AWS ran its European cloud with the global backbone switched off. What Azure and Google contract for

What AWS switched off on 24 October

On Saturday, 24 October 2026, the AWS European Sovereign Cloud ran for several hours with no connection to the AWS Global Network backbone, the private network that normally carries AWS operational data between its locations. AWS says traffic rerouted over the public internet through European internet service providers, that service availability inside the region was not affected, and that customers could see brief disruption at the start and end of the exercise as other networks converged on the new route.

The operational team, described as entirely EU residents working inside the EU, executed the exercise using only the region's own hardware and software. The team also disabled the second global system the region uses in normal operations, a dedicated link that manages limited, controlled transfers of AWS operational data out of the EU. AWS committed to sharing the results with regulators and customers rather than publishing them at the time. Private connectivity through AWS Direct Connect was excluded from the exercise, so a customer whose traffic enters the region over a dedicated circuit had nothing to test.

The claim is narrower than sovereignty

What the exercise tests is operational independence, not legal immunity. The Brandenburg region is described as physically and logically separate from other AWS Regions, with a local in-EU copy of the source code, all customer content and customer-created metadata kept in the EU, and no critical dependencies on non-EU infrastructure. AWS states that the two global systems it uses in standard operations are neither operation-critical nor a threat to its sovereignty assurance, and that the region can run without either at any time.

The framing matters for procurement. AWS designed the test to be consistent with the objectives of the European Commission's EU Cloud Sovereignty Framework and the criteria of the C3A framework from Germany's Federal Office for Information Security, and it maps evidence through a document it calls the Sovereign Reference Framework. A regulator asking whether a provider can operate autonomously wants a control-by-control mapping, so the artefact to ask for is that mapping plus the test results, attached as a schedule rather than quoted in a slide. Ask also which controls the exercise did not touch, because a test that switches off two named systems says nothing about the other dependencies a service team relies on.

Microsoft sells control inside Azure

Microsoft's position, published on 20 October 2026, is a continuum rather than a second cloud: Sovereign Public Cloud for workloads that need more control but can stay in Azure, Sovereign Private Cloud on Azure Local for workloads that must run in a customer-operated environment, and a partner ecosystem for locally operated services. The new controls are mostly about who can touch the system and what record is left.

A Sovereign Control Panel, entering public preview at Ignite, reports sovereignty posture, gaps and workload dependencies in the Azure portal. Data Guardian, rolling out in the EU and EFTA, adds European-resident oversight by a partner for qualifying Microsoft engineer remote access, and the Production Touch Ledger writes tamper-evident records of that access. External key management with Azure Managed HSM is generally available, so the root encryption key can sit in hardware the customer operates on-premises or through a third party. The transition from connected to disconnected operations on Azure Local enters public preview, and GitHub Enterprise Local becomes generally available at Ignite as an appliance on Azure Local.

Google puts the model inside the fence

Google's update the same day went the other way: more capability inside a deployment with no connection to a public cloud. On Google Distributed Cloud air-gapped, Gemini 3.5 Flash is generally available, Gemma and Mistral are supported as open models, Gemini enterprise search is in preview with retrieval kept inside the perimeter, and NVIDIA HGX B300 systems are generally available. Google also raised VM networking to 32 Gbps per VM from 25 Gbps and added hardware-offloaded IPsec at up to 100 Gbps.

The buyer for this is defence, intelligence or critical infrastructure, where the data cannot leave the building at all. The cost is operational: the customer runs the platform, patches it and owns capacity planning, with Google's stack on its own floor and a marketplace of ISVs including Teradata, Solo.io and Mattermost running inside the same perimeter.

What each contract can actually carry

The three offers answer three different questions, and the schedule of evidence follows the question rather than the brand.

The decision rule

If a regulator or a board asks whether a named operator outside the EU can reach production systems, the AWS European Sovereign Cloud is the only one of the three that has published an isolation exercise with the backbone and the transfer system switched off, and its framework maps the result to the two frameworks European assessors are using. If the estate already runs on Azure and the requirement is per-workload control over access, keys and records, Microsoft's approach avoids moving the workload and puts the evidence in the portal. If the workload cannot be connected to anything, Google's air-gapped stack is the fit, and the budget should carry the operations team that comes with it.

In all three cases, name the evidence in the contract. A sovereignty claim that is not listed in the schedule of evidence is a claim the buyer cannot audit, and the difference between the three vendors is less about what they can do than about what they will put in writing.