:
- Pricing: TransAtlas real-time price database (auto-updated every 4h)
- Performance: MLPerf 4.0 + in-house benchmarks (Llama 3 70B + SD3)
- Collection period: September 1-23, 2026
- Sample: 15+ platforms, 500+ price data points
Performance scores are relative (H100 SXM5 = 100). Actual performance varies by driver/CUDA version and framework optimization.