A procurement officer at a U.S. federal agency logs into a cloud console that, until recently, offered only a narrow menu of approved artificial intelligence tools. The new menu includes models from Anthropic, Meta, OpenAI, xAI and Nvidia, plus Amazon’s own Nova family, all running inside AWS GovCloud.
Amazon Web Services said Monday that it had expanded the availability of AI models in its GovCloud regions, giving government agencies access through the Amazon Bedrock platform to a broad set of third-party large language models. The services run in isolated cloud environments built to federal compliance requirements, the company said.
The move opens the door for agencies to use frontier models without building bespoke integrations or negotiating separate contracts with each developer. Previously, agencies seeking access to the most capable models often had to work through individual agreements, a process that slowed adoption, according to people familiar with government procurement. The Bedrock route standardizes the process: one console, one billing relationship, and a catalog of models from multiple vendors side by side.
GovCloud regions are physically and logically separate from AWS’s commercial regions, designed for workloads subject to FedRAMP, ITAR and other U.S. government compliance regimes. The expansion means agencies can now compare models from competing developers against each other on the same infrastructure, which procurement officials said could put downward pressure on the prices model companies charge the government.
The timing reflects a broader push to accelerate AI adoption across the federal government, and a scramble among cloud providers for the fastest-growing slice of government IT spending. Microsoft’s Azure Government and Google’s public-sector cloud have been courting the same agencies, and each of the three hyperscalers has been pairing infrastructure offerings with access to frontier models. Analysts said AWS’s decision to carry competitors’ models inside GovCloud is an acknowledgment that agencies want choice, not a single vendor’s stack.
Amazon’s own Nova models are positioned as a lower-cost option in the catalog, and analysts said the availability of open-weight Meta models matters for agencies that want to fine-tune systems on classified data without sending it to a third party. OpenAI’s and Anthropic’s models are offered as managed services, with the underlying inference running in the government cloud environment.
The announcement is part of a broader pattern: model developers have spent the past year building government sales channels, and cloud providers have become the distribution layer. For Anthropic, OpenAI and the others, a seat inside GovCloud is a government customer reached without building a federal sales force of their own. For AWS, it is a way to make its infrastructure indispensable to the largest institutional buyer in the world.
Security researchers caution that putting frontier models into government workloads raises new data-handling questions. Inference logs, prompt histories and fine-tuning data must be stored and managed under the same compliance regimes as the workloads themselves, and the models’ developers must not be able to see what agencies ask. AWS said the environment is designed so that model providers operate the inference while Amazon controls the data plane, a separation it says satisfies federal requirements.
The commercial stakes are large. Government agencies are among the few buyers whose AI budgets are growing without pause, and multiyear enterprise agreements with the federal government run into the hundreds of millions of dollars. Winning an agency’s default console tends to lock in years of workload growth, which is why the three cloud giants treat every government AI announcement as a competitive event.
Agency adoption of large language models has been slower than commercial uptake, held back by security reviews, procurement rules and the difficulty of getting approval to send data to a model operated by a third party. The GovCloud expansion attacks all three obstacles at once: the infrastructure is pre-certified, the procurement route is standardized, and the data stays inside the government cloud boundary.
Early agency use cases, according to people involved in the deployments, include intelligence analysts summarizing large document sets, procurement officers drafting and reviewing contract language, and mission planners stress-testing logistics schedules. The common thread is that agencies want models that can read their data without the data leaving their environment, which is precisely the capability GovCloud is designed to provide.
The model companies are racing to be included in such catalogs because government contracts, once won, tend to renew for years and grow in value. For the smaller developers in the catalog, a place inside GovCloud is a credibility signal that can unlock commercial enterprise deals as well. AWS is expected to announce additional model partners in the coming quarters, and the company said the catalog will be evaluated against what agencies actually use, with underused models dropped over time.
AWS said the expanded catalog is available now in its GovCloud regions, with additional models expected to be added over the coming quarters. The company framed the expansion as an infrastructure decision rather than a product launch, but the market read it differently: for the model companies, the deal is a distribution win; for the agencies, a menu that did not exist a year ago; and for the cloud competition, a reminder that in government AI, the race is increasingly won at the console where the workloads land.
