Methodology & Transparency

We focus on on-demand cloud GPU instances and try to present prices in a consistent way so providers can be compared side-by-side.

Prices and availability can change quickly. GPUs.io is best-effort and provided as-is. Always verify the final price on the provider site before purchasing.

We ingest pricing and configuration data from publicly available sources such as provider documentation, pricing pages, and provider APIs where available.

If a provider publishes multiple price points (regions, commitment discounts, spot pricing, promo credits), we generally prioritize the most comparable on-demand prices.

Providers describe SKUs differently (GPU naming, VRAM, CPU/RAM bundles, per-hour vs per-month). We normalize where possible to support consistent filtering and sorting.

• Units: Prices are typically shown as hourly rates. When providers publish monthly pricing, we convert to hourly using the provider billing model when available.
• GPU matching: We map vendor/model names to a canonical GPU type to avoid same-GPU different-label issues.
• Bundles: Many instances bundle GPU + CPU + RAM. GPUs.io compares the instance as a whole; we do not attempt to allocate a GPU-only price.

Sorting is typically based on the selected price metric (for example hourly price). Filters narrow visible instances/providers by attributes like GPU type and key specs.

If two entries appear tied, there may be hidden differences (region, network, storage, billing minimums, taxes/VAT).

GPU pages list popular open-weight LLMs and the VRAM we estimate they need. The weight size for each model is measured from the files of its released checkpoint on Hugging Face, not derived from its parameter count — parameter-based math is unreliable for models published in newer quantized formats. A checkpoint whose measured size matches no precision we recognize is held back rather than published with a guess.

On top of the weights we add 20% for KV cache, activations and framework overhead, assuming short context lengths. Long contexts, large batches, speculative decoding and multi-model serving all need more. Sizes for a precision a model was not released at (for example a 4-bit version of a 16-bit release) are estimated from measured community quantizations and rounded up. A model whose weights fit but whose overhead does not is shown as a tight fit, with the margin that configuration actually leaves: it will load, but a long context or a large batch can still exhaust the card.

A "×N" figure means the model fits in the total VRAM of N of these GPUs. It says nothing about interconnect bandwidth, which we do not track and which determines whether splitting a model across cards is actually fast. Treat every figure as a starting point and verify before committing to a rental.

We refresh provider data on a rolling basis. Some sources update instantly, others lag behind. The site may display a last-updated timestamp where available.

• Regional pricing: The same SKU can differ by region/zone.
• Availability: Prices may be published even when capacity is temporarily unavailable.
• Taxes/fees: VAT, local taxes, and transfer fees are often not included in published hourly rates.
• Spot / preemptible: We generally avoid mixing spot pricing into on-demand comparisons unless clearly labeled.
• Serverless: Per-second, scale-to-zero rates are shown as an hourly equivalent for comparison. They are not an hourly pod rate — you pay only while a request runs, and cold starts add time.

Want something corrected? Email [email protected] with a source link.