The price of a GPU-hour, in context.

ACROSS SOURCES · SAME CARD, MORE THAN ONE SELLER

Where two admitted sources price the same GPU

Every source below is recorded separately and none are pooled. But 4 cards appear in more than one of them, and reading four tables to notice that helps nobody. Each provider’s own published figure is repeated here, unchanged, beside the other — nothing is computed across them: no median, no ranking, no cheapest.

1 card is listed without a comparison. Its sources publish in different currencies and billing units, and GPUDeck does not convert between them: an hours-per-month figure and an exchange rate would produce a number neither provider stated, printed under their names. Both figures appear as published and the comparison is left undone rather than invented.

Two sources pricing the same card is not two quotes for the same product. Segment, region, commitment, contract length and what is bundled all differ, and a lower number can buy a worse thing. This table says only that both sources publish a price for that card, and what each one says.

One of these sources is a marketplace, and it may not be independent of the others. Lium lists machines contributed by third-party operators, and its public feed does not identify which operator runs any given node. So where a card appears on Lium and on a direct provider, GPUDeck cannot rule out that they are the same physical machine counted under two names. That is also why no figure on this site is ever computed across sources: pooling them could count one GPU twice. The rows are shown side by side, and this limit is stated rather than assumed away.

SOURCE 1 · DIRECT PROVIDER FEED

Lium — live marketplace nodes

Read directly from Lium’s public nodes feed with archived written permission (methodology §8, dated entry 2026-08-26). Prices are Lium’s own current asks in native USD per GPU-hour — a separate record from the tracker above, never pooled with it, and no cross-source comparison is computed. One node contributes one observation; a node is counted only while it is actually rentable (available ≥ its minimum). Secure tier only in the on-demand tables; spot is separate below.

Current day 2026-09-23 · feed timestamp 2026-09-23 05:21 UTC · 98 rentable nodes · 11 day(s) archived · single-GPU cells 11 · multi-minimum cells 10 · spot cells 1

On-demand, single-GPU — secure tier, minimum rentable unit 1 GPU

On-demand, multi-GPU minimum — a price you can only obtain N GPUs at a time is not a single-GPU price; these rows never enter any single-GPU figure

Spot / preemptible — interruptible capacity, its own segment, never merged into on-demand

Attribution: Lium (lium.io), public nodes feed /api/public/v1/nodes. Daily median (p50) and floor per exact (model, VRAM) from unique currently-rentable nodes. A missing day is a gap, never forward-filled. Full rows: lium-price-context.csv.

SOURCE 2 · PUBLISHED VENDOR TARIFF

Hetzner — published GEX tariff

The first source GPUDeck publishes without permission. Not a grant — a recorded conclusion that none is required: Hetzner created these prices, so the investment is in creating the data rather than obtaining it, and no qualifying database right attaches to a four-row vendor tariff. The full analysis, including what would void it, is in the source registry.

Read from Hetzner’s own published price document (GE-D9256), not from the product page — that page renders its prices client-side, and reaching the endpoint behind them would be the configurator route this source explicitly excludes. Monthly, excluding VAT and IPv4, exactly as Hetzner states it. No hourly figure is derived: Hetzner does not publish one, and inventing it would put GPUDeck’s number on the page under Hetzner’s name.

Tariff effective 15 June 2026 · document GE-D9256 · archived sha256 2eff766c93c7f254 · mirrored 2026-09-23T05:21:14Z · 4 SKUs

Hetzner's own two documents disagree about where the GEX131 runs. The price document files it under “Falkenstein and Helsinki”; the product page says Nuremberg (NBG1) and Falkenstein (FSN1). GPUDeck does not resolve it — picking one would manufacture a certainty neither source supports — so the tariff location is shown and marked disputed. The prices themselves are not in question.

Scope is these four SKUs. Hetzner’s server auction, its configurator and its live stock are not published here and reaching them would void the conclusion this record rests on.

SOURCE 3 · PUBLISHED VENDOR TARIFF (EU)

Leafcloud — published GPU tariff

The second source published without permission, and the first where an EU database right genuinely applies rather than being absent. A tariff table can structurally be a database, so the question is whether substantial investment went into obtaining its contents — and Leafcloud choosing its own prices and naming its own SKUs is creating the offering, which is not the same limb. The recorded analysis is in the source registry, including what would void it.

Only Leafcloud's own rows appear here. The same page carries a comparison table quoting other providers’ prices; none of it is republished, and the mirror refuses to write output at all if a competitor name reaches an extracted row. GPUDeck reads only the server-rendered tariff rows and never calls Leafcloud’s pricing backend. Nor is any hourly figure derived: Leafcloud publishes a monthly tariff, full VM included.

Mirrored 2026-09-23T05:21:15Z · archived sha256 fc343cb854b3e2f2 · 12 SKUs · EUR per month, full VM included

Scope is these SKUs. Leafcloud's competitor comparison, its percentage-cheaper claims, its calculator and its pricing backend are not published here, and reaching them would void the conclusion this record rests on.

SOURCE 4 · PUBLISHED VENDOR TARIFF (EU)

Nebius — published GPU price list

Admitted after a consultant reversed its own earlier refusal, and the distinction it drew is worth stating. Nebius was failed as a source when the route ran through an authenticated SDK enumerating platforms, presets, projects and regions — and that refusal still stands for that route. It does not govern this one: an unauthenticated, server-rendered marketing table is a different object and had to be screened on its own. Treating a route-specific objection as a property of every price the provider publishes was the error.

GPUDeck reads only that public table. Never the SDK, the calculator, the console, project metadata or the backend — reaching any of them would put this back under the route that was refused. The two GB-series entries priced “Contact us” carry no published price and are absent rather than guessed, and a from price is shown with its qualifier because it is a range floor for a configurable instance, not a price Nebius stated.

Mirrored 2026-09-23T05:21:15Z · archived sha256 793feb9813fde435 · 7 GPU rows · USD per GPU-hour

CPU-only instances, storage, networking, commitment discounts, regions and availability are all outside this record and are not published here.

SOURCE 5 · PUBLISHED VENDOR TARIFF

Koyeb — published GPU price list

Admitted on a recorded conclusion that no permission is required for this specific document: an unauthenticated, server-rendered public price list, read once a day and mirrored with its digest. The conclusion was cross-checked by two consultants and this session independently, and it governs this page only — it says nothing about Koyeb's API, console or any authenticated surface, none of which GPUDeck reads.

Only numerically priced GPU rows appear. CPU, storage, bandwidth, support tiers and the per-second column are excluded by construction rather than filtered afterwards, so nothing here depends on a judgement made at render time.

Mirrored 2026-09-23T05:21:15Z · read from https://www.koyeb.com/pricing · archived sha256 afa8c29e9c500dca · USD per GPU-hour

Regions, commitment discounts, availability and every non-GPU line item are outside this record and are not published here.