Compare cloud GPU cost per GPU-hour
AWS, Google Cloud, Azure, Oracle, CoreWeave, Lambda & Nebius for H200, H100, A100 and L40S. Model your real workload — GPU count, hours, and term — and see total cost ranked instantly.
| Provider | Instance | $/GPU-hr | Monthly | Total · 12mo | vs priciest |
|---|---|---|---|---|---|
LambdaBEST neocloud | gpu_8x_h100_sxm5 80GB · NVLink (8-GPU node) | $3.99 | $23.3K | $279.6K | 68% cheaper |
Nebius neocloud | HGX H100 80GB · 3.2 Tbps InfiniBand | $4.50 | $26.3K | $315.4K | 63% cheaper |
CoreWeave neocloud | HGX H100 80GB · 3.2 Tbps InfiniBand | $6.16 | $36K | $431.7K | 50% cheaper |
AWS hyperscaler | p5.48xlarge 80GB · 3.2 Tbps EFAv2 | $6.88 | $40.2K | $482.2K | 44% cheaper |
Oracle (OCI) hyperscaler | BM.GPU.H100.8 80GB · 3.2 Tbps RoCEv2 (cluster network) | $10.00 | $58.4K | $700.8K | 19% cheaper |
Google Cloud hyperscaler | a3-megagpu-8g 80GB · 3.2 Tbps RoCE | $11.62 | $67.9K | $814.3K | 5% cheaper |
Azure hyperscaler | ND96isr H100 v5 80GB · 3.2 Tbps Quantum-2 InfiniBand | $12.29 | $71.8K | $861.3K | — |
* No published committed rate for this offering; on-demand shown. Per-GPU-hour derived from public 8-GPU instance list prices. Figures are estimates — verify before purchasing.
Looking for cheaper GPUs?
SponsoredMarketplace and budget providers often undercut the list prices above. These are partner links — how this works.
On-demand & spot GPUs (incl. H100/H200) billed by the second — often the cheapest way to start.
GPU marketplace with the lowest rates anywhere — rent spare capacity from verified hosts.
Dedicated NVIDIA clusters and on-demand instances with no minimums.
Developer-friendly GPU droplets and notebooks — great for fine-tuning and inference.
Estimate spend for Claude Code, Codex, LangGraph & CrewAI agents: cost per task, cost per successful task, monthly waste from retries, and a cheaper-routing suggestion across OpenAI, Anthropic, Gemini, DeepSeek and more.
Go deeper than list prices
Historical price trends, $/effective-FLOP and $/token normalization, 20+ providers and regions, reserved-vs-on-demand break-even modeling, TCO with storage & egress, and CSV / API export for your models.
- ✓ Price history & trend alerts
- ✓ Performance-normalized $/token
- ✓ 20+ providers & all regions
- ✓ Reserved break-even calculator
- ✓ Full TCO (storage, egress, networking)
- ✓ API access & CSV export
Frequently asked questions
How much does an NVIDIA H100 cost per hour in the cloud?+
As of Oct 25, 2026, public on-demand list prices run $6.88–$12.29 per GPU-hour on hyperscalers (AWS, Oracle, Google Cloud, Azure) and $3.99–$6.16 on neoclouds (Nebius, Lambda, CoreWeave). Committed or reserved terms are lower. Use the calculator above for your exact GPU count and runtime.
Why are neoclouds usually cheaper than AWS, Google Cloud and Azure?+
Neoclouds specialize purely in GPU compute, run leaner software stacks, and price aggressively to win reserved-capacity deals. Hyperscaler prices bundle a much broader managed-services ecosystem, global regions, and enterprise SLAs — which many AI workloads don't strictly need. The gap varies by provider: some hyperscalers have cut GPU prices sharply, so check the live table rather than assuming.
Are these the prices I'll actually pay?+
Treat them as directional. Real costs depend on region, reservations, committed-use discounts, savings plans, capacity blocks, egress, storage, and negotiated enterprise contracts. We surface public list prices and published reference rates so you can compare apples to apples — then confirm with each provider.
Why isn't xAI / SpaceX in the comparison?+
xAI (Colossus) and SpaceX don't sell a public, per-GPU-hour cloud product, so any number would be speculation. We only list providers with published rates. If that changes, we'll add them with clear sourcing.