What was announced
Microsoft's Windows Server team, in a post by Priya Satheesh dated November 17, 2026, says the two companies are working to enable MI350P support. The goal is to give customers another option for accelerated AI inference in their own datacenters. Microsoft says AMD has developed Windows Server driver support, and calls it a key step toward running accelerated inference alongside existing enterprise applications.
Read that as the state of the enablement effort. The post gives no driver version, download location, supported server models, firmware requirements or installation procedure. A driver existing is not the same as a production-ready, publicly downloadable one.
The hardware, as vendors describe it
Microsoft's post lists these MI350P specifications:
- It is a PCIe accelerator built on AMD's CDNA 4 architecture.
- It has 128 compute units.
- It has 144 GB of HBM3E memory.
- Memory bandwidth is up to 4.0 TB/s.
- The PCIe form factor is rated at 600 W.
Microsoft says one card can host several small to mid-size models for concurrent users. It also says one card can host a single model of up to 260 billion parameters, with up to 4.6 PFLOP of inference throughput at MXFP4.
Keep the qualifiers in view. The throughput figure is specific to MXFP4, a low-precision format. Microsoft gives no benchmark methodology, workload, batch size, latency or software configuration. It is a vendor peak number, not a promise for your workload. Likewise, "up to 260 billion parameters" does not mean every model of that size will fit or run at a useful speed.
Some extra detail comes from AMD's ROCm documentation. It describes the MI350P as a PCIe card at 600 W that is configurable to 450 W, with support for up to four compute partitions and a single memory partition across its 144GB. A Gigabyte product page lists it as a dual-slot, passively cooled card with a 16-pin power connector. For anyone planning racks, that raises real power and airflow questions. Neither Microsoft nor AMD's announcement addresses those questions for Windows Server deployments.
AIMs and GPU partitioning
AMD's own post says AMD Inference Microservice (AIMs) and GPU partitioning are supported at launch, so on-premises customers can simplify model deployment and allocate GPU capacity across workloads. In practice:
- AIMs is meant to simplify deploying supported models. The announcement doesn't list which ones.
- GPU partitioning lets customers divide an MI350P into supported partitions and allocate resources across workloads.
Partitioning matters because many enterprise inference jobs are small. An embedding model for search doesn't need 144 GB, and dedicating a whole card to it would be wasteful. Splitting the card lets one accelerator serve several modest services. AMD's partition sizes and sharing limits aren't spelled out in the announcement, so don't size a deployment from it.
When can you actually use it?
This is the part to read twice. Microsoft's Windows Server 2028 announcement says the MI350P will enter preview in the first half of 2027 for supported Windows Server versions, including Server 2028. The Server 2028 post phrases it as "in-market supported Windows Servers" plus 2028. That is a preview statement, not general availability, and it gives no specific date.
Windows Server 2028 is also only in public preview. Microsoft says its hybrid intelligence capabilities are available to a limited number of customers through an early access program, and that hardware requirements apply. The core security, performance and management features are in preview through the Windows Server Insider Program. In short, AMD's OS targets (2025 and 2028) are a compatibility direction. Microsoft's schedule is a limited preview in 2027.
The software layer: promises, not guarantees
Microsoft frames the GPU as one piece of a stack with three layers: hardware, platform and runtime. At the runtime layer it says it is working to support open-weight and custom models across technologies including Foundry Local, Windows ML, vLLM, and Llama.cpp. The wording is "working to support." Don't assume every runtime works with this GPU on every Windows Server version today.
On governance, Microsoft points to Agent 365 for agent identity and governance. It also points to Microsoft Execution Containers (MXC) as a security boundary for agent code and tools. These belong to the wider Windows platform story and are not MI350P features. They are a sensible response, though. Agentic workflows may make repeated model requests and invoke tools, so controlling what agents can do is as important as raw throughput.
Two example workloads
Microsoft offers two application patterns and says availability depends on Windows Server version, hardware, drivers and inference software.
- Coding assistance. A compatible coding tool connects to a model hosted on Windows Server. Developers get help using authorized internal code and documentation without needing an accelerator at every desk. Microsoft calls this an enterprise "token factory." It adds that teams can still use Foundry models on Azure when they need more capability and their data policies allow.
- SQL-connected RAG. An embedding model retrieves records or passages by meaning rather than keyword. The application then passes that content to a language model. Customers choose where embedding and response generation run. Coverage of AMD's post links this to SQL Server 2025's native vector capabilities.
Neither example means a finished coding or RAG product ships with the card. You still have to build and operate the pieces.
Ignite demo
AMD and Microsoft's Windows Server team planned to show a proof of concept at Ignite, running an open-weight model on the MI350P. The post doesn't name the model or give any figures from the demo. A proof of concept is not a customer deployment.
What admins should do now
- Start with the workload, not the peak number. Microsoft's own advice is to weigh model quality needs, expected concurrency, response-time targets and data-handling requirements. Those determine the capacity and operating cost you should evaluate.
- Check preview eligibility. The Windows Server Technology Adoption Program and the Insider Program are the routes Microsoft names for early access to Server 2028 features.
- Ask your server vendor about the physical side. That means slot space, power delivery and cooling for a 600 W card, and whether your platforms are validated.
- Wait for the missing documentation. That includes driver and firmware versions, supported models, partition limits and runtime support. Don't commit budget until it appears.
Analysis
This is a sensible strategic move, with the usual vendor-announcement caveats. Many organizations keep sensitive data on Windows Server and want to avoid sending it to the cloud. Adding another accelerator vendor gives them more choice. The sources are mostly vendor-authored, so the performance and "cost and power efficient" claims remain unverified by independent testing. The main unknowns are preview scope, pricing and how mature the Windows driver and runtime stack is at launch. Watch for the first-half-2027 preview before treating this as something you can plan a rollout around.
Correction (November 19, 2026): The Microsoft Windows Server 2028 announcement post cited here was published November 17, 2026, not November 19, 2026. The date has been corrected.
References
- Expanding AI silicon capabilities on Windows Server with AMD Windows Server News and Best Practices · 2026-11-19T19:44:15+00:00
- AMD Instinct MI350 Series microarchitecture — AMD ROCm 10.0.0 rocm.docs.amd.com
- More AI. More Control. AMD Instinct MI350P Comes to Windows Server. amd.com