Amazon Bedrock adds NVIDIA Nemotron and GPT OSS to AWS GovCloud

Amazon Bedrock now hosts OpenAI’s gpt-oss-120b and gpt-oss-20b and NVIDIA Nemotron models inside AWS GovCloud (US), so government agencies and their contractors can run inference on these open-weight foundation models without moving sensitive data outside the GovCloud compliance boundary. In-region inference is available in us-gov-west-1, and Geo cross-Region inference can route requests across us-gov-west-1 and us-gov-east-1 while keeping traffic inside AWS GovCloud (US).

What models are available and what are their key specs?

OpenAI’s GPT OSS and NVIDIA Nemotron families are both present, with concrete parameter counts and context windows defined. OpenAI supplies gpt-oss-120b (120 billion parameters) and gpt-oss-20b (20 billion parameters), both offering a 128K-token context window and up to 16K output tokens. NVIDIA Nemotron on Bedrock includes Nemotron 3 Super 120B and Nemotron 3 Nano models (Nano 9B v2, Nano 12B v2, Nano 30B). Nemotron 3 Super is a 120B hybrid mixture-of-experts model that activates 12 billion parameters per token and provides a 1-million-token context window. Nemotron 3 Nano is a 30B model that activates approximately 3 billion parameters per token, also with a 1-million-token context window.

NVIDIA describes the Nemotron upgrades as delivering throughput and generation improvements: Nemotron 3 Super’s MoE design delivers up to 5 times higher throughput than the previous generation, and Nemotron 3 Nano delivers 4 times higher throughput while reducing reasoning-token generation by up to 60 percent.

How does inference run and how is data residency preserved?

Bedrock serves these models on a next-generation inference engine inside the AWS GovCloud (US) isolation boundary, with two API endpoints for invocation: the bedrock-mantle endpoint, which is OpenAI-compatible, and the bedrock-runtime endpoint, which exposes native Converse and InvokeModel APIs. The bedrock-mantle endpoint accepts calls using the OpenAI Python and TypeScript SDKs and uses the Chat Completions and Responses APIs. The inference engine separates the serving infrastructure (engine) from the endpoint, and Bedrock’s design enforces Model Deployment Account isolation and zero operator access so that no operator can access customer inference prompts or completions.

In-region inference is available in us-gov-west-1. Geo cross-Region inference uses a dedicated AWS GovCloud (US) cross-Region inference ID to route requests across us-gov-west-1 and us-gov-east-1, keeping traffic inside the AWS GovCloud (US) boundary. Global cross-Region inference across commercial AWS Regions is not available in AWS GovCloud (US).

What service tiers and operational options exist?

Amazon Bedrock offers Standard, Priority, and Flex tiers for these models; the Reserved tier is not currently available. Standard is pay-per-token without commitment, Priority provides higher throughput for latency-sensitive traffic, and Flex is a lower-cost option for non-time-sensitive workloads. By default requests use on-demand inference on the Standard tier. For authentication to the bedrock-mantle endpoint customers can use Amazon Bedrock API keys or standard AWS credentials; short-term API keys that expire automatically (maximum 12 hours) are recommended for production.

The console includes a Playground for interactive testing where users can select provider and model (for example NVIDIA Nemotron 3 Super or 120B gpt-oss-120b), apply the model and enter a prompt. The post supplies sample IAM policy statements required to call bedrock-mantle, and points to controls like SCPs to restrict which models an organization may use.

Why it matters

Government teams gain access to frontier open-weight models while keeping inference and data inside compliant GovCloud regions, which supports regulated workloads that require FedRAMP High, DoD SRG Impact Levels 2, 4 and 5, ITAR, and CJIS. The combination of OpenAI’s transparent open-weight models and NVIDIA’s MoE efficiency claims gives agencies options for higher-reasoning workloads and for more compute-efficient agentic systems. The platform-level protections, including Model Deployment Account isolation and the stated "zero operator access" design, reduce the operational risk of using external model families on managed infrastructure.

What to watch

Monitor latency and throughput behavior in your chosen tier when you move from evaluation to production, and test the Geo cross-Region inference path across us-gov-west-1 and us-gov-east-1 for resilience. Also track whether Amazon Bedrock adds Reserved tier support for these GovCloud models and whether additional model variants are added to the NVIDIA and OpenAI listings for AWS GovCloud (US).