Microsoft has reached an agreement to use Amazon Web Services infrastructure to support GitHub operations, a move that was confirmed in mid-June 2026 after months of reliability problems on the platform. The decision is operationally significant in its own right, but it also carries a broader architectural message that is relevant to every engineering team managing cloud infrastructure in 2026: AI workloads are scaling faster than cloud capacity planning models were built to handle, and the consequences are visible even at the scale of one of the world’s largest technology companies.

What happened

GitHub’s COO Kyle Daigle confirmed in a public statement that GitHub is now hitting 275 million commits per week, on pace for 14 billion in all of 2026, compared with 1 billion across the entirety of 2025. AI agent-opened pull requests surged from 4 million in September 2025 to more than 17 million by March 2026. GitHub Actions weekly compute minutes grew from 500 million in 2023 to 2.1 billion in a single week in early 2026.

This growth was not evenly distributed. GitHub Copilot, which became GitHub’s dominant product by revenue and usage in late 2025, drives AI inference and training workloads that are compute-intensive and latency-sensitive in ways that GitHub’s conventional version control workloads are not. By early 2026, GitHub’s AI workloads accounted for over 60 percent of the platform’s total compute consumption, up from 15 percent in early 2023.

Microsoft’s Azure capacity provisioning did not keep pace. GitHub recorded nine service incidents in May 2026. Platform availability dropped to approximately 88.4 percent in June, well below the 99.9 percent or higher that engineering teams depend on for CI/CD pipelines and collaborative development workflows. After those incidents, Microsoft provisioned AWS capacity as an operational measure to restore reliability.

What Microsoft is using AWS for

The AWS agreement covers AI inference and training workloads specifically. GitHub’s core version control functions, repository storage, and authentication remain on Microsoft’s own infrastructure. The AI layer, which includes Copilot inference, the AI features driving the surge in agent-opened pull requests, and the compute-intensive elements of GitHub Actions triggered by AI coding workflows, is the workload that exhausted Azure’s provisioned capacity.

This is not a permanent architectural decision. Microsoft had GitHub on a path toward full Azure migration by 2027, and the company has framed the AWS arrangement as a temporary operational measure to address the capacity gap while Azure provisioning catches up. That framing is credible. Microsoft has strong commercial incentives to host GitHub on Azure and no long-term interest in paying AWS for capacity that its own platform should supply.

But the framing does not change the practical implication: in June 2026, the reliability of a platform that 100 million developers depend on for their daily work required capacity from a competing cloud provider.

The architectural lesson

The GitHub situation is an extreme example of a problem that engineering teams encounter at much smaller scale: AI workloads do not scale like conventional application workloads, and the assumptions that informed cloud capacity planning before the AI adoption inflection point of 2023 to 2024 are no longer reliable.

AI inference is compute-intensive, burst-heavy, and latency-sensitive. A Copilot code completion request has a different resource profile than a git push. A GitHub Actions workflow triggered by an AI agent opening a pull request has a different duration and compute footprint than a workflow triggered by a human commit. When the proportion of AI-driven activity on a platform grows from 15 percent to 60 percent of total compute, the platform’s resource requirements change in ways that require updated capacity models, not just more of the same infrastructure.

For engineering teams running cloud infrastructure, the practical implication is that AI integration into developer workflows and production systems requires a deliberate reassessment of cloud architecture. Questions that were settled before AI adoption need to be revisited: Is the current instance mix appropriate for AI inference workloads? Are reserved capacity commitments still accurate? Does the architecture need burst capacity from a secondary provider to absorb spikes that the primary cannot absorb quickly enough?

What this means for European engineering teams

European organisations using GitHub for their development workflows should be aware of the service reliability context. The nine incidents in May 2026 and the June availability drop affected all GitHub users, including European teams whose CI/CD pipelines, code review workflows, and deployment automation depend on GitHub availability. The AWS capacity addition is intended to address this, but it also means that GitHub is now a multi-cloud platform in a way it was not six months ago.

For teams evaluating their dependency on GitHub, the incident is a useful data point for business continuity planning. A platform outage that runs at 88 percent availability for several weeks is a material disruption for engineering teams that deploy continuously. European teams that have not reviewed their fallback procedures for CI/CD pipeline failures or source control unavailability should do so.

More broadly, the Microsoft-GitHub-AWS situation illustrates a point that applies directly to European cloud architecture decisions. The assumption that a single cloud provider’s services are reliably self-consistent breaks down as AI workloads are added to existing infrastructure. Multi-cloud architecture, or at minimum cloud-agnostic infrastructure tooling, provides resilience against the capacity constraints that even the largest providers can encounter when workload profiles shift faster than capacity planning models anticipate.

European organisations that are working through the data residency and sovereignty implications of multi-cloud architectures also face an additional layer of complexity. GitHub’s use of AWS capacity raises questions about where data is processed that are relevant to GDPR compliance for organisations handling personal data through their development pipelines. The answer depends on what data flows through the workloads now running on AWS infrastructure and which AWS regions are involved, neither of which Microsoft has disclosed in detail.

If your engineering team is evaluating cloud architecture resilience, assessing your exposure to provider-level capacity constraints, working through multi-cloud strategy, or reviewing the data residency implications of using GitHub-hosted CI/CD for GDPR-regulated workloads, contact Excello Digital. We help European engineering teams build infrastructure that is resilient, compliant, and appropriately distributed across cloud providers.