Frequently asked questions

No. We do not sell, resell or broker compute and we never take payment. We publish normalised pricing collected from providers' public pages and link you to the provider so you rent directly from them.

A scheduled collector runs every day at 02:00 UTC. Every table shows the timestamp of the last successful collection. If a source fails, the run is logged and an operational alert is raised rather than silently reusing stale data.

Through affiliate partnerships such as the Runpod PartnerStack programme. If you click a provider link and later pay for compute, the provider may pay us a commission. You never pay more, and commissions never change the prices, ordering or verdicts on this site.

On-demand (also called uninterruptible or secure) means the instance stays yours for as long as you pay. Spot, interruptible or community capacity can be reclaimed by the host and is therefore cheaper. Prices are normalised to a single GPU per hour so different configurations can be compared fairly.

No. Both calculators run entirely in your browser and the results are stored in the URL fragment, so you can bookmark or share a configuration. Nothing is sent to our servers.

Either the provider does not currently list that model, or it was listed but the offer was withdrawn before the last collection. We would rather show a gap than an invented number.

For every GPU model both providers list, we compare the normalised single-GPU hourly price. The average of those unit prices is shown for each side, and the lower one is labelled cheaper. We deliberately exclude models only one of the two offers, because that is not a like-for-like comparison.

Not always. Marketplace capacity can be reclaimed at short notice and its hosts vary in network and disk performance, while a dedicated cloud gives more predictable throughput. For multi-node training, interconnect and NVLink support usually matter more than a few cents per hour.

Only roughly. Each platform defines interruption rules, notice periods and reclaim behaviour differently. Treat spot prices as an indication of the cheapest achievable rate, not as a guaranteed price.

Most providers bill by the second or the hour, and per-hour figures are the only ones that can be compared without guessing how many hours you will actually run. Use the GPU cost estimator to convert an hourly rate into a monthly budget.

For models that fit in memory, memory bandwidth matters more than raw FLOPS, so cards like the L40S, RTX 4090 or A100 often give better tokens per dollar than an H100. Once a model needs multi-GPU tensor parallelism, NVLink-equipped H100 or H200 instances usually win on throughput.

Full fine-tuning needs roughly the parameter count times the bytes per parameter for weights, plus optimiser state which can be several times larger again. LoRA and QLoRA reduce the optimiser footprint dramatically. Use the LLM VRAM calculator to size a specific configuration.

Renting usually wins unless you can keep a device busy for a large fraction of the year. Cloud rental also shifts the cost of hardware refreshes onto the provider, which matters when new generations arrive every twelve to eighteen months.

Configuration differs: number of GPUs per instance, system RAM and vCPU allocation, local disk, region, and whether the capacity is interruptible. We normalise to a single GPU so you can see through those differences.

It multiplies the hourly rate we collected by your expected running hours and adds storage and egress at the stated assumptions. It is a budgeting tool, not a quote: real invoices include taxes, idle time and any minimum billing increments the provider applies.

It computes weight memory from parameter count and precision, KV cache from the attention geometry (2 x layers x KV heads x head dimension x context x batch, in FP16), and adds 15 percent of the weights for activations and framework overhead.

Real usage depends on the serving framework, whether KV cache is paged or preallocated, CUDA graph capture, and how aggressively the engine reserves memory. Treat the number as a sizing guide and leave headroom.

Yes. The share button copies a link containing your inputs in the URL fragment (after the # symbol). The fragment never leaves the browser, so the shared link reveals nothing to our servers.

Almost all cloud GPU providers bill by the second or the hour for the compute, then separately for persistent storage and outbound data transfer. The figures on this page are the compute rate for one GPU, normalised so you can compare different instance shapes without doing the arithmetic yourself.

That depends on the provider and is shown in the availability column. Marketplace-style platforms usually offer a bid-based interruptible tier, while dedicated clouds tend to sell only on-demand capacity. If spot is available it is almost always cheaper, provided your workload can checkpoint and restart.

For on-demand and spot capacity there generally is not: you pay per hour or per second and can stop at any time. Reserved contracts exist at some providers and can cut the effective rate, but they are not required for the prices shown on this page.

Persistent volumes are billed per GB per month, and outbound transfer is charged per GB beyond whatever allowance the provider includes. Those costs are additive to the GPU hour and can dominate the bill on data-heavy pipelines, so budget for them explicitly.