GPU Cloud Cost Calculator
Compare on-demand hourly pricing for AI GPUs — H100, A100, L40S, RTX 4090 and more — across major clouds and specialized GPU providers.
Flagship GPU for training and high-throughput inference — 80GB VRAM
| Provider | Price / hour | Monthly Cost |
|---|---|---|
| bestVast.ai | U$ 1.65 | U$ 264 |
| RunPod | U$ 2.89 | U$ 462.4 |
| Lambda Labs | U$ 3.99 | U$ 638.4 |
| CoreWeave | U$ 4.25 | U$ 680 |
| Paperspace | U$ 5.95 | U$ 952 |
| AWS | U$ 6.88 | U$ 1,100.8 |
| Azure | U$ 6.98 | U$ 1,116.8 |
| Google Cloud | U$ 11.06 | U$ 1,769.6 |
GPU cloud prices change often and vary by region, availability and commitment level. Figures above were researched in July 2026 from each provider's pricing page — always confirm the current rate before committing to a workload. Unlike our LLM cost calculators, this data is not fetched live.
How GPU cloud pricing works
Renting a GPU in the cloud is priced per hour (sometimes billed per second or per minute), and the rate depends on the exact GPU model, how much VRAM it has, and which provider you use. Big hyperscalers like AWS, Google Cloud and Azure bundle GPUs into larger virtual machines with fixed CPU and RAM allocations, which tends to make them more expensive per GPU-hour but comes with broader ecosystem integration, compliance certifications and enterprise support. Specialized GPU clouds like RunPod, Lambda Labs, CoreWeave, Vast.ai and Paperspace focus purely on GPU compute, often at a fraction of hyperscaler prices, trading some of that breadth for lower cost. This calculator lets you pick a GPU tier, set how many hours per month you expect to use it, and compares the resulting monthly cost across providers.
Renting a GPU vs paying per token: when self-hosting wins
The hourly prices above become comparable to API pricing once you factor in utilization. A GPU at $2 per hour running around the clock costs roughly $1,500 a month whether you use it or not — money that buys a very large volume of tokens on a pay-per-use API (run your numbers in the OpenAI or Claude cost calculators for comparison). Self-hosting starts to win when utilization is consistently high, when data cannot leave your infrastructure, or when you're serving a fine-tuned or open-weights model the APIs don't offer.
As a rule of thumb, if your GPU would sit idle more than about 70% of the time, the API is almost certainly cheaper once you add the engineering hours that keeping inference servers healthy actually costs.
The break-even point against API pricing, model by model
The only way to know whether renting beats calling an API is to convert GPU-hours into tokens. Take a GPU at $2.00 an hour running full time: about $1,460 a month. Now ask how many output tokens that money buys at August 2026 list prices, per million tokens (input / output).
On Llama 4 Maverick ($0.20 / $0.60) it buys roughly 2.4 billion output tokens. On GPT-5.6 Luna ($0.20 / $1.20), about 1.2 billion. On Nemotron 3 Ultra ($0.50 / $2.20), about 660 million. On GLM-5.2 ($1.40 / $4.40), around 330 million. On Claude Haiku 4.5 ($1.00 / $5.00), roughly 290 million. On the flagships — GPT-5.6 Sol ($5.00 / $30.00) and Claude Opus 4.8 ($5.00 / $25.00) — under 60 million.
That is the whole argument in one paragraph, and the 30 July repricing sharpened it: Luna moved from about 240 million tokens to 1.2 billion, so a hosted proprietary tier now sits inside the range that used to belong only to open weights. Self-hosting competes with the cheap end of the market — where you would run those same open models anyway — and almost never with a flagship you cannot host at all. Price your real volume in the OpenAI or Claude calculator, then compare it to the monthly figures above.
Three real GPU scenarios, priced
A fine-tuning run. An 8-billion-parameter model on a single H100 for 40 hours. At a hyperscaler rate near $10 an hour that is about $400; on a specialized GPU cloud closer to $2.50 an hour, roughly $100. One-off training jobs are exactly where the specialized clouds pay for themselves, because you are buying raw compute and nothing else.
