Azure OpenAI, now sold as Foundry Models sold by Azure, runs OpenAI and other models in Microsoft's cloud. Pages here cover replacing it with open-weight models self-hosted on vLLM behind your own gateway.
Comparing Azure OpenAI with open-weight models served on vLLM in your own Azure subscription, datacenter or another cloud: data processing location, model choice and version control, throughput guarantees against PTUs, operating effort, and how applications switch with the OpenAI-compatible API.
Moving workloads from Azure OpenAI deployments to self-hosted open-weight models: inventorying deployments and API versions, standing up vLLM behind a LiteLLM gateway, evaluating replacement models on your own prompts, and migrating provisioned workloads before their contract or retirement date.
Cost of ownership by component: pay-per-token Standard deployments, reserved PTUs and 50% Batch pricing against GPU virtual machines or servers, the gateway and monitoring stack, implementation and on-call engineers, with the utilization that makes reserved capacity or self-hosting pay off.
Azure OpenAI GA models retire 18 months after launch, retirement dates cannot be extended, and provisioned deployments must be migrated by hand. This page covers what that cadence costs in revalidation work and how pinning open-weight versions on your own infrastructure changes it.
Global deployments may process prompts in any Azure region and Data Zone EU keeps processing inside the EU Data Boundary, which Microsoft can extend. This page compares those options and Standard regional deployments with models served in an EU location you choose and control.
Azure OpenAI may store flagged prompts and completions for human review unless the customer is approved for modified abuse monitoring. This page explains what is stored, where and for whom, what the approval changes, and when serving models yourself is the cleaner answer for regulated content.
Azure Government clouds lack Global Standard deployments, offer fewer models and usually one model version at a time. This page covers running open-weight models on vLLM in classified, disconnected or sovereign environments instead, including weight transfer, accreditation evidence and upgrades.
Tell us what you're working on, whether it's a deployment, an audit, a security test or a cyber range. You'll speak with an engineer who can help you scope it.