Back to Dashboard

Cloud GPU Price Index

A daily record of what secure on-demand cloud GPUs cost per GPU-hour, computed from the current offers GPUPerHour tracks across cloud GPU providers. Published as a table and as CSV; no charts yet.

What this index measures

Every day at 00:05 UTC the index records, for each of the 13 GPUs in the basket, the minimum, median and maximum price per GPU-hour across every current offer on the site. An offer is current only when all of the following hold:

  • the provider reports it as available right now;
  • the provider has live stock tracking, so "available" is a real-time observation, not a catalog entry;
  • it was seen within the last 15 minutes;
  • it is secure on-demand pricing on provider-operated data-center hardware, billed per GPU-hour;
  • spot, interruptible, reserved and peer-to-peer or marketplace listings are excluded.
Basket version
v2026-09
History
1 day of data since
Last snapshot

Latest snapshot by GPU

Prices are US dollars per GPU-hour. The median is taken across every current offer for the GPU; the minimum is the cheapest current offer. Providers and offers count the current set on the snapshot day.

GPUMedian $/GPU-hrMin $/GPU-hrProvidersOffersSnapshot
NVIDIA H100 SXM5 80GB$3.27$2.79Lyceum822
NVIDIA H100 PCIe 80GB$2.88$2.50Hyperstack914
NVIDIA H200 SXM 141GB$4.27$3.50Ori58
NVIDIA B200 SXM 192GB$6.99$3.75Packet.ai27
NVIDIA B300 SXM6 262GB$7.89$7.89RunPod11
NVIDIA A100 SXM4 80GB 80GB$1.50$0.82Vast.ai721
NVIDIA A100 PCIe 80GB 80GB$1.48$0.68LeaderGPU723
NVIDIA L40S 48GB$1.03$0.80Vast.ai416
NVIDIA RTX PRO 6000 Blackwell 96GB$2.17$0.66Packet.ai827
NVIDIA GeForce RTX 5090 32GB$1.23$0.53Vast.ai310
NVIDIA GeForce RTX 4090 24GB$1.80$0.40Vast.ai37
NVIDIA RTX A6000 48GB$0.57$0.44LeaderGPU753
AMD Instinct MI300X 192GB$2.99$2.39RunPod23

Daily series

The index has 1 day of history. Daily tables appear for each GPU once it has 3 snapshots, and trend commentary starts once 30 daily snapshots exist; the first monthly report is planned for . Until then, the full series is available in the CSV downloads.

No GPU has reached 3 daily snapshots yet, so there are no daily tables to show.

Download the data

One row per snapshot date and GPU, for every GPU with at least one current offer that day (the basket and the rest of the catalog). Columns: snapshot_date, gpu_slug, gpu_name, basket_version, min_price_per_gpu, p25_price_per_gpu, median_price_per_gpu, p75_price_per_gpu, max_price_per_gpu, offer_count, provider_count, cheapest_provider, cheapest_region. Prices are USD per GPU-hour; quartiles are blank below four offers.

Licence: the data may be reused, republished and charted with attribution to GPUPerHour and a link to https://gpuperhour.com/gpu-price-index (Creative Commons Attribution 4.0).

Methodology

Definition

Price field
pricePerGpu (USD per GPU-hour, on-demand list price)
Offer filter
available = true; security tier = secure; pricing type = on-demand; peer-to-peer = false; provider has live stock tracking = true; observed within 15 minutes
Statistics
min, p25, median, p75 and max of pricePerGpu across current offers per GPU model; quartiles omitted below 4 offers; percentiles use linear interpolation between order statistics (median of an even count is the mean of the two middle values)
Cadence
captured daily at 00:05 UTC; a re-run on the same UTC day replaces that day's row with the latest observation
History
series starts on the first capture; no back-filled or synthetic history

The basket (v2026-09)

13 GPUs chosen to cover training, inference, workstation and consumer segments across more than one provider. Membership only changes with a new basket version, so series are never compared across versions silently.

