European technology leaders face three connected decisions: where to run AI workloads, how to keep critical services available and how much flexibility to preserve when choosing suppliers.
October’s developments bring those decisions into focus. Scaleway announced new AI infrastructure in Sweden, a Google Cloud outage tested recovery arrangements in London and the UK’s Competition and Markets Authority proposed action concerning Microsoft’s business software ecosystem.
Our assessment is that these developments should inform the same investment conversation. Infrastructure choices need to be evaluated against the organisation’s ability to deploy applications, recover services and change direction at an acceptable cost.
1. Evaluate new AI infrastructure against production workloads
On 13 October 2026, Scaleway announced an AMD-powered AI supercomputer in Sweden, marking its first supercomputer deployment outside France. The planned system features AMD Instinct MI355X GPUs and AMD’s ROCm software ecosystem.
The announcement broadens the infrastructure options European organisations can consider. However, an infrastructure announcement is the start of an evaluation. Buyers still need to confirm capacity availability, commercial terms and suitability for their applications.
Software compatibility deserves early attention
Moving an AI workload between hardware platforms can involve engineering work. AMD explains that applications using supported frameworks may be able to start with a ROCm-enabled framework version, while code that depends on CUDA-specific APIs or libraries may require changes. Supported configurations also depend on the software release and system environment. AMD’s ROCm guidance provides a starting point for that assessment.
For buyers, the practical implication is to test the complete application before making a substantial capacity commitment.
A representative pilot should assess:
- Application performance: Does the workload meet its requirements for response time, throughput and output quality?
- Engineering effort: Which dependencies need changes and who will maintain them?
- Data movement: How will information reach the platform and connect with existing systems?
- Operational support: Who owns deployment, monitoring and incident resolution?
Compare the cost of delivering the application
We recommend comparing platforms using a business-relevant measure, such as cost per completed training run or per thousand successful requests at a defined response time.
That comparison should include the team’s migration and operating effort. A lower infrastructure price may offer limited value if deployment requires extensive rework. Equally, an alternative platform may justify that initial effort if it performs well for an important, recurring workload.
A useful next step is to select one realistic application and document the evidence required to approve production deployment. This turns additional infrastructure choice into a decision the organisation can defend.
2. Test cloud recovery across the whole service
The Google Cloud incident in London on 20 October 2026 demonstrates how infrastructure failures can extend beyond their initial location.
According to Google’s incident report, an external power disruption affected the facility hosting the europe-west2-b availability zone. A control-system misconfiguration prevented emergency generators from starting automatically. Redirected workloads then contributed to capacity pressure and service errors elsewhere in the region. Google reported that the primary impact was mitigated after six hours and eight minutes.
For technology leaders, the incident offers a concrete scenario against which to examine their own recovery assumptions.
Define what recovery must achieve
Google’s disaster recovery planning guidance distinguishes two essential targets:
- Recovery time objective: The maximum acceptable period an application can remain unavailable
- Recovery point objective: The maximum acceptable period of data loss, measured backwards from the disruption
These targets should reflect business needs. A customer transaction service may require a different recovery approach from an internal reporting tool. More demanding targets generally bring additional cost and operational complexity.
Exercise the customer journey
We recommend testing a complete business process rather than checking only whether replacement infrastructure starts.
For example, a recovery exercise for an online service should establish whether a user can sign in, retrieve information and complete a transaction. The exercise should also verify that monitoring and support teams can see what is happening.
Useful scenarios include an unavailable zone, a failed database connection or loss of access to deployment tools. Teams should record the actual restoration time, check data consistency and identify any steps that required undocumented knowledge.
The London incident also makes capacity a valuable test condition. An alternative environment should be assessed under the additional demand it is expected to absorb.
Before the next exercise, assign responsibility for declaring an incident, authorising recovery actions and communicating with customers. Technical recovery and operational decision-making need to work together.
3. Examine cloud flexibility before procurement commitments
On 16 October 2026, the UK Competition and Markets Authority proposed designating Microsoft with strategic market status in its business software ecosystem.
The CMA identified areas for further investigation including software licensing, security-product bundling and interoperability. Separately, it is working with Microsoft and AWS on cloud choice, including portability and identity and access management across providers.
The proposed designation remains subject to consultation. It is a UK process and should not be presented as an EU-wide requirement or a final finding of wrongdoing.
Make flexibility a procurement question
For European buyers, our recommendation is to use these developments as a prompt to examine dependencies before the next contract renewal or platform commitment.
Procurement and engineering teams should establish where licences can be used, which integrations depend on proprietary services and what assistance would be available during migration. They should also estimate the internal effort required to move a representative application.
An integrated platform may offer substantial operational benefits. The decision becomes stronger when those benefits can be weighed against a documented cost of changing direction.
Choose where portability matters
We recommend prioritising portability according to business value. A critical application with a long expected lifespan may justify investment in an alternative deployment path. A smaller service may be better served by accepting a documented dependency.
A practical procurement review should produce an estimate of migration time, cost and service interruption for the most important workloads. That gives decision-makers a concrete basis for assessing supplier commitments and future flexibility.
Bring the three decisions into the next investment review
These topics belong together in planning. An AI platform evaluation should include operating responsibilities and recovery requirements. A resilience review should examine access to alternative capacity. A procurement decision should account for the engineering work needed to deploy, restore or migrate the service.
For the next investment review, ask teams to bring three pieces of evidence: results from a representative workload test, measured recovery performance and an estimate of the effort required to change providers. Each makes a different assumption visible before it becomes an expensive commitment.
Follow for further updates relevant to European technology and delivery decisions!