How GPU Finder Works
Data sources, collection methodology, and who built this.
Last updated: August 24, 2026 (changelog)
What this is
GPU Finder indexes GPU cloud pricing and availability across 30 providers so you can find the cheapest instance without opening 30 browser tabs. Prices update daily. Availability from 17 provider APIs updates hourly. No ads, no sponsored placements. Rankings based on price and availability data only.
How the data works
| Source | What it covers | Update frequency |
|---|---|---|
| SkyPilot catalog (open source) | On-demand & spot prices for 20 providers | Daily |
| Custom API fetchers | Hot Aisle, Vultr, Theta EdgeCloud, AceCloud, Gcore, EcoHash, Lium, Packet (public pricing page), and other direct pricing integrations | Hourly |
| 17 provider APIs | Real-time GPU stock / availability | Hourly |
| Manual research | Egress costs, GPU specs | Updated on change |
Built on open data
On-demand and spot pricing for 20 of the 32 providers we collect comes from the SkyPilot catalog (Apache 2.0). The same upstream feed powers the SkyPilot GPU Compass; their team links to GPU Finder as a sibling dashboard built on the same open catalog.
What GPU Finder adds on top of the catalog: real-time stock checks via direct provider APIs (17 today), 30-day reliability scores derived from those checks, manually researched egress costs, and per-provider / per-GPU landing pages. The catalog itself doesn't track availability or egress.
Frozen, citation-ready reliability research and availability-only downloads are published in the GPU Cloud Availability Index. Live pages keep moving; Index editions retain an absolute reporting period and generation timestamp.
Methodology
Name normalization
Provider APIs use inconsistent GPU names — “A100-80G”, “a100-80gb”, and “A100 SXM 80GB” all mean the same chip. We normalize names to a canonical model list before matching. If a provider's GPU name doesn't match cleanly, it's flagged for manual review rather than silently misclassified.
Ranking
Results sort by on-demand price per hour, ascending. No provider pays for placement. Sorting is deterministic — same data, same order, every time. Some provider links are referral links — we may earn a small commission if you sign up. This does not affect rankings or which providers are shown. See how this works in practice on our H100 pricing page or AWS vs Lambda comparison.
Availability
“Available” means the provider's API reported stock at the last check. It is not a reservation — high-demand GPUs can go from available to exhausted between our hourly poll and your request. Providers without availability APIs show pricing only. Signals older than six hours are treated as stale and therefore unknown for current availability.
Limitations
- Prices shown are published list prices. Committed, reserved, and negotiated enterprise rates are not compared.
- Egress costs are manually researched from provider pricing pages. Verify before large data transfers.
- We don't run test workloads, benchmark throughput, or verify SLAs. We show what providers publish.
- SkyPilot catalog updates daily. If a provider changed prices in the last 24 hours, there may be a lag.
Who built this
Alexander Salikov. Head of External Solutions at Nebius (one of the listed providers), based in London. Built this to scratch my own itch comparing GPU prices across providers. Nebius data comes from the same SkyPilot source as every other provider — no special treatment.
FAQ
Is GPU Finder the same as GPUFindr?
No. GPU Finder (gpufinder.dev) is a different tool from GPUFindr (gpufindr.com), which is run by a different team. GPU Finder tracks 30 cloud GPU providers with real-time availability from 17 provider APIs, egress costs, and 30-day reliability scores. Built and maintained by Alexander Salikov in 2026.
Why is a provider missing from GPU Finder?
Either they do not publish machine-readable pricing, or the fetcher has not been built yet. Email hello@gpufinder.dev to request a provider.
Why do GPU prices not match the provider website?
Pricing data updates daily via SkyPilot and the custom fetchers. If you spot a discrepancy, the provider likely changed prices recently and the next collection has not caught it yet. Email hello@gpufinder.dev and we will check.
Why does availability say Available but I can't launch an instance?
API-reported availability is not a reservation. High-demand GPUs can go from available to exhausted between the hourly poll and your request. Check the provider console directly for current status.
How do I report bad data on GPU Finder?
Email hello@gpufinder.dev with the incorrect pricing, missing provider, or data issue. We typically investigate within 24 hours.
GPU price digest
Big price moves, new providers, and feature updates. Occasional, not on a schedule. One-click unsubscribe.