A part-time inference server. A single L40S at around $1.00 an hour, but only needed 8 hours a day on weekdays — about 170 hours a month, or $170, provided you actually shut it down. Left running around the clock the same card bills roughly $730. Idle hours are the single biggest source of waste in rented GPU budgets.
A production inference cluster. Four A100s running continuously at $1.50 an hour each is about $4,400 a month, before storage, egress and the engineer keeping it healthy. At that point the comparison above matters: $4,400 buys around 7 billion Llama 4 Maverick output tokens, or 175 million on GPT-5.6 Sol.
The costs the hourly rate doesn't show
The per-hour number is the beginning of the bill, not the end. Storage is billed separately and persists whether the GPU runs or not — model weights, datasets and checkpoints add up quickly on multi-terabyte workloads. Egress is where hyperscalers earn quietly: moving data out can cost more than the compute for data-heavy jobs, while specialized clouds often bundle it or charge far less.
Idle time is billed at the full rate. A GPU reserved to guarantee availability during a traffic spike costs the same at 3am as at peak. Cold starts matter too: pulling a large image and loading weights can take minutes, so autoscaling to zero is rarely as clean as it sounds.
And then there is engineering time — serving stacks, batching, quantization, monitoring, on-call. It never appears in a pricing table and is routinely the largest line item for small teams. A useful sanity check: if adding the cost of one engineer's attention flips your comparison, the API was probably the right answer.
Frequently asked questions
Why is the same GPU so much cheaper on specialized clouds?
Hyperscalers price GPU instances as part of a broader, heavily audited enterprise platform with global infrastructure, compliance certifications, and dedicated support — costs that get folded into the hourly rate. Specialized GPU clouds run leaner operations focused only on GPU compute, and some, like Vast.ai, resell spare capacity from smaller data centers at marketplace prices, which is why the same physical GPU can cost several times less there.
Should I always pick the cheapest option?
Not necessarily. Consider reliability guarantees, region availability, network performance, storage options, and how easy it is to scale to multiple GPUs or nodes. A marketplace provider like Vast.ai can be dramatically cheaper but comes with more variable reliability than a dedicated provider.
Does this include spot or preemptible pricing?
No, these are on-demand, non-interruptible rates. Spot or preemptible instances on AWS, GCP and Azure can cut costs significantly, but the provider can reclaim the instance with little notice — not ideal for long training runs without checkpointing.
Why don't all providers appear for every GPU?
Not every provider offers every GPU model. Older or more specialized cards are usually only available through GPU-focused clouds, while hyperscalers tend to standardize around a smaller set of current-generation options.
What utilization do I need for a rented GPU to beat API pricing?
There's no universal number — it depends on the model you'd self-host and the API rate you'd otherwise pay — but sustained, predictable load is the prerequisite. Spiky or low-volume traffic almost always favors per-token APIs, because idle GPU hours are billed just the same as busy ones.
How current are these GPU prices?
They were researched in July 2026 from each provider's public pricing page and are stored, not fetched live — unlike our LLM cost calculators, which query a live rate. GPU cloud pricing moves with supply, region and commitment level, and marketplace providers change fastest of all. Treat the table as a snapshot for comparison and confirm the current rate with the provider before committing a workload.
Do these rates include storage and bandwidth?
No. The figures are the compute rate for the instance only. Persistent storage, snapshots, and data egress are billed separately by essentially every provider, and on data-heavy workloads they can add a meaningful percentage to the bill. Hyperscalers in particular charge for egress at rates that can exceed the compute cost when you move large datasets out frequently.
Can I run a model like GPT-5.6 or Claude on a rented GPU?
No — those are closed-weights models available only through their providers' APIs. What you can self-host are open-weights models such as Llama 4, GLM-5.2, Kimi K3 and Nemotron 3. That is why the honest comparison is renting a GPU to serve an open model versus paying per token for that same open model through a hosted API, not versus a flagship you have no way of running yourself.
Curious where these numbers come from? Read our pricing methodology.