UK’s trusted IT infrastructure partner since 2003
Servnet
FinanceToolsConfiguratorGet in Touch
AI / GPU

Windows Server 2028 AI Inference: UK Guide 2026

London · Servnet News Desk · IT infrastructure analysis5 min read
Share

Microsoft used Ignite 2026 to put Windows Server 2028 into public preview, pitching it as the foundation for a hybrid AI deployment model that keeps inference on infrastructure enterprises already control. For UK buyers juggling data sovereignty AI obligations and rising cloud bills, the pitch lands at the right moment — but the preview label matters.

Windows Server extended support end dates
2040 year1530 year1020 year510 year0 year2029 yearServer 20192031 yearServer 20222034 yearServer 2025Support ends
View the data behind this chart
Windows Server extended support end dates
Server 2019Server 2022Server 2025
Support ends2029 year2031 year2034 year

What Microsoft actually announced

Windows Server 2028 was unveiled at Microsoft Ignite on 17 November 2026. Microsoft describes Windows Server 2028 as its largest platform update since Windows Server 2025. The preview is distributed through the Windows Server Insider Program, while eligible customers can apply for the Technology Adoption Program (TAP) or early access to selected hybrid-intelligence capabilities; hardware requirements apply.

The headline idea is what Microsoft calls "enterprise hybrid intelligence": new AI accelerator support, storage tooling for faster data transfer, a hardened security baseline out of the box, and simplified admin dashboards, which can let enterprises run selected inference workloads close to their data instead of sending them to a public-cloud endpoint.

Hybrid AI deployment: the promise behind the preview

Windows Server product lead Neil Hinnant has written about avoiding, in his words, a "separate AI stack for every location" — instead, organisations deploy and configure inference environments centrally and route model requests to the right hardware, wherever it sits (Microsoft, Windows Server blog). Microsoft says it is beginning a longer-term effort to extend Microsoft Foundry to local "token factories", with intended support for Foundry Local, Windows ML, vLLM and Llama.cpp, covering both open-weight and custom models.

On the hardware side, Microsoft says AMD's Instinct MI350P will enter preview in H1 2027 for supported in-market Windows Server systems, including Server 2028. A broader hardware compatibility list, performance benchmarks and minimum system requirements have not yet been published, so buyers evaluating an AI GPU sizing calculator approach to capacity planning will need to wait for firmer specifications before committing budget.

Security and storage: the parts that are furthest along

Among the capabilities currently described in detail are security and storage changes, though Microsoft also lists Intune and other management features as available for testing, with availability and maturity varying across capabilities. The preview's Server Core configuration enables hundreds of security settings by default on first boot; administrators can relax selected settings when compatibility requires it. Microsoft's own framing is that "the secure configuration is the starting state, not a project you run after setup." The release also introduces containment and Agent 365 governance, extends Trusted Launch to VMs, and begins confidential-VM support on supported hardware — alongside automatic certificate management and hybrid post-quantum key exchange using ML-KEM.

Microsoft says native NVMe is enabled by default; the roadmap also includes NVMe over TCP, NVMe over RDMA and planned boot-from-ReFS support. These changes matter for anyone planning to keep model files and vector stores on local infrastructure, and are worth reviewing alongside existing server configuration standards before the preview moves towards general availability.

The hardware bill UK buyers should expect

Analysts are blunt about what local inference actually costs. For demanding local-inference deployments, Info-Tech Research Group associate vice-president John Annand says customers may need accelerators such as Nvidia or AMD GPUs — with strong interest reported in AMD's ROCm open-source GPU stack — plus considerably more RAM, and faster networking than 10GbE; the actual requirements depend on workload and architecture. Annand says existing 10GbE links are likely to bottleneck some AI deployments, with upgrades to 50 or 100GbE or faster among the options being considered. As Annand puts it, "to the surprise of no one, if you want to use Server 2028 for AI, you'll need to invest in a new kit."

That reality check sits well alongside a AI servers data study view of the market: the operating system upgrade is necessary but not sufficient, and procurement teams should model accelerator, memory and networking costs separately from the Windows Server licence itself. Teams weighing total cost of ownership against continued cloud spend may find an IT finance calculator useful for comparing the two paths before signing off capital spend.

Why this matters for data sovereignty and edge computing strategy

Running inference on enterprise-controlled infrastructure may support data-residency objectives and some edge use cases, but UK compliance and latency benefits must be assessed per workload. Annand frames the broader market signal plainly, in his own interpretation: rising public-cloud AI costs amount to a "tacit acknowledgement that some Microsoft customers want to return to on-prem from the blue skies of their Azure environments."

