What changed
The preview ran from Build in June 2026 until now. Documentation as recently as a month ago still described the service as public preview, with no SLA and not recommended for production. That caveat is the main practical difference at GA. A service that was fine for experiments can now be considered for production planning.
Microsoft's pitch is simple: you pick the model, and Microsoft provisions the GPUs, maintains the containers and runs the serving runtime. The GA post lists four sets of changes since Build:
- More serving capacity and better prompt reuse.
- More GPU and deployment options, including US and EU Data Zones.
- Clearer visibility into how deployments are running.
- A hardened software supply chain and a larger model catalog, now 150+ models.
Where it sits in Foundry
Microsoft Learn describes three Foundry deployment types.
| Type | Serves | Billing |
|---|---|---|
| Standard pay-per-token | Foundry Models sold by Azure | Per input and output token |
| Provisioned throughput | Foundry Models sold by Azure | Reserved throughput units |
| Managed Compute | Open-source and community models from the catalog | Hourly per accelerator |
All three share one endpoint, the same authentication patterns (Entra ID and key), the same SDKs, the same observability and one bill. You can mix them in a single project. Once a model is deployed, your application code is the same as for any other Foundry model, and only the deployment name changes.
Deploying a model
Prerequisites
- A Foundry account and project.
- Permission to create deployments. Learn lists Cognitive Services Contributor, Foundry Owner or Foundry Account Owner for this.
- Foundry User on the account scope to call the deployment with Entra ID.
- Approved Managed Compute GPU quota for your accelerator family and scope.
Portal path
- Open the Foundry model catalog and filter by "Managed Compute" under deployment options.
- Select a model and choose a deployment template. The template fixes the serving runtime, accelerator family and count, and context length. Depending on the model it can also fix quantization and tuning.
- Choose NVIDIA A100, H100 or H200, or AMD MI300X, depending on what the model offers.
- Start with one instance and check the hourly cost in the wizard.
- Acknowledge the cost and deploy. Learn says provisioning typically takes 10 to 15 minutes, and larger models can take longer.
- When the state reads Succeeded, test in the Playground.
Python path. Microsoft's example deploys DeepSeek V4.1 Flash (1017) using these settings:
- SKU
GlobalManagedComputewith capacity 1. - Model version 5 from the
azure-huggingfaceregistry. - A template for a one-million-token context on four MI300X GPUs.
- Accelerator type
MI300_192GB.
The code raises an error unless provisioning ends in Succeeded. That makes one billable instance on four GPUs. Inference uses the OpenAI-compatible /openai/v1/ route on the account's services.ai.azure.com endpoint. The model field takes the deployment name, not the registry model ID.
The GA post authenticates with an Entra token for the [Microsoft Foundry](https://ai.azure.com/.default) scope. The documentation is not entirely uniform on this point. Its how-to page shows a different token scope in one sample, and it names the required inference role inconsistently: Foundry User in some places and Azure AI User in another. If you get authorization errors, check both the scope and the role assignment on the Foundry account.
Billing and quota
Cost is accelerators per instance × instances × hours × hourly rate. Rates vary by accelerator family and scope. Microsoft directs readers to the Foundry pricing page for current numbers, and this article does not quote any.
- Instance versus GPU. An instance is the full GPU footprint defined by the template. It is not necessarily one GPU. Raising the DeepSeek deployment's capacity from one to five takes it from 4 to 20 GPUs.
- Quota. Managed Compute quota is measured in GPU count, per subscription, accelerator type and scope. It is separate from Azure VM quota, so existing VM quota cannot be reused.
- Scope. The SKU name sets the scope:
GlobalManagedComputefor the broadest capacity, orDataZoneManagedComputeto stay inside a US or EU boundary. The zone is derived from the Foundry account's region, and no separate zone selector exists. - Stopping charges. Deleting a deployment releases the accelerators and stops billing immediately.
Scaling and routing
You scale by adding instances behind the same endpoint. Microsoft reports that it sent the same load to a one-instance and a five-instance DeepSeek deployment. The five-instance deployment processed about five times the tokens and returned fewer HTTP 429 responses. That is a vendor test and not a guarantee. Microsoft advises running your own representative workload and comparing token throughput, status codes and streaming latency in Azure Monitor.
