Microsoft has made Foundry Managed Compute generally available, giving companies a managed way to serve open models on dedicated GPUs. The catalog includes more than 150 models, and deployment templates specify the context length, runtime and hardware configuration.

Customers can choose Nvidia A100, H100 or H200 GPUs, or AMD MI300X accelerators, depending on the model. Microsoft provisions the hardware, maintains serving containers and updates runtimes. It also reviews model licenses, checks weights and untrusted code, and stages signed serving images in Azure rather than loading them directly from an external hub.

The service scales by adding complete model instances behind one endpoint. Each instance consumes the GPU count defined by its template, so a sample DeepSeek deployment grows from four to 20 GPUs when capacity rises from one to five. Billing is hourly per accelerator, making representative load tests important before scaling.

Prompt-aware routing can keep requests with matching prefixes on the same instance to improve cache reuse, while load-aware routing can move traffic when an instance is busy. Microsoft also opened previews for deploying LoRA adapters and customer-supplied SafeTensors weights. Teams need a Foundry project, deployment permissions and enough regional GPU quota before they can use the service.