Skip to main content
InfraGap.com Logo
Home
Getting Started
Core Concept What is a CDE? How It Works Benefits CDE Assessment Getting Started Guide Inner Loop vs Outer Loop Environment Drift Local vs Cloud CDEs for Startups
AI & Automation
AI Coding Assistants Agentic AI AI-Native IDEs Agentic Engineering AI Agent Orchestration AI Governance AI-Assisted Architecture Shift-Left AI LLMOps Autonomous Development AI/ML Workloads CDEs for Data Science GPU Computing
Agent Infrastructure
Agent Experience (AX) Agent Egress Control Computer Use Agents Agent Evals Agent Runbooks Agent Client Protocol AGENTS.md MCP Servers Git Worktrees Kubernetes Agent Sandbox Agent Fleets Agent Identity Prompt Injection Defense Agent Observability Context Engineering AI Code Review Bottleneck Headless Agents in CI Spec-Driven Development Agent Readiness Code Provenance
Implementation
Architecture Patterns DevContainers Advanced DevContainers Language Quickstarts IDE Integration CI/CD Integration Platform Engineering Developer Portals Container Registry Multi-CDE Strategies Remote Dev Protocols Nix Environments Hermetic Builds OpenTofu for CDEs Kubernetes Development
Operations
Performance Optimization High Availability & DR Disaster Recovery Monitoring Capacity Planning Multi-Cluster Development Troubleshooting Runbooks Ephemeral Environments Sandbox Environments Workspace Snapshots Database Branching
Security
Security Deep Dive Zero Trust Architecture Secrets Management Vulnerability Management Network Security IAM Guide Supply Chain Security Air-Gapped Environments AI Agent Security MicroVM Isolation Compliance Guide EU AI Act Cyber Resilience Act Data Residency Governance
Planning
Pilot Program Design Stakeholder Communication Risk Management Migration Guide Cost Analysis FinOps GreenOps Vendor Evaluation Training Resources Developer Onboarding Team Structure Platform Maturity Model Open Source CDEs AI Productivity Paradox Build vs Buy DevEx Metrics Productivity Engineering Industry Guides CDEs for Healthcare CDEs for Financial Services CDEs for Government Edge Development WebAssembly in CDEs
Resources
Tools Comparison State of CDEs 2026 Isolation Decision Tool Template Library
Learning Paths
All Paths Platform Engineer Security and Compliance Engineering Manager
Vendor Reviews
GitHub Codespaces Coder Ona Google Workstations Microsoft Dev Box Okteto Eclipse Che DevPod Daytona E2B
Head to Head
Coder vs Codespaces Coder vs Ona Ona vs Codespaces Coder vs Okteto Self-Hosted vs Managed E2B vs Daytona CDE Market Guide CDE vs Alternatives Case Studies Lessons Learned Glossary FAQ Sources & Citations

Data Residency and Sovereignty for Source Code

Knowing where your repository is hosted answers about a fifth of the question. The working copy, the build cache, the prompt and the model inference all live somewhere too.

Not legal advice

This page describes how data location works in practice for development environments, and which regulatory regimes ask about it. It is not legal advice, and transfer analysis in particular is genuinely fact-specific. If you are subject to DORA, sectoral banking rules, or a national sovereignty requirement, take advice. Product facts and regulatory dates were verified against primary sources in July 2026 and move quickly.

The question is not where the repository is

Most residency conversations stop at repository hosting. Pick an EU region for the git remote, tick the box, move on. That answers where the code sits at rest and almost nothing about where it is actually processed.

A cloud development environment spreads your source across several distinct locations, and they are frequently in different jurisdictions from one another. The working copy is checked out onto a compute node somewhere. The build cache holds compiled artifacts derived from that source, often in a different service and sometimes a different region. Secrets are resolved somewhere. And an AI assistant sends fragments of the file you have open, plus whatever context it gathered, to a model endpoint.

That last one is the gap. An agent sending your proprietary source to an inference endpoint in another jurisdiction is a transfer, whatever your repository hosting says on its status page. It is also the leg most likely to be undocumented in an architecture diagram, because it was added by a plugin rather than by a platform decision.

Repository at rest

The easy one, and the only one most vendors advertise. Usually configurable by region.

