US Data Residency for AI: Which Rules Apply to Which Data

NoraLin 8 2026-10-15 23:11:07 Edit

US data residency for AI means keeping the data used and produced by AI systems, including training sets, embeddings, model weights, prompts, and logs, stored and processed only in U.S. locations, and usually administered only by authorized U.S.-based personnel. There is no general federal law requiring all data to stay in the United States. Residency obligations come from specific rules for specific data types, from federal programs, and from customer contracts. The work is mapping each data type to its rule, then proving that every copy an AI pipeline creates stays where it should.

Which Rules Actually Require U.S. Residency

The first step is knowing which obligation applies, because each rule asks for something different: some restrict location, some restrict who can access data, and some only require a risk decision.

Data type Rule What it requires
Federal tax information IRS Publication 1075 No access by offshore equipment or personnel; cloud environments hosting it must be FedRAMP authorized
Criminal justice information FBI CJIS Security Policy Screening of personnel with access, plus security controls agreed with the criminal justice agency
Export-controlled technical data ITAR and EAR Access by unauthorized foreign persons can be an export; ITAR has allowed certain end-to-end encrypted data to be stored or sent abroad since a 2020 rule change, under specific conditions
Federal agency data FedRAMP and agency contracts Authorized cloud services, often with U.S. location and personnel terms in the contract
Bulk U.S. sensitive personal data DOJ Data Security Program, 28 CFR Part 202 (effective April 8, 2025) Prohibits or restricts access by countries of concern and covered persons, including through vendor and employment agreements
Protected health information HIPAA No explicit location requirement; HHS guidance permits offshore storage but expects the risks to be weighed in the risk analysis
Texas public-sector cloud data TX-RAMP Certified cloud services for state agencies and public universities
Customer data under contract Your customer agreements Often explicit U.S. location and subprocessor terms

Two points from this table shape most AI projects. First, HIPAA does not require U.S. residency, but many healthcare organizations adopt it anyway because offshore storage complicates risk analysis and incident response. Second, the DOJ rule is about access, not only location. Data held in the U.S. can still breach it if a vendor's offshore staff in a country of concern can reach it. Its thresholds start at more than 100 U.S. persons for human genomic data and more than 10,000 for personal health or financial data, measured over 12 months, and it covers anonymized and encrypted data too.

How AI Pipelines Break Residency

Residency usually holds for the primary database. It breaks in the copies that AI pipelines create and in managed services that route traffic for capacity.

Leak point How it happens Control
Global or cross-region model endpoints Some managed AI services offer deployment types that process requests in any region, or route across regions within a geography, for capacity Choose deployment types limited to U.S. regions, and record the choice
Staging buckets Training data copied to a bucket created in a default region Organization-level region restriction policies
Vector databases Managed vector service hosted outside the approved region Self-host the index in-region, or verify the service's location
Model hubs and registries Fine-tuned weights pushed to an external hub Private registry in-region with access control
Observability tools Prompts and responses sent to third-party logging or tracing services Redaction before export, or in-region logging
Backups and disaster recovery Replication to a secondary region by default Pin DR regions to U.S. locations
Support access Provider engineers outside the U.S. can reach the environment Contractual access restrictions and privileged access logs
Labeling and evaluation vendors Sample data sent to offshore annotation teams Vendor review under the DOJ rule and contracts

Worked Example: Residency Map for a RAG Assistant

Consider an internal assistant that answers questions over customer contracts. Mapping each component shows where residency is at risk.

Component Data it holds Where it runs Residency status Fix
Document store Original contracts U.S. region Compliant None
Embedding job Contract text in transit U.S. GPU nodes Compliant None
Vector index Embeddings and text chunks Managed service, region unverified At risk Verify location or self-host in-region
LLM endpoint Prompts with retrieved contract text Global deployment type At risk Switch to a U.S.-only deployment or a self-hosted model
Conversation logs Prompts and answers Third-party tracing tool At risk Redact or move logging in-region
Backups Index and logs Default DR replication Unknown Pin to U.S. regions

Four of six components needed attention, and none of them was the original database. This pattern is typical.

