Your Practice's Data Is Not Your Own — Unless You Make It So

How open-source tools, European infrastructure, and local AI can give architectural practices genuine digital sovereignty

There is a question that most architectural practices have never been asked, and perhaps never thought to ask: who has legal access to your project data?

Not who you have given permission to. Who has the legal right to demand it — from your cloud storage provider, your email platform, your AI assistant — regardless of your wishes, and potentially without your knowledge?

The answer, for most UK and European practices currently relying on US-based software and cloud services, is: the United States government.

This is not a theoretical risk or a piece of privacy-activist hyperbole. It is a straightforward consequence of US federal law, and understanding it is increasingly a matter of professional responsibility — particularly for practices working on public sector commissions, critical infrastructure, defence-adjacent projects, or for clients with strong confidentiality requirements.

The CLOUD Act: What It Actually Means

The Clarifying Lawful Overseas Use of Data Act — universally known as the CLOUD Act — was signed into US law in March 2018.¹ It amends the Stored Communications Act of 1986 and resolves a long-running legal dispute between the US government and Microsoft over access to data stored on servers in Ireland.²

The key provision is this: any electronic communication service provider or remote computing service provider that operates or has a legal presence in the United States can be compelled by US law enforcement to produce stored data — regardless of where that data is physically located.³

The implications are significant:

Google Drive, Dropbox, OneDrive — all US companies. If your project drawings, client briefs, or structural calculations are stored there, a US federal warrant can compel their production.
Microsoft 365, Google Workspace — your emails, documents, video calls, and shared files are similarly reachable.
US-based AI platforms — when you paste a client's confidential brief into ChatGPT or Claude, that content passes through US-jurisdiction servers.
Autodesk, Adobe, Bentley Systems — cloud-connected, US-incorporated design tools may hold cached data, telemetry, or project metadata on US infrastructure.

The CLOUD Act applies not only to companies headquartered in the US, but to any provider that has 'operations or offices in the US or enters into contracts with US customers.'⁴ A European subsidiary of a US parent company does not escape this reach. The act follows corporate control, not geography.⁵

The UK signed a bilateral Data Access Agreement with the US under the CLOUD Act framework in 2019, which came into force in 2022.⁶ This agreement facilitates reciprocal law enforcement data access between the two countries — meaning the exposure runs in both directions.

It is worth being clear about what the CLOUD Act is not. It is not a surveillance dragnet. Warrants require judicial approval and probable cause; there is no bulk collection.⁷ But for practices with clients in sensitive sectors, the theoretical possibility of compelled disclosure — combined with the structural reality that your data is held by entities subject to a foreign jurisdiction — is a professional risk that warrants serious consideration.

The Open-Source Alternative: A Sovereign Stack Already Exists

Here is what may surprise many practitioners: a complete, professional-grade digital workflow for architectural practice can be built entirely from open-source tools running on European-owned infrastructure. The technology is mature, the tools are capable, and in most cases they are free.

The Design Tools

Blender is the most capable free 3D modelling and visualisation environment available. The Blender Foundation is a Dutch public-benefit corporation — Stichting Blender Foundation — headquartered in Amsterdam.⁸ Its terms of service are governed exclusively by Dutch law.⁹ It has no US parent company, no US incorporation, and no US operational presence. The CLOUD Act has no reach over it.

Bonsai (formerly BlenderBIM) is a Blender extension for native IFC-based Building Information Modelling, built on the IfcOpenShell library.¹⁰ It enables full BIM authoring — structural elements, spaces, properties, classification systems, drawing generation — in the open IFC standard. Like Blender, it runs entirely locally. No data leaves your machine unless you choose to send it.

FreeCAD with the BIM workbench provides an alternative parametric modelling environment. LibreOffice replaces the Microsoft Office suite for documentation and reports. All are open source, all run locally, all are free of US corporate jurisdiction.

