AI compute market signals and learning

ComputeTape Capacity Board · CT-CAP

Where can you actually get GPU capacity?

A weekly board of where H100, H200, B200/GB200 and MI300X capacity can actually be requested across major providers — distilled into one GPU Availability Score. Directional, not a guarantee of stock.

78/ 100 · GPU Availability Score

Normalized from qualified provider observations · last refreshed July 21, 2026 · methodology

Changed this week

What moved since the last refresh

  • Oracle MI300X: No Some
  • CoreWeave B200 / GB200: Emerging Yes
  • Lambda B200 / GB200: No Yes
  • RunPod B200 / GB200: No Yes
  • Crusoe MI300X: No Yes

The board

Availability by provider and chip

Updated weekly. Per-provider on-demand prices live in the Provider Directory and the H100 Index.

ProviderH100H200B200 / GB200MI300XLarge ClusterQuota FrictionStatus
AWSYesYesEmergingNoStrongHighConstrained
AzureYesYesEmergingSomeStrongHighConstrained
OracleYesYesEmergingSomeStrongMediumLimited
CoreWeaveYesYesYesNoStrongMediumAvailable / tight
LambdaYesSomeYesNoModerateLow/MedAvailable
RunPodYesSomeYesSomeLowerLowVariable
CrusoeYesSomeEmergingYesStrongMediumExpanding
NebiusYesSomeEmergingNoStrongMediumExpanding

Use it

Embed, cite, or download the board

Embed the board

A card refreshed with the weekly data build and attributed back to this page.

Embed code

Cite this

Citation

Download the data

The full board as CSV — provider × chip availability and access friction.

How it's built

How the board is built

H100 and H200 availability is auto-confirmed from fresh provider observations; the remaining columns are editorially reviewed each week. The GPU Availability Score is a normalized 0–100 estimate derived from qualified provider observations — a directional signal, clearly an estimate. See the full methodology.

Availability reflects what providers advertise and what ComputeTape can observe — not a guarantee of stock at any moment. Power and grid constraints behind capacity are tracked on Power Watch.