Flopper Now Tracks GPU Rental Prices Across 24 Cloud Providers

What the Same GPU Costs, Cloud to Cloud

The spread is the story. On the day of publication, an on-demand H100 SXM5 80 GB (2022) is listed by 16 providers, and the most expensive listing costs nearly seven times the cheapest. The B200 (2024) shows the same pattern with a three-fold gap across eight clouds. Some of that gap is real: interconnect, storage, support and region all cost money. Some of it is simply price discovery that has not happened yet, because nobody was lining the numbers up.

On-demand, per-GPU, USD, as recorded on 9 September 2026. Spot, community and reserved listings are excluded from the range. Each GPU links to its live pricing page, which will have moved on by the time you read this.

The 24 Providers

Coverage spans three kinds of seller. Hyperscale and general clouds publish rate cards: Microsoft Azure, Oracle Cloud, DigitalOcean, OVHcloud, Scaleway and Vultr. AI-first clouds, the neoclouds, sell the dense H100, H200 and Blackwell capacity that training runs need: CoreWeave, Lambda Labs, Nebius, Crusoe, DataCrunch, Hyperstack and their peers. Marketplaces, RunPod and Vast.ai, list third-party hosts and set the floor for most consumer and workstation parts.

A listing is one GPU at one pricing type from one provider. Vast.ai and RunPod count high because their marketplaces publish hosts for consumer GPUs the dedicated clouds do not carry.

How the Numbers Are Normalized

No two providers quote the same way. One prices an eight-GPU node, another a single accelerator, a third a monthly contract, and Scaleway prices in euros. Every listing goes through the same four steps before it reaches a page.

• Per GPU, per hour, in USD. Node prices are divided by the GPU count. Monthly rates are divided by 730 hours. Non-USD prices are converted at the live exchange rate on the day of the sync.
• Pricing types are kept apart. On-demand (191 listings), spot or preemptible (31), marketplace community hosts (111) and reserved commitments (4) never share a row. The headline cheapest price on each GPU page is always on-demand, so an interruptible instance cannot masquerade as the market rate.
• Every listing maps to one canonical part. A provider's "H100" is matched to the exact variant it sells, SXM5 or PCIe, 80 GB or 94 GB, so a PCIe price never dilutes the SXM figure. Names the matcher cannot place are logged and dropped rather than guessed.
• Bad data is quarantined, not published. A listing that lands far outside its GPU's price band is held back and flagged for review. Median, dispersion and FP8-per-dollar figures on the pricing index are calculated so that a single bad listing cannot move them.

The full method, including which provider feeds are official APIs and which are published rate cards, is on the methodology page. Curated listings from providers who work with Flopper directly sit in the same table but are never overwritten by the automated sync.

Price History, Not Just a Snapshot

A price comparison that only shows today cannot tell you whether $2.99 for an H200 is a bargain or the new normal. Since 14 July 2026 the pipeline has recorded 2,297 price points across 58 days and 101 GPUs, appending a row every time a listing changes. Each GPU pricing page plots the last 90 days with one line per provider, so you can see who cut, who held, and who quietly raised.

The index also reports a 30-day change per GPU. That figure compares the same providers' listings at both ends of the window, so a new cloud joining coverage registers as more data, not as a price move.

Where the Prices Show Up

• The pricing index: cheapest live rate per GPU, provider count, median, 30-day change and FP8 TFLOPS per dollar, sortable and filterable.
• Per-GPU pricing pages: every provider's rate for one accelerator, the 90-day history chart, and the sibling variants.
• GPU detail pages: the "Rent" call to action now leads straight to that GPU's live rates.
• Provider pages: each cloud's full price list, regions and pricing models in one place.
• The LLM GPU calculator: size a model, then see what the recommended hardware costs per hour today.

What It Does Not Cover Yet

Public list prices only. Negotiated enterprise rates, reserved-capacity contracts and volume discounts are not published and are not tracked. Availability is shown where a provider publishes it, and a listed price does not guarantee capacity in the region you need. AWS and Google Cloud are the largest gaps in coverage; both publish pricing in formats that need heavier parsing and validation, and both are next on the list.

Confirm the live rate with the provider before committing. Some outbound links are affiliate links, which fund the pipeline and never change which provider appears cheapest.

For GPU Clouds

If you sell GPU capacity and you are on the list above, your listing was built from your public pricing. You can claim the listing to add regions, interconnect details and a direct link. If you are not on the list, the same page is where to start. Density is the point: the comparison is only useful when it includes everyone.

Source Confidence

Pricing data is held to the same sourcing standard as spec data. Every listing carries its source and the time it was last synced, and both are visible on the pricing pages.