The Infrastructure

For anything requiring external hosting — client portals, collaborative file sharing, remote access — EU-sovereign providers exist and are mature:

Hetzner (Germany) offers dedicated servers, cloud VPS, and storage products on infrastructure based entirely in Germany and Finland, governed by German law and the GDPR.¹¹

OVHcloud (France) is Europe's largest cloud provider by server count. It offers services across EU data centres and explicitly addresses its obligations under the CLOUD Act, noting that it operates 'the availability of legal mechanisms to quash or modify requests.'¹²

Infomaniak (Switzerland) operates under Swiss data protection law, which is robust and independent of both EU and US frameworks.

For file synchronisation and collaboration, Nextcloud — an open-source platform that can be self-hosted or hosted on a European provider — replaces Google Drive or SharePoint entirely, with no US corporate parent.

Version Control

Both Blender and Bonsai host their source code repositories on GitHub, which is owned by Microsoft — a US company and therefore subject to the CLOUD Act. This affects the source code repository, not your project files. But for practices wishing to version-control their own IFC models, drawings, or project data, Gitea or Forgejo can be self-hosted on your own server, keeping project history entirely under your control.

The Local LLM: The Piece That Makes Sovereignty Practical

For many practices, the most compelling reason to rely on US-based platforms has been AI assistance. ChatGPT, Copilot, and similar tools have become integral to specification drafting, client communication, code assistance, and document analysis. Migrating away from them has seemed to mean sacrificing AI capability entirely.

This is no longer true.

Ollama is an open-source runtime for large language models that runs entirely on local hardware.¹³ Once installed, it allows you to run models like Llama 3, Qwen, Phi-4, and — critically — Mistral, directly on an office server or workstation. No internet connection is required after the initial model download. No data is transmitted to any external server. The model processes your prompts locally and returns its response locally.

Mistral AI is a French company founded in April 2023 by researchers from Google DeepMind and Meta.¹⁴ It is headquartered in Paris, incorporated under French law, and has positioned itself explicitly as a European sovereign alternative to US AI providers. Its open-weight models — Mistral 7B, Mixtral 8x7B, and others — are released under the Apache 2.0 licence, meaning they can be downloaded, self-hosted, and operated with no commercial constraints and no data leaving your infrastructure.¹⁵ In January 2026, France's Ministry of the Armed Forces selected Mistral's models for deployment across all branches of the military, running on French-controlled infrastructure.¹⁶

A local Mistral model running on a capable office server — a workstation-class machine with a modern multi-core processor and a reasonable GPU — can assist with:

Specification drafting and editing
Planning statement preparation
Client report generation
Analysis of uploaded documents and drawings
Code assistance for parametric modelling scripts
BIM data queries against local IFC files

All of this without a single piece of client data leaving the building.

The Air-Gap Option

For practices working on projects requiring the highest level of data security — Ministry of Defence commissions, critical infrastructure, high-security private clients, legally privileged matters — the logical extension of sovereign infrastructure is the air-gapped workstation.

Air-gapping means the project machine has no network connection at all. Data enters and leaves only via physically controlled media under strict procedural protocols. This approach has been used in defence and government contexts for decades. What has historically made it difficult for knowledge-work practices is that it meant sacrificing all AI-assisted capability.

A local LLM removes this constraint. The model weights, once downloaded, are simply files on a drive. They require no network connection to operate. A fully air-gapped workstation running Blender, Bonsai, and a local Mistral or Llama model can produce BIM models, generate drawings, draft specifications, and answer questions about the project — all without ever connecting to the internet.

The honest caveats are worth stating clearly:

Hardware supply chains remain partially outside European control. CPUs, GPUs, and network interface firmware sit below the software layer. For government-grade security requirements, hardware provenance is subject to its own audit processes — but for professional practice purposes, this is a largely theoretical concern.

