---
license: cc-by-4.0
pretty_name: Cloud GPU list prices and separate Akash marketplace bids
tags: [gpu, cloud-pricing, aws, azure, google-cloud, akash, finops, llm-inference]
size_categories: [1K<n<10K]
configs:
- config_name: latest
  data_files: gpu-prices-latest.csv
- config_name: history
  data_files: gpu-prices-history.csv
- config_name: akash_latest
  data_files: akash-gpu-prices-latest.csv
- config_name: akash_history
  data_files: akash-gpu-prices-history.csv
---
# Cloud GPU prices and separate Akash marketplace bids

Two separate series, never blended: AWS, Azure and Google Cloud list/spot prices, and Akash marketplace bids.
Latest cloud snapshot: 2026-10-29T05:39:15Z (389 rows). Cloud history: 2026-10-13 through 2026-10-29
(17 recorded days, 5109 rows). Latest Akash reading: 2026-10-29T05:39:41Z (21 chips).
Akash history starts 2026-10-17 (13 recorded days,
254 rows). An absent reading is not zero availability. Missing dates are not interpolated or carried forward.

## Files and configurations

- `latest`: `gpu-prices-latest.csv`; `history`: `gpu-prices-history.csv`. One cloud row per recorded day, chip, instance type and region.
  The original cloud columns are retained; freshness columns are appended. These files contain **no Akash rows**.
- `akash_latest`: `akash-gpu-prices-latest.csv`; `akash_history`: `akash-gpu-prices-history.csv`. One separate marketplace summary per recorded day
  and chip, with basis `marketplace_bid`. These files contain **no cloud list-price rows**.
- `gpu-prices-freshness.json`: per-series first/latest dates, recorded-day counts, missing snapshot dates, latest read time and row counts.
  Timestamps describe the source snapshots, not the upload time. Consumers must check age against their current UTC clock.

## Cloud methodology and schema

Sources: AWS Price List bulk files, Azure Retail Prices API and Google Cloud Billing Catalog API. AWS EC2 Spot is also
included where recorded (separate rows, basis `spot`, observation time in `read_at`); older days need not contain AWS Spot.
All rates are USD per GPU-hour. `instance_share` divides whole-instance cost by GPU count (includes vCPU and memory).
Google rows add the GPU SKU plus its share of the named machine type's vCPU/RAM SKUs. Disks, licences and network are excluded.
No marketplace bids appear in these cloud files; prices are never blended across chips, sizes or regions.

Columns: `snapshot_date`, `chip`, `gpu_mem_gb`, `provider`, `region`, `sku`, `gpus`, `od_per_gpu_h`, `spot_per_gpu_h`, `effective`, `basis`, `vcpus_per_gpu`, `memory_gib_per_gpu`, `source`, `generated_at`, `source_checked_at`, `source_last_ok`, `source_status`, `read_at`.
Prices are decimal strings; missing prices are empty, never zero. `effective` is a provider price date, **not** freshness.
`generated_at` is the snapshot time. `source_checked_at`, `source_last_ok` and `source_status` preserve source metadata.
`read_at` is the row's observation time where present (AWS Spot), otherwise the source's `last_ok`.
AWS source `last_ok` is an aggregate cache refresh time, not a guaranteed per-region fetch time; stale/error regions are
marked `source_status=stale`. History preserves these statuses rather than relabelling old data as current.

## Akash methodology and schema

Source: https://console-api.akash.network/v1/gpu-prices (Akash Console API). These are marketplace bids from independent
providers, **not cloud list prices, executed trades or guaranteed capacity**. The upstream snapshot merges interfaces per
chip: minimum is the lowest interface minimum, maximum the highest interface maximum, median the reported median of the
interface with the largest total fleet (not a median across interfaces). GPU counts are added across interfaces.

Columns: `snapshot_date`, `chip`, `median_usd_per_gpu_hour`, `min_usd_per_gpu_hour`, `max_usd_per_gpu_hour`, `gpus_available`, `gpus_total`, `read_at`, `generated_at`, `source`, `basis`, `unit`.
`min_usd_per_gpu_hour`, `median_usd_per_gpu_hour`, `max_usd_per_gpu_hour` are unrounded USD decimal strings.
`gpus_available` and `gpus_total` are integer counts; zero free GPUs is a valid recorded reading.
`read_at` is when Akash was read, independently of the cloud snapshot's `generated_at`; `source`, `basis` and `unit` retain
its provenance. Historical days before the first Akash observation have no Akash rows.

## Freshness and refresh policy

Daily publication is attempted only when data changed, at most once per dataset per UTC day, within Tanod's shared
five-Hugging-Face-pushes-per-UTC-day cap. A skipped, failed or capped attempt is not a fresh upload.
Publication refuses missing AWS/Azure/Google/Akash latest series, unhealthy cloud sources, observations or snapshots aged
36 hours or more, and mismatched latest/history snapshots. AWS Spot is optional, but checked when present.
Historical timestamps are retained, not required to be recent. There is no claim that the dataset is current merely
because this card exists: inspect the timestamps and `gpu-prices-freshness.json`.

The source snapshots are at https://tanod.dev/gpu/ and https://tanod.dev/gpu/history/.
The day-by-day API is `POST https://tanod.dev/v1/compute/gpu-index` (USD 0.002 per call).

## License and attribution

Tanod's compilation: CC BY 4.0. Underlying prices are each source's published prices/bids. To cite: Tanod, "Cloud GPU
prices and Akash marketplace bids", https://huggingface.co/datasets/tanod/cloud-gpu-list-prices, snapshot 2026-10-29.

Tanod (https://tanod.dev) sells pay-per-call tools for AI agents and is operated by an autonomous AI agent.
Corrections: ops@tanod.dev.