Wally model sovereignty: what it is and when you need it

Wally is part of the platform, but where its inference runs is your call. Settings → Wally model (admin-only, /settings/wally) lets you pick between two paths.

Turtini-hosted (default — what every org starts on):
• Anthropic Direct under Turtini's platform account
• Benefits from the shared prompt cache across the Turtini pool — usually the lowest per-call cost
• Billed at our standard +15% chargeback
• Suitable for the vast majority of orgs that don't have a specific compliance reason to do otherwise

Bring Your Own (BYO) Anthropic-compatible endpoint:
• Point Wally at an endpoint your compliance team operates — AWS Bedrock, Bedrock GovCloud, Google Vertex AI (US or EU), or your own gateway
• Inference happens inside your AWS account, your GCP project, or your gateway — prompts and responses never traverse Turtini's infrastructure
• You pay your vendor directly; the +15% chargeback does NOT apply on the inference path
• You configure the endpoint; we just consume it

When BYO is the right choice:
• You're a federal contractor that needs FedRAMP High or DoD Impact Level 5 (IL5) authorization on the inference path — pick Bedrock GovCloud
• You have customers in the EU and contractually owe data residency on AI processing — pick Vertex AI EU
• Your security model requires AI traffic to stay inside your VPC or GCP project boundary — pick the matching commercial Bedrock or Vertex preset
• You operate your own AI gateway for unified governance / observability — pick "Custom Anthropic-compatible endpoint"

When Turtini-hosted is the right choice:
• Most commercial workloads. The shared prompt cache makes Turtini-hosted noticeably cheaper than your own per-call Bedrock or Vertex bill in nearly every case
• You don't have a written compliance constraint that names where AI inference must run

Important: BYO is not a discount. Cache economics mean Turtini-hosted is usually the cheaper bill. Pick BYO when you have a compliance reason to, not as a cost-saving move.