What actually opened
This is not three new regional deployments. It is a routing arrangement. A team in Bangkok or Taipei calls its local Bedrock endpoint, passes a profile ID such as global.openai.gpt-6-sol instead of a model ID, and Bedrock sends the request to a supported commercial Region where GPT-6 capacity exists. AWS says the point is straightforward: single-Region inference throttles when local demand exceeds local capacity, and a global profile gives the request somewhere else to go.
All three models take text and image input, client-side tool use and prompt caching. Sol and Luna accept reasoning effort from none to max; Astra accepts the same range except none. Existing OpenAI integrations can move by changing a base URL and a credential, and the ChatGPT desktop app and Codex CLI can be pointed at the same profiles, though AWS notes Codex needs version 0.148 or later and reports a 272,000-token window that is the client's own bundled figure, not a service limit.
Local bill, roaming request
The operational surface stays reassuringly local. You pay your source Region's rates, your metrics and logs land there, and CloudTrail records the Region that actually processed each call in additionalEventData.inferenceRegion. Read that field and you can find out where a request went. After it went there.
The retention clause is the part to read twice. Traffic flagged by AWS's automated abuse detection is kept for up to 30 days and processed programmatically, and under cross-Region inference that retention happens in the destination Region, which for a global profile can be any supported commercial Region where the model is deployed. AWS itself tells teams with data residency requirements to take the question to their compliance people early. That sentence is doing more work than the headline.
Quota burns output ten to one
Quota is measured in tokens per minute against the source Region, and the arithmetic favours short answers. Input and cache-write tokens count one for one; each output token consumes ten. AWS's own example: 2,000 input tokens plus 1,000 output tokens burns 12,000 tokens of quota. Bedrock also reserves the full max_tokens allowance up front, so a generous ceiling set out of habit shrinks your effective throughput before a word is generated. AWS advises setting it close to the completion you actually expect, and alarming at 70 to 80 percent of allocation.
The setup is not a switch you flip
Four of the five source Regions are opt-in Regions that must be activated first; Singapore is enabled by default. The IAM policy needs four statements, including one for the Region-agnostic model ARN with aws:RequestedRegion set to unspecified. That word is also why enterprise accounts trip at the first call: a global request carries no Region name, so a Region allowlist in a service control policy denies it, and the error names the policy type without saying which policy. AWS recommends a scoped exception for these profiles rather than adding unspecified to the allowlist for every service.
Practical checks before production, in order: confirm your residency position tolerates processing in an unnamed commercial Region; scope the SCP exception to the global profiles; size quota at the ten-to-one output rate; and verify after a test call that inferenceRegion appears in CloudTrail. None of that is difficult. All of it is easier before the first real workload than after.
Sources
- [1] AWS Machine Learning Blog, "Accessing OpenAI GPT models on Amazon Bedrock with global cross-Region inference in Thailand, Malaysia, Singapore, Indonesia, and Taiwan"Read source
- [2] Unite.AI, "AWS Expands GPT-6 Access on Amazon Bedrock to Five Asia Pacific Regions"Read source