It's worth being precise about what's proven and what isn't. The announcement does not quantify latency reductions, bandwidth savings, or UK-specific compliance outcomes — those depend on model size, accelerator capacity, network design and application architecture, and will need validating case by case rather than assumed from the marketing.

Where to run AI inference: three live options
Cloud-only AIServer 2028previewStay on 2025Data residencyData leaves UK premisesInference stays on-premNo new AI capabilityHardware spendNone, pay-as-you-goNew GPU, RAM, 50/100GbENo change neededPlatform maturityMature, GA todayPublic preview,evolvingMature, GA,long supportedSupport runwayOngoing subscriptionPreview; GA date unsetSupported to 2034
View the data behind this chart
Where to run AI inference: three live options
Cloud-only AIServer 2028 previewStay on 2025
Data residencyData leaves UK premisesInference stays on-premNo new AI capability
Hardware spendNone, pay-as-you-goNew GPU, RAM, 50/100GbENo change needed
Platform maturityMature, GA todayPublic preview, evolvingMature, GA, long supported
Support runwayOngoing subscriptionPreview; GA date unsetSupported to 2034

Compatibility risk and the decision buyers face today

Because the preview introduces hardened security and stricter memory-related defaults, some workloads that run on Windows Server 2025 may require configuration changes — Microsoft flags older protocols, legacy authentication, and some management tools and scripts as likely friction points. Testing applications and operational tooling against the new baseline before any deployment is essential, and Microsoft is explicit that parts of the prerelease product may still be "substantially modified" before general availability.

There is no immediate pressure to move. Windows Server 2025 remains under extended support until 2034 (year shown in Microsoft support information), Windows Server 2022 until 14 October 2031, and Windows Server 2019 until 9 January 2029 — giving most UK enterprises a long runway to test Server 2028 in preview, validate accelerators and drivers, and plan hardware refreshes on their own timeline rather than Microsoft's. For organisations already rationalising their virtualisation estate, this is also a sensible moment to revisit platform choices more broadly, including options explored in our VMware alternatives coverage, before new AI workloads add another layer of dependency. Teams planning accelerator procurement should also keep an eye on the wider GPU accelerators market as supply and specifications firm up. Enterprises wanting a structured view of their own readiness can talk to a Servnet engineer before joining the preview programme.

Share
Key takeaways
  • ✓Windows Server 2028 is in public preview only, announced at Ignite on 17 November 2026 — treat it as a roadmap, not a production platform, until general availability is confirmed.
  • ✓Local AI inference will usually require new accelerators (AMD Instinct MI350P previews in H1 2027), substantially more RAM, and a networking upgrade from 10GbE to 50/100GbE or faster.
  • ✓Hardened-by-default security and native NVMe are the most mature elements today; expect some existing apps, legacy protocols and scripts to break without configuration changes.
  • ✓No upgrade urgency exists: Windows Server 2025 (to 2034), 2022 (to 2031) and 2019 (to 2029) extended support dates give UK buyers time to test before committing budget.
Frequently asked

FAQs — Windows Server 2028 AI Inference

Is Windows Server 2028 available to buy now?

No. It is in public preview via the Windows Insider programme, with earlier access through the Windows Server Technology Adoption Program. Microsoft has not published a general-availability date, and parts of the prerelease may still change.

What hardware do we need for on-premises AI inference?

Most existing servers will need an added accelerator such as AMD's Instinct MI350P (preview in H1 2027), significantly more RAM, and a network upgrade beyond 10GbE to avoid bottlenecks. A AI GPU sizing calculator approach helps scope this before procurement.

Does this actually improve UK data sovereignty?

Running inference on enterprise-controlled infrastructure keeps data from being sent to public-cloud endpoints, which supports data sovereignty aims, but Microsoft's announcement does not quantify UK-specific compliance benefits — these need validating per workload.

Should we upgrade from Windows Server 2025 now?

There is no forced timeline. Windows Server 2025 is supported until 2034, so most organisations can evaluate the Server 2028 preview, test compatibility, and plan accelerator and networking investment before migrating.

Related

Continue reading

More in AI / GPU →

Turning this into a buying decision?

One conversation with an engineer who's specced this before. No sales script.

Talk to Servnet →

Talk to a UK specialist

Get expert advice or a no-obligation quote — servers, storage, networking, maintenance, finance and cloud. We reply the same working day.

or call 0800 987 4111