Microsoft Launches Windows Server 2028 Preview for On-Premises AI

Microsoft opened the public preview of Windows Server 2028 this week, designed in anticipation of increased enterprise use of on-premises AI. This version introduces protections and administration tools, but several model execution capabilities remain announced or in development.
Protections for Agents and Virtual Machines
For Microsoft, containers and virtual machines do not serve the same function in AI protection and resilience. Introduced last month, Microsoft Execution Containers (MXC) confine agent actions within a sandboxed environment that restricts their interactions with external resources. Their role is to prevent these agents from encroaching on company resources.
Confidential virtual machines rely on a trusted hardware environment designed to prevent tampering and protect sensitive processing from intrusions. Neil Hinnant, director of product management for Windows Server, describes this as an early-stage project concerning compatible CVM hardware. Microsoft plans to work with processor manufacturers, including AMD and Intel, to ensure isolation directly at the silicon level.
The integration of ML-KEM, or Module-Lattice-Based Key-Encapsulation Mechanism, also helps secure communications. On a public network, this major post-quantum cryptography standard from the National Institute of Standards and Technology (NIST) allows for the generation of shared encryption keys. According to NIST, attackers could already collect data with the intent to decrypt it later, once current protections are compromised, although quantum computers may still be a few years away.
For Server Core installations, initial settings now impose more restrictions on network communications, authentication, and remote connections.
Server Boot, Recovery, and Management
The availability of infrastructure was recently tested at Microsoft: a prolonged outage disrupted Microsoft 365, including Outlook, Teams, Copilot, SharePoint, and OneDrive in their cloud versions, and recovery was challenging. Don Gentile, an analyst at HyperFrame Research, described the previous week as difficult for the company amid cloud challenges. During his keynote at Ignite, Satya Nadella, CEO of Microsoft, acknowledged the incident and reiterated the critical nature of digital infrastructure. For the analyst, this episode also serves as a reminder to companies operating their own systems that they must manage their resilience.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Absent from this speech, the preview includes several features that Neil Hinnant presents as means to maintain system continuity. The ReFS file system now extends to the boot volume: it checks checksums during reads and writes and periodically scans for corruption in the background. File systems are essential for the reliability of the operating system as well as for the protection and availability of data. However, Don Gentile notes that protecting the system volume with ReFS does not replace file system integrity requirements or application recovery mechanisms.
A quick recovery feature also aims to bring a server back online when a critical error prevents it from booting. Administrators can configure retry intervals and delays between attempts.
Management of the infrastructure also involves Intune, which has primarily been used for Windows computers and mobile devices until now. For the first time, this tool supports Windows Server: it can automate patches, verify compliance with configurations, and manage security policies. Microsoft specifies that this server support remains in preview.
Accelerators and Planned Integration with Foundry
At Ignite, Satya Nadella urged leaders to maintain control over their intellectual property and advocated for an architectural separation between the provision of intelligence and authority over it. Conducting part of the AI modeling in-house can help preserve this intellectual property and reduce costs. Microsoft anticipates an increase in AI processing performed within enterprise infrastructures.
Windows Server offers native support for accelerators designed for heavy inference workloads. However, for the AMD Instinct MI350P, Microsoft announces a preview in the first half of 2027. NVMe storage is now enabled by default to speed up data feeding to these accelerators; NVMe over Fabrics allows for exchanges with fast SSDs in place of SCSI/iSCSI or Fibre Channel.
Neil Hinnant announces the continuation of collaboration with Foundry AI, while Don Gentile states that no integration between Foundry and Windows Server currently exists. Microsoft plans to offer out-of-the-box execution of open-weight or custom models with Foundry Local, Windows ML, vLLM, and the llama.cpp inference engine. According to the analyst, AI services based on Foundry will have security and resilience mechanisms, but clients maintaining their processing on-premises will remain responsible for backups and security, necessitating reliance on third-party services for these tasks.
Beyond public access, the Windows Server Technology Adoption Program is accepting applications to test upcoming features and engage directly with Microsoft engineers. Hybrid AI capabilities, however, require participation in a separate access program.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.