Model provenance: Llama is a product of Meta, a US company. If even this matters for a particular project's threat model, Mistral's open-weight models are the European-origin alternative. For the vast majority of architectural practice use cases, Llama's US origin presents no meaningful operational risk.

Human factors are the dominant vulnerability. An air-gapped server is defeated the moment a team member emails a drawing to a personal Gmail account or opens a project file on a personal laptop with cloud sync enabled. The technology can only take you as far as the discipline to use it consistently.

Collaboration with external parties is the hardest boundary to control. When you email a PDF to a client using Google's servers, that document passes through US infrastructure. You can only guarantee sovereignty at your own end.

Why This Matters Now

The geopolitical context of 2025-26 has made digital sovereignty a live concern for European governments and institutions in a way that it was not three years ago. The European Commission's repeated attempts to agree data-sharing frameworks with the United States — Safe Harbour, Privacy Shield, the EU-US Data Privacy Framework — have all faced legal challenge precisely because US surveillance law is structurally incompatible with EU data protection principles.¹⁷

The EU's Gaia-X initiative, launched in 2019 by Germany and France, represents the institutional response: an attempt to create a federated, interoperable European data infrastructure with sovereignty as a design principle.¹⁸ While Gaia-X has faced well-documented difficulties with governance and pace of implementation,¹⁹ its existence signals a broad European political consensus that dependence on US-controlled digital infrastructure is a strategic vulnerability.

For architectural practices, this translates into a concrete professional opportunity. A practice that can credibly state:

'Your project data is held on EU-sovereign infrastructure, processed only by software with no US jurisdiction exposure, and never passed to any US-based AI platform'

has a genuine and verifiable differentiator in public procurement, defence-adjacent work, healthcare estate commissions, and high-net-worth private clients with legitimate confidentiality concerns.

The technology to deliver that statement exists today, is proven, and is largely free of cost. The primary investment is in setup, server hardware, and the professional discipline to maintain the workflow consistently.

A Practical Starting Point

For a practice interested in exploring this direction, a phased approach is sensible:

Phase 1 — Local AI first. Install Ollama on an existing workstation and run a local Mistral or Llama model alongside existing tools. Experience what local AI assistance feels like without changing anything else. This has no cost beyond electricity and the initial download.

Phase 2 — File storage sovereignty. Move project file storage to a self-hosted Nextcloud instance on a Hetzner or OVHcloud server, or to a European provider's managed storage. This immediately removes the most significant cloud dependency.

Phase 3 — Design tool migration. For new project types or team members, begin exploring Blender and Bonsai for BIM authoring alongside existing tools. The learning curve is real but the capability is substantial.

Phase 4 — Full stack for qualifying projects. For projects where data sovereignty is a contractual or professional requirement, operate a complete sovereign stack: local workstations, European hosting, local LLM, self-hosted version control.

None of this requires abandoning existing tools overnight. The sovereign stack can co-exist with legacy workflows while capability and confidence develop.

Conclusion

Now, I admit that I'm not an expert in all these technical, financial and legal matters, and I'd be interested to hear from anyone who is. But my own research and reading leads me to the inescapable conclusion that the question of who has legal access to your practice's data is not abstract. It has, I believe, a concrete answer and for most practices that answer involves US federal law. The professional and commercial implications of that answer are becoming harder to ignore.

The good news is that the alternative is fully formed. Open-source tools, European cloud infrastructure, and locally-run AI models together constitute a complete, capable, sovereign digital environment for architectural practice. The Blender Foundation is Dutch. Bonsai is a community open-source project. Mistral is French. The servers are in Frankfurt and Helsinki. The model weights sit on your own drive.

Your data can be yours. The tools to make it so are already here.

Reference Appendix

All references in Oxford (footnote) format. Web sources accessed April 2026.

¹ United States Congress, Consolidated Appropriations Act, 2018, H.R.1625, 115th Congress, signed into law 23 March 2018, Division V: 'Clarifying Lawful Overseas Use of Data Act' (CLOUD Act). Available at: https://www.congress.gov/bill/115th-congress/house-bill/1625 [accessed 24 April 2026].

