Microsoft put Windows Server 2028 into public preview at Ignite, and the pitch is narrower and more interesting than a usual server refresh. The company is not selling faster virtual machines. It is selling servers as the place where AI runs next to the data it needs, and servers as a workload your endpoint team can manage with the same tools it already uses for laptops.

Key Takeaways

- Windows Server 2028 is in public preview, reachable through the Windows Insider Program for Windows Server or the Windows Server TAP for earlier access.
- New security hardening includes Trusted Launch virtual machines, post-quantum cryptography and a hardened OS baseline.
- Servers and PCs become one Windows fleet for the first time: Intune management for Windows Server is now in preview, with Azure Arc continuing for cloud-native management.
- Native NVMe storage is on by default, with NVMe over TCP and RDMA transports and the ability to boot from ReFS.
- Microsoft Foundry is extended to Windows Server 2028, so AI inference can run alongside business-critical applications and data.
- Hardware choice expands, with AMD's MI350P accelerator in preview in the first half of 2027 and NVIDIA's DGX Station for Windows supporting Windows Server 2028 at general availability.
- No GA date and no licensing detail were announced.

What the preview actually contains

Three hardening items carry the security story, and Microsoft's own announcement adds two more:

- Trusted Launch virtual machines, guaranteeing boot integrity across the VM fleet.
- Post-quantum cryptography, which matters to any customer with a long-lived data obligation and a compliance team already asking the question.
- A hardened Server Core configuration by default plus a new security baseline, which is the most operational of the items because it moves the default rather than offering an option.
- Automatic certificate management, aimed at the unglamorous day-to-day task of keeping secure connections valid.
- Confidential VM support, which Microsoft describes as beginning a journey rather than arriving, so treat it as a direction and not a capability you can sell today.

Microsoft frames the server as a place where the trusted foundation extends to the datacenter, which is a more accurate description than a feature list. The interesting question for a partner is not what is new, but what a customer must change to get it.

The management change is the real story

For the first time, endpoint teams can manage servers and PCs as one Windows fleet: Intune management for Windows Servers is now in preview, covering configuration, compliance, applications and updates for eligible persistent servers, while infrastructure teams keep using Azure and Azure Arc for cloud-native management.

That sentence is small and load-bearing. In most organisations, servers and endpoints are two different teams, two toolchains and two change processes. Bringing them under one management plane is an organisational change disguised as a product feature, and it is where partner services are easiest to justify: inventory, policy design, update rings and the conversation about who owns what after the merge.

It also means the Autopatch argument now applies to servers. Microsoft reports that Autopatch customers deploy updates more than 40 percent faster than with on-premises management, and that enrollment passed 150 million devices in the past year. Those numbers were about PCs. Extending the same mechanism to servers is a different risk conversation, and a customer will want it framed that way.

AI inference next to the data

Microsoft's own summary is blunt: your servers will become token factories, built for intelligence you can own.

Two capabilities make that concrete. Microsoft Foundry is extended to Windows Server 2028, so inference runs on the same server as the applications and data rather than being shipped to a region. And the hardware list widens:

- AMD MI350P accelerator, available in preview in the first half of 2027 for Windows Server 2025 and later.
- NVIDIA DGX Station for Windows, with the GB300 Grace Blackwell Ultra Desktop Superchip and up to 748 GB of coherent memory, supporting Windows Server 2028 at general availability.

Microsoft describes one path for teams building on open models such as NVIDIA Nemotron: build on a Windows PC, run on Windows Server next to the data, and scale to Azure when more compute is needed.

Underneath the hardware, Microsoft is building three layers: platform capabilities to deploy and configure inference environments, manage accelerated hardware and route model requests to the right environment; runtime support for open-weight and custom models across Foundry Local, Windows ML, vLLM and Llama.cpp; and governance through Microsoft Execution Containers for isolation, Foundry for model catalogs, and Agent 365 for agent identity and governance.

That is a coherent story for regulated customers and for anyone whose data cannot leave a building. It is also, honestly, a licensing and support question that has not been answered yet. Nobody has published what it costs to run inference on Windows Server 2028, and until that lands, the total cost comparison against a cloud region is a spreadsheet with a hole in it.

Storage gets the least attention and may matter most

The performance items are easy to skip and hard to argue with later. Windows Server 2028 adds native NVMe storage enabled by default, NVMe over TCP and RDMA transports, and the ability to boot from ReFS.

Microsoft's framing is that faster hardware only pays off if the rest of the stack keeps up, which is why these land in the base platform rather than as add-ons. For a storage-heavy estate, changing boot file system and storage protocol defaults is a refresh trigger, and it is the kind of change that needs a pilot rather than a press release.

What was not announced

No general availability date. No SKU or licensing structure. No pricing for AI inference on the server. The preview is open through the Windows Insider Program for Windows Server and the Windows Server TAP, and that is the extent of availability information in the announcement.

For a customer planning a refresh in 2027, that means the architectural direction is clear and the commercial terms are not. Anyone who signs a three-year server plan this quarter is doing so without knowing what the AI half costs.

Partner Play

Monday morning, in this order:

  1. Find the servers that cannot reach a cloud region. Regulated data, latency-sensitive workloads, factory floors. Those are the accounts where running inference next to the data is not a preference but a requirement.
  2. Open the fleet management conversation. Servers and PCs in one Intune and Autopatch estate touches two teams with two budgets. Price the inventory, policy design and ownership model, not the licence.
  3. Get customers into the preview deliberately. Windows Insider Program for Windows Server, or the TAP for earlier access. Test the hardening baseline against real workloads before it becomes the default.
  4. Do not quote AI server economics yet. No SKU, no price, no GA date. Model the scenario, label the assumptions, and revisit when the price list lands.
  5. Watch the storage claims. Native NVMe and ReFS boot, if confirmed, are refresh triggers for storage-heavy estates.

Sources

- Windows Experience Blog: Microsoft Ignite 2026, building Windows for the next era of work - Microsoft Azure Blog: Introducing Azure Catalyst agentic modernization and new AI infrastructure innovations - Microsoft Tech Community: Windows Server 2028, a trusted foundation for an evolving enterprise - Microsoft: Windows Server 2028 announcement

Deep Microsoft licensing, incentive, and Partner Center guidance for resellers is published on Cloud Factory.


Ready to grow your Microsoft practice? Explore Cloud Factory's partner platform →

Looking for your local Cloud Factory team? Find your country's contact →