NVNVGPU

How NVGPU compares cloud GPUs

Editorial methodology reviewed September 26, 2026. Published by NVGPU. Price source dates below are maintained separately from editorial review dates.

Provider inclusion and coverage

Our automated coverage is a maintained list of 9 providers whose versioned gpuhunt feeds we can parse and validate. A provider needs a public pricing destination and valid NVIDIA compute listings with a positive hourly price, an integer GPU count and an explicit spot/on-demand flag. Inclusion does not require an affiliate arrangement. A provider outside this list is not presumed more expensive or less reliable.

We currently focus on NVIDIA cloud compute, not AMD GPUs, TPUs, every marketplace seller, reserved contracts, negotiated enterprise quotes or serverless per-request billing. DataCrunch consumes the Verda feed under its existing site key; it is counted once. Provider selection and incomplete upstream coverage limit every “lowest listed” claim.

Follow a price back to its source

Rental prices are imported from gpuhunt's versioned provider catalogs, a third-party aggregation project. NVGPU normalizes those records; it does not claim to have collected each price directly from the provider. Each offer links to its immutable upstream ZIP and the provider's pricing destination. The ZIP's provider CSV contains the instance name, region, price and configuration.

Download the normalized catalog (JSON). It changes with refreshes; cite the immutable source version when you need reproducibility.

Normalization and billed cost

USD per GPU-hour = USD per instance-hour ÷ reported GPU count. An eight-GPU instance at $16/hour is $2/GPU-hour, but you pay $16/hour for that configuration. These are illustrative numbers, not a current offer. A row describes a rentable configuration in the feed; it does not prove the provider has no smaller instance.

Compute budget = full-instance hourly price × billed hours. Monthly illustrations use 730 hours. We retain source price precision to four decimal places and round display values; rounding is not a discount. We do not convert currencies. Unknown storage is not free storage, and GPU memory is separate from host RAM.

Storage, egress, public IPs, software licenses, support, taxes and billing minimums are not consistently described by the feed and may add charges. The calculator accepts your own budget for extras; zero means excluded. Commitment and reservation prices need a separate contract comparison. We do not assume a spot discount, duration or available capacity.

Ranking, ties and grouping

The homepage and comparison tool default to per-GPU hourly price ascending. Advanced search defaults to full-instance hourly price; its text search uses relevance. Ties use lower whole-instance cost, then provider key in alphabetical order, then the stable offer identifier. A user-selected sort applies to its named field, with the same price rules breaking ties. Affiliate links, commission amounts, enrollment status and advertising are not ranking inputs.

The homepage defaults to H100 on-demand offers with any GPU count. The comparison tool defaults to H100 80GB and A100 80GB, one GPU each, on-demand. Both include all listed regions until you filter. These different defaults can produce different minima; matching conditions must produce the same price.

We group raw records by provider, GPU model, GPU count, reported VRAM, explicit hardware variant, spot status and continent. Within a group we retain the cheapest row and up to five additional distinct configurations. Raw-price ties use region, instance name, CPU, RAM and disk in that order. GPU/provider summary pages select the cheapest continent for each configuration group. A stable offer link identifies that group; its price can change. Source snapshot links preserve the original evidence.

Filters operate on those retained headline groups, not every underlying region. A continent match does not prove availability in your exact country. Different GPU variants and providers are alternatives, not equivalent measured performance. “Cheapest” means cheapest among listed offers under the stated filters; it is never a claim about the whole market.

Missing data and hardware evidence

Invalid or non-positive prices, fractional or missing GPU counts, non-NVIDIA rows and ambiguous spot flags are excluded. Missing host resources or GPU memory remain unknown; we do not fill them from a full-card specification. PCIe, SXM and NVL labels are extracted only when explicitly present in the source GPU or instance name. “Unspecified” is not evidence of a particular form factor. Region labels are retained as supplied.

Manufacturer reference specifications and theoretical peaks are separately labeled and sourced on the hardware comparison. They are not application benchmarks. NVGPU has not conducted throughput, uptime or reliability tests for these listings. Third-party benchmark claims require an identifiable report and test conditions; we do not infer tokens per second from TFLOPS.

Source age and failed refreshes

Source publication dates in this view: Sep 24, 2026–Sep 26, 2026. Publication is date-level; the time each price was observed is not supplied.

Catalog generated: . This is not a source or deployment timestamp.

The scheduled GitHub workflow attempts an upstream refresh daily, validates every configured provider, runs tests and builds the site before publishing data. Ordinary site builds validate the committed catalog without silently downloading different prices. The separate cache-revalidation endpoint does not download pricing.

A provider snapshot date comes from its version string; its exact publication time and each price's observation time are not supplied. Midnight UTC in machine-readable dates represents date-level precision. Catalog generation records normalization time. Deployment completion is a third, separate event.

A failed scheduled refresh fails the workflow and leaves production's previous catalog intact. An explicitly requested stale fallback retains the previous offer values and source dates and records a failed refresh attempt. We flag sources more than 72 hours old relative to catalog generation; this is an editorial review threshold, not proof of incorrect pricing. Source dates remain visible even below that threshold, and old prices remain in the order rather than being silently described as current. Check today's date and the provider quote before acting.

Authorship and corrections

NVGPU publishes these comparisons using the source and calculation policy above. Our existing public contact is @nvgpu on X. For a correction, include the page and offer link, provider, GPU variant, region, pricing type, source version and supporting public URL. Do not send credentials or account details.

Corrections are checked against the cited evidence, then reflected in the relevant content and change record. Editorial dates change when the substantive content is reviewed or corrected; a deploy alone does not update them. No personal credentials or hands-on testing experience are implied by the NVGPU byline.

Change record

September 26, 2026: replaced the comparison tool's static prices and unsupported benchmarks with the shared catalog; separated explicit hardware variants; added per-offer provenance; corrected full-instance budgeting; revised older guides and documented ranking and commercial independence.

Commercial disclosure · Compare offers · Provider coverage