Anthropic PBC plans to let business customers store data from its top-tier artificial intelligence models on their own cloud infrastructure instead of Anthropic’s servers. The change applies to the company’s most advanced models and is scheduled for rollout later this year through a new safety system. Customers will still face a 30-day retention requirement, but they gain the option to meet it inside their own environments.
The prior policy and its rationale
The move reverses an earlier policy that kept all such data on Anthropic’s systems. That prior approach was designed to reduce exposure to cyberattacks. Under the updated arrangement, enterprise users receive greater direct control over how and where their data resides during the required retention period. The policy applies specifically to business accounts using the highest-capability models.
Anthropic developed the original 30-day rule after weighing risks of external breaches against operational needs for model oversight. Keeping data on company-controlled servers limited the number of parties that could access interaction logs from advanced models. The new option relocates that storage without shortening or lengthening the window itself.
Implementation details
The new system maintains the 30-day retention rule but relocates the storage location. Anthropic expects the update to become available before the end of the year. No other changes to the retention duration or to the models covered have been stated. The company frames the adjustment as a response to customer demand for more control while preserving the original security rationale behind the 30-day period.
The safety system that enables the shift will handle the handoff between Anthropic’s inference infrastructure and customer cloud accounts. Enterprise teams must still preserve logs for the full 30 days, yet they can now apply their own encryption keys, access policies, and audit trails during that interval. The Bloomberg report notes the change applies only to the most capable models, leaving lower-tier offerings under the previous centralized arrangement.
Why it matters
Enterprise teams that handle sensitive workloads now have a concrete path to keep model interaction data inside their own compliance boundaries rather than sending it to a third-party provider for the full retention window. This reduces one vector of external exposure, though the 30-day requirement itself remains unchanged. Security teams will still need to manage the data on their side for that period, which shifts operational responsibility without eliminating the underlying storage obligation.
The adjustment also signals that Anthropic is willing to modify infrastructure choices when customer control becomes a competitive factor. Rivals offering similar model access may face pressure to match the option. For organizations already running large cloud estates, the change lowers the friction of adopting the most advanced models because data residency concerns can be addressed internally.
At the same time, the policy keeps a fixed retention length rather than moving to zero-day deletion or customer-chosen durations. Companies that wanted shorter or no retention will not receive that flexibility. The net effect is incremental rather than transformative: better location control for those who already operate secure clouds, but no reduction in the time window during which the data must be kept.
The rollout timing later this year gives customers a defined deadline to prepare their own storage and access controls. Teams evaluating Claude for production use can now factor the customer-cloud option into procurement and compliance reviews instead of assuming all data will land on Anthropic infrastructure. Procurement cycles that previously stalled over data-sovereignty questions may advance once internal cloud teams confirm they can meet the 30-day obligation on existing accounts.
Security architects will still perform the same core tasks—logging, encryption at rest, and access reviews—but they will perform them inside environments they already monitor rather than inside Anthropic’s perimeter. That consolidation can simplify incident-response playbooks and reduce the number of external parties that hold copies of prompts and completions from frontier models.
The decision also highlights a practical limit in current enterprise AI adoption. Many regulated industries already maintain strict data-residency rules for other categories of logs; extending those rules to model interactions removes one more external dependency. Anthropic’s move does not eliminate the retention obligation, yet it aligns the storage location with rules those industries already enforce on their own infrastructure.
---
Sources:
{"word_count": 682, "sources_used": 1, "expanded_sections": ["context", "why_it_matters"]}
No comments yet