Controls That Enforce Residency

Policy documents do not keep data in place. These controls do:

  1. Region restriction at the organization level. Major clouds provide policies that block resource creation outside approved locations. Apply them to every account or project that touches AI data.
  2. Keys in-region. Encrypt AI data with keys held in a key management service or HSM in an approved location, so copies elsewhere cannot be decrypted.
  3. Egress control. Restrict outbound network traffic from GPU and data environments to approved destinations, which blocks accidental uploads to external hubs and APIs.
  4. Self-hosted models for regulated data. Running open-weight models on infrastructure you control removes uncertainty about where a model provider processes prompts.
  5. In-region logging and tracing. Keep prompts, responses, and traces in approved locations, with redaction for sensitive fields.
  6. Access restrictions. Limit administrative access to authorized personnel, log privileged sessions, and include location and personnel terms in provider contracts.

Dedicated infrastructure simplifies several of these controls at once. On single-tenant hardware in a named U.S. facility, there is no cross-region routing to configure, and the access model can be defined in the contract.

Evidence an Auditor or Customer Will Ask For

Evidence What it shows
Data flow diagram per AI workload Every component and copy, with location
Artifact register Datasets, embeddings, weights, logs, backups, with location and retention
Region policy configuration Resources cannot be created outside approved locations
Key management records Key location and who can use the keys
Provider contracts and subprocessor lists Location and personnel commitments
Privileged access logs Who accessed the environment and from where
Deletion records Data and model copies removed at the end of a project or contract

Where OneSource Cloud Fits

OneSource Cloud provides U.S.-based AI infrastructure designed for data residency requirements. Its Private AI Infrastructure runs single-tenant GPU environments in U.S. data centers, including Richardson, Texas. AI Storage Architecture keeps datasets, checkpoints, and vector indexes on in-facility storage, and the OnePlus Platform, OneSource Cloud's AI orchestration platform, provides model serving and developer tools inside the same environment, so prompts and logs do not need to leave it. For the jurisdiction and operator-control side of the topic, see our guide to sovereign AI cloud.

FAQ

Is there a U.S. law that requires data to be stored in the United States?

There is no general federal data localization law. Specific rules apply to specific data, such as IRS Publication 1075 for federal tax information, CJIS for criminal justice information, ITAR and EAR for export-controlled data, and contract terms in federal and commercial agreements.

Does HIPAA require data residency in the U.S.?

No. HIPAA has no explicit location requirement, and HHS guidance allows offshore cloud storage if the risks are addressed. Many healthcare organizations keep PHI in the U.S. as a risk decision.

Can I use a hosted LLM API and keep U.S. data residency?

Only if the provider commits to U.S.-only processing for the deployment type you use. Some managed AI services offer global or cross-region processing options, so check the specific deployment type and record it.

How does the DOJ bulk data rule affect AI vendors?

It restricts vendor, employment, and investment agreements that could give countries of concern or covered persons access to bulk U.S. sensitive personal data. AI vendors with offshore support or labeling teams should review who can access training data and logs.

What is the easiest way to prove data residency for AI?

Keep a data flow diagram and artifact register per workload, enforce region restrictions and in-region keys, and keep privileged access logs. Running the workload on dedicated infrastructure in a named U.S. facility reduces the number of components to prove.

Summary

U.S. data residency for AI is a mapping exercise. Identify which rule applies to each data type, then follow every copy the AI pipeline creates: staging data, embeddings, weights, prompts, logs, traces, and backups. Enforce residency with region policies, in-region keys, egress controls, self-hosted models for regulated data, and access restrictions. Then keep the evidence that shows it works.

To run AI workloads with U.S. data residency on dedicated infrastructure, explore OneSource Cloud's Private AI Infrastructure.

Tags:
Previous: AI Infrastructure for Healthcare: How to Build HIPAA-Ready Private AI Environments
Related Articles
新模块