Prompt caching works where the model and runtime support it. Foundry's prompt-aware routing prefers to send matching prefixes to the same instance. Load-aware routing can send requests elsewhere when that instance is busy. You can pass prompt_cache_key as an affinity hint. When the runtime reports it, usage.prompt_tokens_details.cached_tokens shows how many input tokens were reused. Speculative decoding is enabled for leading models that have a good draft model. It is not universal.
Supply chain and security
Microsoft says it takes these steps for catalog models:
- Reviews each model's license.
- Screens for untrusted code.
- Validates SafeTensors weights.
- Stages the weights in Azure, so deployments don't need to reach Hugging Face Hub.
- Builds, scans and signs the serving images, and rolls runtime updates out to existing deployments.
Learn adds that repositories needing third-party Python at load time, the trust_remote_code pattern, are remediated or excluded. It also says private networking with no egress is possible. Canonical says Ubuntu Pro Expanded Security Maintenance covers the Universe packages in the images.
These controls cover the model and runtime supply chain. They do not cover everything:
- Content filtering. Learn says built-in Azure AI Content Safety filters are not part of the Managed Compute data path. If you need request or response filtering, call the Content Safety APIs yourself.
- Model behavior. Curation does not vouch for how a model behaves.
Benchmark and customer claims
Microsoft partnered with Signal65 for a benchmark on AMD MI300X. One DeepSeek V4.1 Flash (1017) instance on four MI300X GPUs supported 128 concurrent agents. Microsoft says it matched Signal65's tuned self-hosted environment on MI300X virtual machines. The result applies to the study's workloads and setup.
AT&T runs Managed Compute alongside a serving platform it built on Azure Kubernetes Service. Microsoft says AT&T used Phi-4, GPT-OSS-120B and Gemma 4 on about a trillion tokens of telecom data for OTel 2.0 training. The post gives no duration, configuration or accuracy results.
What is still preview
- LoRA adapters. Adapters from Foundry fine-tuning can be deployed on a matching base-model version and a LoRA-enabled template. Several adapters share the target's GPU capacity.
- Bring Your Own Weights. This is a limited public preview for custom full-weight SafeTensors models, including ones trained outside Foundry.
Fine-tuned custom models are therefore not yet a generally available path. Commentary during the preview period made the same point. Techjack Solutions argued that the missing bring-your-own-weights path left teams without a route for domain-specific fine-tunes. The preview status of BYOW means that gap is only partly closed.
Analysis
The strongest case for Managed Compute is operational. Many teams want open models for cost, control or data reasons but don't want to run GPU nodes, tune vLLM or patch images. Hourly billing per accelerator also means dedicated capacity costs money whether or not traffic arrives. A pay-per-token endpoint is cheaper for bursty or low-volume use. Dedicated GPUs make more sense for steady load, large contexts or agent fleets.
There is also a lock-in argument. Techjack Solutions says Microsoft gains influence over which models are available, in which runtime configuration, on which hardware and on what refresh cycle. That is a fair trade-off to weigh. Curation is what delivers the security promises, but it also limits choice. Because the endpoint and SDK are OpenAI-compatible, the app-level code is relatively portable.
Practical checklist
- Request GPU quota for the right accelerator family and scope before you try to deploy.
- Start with capacity 1 and measure with your real prompts.
- Use the Data Zone SKU if residency matters, and confirm the account's region.
- Add your own content filtering.
- Delete test deployments, since billing runs hourly.
- Check the template and model version are compatible if a deployment fails. The Learn troubleshooting section names this as a common cause, along with quota and capacity problems.
Which accelerators and regions are available varies by model and template. Check the wizard for the exact combinations before you plan.
References
- Serve open models in production with Foundry Managed Compute Microsoft Foundry Blog · 2026-11-20T15:00:44+00:00
- Deploy open-source models with managed compute in Microsoft Foundry - Microsoft Foundry | Microsoft Learn learn.microsoft.com
- Managed compute in Microsoft Foundry - Microsoft Foundry | Microsoft Learn learn.microsoft.com