Working copy

The full checkout on a live workspace node, plus whatever the developer generates alongside it.

Build and dependency cache

Derived artifacts are still your source code in a different shape. Remote caches are easy to forget.

Prompt and inference

Where the model runs, and whether the request was routed to a fixed geography or to whichever region had capacity.

Provider-side logs

Abuse monitoring and safety review stores retain prompts and completions independently of your retention settings.

Control plane metadata

Repository names, branch names and workspace telemetry can themselves be sensitive in a regulated setting.

Model endpoint residency and retention

Three questions decide whether an AI assistant is compatible with a residency requirement: where inference physically runs, how long the provider keeps the prompt and completion, and whether either of those guarantees survives your choice of model. The third catches people out.

ProviderGeographic controlRetention position
OpenAI APIData residency offered across roughly ten regions including the EEA, UK, Canada, Japan and AustraliaAPI data not used for training since March 2023. Abuse logs retained up to 30 days by default. Zero data retention requires prior approval and is configured per project
Anthropic APIInference routing can be pinned to the US, Europe, Asia or Australia, but Anthropic states data is stored in the USCommercial terms state customer content is not used to train models; API inputs and outputs deleted within 30 days by default. Zero retention is a contractual arrangement. Note that "covered models" retain prompts and outputs for 30 days even for zero-retention organizations
Amazon BedrockGeographic cross-region inference stays within US, EU or APAC. Global inference profiles may process in any supported commercial Region worldwideRetention is mode-based and can be set to none. Some frontier models require a provider data sharing mode with 30-day retention and are unavailable under zero retention
Azure OpenAIProcessing stays within the customer geography unless the deployment type is Global, or DataZone which stays within the EU or US zonePrompts and completions are not used to train foundation models without permission. The abuse monitoring store sits in the customer geography, with EEA-located reviewers for EEA deployments. Modified abuse monitoring removes human review, subject to approval

Zero data retention is not absolute

Two of the four positions above carve specific frontier models out of zero retention. A team that negotiated a zero-retention agreement and then adopted the newest model may have quietly left the arrangement. Treat model selection as a compliance control, not a preference, and re-check when you upgrade.

Watch the routing default

Global or capacity-routed inference profiles exist because they are cheaper and faster. They are also the setting most likely to be on by default and most likely to defeat a residency commitment. Pin the geography explicitly and verify it in the request path rather than in the contract.

Provider positions above are drawn from each vendor's own current documentation and change frequently. Google Cloud's Vertex AI position is documented in its service terms and data residency commitments; confirm the specifics directly rather than relying on a summary.

Who controls the control plane

"Runs in your VPC" and "self-hosted" are frequently used as synonyms and are not the same claim. In the first, your compute and your source stay inside your account while the vendor still operates the orchestration. In the second, you operate everything and the vendor cannot reach it. For a sovereignty requirement written by a regulator, that distinction is usually the whole point.

PlatformDeployment modelNotes
CoderFully self-hostedCommunity edition is free and open source; the root license is AGPL-3.0, with enterprise directories under a separate proprietary license. License keys validate locally, so air-gapped operation is supported
OktetoSaaS, self-hosted, or bring your own cloudSelf-hosted ships as a Helm chart into your own cluster. A free Starter tier exists; the CLI is Apache-2.0 with reduced features
Eclipse Che / OpenShift Dev SpacesFully self-hosted on KubernetesChe is EPL-2.0 and actively developed. Red Hat OpenShift Dev Spaces is the supported downstream, included with an OpenShift subscription
Ona (formerly Gitpod)Bring your own cloud, on EnterpriseEnterprise runners deploy into your own AWS or GCP VPC, so source and credentials stay inside your boundary. Ona describes this as self-hosted and Ona-managed. No customer-operated edition exists at any tier - the control plane stays Ona's. Ona's own pricing page and runners documentation also disagree on who manages the runner, so settle that contractually. The Core plan is multi-tenant on Ona Cloud, AWS only, in eu-central-1 and us-east-1. No Azure option
GitHub CodespacesVendor-hosted with data residencyGenerally available for GitHub Enterprise Cloud with data residency since April 1, 2026 across Australia, EU, Japan and US. Enterprise or organization-owned codespaces are required; user-owned codespaces are not supported

Corporate control is part of sovereignty

OpenAI announced an agreement to acquire Ona on June 11, 2026. Ona's own statement says the transaction is subject to customary closing conditions including required regulatory approvals, and that until close the two remain separate and independent companies. If your sovereignty case rests on who ultimately controls a vendor rather than on where bytes sit, an unclosed acquisition is a live item on the risk register, not a footnote.