² 'CLOUD Act', Wikipedia, updated April 2026, https://en.wikipedia.org/wiki/CLOUD_Act [accessed 24 April 2026].

³ BSA | The Software Alliance, What Is the CLOUD Act? (Washington, DC: BSA, 2021), p. 1, https://www.bsa.org/files/policy-filings/09012021whatiscloudact.pdf [accessed 24 April 2026].

⁴ BSA | The Software Alliance, What Is the CLOUD Act?, p. 1.

⁵ Claromentis, 'Understanding the Implications and Risks of the US Cloud Act', Claromentis Blog, 23 October 2025, https://www.claromentis.com/blog/understanding-the-implications-and-risks-of-the-us-cloud-act [accessed 24 April 2026].

⁶ Wiley Law, 'The CLOUD Act Data Access Agreement — 10 Things That US Telecommunications Companies Need to Know Now', Wiley Alert, 2022, https://www.wiley.law/alert-The-CLOUD-Act-Data-Access-Agreement-10-Things-That-US-Telecommunications-Companies-Need-to-Know-Now [accessed 24 April 2026].

⁷ BSA | The Software Alliance, US CLOUD Act: Myths vs. Facts (Washington, DC: BSA, April 2019), https://www.bsa.org/files/policy-filings/04112019uscloudactmyth.pdf [accessed 24 April 2026].

⁸ Bloomberg Markets, 'Blender Foundation — Company Profile', Bloomberg Markets, https://www.bloomberg.com/profile/company/1411166D:NA [accessed 24 April 2026].

⁹ Blender Institute B.V., 'Terms and Conditions', Blender Studio, https://studio.blender.org/terms-and-conditions/ [accessed 24 April 2026].

¹⁰ IfcOpenShell, 'Bonsai', Blender Extensions, https://extensions.blender.org/add-ons/bonsai/ [accessed 24 April 2026].

¹¹ Hetzner Online GmbH, About Hetzner, https://www.hetzner.com/unternehmen/ueber-uns/ [accessed 24 April 2026].

¹² OVHcloud US, 'Cloud Act — Clarifying Lawful Overseas Use of Data', OVHcloud Legal FAQs, https://us.ovhcloud.com/legal/faqs/cloud-act/ [accessed 24 April 2026].

¹³ Ollama, Ollama: Run Large Language Models Locally, https://ollama.com [accessed 24 April 2026].

¹⁴ Mistral AI, 'About Us', Mistral AI, https://mistral.ai/about [accessed 24 April 2026].

¹⁵ EuropeanStack, 'Mistral AI Review 2026', EuropeanStack, March 2026, https://europeanstack.com/software/mistral-ai [accessed 24 April 2026].

¹⁶ D. Nani, 'Mistral AI and Europe's Push for Autonomous AI Systems', Sovereign Magazine, 18 January 2026 (updated 26 February 2026), https://www.sovereignmagazine.com/article/mistral-ai-and-europe-s-push-for-autonomous-ai-systems [accessed 24 April 2026].

¹⁷ Cross-Border Data Forum, 'Frequently Asked Questions about the US CLOUD Act', Cross-Border Data Forum, https://www.crossborderdataforum.org/cloudactfaqs/ [accessed 24 April 2026].

¹⁸ Gaia-X Association AISBL, Home — Gaia-X: A Federated Secure Data Infrastructure, https://gaia-x.eu/ [accessed 24 April 2026].

¹⁹ S. Abram, 'Gaia-X: Has Europe's Grand Digital Infrastructure Project Hit the Buffers?', Data Centre Dynamics, March 2026, https://www.datacenterdynamics.com/en/analysis/gaia-x-has-europes-grand-digital-infrastructure-project-hit-the-buffers/ [accessed 24 April 2026].