Background
Artificial Intelligence computing capacity, known as “AI compute”, has become an important consideration within the AI ecosystem.
Research by the OECD in collaboration with Oxford University Innovation identified the availability of public cloud-based compute resources relevant to AI across cloud regions. Data presented on OECD.AI represents an updated approach, automating the extraction of data on GPU availability at a more granular level using availability zones.
A cloud region is a geographic area designated by the cloud provider. In practice, each cloud region contains one or multiple cloud availability zones, which in turn may host one or more physical data centres.
Scope of the data
The analysis covers compute resources offered by major cloud service providers: Google, Microsoft Azure, AWS, Exoscale, Baidu, Tencent, and Alibaba. Additional providers (e.g., Huawei, CoreWeave, Oracle) will be included in future updates.
The unit of analysis is the cloud availability zone, the smallest deployable compute location unit exposed by a cloud provider’s API. Each zone is associated with a physical location and a set of compute offerings, including those with GPUs capable of AI workloads.
Data collection
Compute availability data are obtained through programmatic queries to the publicly accessible APIs of each cloud provider. An API key or account may be required by the provider as part of their standard access procedures; however, no privileged or confidential access beyond standard API usage is used. The compute offerings come in the form of specific instances, or sets of compute specifications that include memory, CPU, and if applicable, GPU type.
For each provider, we then retrieve:
- the list of cloud regions and cloud availability zones
- the set of compute instances available in each zone
- the capabilities of each instance, including GPU model and attributes
All analyses are based on information accessible through standard provider APIs. No proprietary, confidential, or provider-specific contractual data are included.
The dataset is updated on a regular schedule to incorporate changes in provider offerings and new availability zones.
Classification of GPU compute capability
GPU tiers
Each GPU model is classified into one of four broad tiers based on its capability to support AI workloads:
| Training | Fine-tuning | Inferencing | |
| Tier 1 (high AI capability) | ✔ | ✔ | ✔ |
| Tier 2 (medium AI capability) | ✘ | ✔ | ✔ |
| Tier 3 (low AI capability) | ✘ | ✘ | ✔ |
| Tier 4 (no AI capability) | ✘ | ✘ | ✘ |
This tiered classification draws on documented capabilities of GPU models commonly used for AI workloads, as well as expert input. Many availability zones do not contain GPUs at all, and are categorised as “no GPU available.”
See below for the full list of GPUs discovered and their classification.
Zone classification
Each availability zone is assigned the highest tier GPU it offers. For example, an AZ offering at least one Tier 1 GPU is classified as Tier 1, regardless of the presence of lower-tier GPUs. Note that a fifth category of zones are those that contain no GPU at all.
This approach allows zones to be compared based on the most advanced compute capability available at that location.
Indicators, aggregations, and limitations
The methodology yields several key indicators:
- Shares of availability zones by country and tier
- Counts of specific GPU models across availability zones
Where appropriate, results are aggregated to provide high-level summaries (e.g., total zones by country) by summing across zones and providers. Shares are calculated at the number of availability zones in a given tier divided by the total number of zones in that country or territory.
Limitations of the data
This methodology is subject to the following limitations:
- Availability vs. scale:
The presence of a GPU model in an AZ indicates availability but does not quantify the number of GPUs deployed. Providers do not disclose detailed capacity metrics via public APIs. - Temporal dynamics:
Cloud provider offerings change frequently. Periodic API snapshots may capture changes with delay. Data therefore provide a snapshot in time, and caution is advised when comparing different versions of the data.
Future work aims to complement these data with additional measures such as energy consumption and data centre capacity estimates to provide more comprehensive assessments of compute scale.
Full GPU classification
The following GPU classifications were used for all GPUs made available via the providers’ APIs.
| Tier 1 (Capable of training, finetuning, and inference) | Tier 2 (Capable of fine-tuning and inference) | Tier 3 (Capable of Inference) | Tier 4 (Not capable of AI workloads) |
| Nvidia A100 | NVIDIA A10 | NVIDIA T4 | FPGA (unspecified) |
| Nvidia A800 | NVIDIA A10G | NVIDIA T4g | Alibaba Cloud ANT1 |
| Nvidia A40 | NVIDIA A30 | AWS Inferentia | Alibaba Cloud GI6S |
| NVIDIA B200 | NVIDIA V100 | AWS Inferentia2 | NVIDIA Tesla P4 |
| NVIDIA B300 | NVIDIA L4 | AMD Radeon Pro V520 | NVIDIA Tesla P40 |
| NVIDIA H100 | NVIDIA GeForce RTX 3080 Ti | AMD Radeon Pro V710 | NVIDIA Tesla P100 |
| NVIDIA H200 | NVIDIA RTX A5000 | Intel Xe SG1 | NVIDIA Tesla M60 |
| NVIDIA L40S | Google TPU v3 | NVIDIA Tesla M40 | |
| NVIDIA RTX 6000 Ada Generation | AMD Radeon Instinct MI25 | ||
| AMD Instinct MI300X | |||
| Intel Gaudi HL-205 | |||
| AWS Trainium | |||
| AWS Trainium2 | |||
| Google TPU v7 | |||
| Tencent Zixiao C100 | |||
| Google TPU v3 | |||
| Google TPU v5 Lite Pod | |||
| Google TPU v5p | |||
| Google TPU v6e |
References
Lehdonvirta, V. et al. (2025), “Measuring domestic public cloud compute availability for artificial intelligence”, OECD Artificial Intelligence Papers, No. 49, OECD Publishing, Paris, https://doi.org/10.1787/8602a322-en.