What drives the L40S price gap
The NVIDIA L40S is identical silicon everywhere — the 1.2× gap between Spheron at the bottom and the hyperscalers at the top is about packaging, not performance. Specialist and marketplace clouds compete on raw $/hr; AWS, Azure and Google bundle the card with their platform, networking and support and price accordingly. For a sustained training run or a busy inference fleet, settling on the cheapest reliable provider is one of the largest single levers on your compute bill.
On-demand vs spot
On-demand guarantees the GPU is yours; spot and marketplace supply is cheaper but can be reclaimed, so it suits fault-tolerant or checkpointed workloads. Toggle "Best (incl. spot)" in the table above to see the cheapest available rate including interruptible supply.
Should you rent, or use an API?
If your goal is running an LLM rather than training one, renting a L40Sonly beats a managed API above a breakeven volume — and the maths flips once you count idle hours. Check it first with the self-host vs API breakeven calculator.
Frequently asked questions
What is the cheapest cloud L40S price?
The lowest on-demand rate we track for the NVIDIA L40S is $0.72/hr at Spheron, across 2 providers (verified 2026-06-05). Spot and marketplace rates run lower with variable availability.
How much does an L40S cost per hour?
On-demand L40S pricing ranges from $0.72/hr to $0.86/hr per GPU depending on provider — about a 1.2× spread for the same card. Specialist clouds are cheapest; hyperscalers (AWS/Azure/GCP) sit at the top.
Is it cheaper to rent an L40S or use an LLM API?
Renting only wins above a breakeven volume, because a GPU bills every hour it exists while an API bills per token. Model your own crossover with the self-host vs API breakeven calculator before committing.
Independent comparison, no vendor influence. Standard on-demand pricing per single GPU, re-verified weekly (last 2026-06-05); negotiated and committed-use rates differ. Published under CC BY 4.0.