  • NVIDIA H100 SXM5: Reference training GPU; deepest on-demand market and the price most buyers benchmark against.
  • NVIDIA H100 PCIe: PCIe H100 trades separately from SXM and tracks the single-GPU inference segment.
  • NVIDIA H200 SXM: Current Hopper part with 141GB; captures the memory-bound inference premium.
  • NVIDIA B200 SXM: First Blackwell datacenter GPU with multi-provider on-demand availability.
  • NVIDIA B300 SXM6: Blackwell Ultra; thin early market worth recording from its first listings.
  • NVIDIA A100 SXM4 80GB: Previous-generation workhorse with the largest installed base; clearest depreciation signal.
  • NVIDIA A100 PCIe 80GB: PCIe A100 sits at a distinct discount to SXM; needed to describe the A100 market honestly.
  • NVIDIA L40S: Ada datacenter inference GPU; the most common mid-tier enterprise part.
  • NVIDIA RTX PRO 6000 Blackwell: Blackwell workstation GPU displacing L40S in new deployments.
  • NVIDIA GeForce RTX 5090: Current consumer flagship; proxy for the budget inference segment.
  • NVIDIA GeForce RTX 4090: Most-listed consumer GPU across secure providers; long comparable history.
  • NVIDIA RTX A6000: Ampere workstation GPU still widely offered; floor for 48GB-class pricing.
  • AMD Instinct MI300X: Only AMD part with multi-provider on-demand availability; keeps the index from being NVIDIA-only.

Exclusions

  • Spot, interruptible and preemptible pricing.
  • Reserved, committed and cluster pricing (see the GPU cluster page).
  • Peer-to-peer and marketplace hosts (community machines listed through a marketplace).
  • Providers without live stock tracking, whose "available" is a catalog entry rather than an observation.
  • Observations older than 15 minutes at snapshot time.
  • Offers with a missing or zero per-GPU price.

Known limitations

  • SKU-weighted, not capacity-weighted. Every current offer counts once, whether it is a single instance type or a fleet. A provider that lists six H100 SKUs moves the median more than one that lists one.
  • Thin GPUs can be dominated by one provider. When a GPU has two or three current offers, the median is that provider's price; check the provider and offer counts before reading the number as a market price.
  • One observation per day. Snapshots are taken at 00:05 UTC; intraday moves and stock-outs between snapshots are not captured. Prices can differ from a provider's checkout by a few minutes of lag.
  • List prices only. Storage, egress and negotiated discounts are not included.

Changelog

  • : series started, basket v2026-09 (initial capture dated 14 September 2026)

Frequently asked questions

How is the cloud GPU price index computed?

Once a day at 00:05 UTC, for each GPU in the basket, we take every current offer on GPUPerHour (available now, from a provider with live stock tracking, seen within 15 minutes, secure on-demand, priced per GPU-hour) and record the minimum, 25th percentile, median, 75th percentile and maximum price, plus the provider and offer counts. Quartiles are omitted when a GPU has fewer than four offers.

Why do index prices differ from a provider's list price?

The median is taken across all providers and SKUs, so it sits between the cheapest and most expensive current offers rather than matching any one price list. The minimum column is the single cheapest current offer. Prices are collected from public provider pages and can lag a provider's checkout by a few minutes.

How often does the index update?

One snapshot per UTC day, captured at 00:05 UTC. This page re-renders at most once an hour, and the CSV exports are regenerated on request with a five-minute cache. The live per-provider prices on the rent pages update every minute.

Can I use the data?

Yes. The CSV exports and the numbers on this page are licensed under Creative Commons Attribution 4.0: reuse, republish and chart them with attribution to GPUPerHour and a link to https://gpuperhour.com/gpu-price-index.

Why is there so little history?

The series started in September 2026 with basket v2026-09 and has no back-filled or synthetic history; every row is a real daily observation. Daily tables appear per GPU after three snapshots, trend commentary after thirty, and the first monthly report is planned for 15 October 2026.

How prices are collected and what counts as a current offer is described in full on the About page. Corrections: [email protected].