An inference profile is a model ID with a routing table attached. You pass the profile ID where you would pass a model name, and the profile defines both the model and the Regions Bedrock may route the request to [7]. For the India geographic profiles that list is Asia Pacific (Mumbai) ap-south-1 and Asia Pacific (Hyderabad) ap-south-2, and AWS says requests route only between them [4]. Nothing in your application code enforces that; the enforcement is a prefix on a string that, in most codebases, arrives from an environment variable.
The same India endpoints accept `global.` profiles, which route to supported commercial Regions worldwide, cover Sol in addition to Terra and Luna, and come with AWS's own advice to use the `in.` profiles instead if you have local data processing requirements [12]. That gives you two compliance postures behind one identical call shape, which is a good argument for asserting the prefix in a test rather than trusting a deploy-time variable.
AWS describes cross-Region inference as primarily a capacity mechanism: instead of being bound to one Region's capacity, requests draw on a broader pool so throughput holds under load [6]. The India profile runs that machinery over a pool the post enumerates as two Regions [15]. That is better than one Region, but nowhere near the worldwide pool the global profile reaches [12]. For the throughput argument to transfer to your workload, both Mumbai and Hyderabad need to be carrying your chosen model with headroom during your peak, and AWS points readers at the Regional availability by models table for current per-Region availability [14]. That table is what tells you whether your two-Region pool is really two Regions.
Observability lands in one place by design. Billing and quota consume against your account in the source Region regardless of which backend Region served the request [8], and CloudWatch and CloudTrail record entries in the source Region only [9]. That is convenient for dashboards, but it also means your own audit trail will not show which Region handled a given call, so the evidence that processing stayed in India is the profile definition and AWS's statement about it [4], not a log line you hold.
The retention detail matters too. Bedrock's default is zero data retention: it does not store model inputs or outputs. For certain models, including GPT-5.6, content flagged by the automated abuse-detection classifiers is retained for offline abuse detection [10]. The data residency section states that exception without saying where the retained content is held [16]. If the reason you reached for the `in.` profile is that a prompt can carry patient or account data, that gap is the item to close before sign-off, not after. Input prompts and outputs may also move between the two Regions, encrypted in transit across the Amazon network [11].
One more adoption cost. AWS recommends the bedrock-runtime endpoint for new applications and says that is where features including Guardrails, intelligent prompt routing and cross-Region inference are available [13]. Anything still built on Mantle endpoints pays that migration first.
In a regulated Indian workload I would take the narrower pool. The 1-million-token window with text and image input is what makes the boundary worth the engineering, because a single request can carry a long document or a large code base [5], and that is the class of request you cannot casually split across geographies.
