Windows AI Bloat Is Making Installations Unsustainable for Smaller Drives

Windows AI Bloat Is Now a Real Infrastructure Problem

Windows AI bloat installation size has become a growing concern for IT administrators, developers, and privacy-conscious professionals who manage lean, controlled environments. Microsoft's Windows installation files have more than doubled in size over the past decade — and the primary culprit, according to Windows deployment experts, is a growing payload of AI-related code embedded in Copilot+ PC features that gets pushed to every machine, regardless of hardware capability.

Windows deployment specialist Michael Niehaus, who spent 16 years at Microsoft working on Windows deployment and now works for Swedish software company 2Pint Software, has been tracking the growth in Windows installer ISOs. His data reveals that from the initial release of Windows 11 version 24H2 in November 2024 through version 25H2 in July 2026, nearly 6,000 new folders and over 20,000 new files were added to the installer — in less than two years. The multi-edition ISO available through Microsoft's Visual Studio subscription has ballooned accordingly, while even the Media Creation Tool now generates a file of approximately 6.8 GB — nearly twice the size of the original Windows 10 release in 2015, as reported by ZDNET.

How Much Space Does a Windows 11 Installation Actually Consume?

The numbers are stark when laid out against available storage. Microsoft's official minimum system requirement for Windows 11 is a 64 GB system drive. However, when tested on a virtual machine with exactly that configuration, a fresh install of Windows 11 with the latest cumulative updates consumed 28.2 GB at peak installation — roughly half of the entire drive capacity, before a single application is installed or a single user file is saved.

Deployment professionals are now recommending a minimum of 128 GB for any practical Windows 11 installation, with a full terabyte recommended for developers, content creators, or anyone running multiple tools in parallel. This creates a real cost burden — especially at a time when, as Tom's Hardware has documented, memory and storage prices have risen sharply and are expected to remain elevated.

For context, the original Windows 10 ISO in 2015 was just over 3.5 GB. The trajectory has been consistently upward, but the most dramatic acceleration coincides with the rollout of Copilot+ features and embedded AI model packages pushed through Windows Update.

Why AI Files Are Being Pushed to Every PC — Even Those That Can't Use Them

The core technical issue is how Microsoft distributes its AI components. Approximately 3 GB of the cumulative update payload consists of shared AI models used by local AI features on Copilot+ PCs — machines equipped with a dedicated Neural Processing Unit (NPU). These models are bundled into dozens of WindowsWorkload.* files embedded in the update package.

The problem: these files are downloaded and processed by every Windows PC during cumulative updates, including machines with no NPU and no Copilot+ capability. The installer downloads them, then simply ignores them. This means users running budget laptops, older workstations, and enterprise machines that will never run a Copilot+ feature are still absorbing gigabytes of AI payload with every major update cycle.

This design decision stands in contrast to how Windows 11 handles some of its other components. A number of system features are now distributed through the Microsoft Store rather than Windows Update. This approach has clear advantages: Store-distributed components are only delivered when relevant, don't require a system restart to complete installation, and can be managed independently of the OS update cycle. The question being asked by IT professionals is why Microsoft hasn't extended this model to its AI packages — which are, after all, device-specific by nature.

What This Means for IT Administrators and Enterprise Deployments

For enterprise IT teams, the implications are practical and immediate. Large-scale Windows deployments — whether on-premises, hybrid, or cloud-hosted — depend on predictable image sizes for orchestration, bandwidth budgeting, and storage provisioning. When Microsoft ships an extra 3 GB of AI model data through a monthly cumulative update, it creates cascading costs: more bandwidth consumed across update distribution networks, larger golden images to maintain, and tighter storage margins on endpoints.

This is particularly acute in environments that still operate with older or cost-conscious hardware. Small and medium-sized businesses, educational institutions, and public sector organisations often run fleets of machines with modest storage configurations. The assumption that 64 GB is a viable minimum — an assumption embedded in Microsoft's own published requirements — is increasingly disconnected from the reality of update sizes on the ground.

From a digital sovereignty perspective — a concern central to European IT policy — this dynamic introduces a new layer of dependency. Organisations pursuing data minimisation strategies, or operating under strict data governance frameworks like the GDPR, must now contend with an operating system that silently installs AI model files they never requested, on hardware that cannot use them. As the European Data Protection Board's guidelines on data protection by design make clear, minimising unnecessary data processing and software components is not merely best practice — in some contexts, it's a compliance obligation.

The broader trend also connects to ongoing European regulatory scrutiny of Microsoft's bundling practices. The European Commission has previously examined Microsoft's integration of features into Windows as a potential competition concern, and the increasing embedding of AI capabilities into the core OS update cycle is likely to attract renewed attention from regulators focused on interoperability and user choice.

Could Microsoft Distribute AI Components More Responsibly?

There is a technically straightforward solution available to Microsoft: move the AI model packages from the monthly cumulative update channel into the Microsoft Store, where they can be distributed conditionally — only to devices with compatible NPU hardware, and only when the relevant feature is actively in use. This mirrors how Windows 11 already handles a growing number of its system components, from certain DirectX packages to inbox app frameworks.

The advantages of this approach are significant. Store-distributed components skip the restart requirement that accompanies Windows Update packages, reducing disruption to users and administrators alike. They can be version-controlled and pulled independently, making rollback easier. And critically, they would only consume bandwidth and storage on machines where they serve a function. According to analysis from Microsoft's own Windows Experience Blog, the Store-as-system-component model is already seen as an evolution path for the OS — the question is whether the AI payload will follow.