Microsoft releases Windows Server 2028 preview, laying groundwork for on-site AI
The platform preview is available now, but AI capabilities have separate access limits. AMD accelerator support is scheduled to enter preview in the first half of 2027.
Loading page…
The platform preview is available now, but AI capabilities have separate access limits. AMD accelerator support is scheduled to enter preview in the first half of 2027.
Listen to this story
Windows Server 2028 is the first step in Microsoft’s longer-term plan to bring its Foundry AI tools to enterprise-controlled servers, connecting local model services with its broader AI stack. The planned deployment and routing tools are intended to support open-weight and custom models without requiring a separate software stack at each site. For customers weighing control of data and inference costs against cloud use, the practical limits are significant: hybrid-AI features are still in development, and access is restricted to a hardware-qualified early-access program.
AMD’s Instinct MI350P is slated for preview in the first half of 2027 on supported Windows Server versions, including 2025 and 2028.
AMD says MI350P launch support will include model-deployment microservices and GPU partitioning for sharing accelerator capacity across workloads.
Server Core now starts with a hardened configuration, but legacy protocols, authentication and some tools may need changes from Server 2025.
Microsoft put Windows Server 2028 into public preview on November 17 at Ignite, beginning a push to run AI alongside the applications and data companies already keep on their own infrastructure. The release combines that AI groundwork with changes to server security, storage and management, rather than delivering a finished local-AI platform.
The question behind the update is where inference—the work of running an AI model—should happen. In Microsoft’s announcement, Windows Server leader Neil Hinnant described customers wanting AI near physical operations, business applications or data they control. He identified cost, response times, intellectual-property protection, data residency and regulation as factors in that choice. The company’s answer is a mix of cloud and local computing, not an all-local replacement.
Microsoft is developing tools to deploy and configure inference environments, manage the hardware that accelerates them, and route model requests to the appropriate environment. Hinnant framed the goal as letting customers add local inference without having to build and operate a separate AI software stack for every location. Those capabilities remain development work, not a claim that every piece is ready.
The software roadmap includes support for open-weight and custom models through Foundry Local, Windows ML, vLLM and Llama.cpp. Microsoft also describes the release as the start of a longer-term effort to extend Microsoft Foundry onto enterprise-controlled servers. Its planned architecture connects local model services to the company’s wider AI tools, rather than treating each server as an isolated installation.
Hardware support is following its own schedule. Microsoft says AMD’s Instinct MI350P accelerator will enter preview in the first half of 2027 for supported Windows Server versions, including Server 2028. AMD’s accompanying announcement names both Server 2025 and Server 2028. AMD says its model-deployment microservices and GPU partitioning will be supported at launch, allowing customers to divide an accelerator’s supported capacity among workloads instead of assigning the whole device to one service.
The platform update also changes the starting point for administrators. Server Core comes with a hardened configuration by default and a new security baseline. Microsoft is adding Trusted Launch protection for virtual-machine boot integrity and beginning support for confidential virtual machines on compatible hardware. Automatic certificate management and post-quantum cryptography are further investment areas, addressing connection security and longer-term changes to encryption.
Running AI locally still brings a hardware bill. Info-Tech associate vice president John Annand told Network World that customers adding local AI to existing servers would need accelerators and substantial additional memory. He also warned that 10-gigabit Ethernet connections could become a bottleneck, requiring faster networking. His assessment separates the appeal of controlling inference from the investment needed to support it.
The stricter defaults create another migration task. Microsoft warns that older protocols, legacy authentication and some management tools or scripts may need configuration changes compared with Server 2025. It advises testing applications and operational tooling before deployment, and relaxing settings only when necessary. Server 2028 remains prerelease software, so features may also change before the final release.
For teams ready to evaluate the platform, Microsoft directs preview testing through the Windows Server Insider Program. Customers seeking early feature access and direct engagement with its engineers can apply to the Windows Server Technology Adoption Program.
Loading discussion...
Join the conversation
Explain when greater control would justify the investment.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.