Vendor continuity is a residency risk

Products in this category change direction quickly. Daytona, once positioned as a self-hostable development environment, has repositioned toward running AI-generated code, and its open-source repository states that core development moved to a private codebase as of June 2026. Check the current state of any platform before building a control around it.

The regulatory frame

DateWhat appliesStatus
10 Jul 2023EU-US Data Privacy Framework adequacy decision adoptedIn force, contested
17 Jan 2025DORA applies to EU financial entities and their ICT third-party providersIn force
12 Sep 2025EU Data Act applies. Some provisions phase in later, including Article 3(1) from September 12, 2026In force
18 Nov 2025The European Supervisory Authorities designate the first critical ICT third-party providers under DORADone
14 Jan 2026AWS European Sovereign Cloud reaches general availability in Brandenburg, GermanyAvailable
3 Jun 2026Cloud and AI Development Act presented by the CommissionProposal only
12 Jan 2027Data Act Article 29: cloud providers may no longer impose switching charges, egress includedScheduled

DORA

The Digital Operational Resilience Act, Regulation (EU) 2022/2554, not to be confused with the DevOps Research and Assessment program that shares the acronym in engineering circles. It has applied to EU financial entities since January 17, 2025.

Article 28(3) requires a register of information covering every contractual arrangement with an ICT third-party provider. Article 28(8) requires exit strategies with transition periods during which the provider must keep serving you. Article 29 requires concentration risk assessment.

AI vendors are ICT providers

DORA has no AI-specific carve-in, and it does not need one. Article 3(21) defines ICT services broadly as digital and data services provided through ICT systems on an ongoing basis. A model API supplied to a financial entity fits that description, which means it belongs in the register and needs an exit strategy like any other dependency.

Designated critical providers already include the major cloud platforms, so a development environment running on one of them may sit under two layers of the same regime.

Transfers to the US

The Data Privacy Framework adequacy decision remains in force. It is not, however, settled. The General Court dismissed the Latombe challenge in September 2025 and that ruling is under appeal to the Court of Justice.

A US Supreme Court ruling in June 2026 on the removal of officials from independent agencies has prompted fresh calls for the Commission to withdraw the decision. Nothing has been adjudicated. Plan as though standard contractual clauses may need to carry the load again.

What regulated teams are actually asked to prove

An auditor rarely asks whether you have a sovereign region. They ask you to demonstrate a specific claim about a specific data flow, with evidence that does not depend on someone remembering.

A complete flow map

Every location source code reaches, including build caches, telemetry and model endpoints. A diagram that stops at the repository will not survive a competent reviewer.

Enforcement, not policy

A written rule that engineers must only use approved models is not a control. Egress restrictions from the workspace network, and a proxy that only reaches approved endpoints, are.

Retention terms in writing

Marketing pages are not contracts. Zero retention that depends on approval, or that excludes certain models, needs to be recorded with those conditions attached.

A tested exit

DORA asks for exit strategies, and the Data Act removes the cost excuse from January 2027. Workspaces defined as code in your own repository are what makes an exit real rather than aspirational.

What a sovereign region does and does not buy you

The AWS European Sovereign Cloud is a separate partition with its own region in Brandenburg, operated by EU residents located in the EU through dedicated European legal entities, with independent identity and billing and an EU-citizen advisory board. The Commission's own cloud sovereignty framework, published in October 2025, scores providers against sovereignty objectives on a four-level scale and was used in a procurement whose contracts were awarded in April 2026.

None of that governs your application layer. A sovereign region will happily run a workspace whose coding assistant calls an endpoint on another continent. The infrastructure choice sets a ceiling on your exposure; the platform configuration decides whether you reach it.

Related reading

Residency questions usually arrive attached to a sector regulator, and rarely travel alone.