Most teams building with large language models don't run the models. They call one through a cloud platform that hosts many: Amazon Bedrock, Google's Gemini Enterprise Agent Platform (which Google renamed from Vertex AI in April 2026), or Microsoft Foundry (formerly Azure AI Foundry). A few run open-weight models on their own GPUs. The AI systems module covers how to build on a model; the LLM serving lab covers what it takes to serve one yourself. This page compares the platforms on the things that decide production designs: how capacity is guaranteed, where your data is processed, what you pay for, and how to stay portable.
Side by side
| AWS | Google Cloud | Azure | Self-hosted | |
|---|---|---|---|---|
| Model platform | Amazon Bedrock; SageMaker AI for your own endpoints | Gemini Enterprise Agent Platform, formerly Vertex AI; Model Garden | Microsoft Foundry, formerly Azure AI Foundry | vLLM or similar on GPUs in EKS, GKE or AKS |
| Default capacity | Per-model quotas, on-demand | Standard pay-as-you-go on shared capacity, no per-project quota to request | Per-deployment token quotas; Global Standard has the highest | Whatever GPUs you hold |
| Guaranteed capacity | Provisioned Throughput in model units, no commitment, 1 month or 6 months | Provisioned Throughput | Provisioned throughput units (PTUs) | Reserved GPU capacity |
| Where data is processed | One Region, or cross-Region inference within a geography or worldwide | Regional or global endpoints | Global, data zone (US, EU, Asia Pacific) or one Azure geography | Wherever your GPUs are |
| Cheap asynchronous mode | Batch inference, 50% cheaper for supported models | Batch prediction | Batch deployments, 50% cheaper, 24-hour target | Fill idle GPUs |
| You pay per | Token, or model-unit hour when provisioned | Token, or provisioned capacity | Token, or PTU hour | GPU hour, used or not |