#1
Vast.ai
Cheapest GPU cloud — peer-to-peer marketplace for budget training
- Absolute cheapest GPU compute available
- Widest GPU variety including consumer cards
Price accurate?
H100 cloud comparison · April 2026
Best H100 Cloud Providers 2026
Where to actually get NVIDIA H100 capacity — 7 clouds compared on on-demand price, availability and cluster size. From $1.99/h.
The H100 market in April 2026
The NVIDIA H100 is the dominant accelerator for serious LLM training and high-throughput inference in 2026. Compared to the A100, it delivers ~3× FP16 throughput and ~6× FP8 throughput thanks to the Transformer Engine — but availability is the bottleneck, not performance.
Across the 7 GPU clouds with on-demand H100s, hourly pricing spans $1.99/h to $4.10/h for identical hardware. The choice is rarely just price — it's where you can actually get H100 capacity right now.
Specialist clouds win on price. RunPod, Lambda Labs and CoreWeave dominate on-demand H100 availability and cost 40–60% less than AWS p5 / GCP A3 / Azure NDA100 v5 for equivalent compute.
| Provider | Starting Price | Top GPUs | Highlights | Rating | CTA |
|---|---|---|---|---|---|
| Vast.ai Editor's Choice | from $0.10/h | RTX 3090, RTX 4090, A100 ≤80GB |
| ★★★★☆ | View pricing |
| RunPod Editor's Choice | from $0.16/h | RTX A5000, RTX 3090, RTX 4090 ≤80GB |
| ★★★★★ | View pricing |
| AWS GPU (EC2) | from $0.526/h | T4, A100, H100 ≤80GB |
| ★★★★☆ | View pricing |
| Azure GPU (NC T4/A100) | from $0.526/h | T4, A100, H100 ≤80GB |
| ★★★★☆ | View pricing |
| Lambda Labs Editor's Choice | from $0.69/h | Quadro RTX 6000, A100 40GB, A100 80GB ≤80GB |
| ★★★★★ | View pricing |
| CoreWeave | from $1.25/h | L40S, H100 SXM, A100 SXM ≤80GB |
| ★★★★☆ | View pricing |
| Google Cloud GPU | from $3.67/h | A100 40GB, A100 80GB, H100 ≤80GB |
| ★★★★☆ | View pricing |
#2
RunPod
Best value GPU cloud — huge selection, community + secure cloud
- Cheapest community GPUs from $0.16/h
- Massive GPU variety including H100
Price accurate?
#3
AWS GPU (EC2)
Largest GPU fleet worldwide — T4 entry, P4/P5 for enterprise
- Most comprehensive ML toolchain (SageMaker)
- Spot instances for massive cost savings
Price accurate?
#4
Azure GPU (NC T4/A100)
Microsoft's GPU cloud — T4 entry, best for Azure ML and enterprise AI
- Deep OpenAI / Azure OpenAI integration
- Best choice for Microsoft-stack enterprises
Price accurate?
#5
Lambda Labs
On-demand H100 clusters — developer-favourite for serious ML
- Reliable on-demand H100 availability
- No complex setup — SSH ready in seconds
Price accurate?
#6
CoreWeave
Enterprise GPU clusters — Kubernetes-native with H100 & L40S
- Best multi-node GPU cluster performance
- High-speed InfiniBand interconnects
Price accurate?
Frequently Asked Questions
Which cloud has the cheapest H100 in 2026?
RunPod Secure Cloud at $1.99/h is the cheapest on-demand H100 80GB. RunPod Community can be cheaper but is interruptible. For reserved/long-term commits, Lambda Labs and CoreWeave can quote significantly lower than the $1.99/h on-demand rate.
Why are H100s often unavailable on AWS?
AWS p5 (8× H100) instances are concentrated in select regions (us-east-1, us-west-2, eu-west-1) and are heavily reserved by enterprise customers. On-demand stockouts are common during US working hours. Specialist clouds like RunPod and CoreWeave have larger free-pool inventories.
H100 vs A100 — which should I rent?
For Llama-3 70B fine-tuning or large-scale training, H100 is 2–3× faster and despite costing more per hour, often cheaper per training run. For inference of <13B models or research workloads, A100 80GB is more cost-effective.
How many H100s do I need to fine-tune Llama-3 70B?
For full fine-tuning: 8× H100 (one DGX-equivalent node) for ~12-24 hours per epoch with 100K samples. For QLoRA: 1× H100 80GB suffices for ~6-8 hours. CoreWeave and Lambda Labs are best for multi-node H100 jobs (InfiniBand interconnect).
H100 SXM vs PCIe — what is the difference?
H100 SXM (used by CoreWeave, AWS p5, GCP A3) has NVLink up to 900 GB/s for multi-GPU jobs, while H100 PCIe (RunPod, Lambda) is limited to PCIe Gen5 ~128 GB/s but is ~10-15% cheaper. SXM is essential for ≥4-GPU training, PCIe is fine for single-GPU inference and ≤2-GPU training.
Get notified when prices change
Get an email when GPU prices drop or availability changes at your preferred provider.
No spam. Unsubscribe any time.