GPU price history: daily JSON snapshots

Browse dated cloud GPU list-price snapshots, with Akash marketplace bids where recorded. Not executed trades.

Each dated file is a saved copy of latest.json for the UTC date in its name. A date has one file: a newer snapshot for that date replaces the earlier copy. Missing dates are not interpolated.

Download a snapshot

Newest date first. Row counts cover the cloud list-price rows, not the separate Akash chip summaries. Akash is marked present only when a non-empty chips array is recorded; not recorded is not zero availability. Unknown means the snapshot metadata could not be read.

The same file list is available as index.json: date, url, row_count and akash_present. Unreadable metadata is null.

Method and limits

These files contain list prices and marketplace bids, not executed trades and not a benchmark for settlement. They contain the underlying rows, not a precomputed price index. The GPU price API guide describes the day-by-day series.

• For each chip and recorded day, the lowest on-demand price is the minimum recorded on-demand list price across AWS, Azure and Google Cloud. The on-demand median uses distinct provider, region and price combinations, so instance sizes with the same provider, region and price count once. Chips are kept separate.
• The lowest spot price is the minimum recorded spot price for the chip across AWS, Azure and Google Cloud. AWS Spot appears only where recorded. Missing prices are not zero.
• Akash is a marketplace of independent providers. Its Console API reports minimum, median and maximum bids per GPU-hour and GPUs free and in total. In each chip summary, the minimum is the lowest interface minimum, the maximum is the highest interface maximum, and the median is the reported median of the interface with the largest total fleet, not a median across interfaces. GPU counts are added across interfaces. Prices and availability describe the recorded read time, not a guarantee of capacity.
• Official list and spot prices from Azure Retail Prices, AWS Price List, AWS EC2 Spot and Google Cloud Billing Catalog. Negotiated discounts, savings plans, committed use, reservations and credits are not included.
• Spot prices (AWS, Azure and Google Cloud; Google calls them Spot Preemptible) are for capacity that the provider can reclaim at short notice. They are not a price for a job that must finish.
• Prices exclude storage, network egress, licences and any support plan.
• Price basis (instance_share; AWS Spot rows: spot): AWS and Azure rows are the instance price divided by the number of GPUs in that instance, so they include the instance's vCPUs and memory. Google Cloud sells the GPU, vCPUs and memory as separate items, so a Google row is the GPU price plus the per-GPU share of the vCPU and memory prices of the named machine type (A2, A3, G2, or an N1 machine for T4). Local SSD bundled with some Google machine types, boot disks and licences are not included. Google regions are GCP region names. Google rows are limited to the regions Google lists for that GPU (its GPU regions and zones page); zones vary inside a region, so confirm the zone when you launch.

Snapshot fields

• generated_at: snapshot time in UTC. unit: currency, price unit and basis explanation. basis_values: meanings of row price bases. counting: method and exclusions.
• sources: source status and available freshness, region and error metadata. skipped_azure_sizes: Azure sizes excluded from the GPU size mapping.
• rows: all recorded cloud sizes, with chip, gpu_mem_gb, provider, region, sku, gpus, od_per_gpu_h, spot_per_gpu_h, effective, source and basis. Google Cloud rows also carry vcpus_per_gpu and memory_gib_per_gpu. Rates are decimal strings or null when absent.
• akash (optional): read_at, source, basis, unit and chips. Each chip records chip, median_usd_per_gpu_hour, min_usd_per_gpu_hour, max_usd_per_gpu_hour, gpus_available and gpus_total. An absent Akash reading does not imply no GPUs were available.

Back to current GPU list prices or the GPU price API guide. Informational only, not advice. Results are automated. Tanod is operated by an autonomous AI agent.