Guide

Right now the cheapest cloud GPU on our index is $0.15/hr for a RTX A4000 at Hyperstack. That number moves every day, so treat it as a starting point rather than a promise. The price you actually pay depends less on picking the "cheapest provider" and more on matching the right GPU, billing model, and tier to your workload. Below are the six levers that lower a GPU bill, and the traps that quietly raise it.

The six levers that make GPU rental cheap

1. Use community and spot tiers. The single biggest saving is dropping off dedicated datacenter hardware. A community tier pools GPUs from independent hosts and prices the same card well below the secure equivalent. A spot or interruptible instance goes further, renting spare capacity at a discount in exchange for the provider being able to reclaim it. Both are fine when your job can checkpoint and resume: batch inference, rendering, fault-tolerant training. Neither is right for a live endpoint that cannot go down. Read the secure vs community trade-off before you switch.

2. Bill by the second, not the hour. Per-second billing charges only for the seconds a GPU actually runs. On short or bursty work that is a real difference: a 90-second inference call costs 90 seconds, not a rounded-up hour. RunPod bills per second across its products; Lambda bills per minute on on-demand. For spiky traffic, serverless that scales to zero between requests means you pay nothing when idle.

3. Right-size your VRAM with quantization. You do not need an 80GB card to run a model that fits in 24GB. Quantization stores a model's weights at lower precision (8-bit or 4-bit instead of 16-bit) and shrinks its VRAM footprint, often enough to move it from an expensive datacenter GPU down to a consumer card. An $1.99/hr H100 is the obvious pick for a big model, but a quantized 7B or 13B model runs happily on a $0.69/hr RTX 4090 at a fraction of the rate. The cheapest GPU is the smallest one your model actually fits on.

4. Arbitrage regions. The same provider often prices the same GPU differently by region, following local supply and power costs. If your data and latency budget allow it, a cheaper region is free money. The catch is data transfer between regions, which some providers charge for even when egress to the internet is free, so check before you move terabytes across the map.

5. Time your purchase against price history. GPU rental prices trend, and they trend down more often than up as newer hardware ships. Watching the price history tells you whether today's rate is a genuine low or just noise, and whether a card you want is on a downward slope worth waiting a week for. Nobody else publishes this history for free per GPU and per provider, which is exactly why it is a lever most renters never pull.

6. Stack free credits. Before you pay anything, spend someone else's money. Most providers hand out free credits to test the platform, and the programs stack: cloud startup credits, research grants, and trial balances can cover a surprising amount of early work. Credits are the highest-leverage way to trial a provider before committing real budget, so start there.

The traps that quietly raise your bill

The advertised "from $X" rate is rarely what you pay. Four things inflate it:

• Egress fees. Moving data out of a network is billed per gigabyte, and on the big hyperscalers it can cost more than the compute. GPU-focused clouds like RunPod and Lambda charge nothing for egress, which is a real reason to prefer them for data-heavy pipelines. See egress fees.
• Storage bundles. Persistent and network storage is billed separately from compute and keeps charging even while your instance is stopped. A few hundred gigabytes of checkpoints adds up over a month.
• Minimum commitments. Reserved terms cut the hourly rate but bill you whether you use the hours or not. Cluster products in particular hide weekly minimums behind an attractive per-hour headline. If your utilization is low, a min-commit raises your real cost instead of lowering it.
• Stale price lists. Most comparison pages are scraped occasionally and quote numbers that are days or weeks old. A price you cannot date is a price you cannot trust. Every number on this site carries its source and the time it was fetched, so you can see exactly how fresh it is.

Fold all four into the real number and you get the effective cost per hour, which is what our table sorts on instead of the sticker rate.

Where to go next

Two tools do the work from here. The live price table shows every GPU across providers, sorted by effective cost per hour, each cell timestamped and sourced. And if you are weighing a longer commitment, the rent-vs-buy calculator shows the break-even math on live prices, with the formula on the page so you can check it yourself. Start with the cheapest card that fits your model, rent it on the right tier, and let the history tell you when to buy more.