What Microsoft shipped at Ignite
On 17 November, Microsoft's Azure Arc team published its Ignite 2026 sovereign private cloud update, and the headline item for anyone running private AI was Foundry Local reaching general availability on Azure Local. Microsoft had flagged the GA date in October. Foundry Local is now a supported production service you can run on hardware in your own server room or colocation rack.
According to Microsoft's product documentation, Foundry Local installs as an Azure Arc extension on a Kubernetes cluster running on Azure Local. Teams can use it to deploy models from Microsoft's catalog or bring their own. Applications reach those models through OpenAI-compatible REST endpoints. The service runs on CPU nodes or GPU nodes, using the ONNX-GenAI engine or, on GPUs only, vLLM. The model catalog and local retrieval features shipped as generally available. The agent service and content safety guardrails are still in preview, and so is moving a connected deployment to a disconnected one. Microsoft also announced support for Nvidia HGX B300 systems from Dell, HPE and Lenovo, but the Ignite post gives no availability date for those.
Why the release matters for data sovereignty
Most NZ organisations that want generative AI without sending prompts offshore have had two choices so far. One is a public cloud service hosted in a region they trust. The other is a self-built stack of open-source inference software that their own team has to patch and support. Foundry Local adds a third option for Microsoft shops: a vendor-supported inference layer that sits on hardware you control and is managed through the Azure tooling your infrastructure team probably already knows.
The disconnected mode is the most interesting part for sovereignty. Microsoft's disconnected operations guide says a disconnected cluster sends no telemetry to Microsoft. Authentication goes through your local Active Directory instead of public Entra ID endpoints, and model evaluation data stays on the cluster. Models arrive as imported "expansion packs" loaded into a local container registry instead of being downloaded from the internet. For a healthcare provider, a council or a firm handling defence or iwi data, a vendor-supported inference service that works with no outside connection at all is a meaningful step.
The NZ detail most coverage misses
The connected version is where NZ buyers need to read the fine print. The Foundry Local overview page lists the Azure regions where the Arc extension is available. Australia East and Australia Southeast are on that list, and New Zealand North isn't. Microsoft opened its Auckland-based NZ North region in December 2024, so a connected Foundry Local deployment managed from Azure will most likely register its management resources against an Australian region.
How much that matters depends on what crosses the link. Inference runs on your own hardware, and the documentation describes processing staying on-premises. Management metadata, extension lifecycle operations and catalog sync go through Azure Arc, though. If you deploy models through the Foundry portal, the docs list extra requirements: extension version 2610 or later, an externally exposed control-plane endpoint and a publicly trusted certificate. An NZ organisation whose policy says "no data or metadata leaves New Zealand" can't assume a connected deployment meets that rule. The disconnected mode probably can, at the cost of manual model imports and running your own certificate tooling.
Costs and maturity
Microsoft hasn't published a separate price for Foundry Local. The underlying platform is billed per physical core: the Azure Local pricing page says fees apply to all physical cores daily while the system is registered, after a 60-day trial. Windows guest licences cost extra. For disconnected operations the page lists no price and tells buyers to contact their account representative. Anyone building a business case today is working from a quote, since there's no rate card to point at.
The release also has rough edges. WindowsForum's review of the Ignite announcements found GitHub Enterprise Local described as GA in the Ignite blog while the matching Microsoft Learn page still said preview. It also found conflicting Azure Local version requirements for confidential VMs. The disconnected guide warns that the default Kubernetes worker size is "usually too small" for caching and serving models, so sizing needs real attention before hardware is ordered.
What this means for NZ businesses
Organisations already running Azure Local, or replacing ageing Hyper-V or VMware clusters, now have a supported way to add private AI inference to infrastructure they're buying anyway. Firms with no on-premises footprint face a bigger decision. Buying validated hardware, Kubernetes skills and per-core fees just to keep prompts on-site is a serious commitment, and an open-source stack on a single GPU server may suit a smaller workload better. We cover that trade-off on our local AI server page.
Guardrails and the agent service are both still in preview. Anyone planning to put Foundry Local in front of staff or customers this year will be relying on preview features for safety filtering, or building their own.
Practical next steps
- Write down which data classes your AI use cases would touch, and decide whether your policy covers management metadata as well as content. That decision tells you whether connected mode is acceptable or whether you need disconnected mode.
- Ask your Microsoft partner or account manager in writing which Azure region a connected NZ deployment registers against, and what data flows to that region.
- Get a quote for disconnected operations before you design around them, since Microsoft doesn't publish the price.
- Size GPU nodes for your target models using Microsoft's capacity guidance, and don't accept default worker sizes.
- Treat guardrails and agents as preview. Plan your own content filtering and logging until they reach GA.
- Compare the three-year cost against a smaller self-hosted stack for the same workload before committing to per-core licensing.
If you're working through these questions as part of a wider AI governance programme, our AI consultancy team can help map use cases to data classes and hosting options.
Sources
- What's new with Sovereign Private Cloud at Microsoft Ignite 2026 (Azure Arc Blog)
- Advancing sovereign AI with Foundry Local and Azure Local (Azure Arc Blog)
- Sovereign Private Cloud: Bringing more choice and control to AI and cloud infrastructure (Azure Arc Blog)
- What is Foundry Local on Azure Local? (Microsoft Learn)
- Foundry Local on Azure Local in Disconnected Environments Overview (Microsoft Learn)
- Azure Local pricing overview (Microsoft Azure)
- Microsoft Ignite 2026: Azure Local Sovereign Private Cloud GA Features, Previews and Caveats (WindowsForum)
- NZ North One Year On: Building the Foundations of New Zealand's AI Future (Microsoft Source Asia)
Want to know what this means for your organisation? Our team in Auckland can show you your top three opportunities in a free 30 minute assessment.
Get my free assessment How we help