How we collect and rank prices.
Where every number on RentGPU comes from, how often it refreshes, how raw listings become the single hourly rate you see, and where the numbers can mislead you.
1. GPU rental prices
Most prices come from each provider's own public API or listings endpoint, collected by a dedicated adapter per provider:
- Marketplaces, every 5–10 minutes — Vast.ai, RunPod, io.net, Clore.ai, Akash, Fluence and Nosana. Prices there are set by individual hosts and change constantly, so they are sampled most often.
- Hourly — Lium.
- Every 6 hours — Thunder Compute and Cyfuture AI.
- Daily — Lambda and DeepInfra, whose list prices rarely move.
- Hyperscalers — AWS, Google Cloud, Azure, Oracle, DigitalOcean and Vultr don't offer a simple price feed, so we record the on-demand list price from each one's public pricing page and refresh it by hand. These are list prices: spot, reserved and committed-use discounts can be much cheaper.
2. From listings to one number
- Normalize the GPU. Providers name the same card many ways ("H100 SXM5 80GB", "H100-80G", "1x H100"). An alias table maps each name to one canonical GPU; variants that genuinely differ (H100 SXM vs PCIe vs NVL) stay separate.
- Per GPU. The rate shown is for one orderable GPU, so an 8× node compares fairly with a single card.
- Summarize. For each GPU on each provider we store the median, minimum and maximum hourly price and the number of offers on every collection run. The headline rate is the median — one lowball listing shouldn't set the price you expect to pay.
- Freshness. GPU pages show only prices collected in the last 24 hours. A GPU without a recent price says so instead of showing a stale number.
- Rank. Providers are sorted by that median, cheapest first. Nothing else — including whether we have an affiliate link — affects the order.
Worked example: if a marketplace lists four RTX 4090 offers at $0.29, $0.34, $0.36 and $0.95 per hour, we record min $0.29, max $0.95 and median $0.35, and show $0.35/hr. The $0.95 outlier doesn't distort it, and neither does the single cheapest offer, which may already be taken.
3. AI models and hosted-API prices
The AI-model catalog — which models exist, their context window, and what each hosted API charges per million tokens — is refreshed daily from the public model catalogs of OpenRouter, Together AI and Fireworks AI. OpenRouter is itself an aggregator of inference providers, so a price shown "via OpenRouter" is OpenRouter's price. When several sources list the same model, each appears as its own row so you can compare them. Open-weights model metadata is supplemented weekly from the Ollama library and from model cards.
Model pages that don't yet have enough information — no usable description, or too few prices or hardware recommendations to compare — stay reachable but are kept out of search engines until they do.
4. Hardware recommendations and calculators
- VRAM needed = model weights (parameter count × bytes per parameter at the chosen precision: FP16 = 2 bytes, INT8 = 1, INT4 ≈ 0.5) + the KV cache for your context length and batch size, times a headroom factor for the runtime. It's a planning estimate; inference frameworks differ by a few GB.
- "Run it yourself" recommendations pair each open-weights model with GPUs whose memory fits it at a given quantization, and attach the live cheapest rental price for that GPU.
- Self-host vs API compares a month of rented-GPU time against hosted-API token prices for the same model at your traffic level, and shows the volume at which one overtakes the other.
5. Caveats
- Marketplace prices move minute to minute and an offer can be rented by someone else before you get there. The provider's checkout is the only binding price.
- Some decentralized providers (for example Akash) settle in tokens; we convert to USD at a recent exchange rate.
- Hourly price isn't total cost. Storage, egress, minimum billing increments and idle time add up — see the individual provider pages.
6. Affiliate disclosure
Some "Launch" / "Sign up" links are affiliate links and we may earn a referral fee. This never affects ranking — see the editorial policy.
Found a wrong number?
Tell us through the contact form. We fix data at the source, so one correction fixes every page that uses it.