US Government to Directly Host AI from OpenAI, Google
OpenAI, Google, and Perplexity are reportedly nearing approval to host their AI models directly for the U.S. government. This development marks a significant shift in public sector AI adoption, moving beyond API calls to dedicated hosting. The move is expected to accelerate government AI projects and establish new standards for security and auditability.
- The approvals for OpenAI, Google, and Perplexity are being expedited under a federal security initiative called FedRAMP 20x, with the initial authorization expected to be for "low impact" pilot programs. - The General Services Administration (GSA) is facilitating these deals through its "OneGov" strategy, which aims to work directly with tech companies to streamline IT acquisition; a recent OneGov deal made Perplexity Enterprise Pro available to federal agencies for just $0.25 per agency for an 18-month term. - This direct-to-government model allows AI providers to bypass intermediaries like Palantir and Microsoft, a significant shift highlighted by recent friction between Anthropic and the Pentagon over the use of its model in military applications, which was provided via Palantir. - The Department of Defense is a key customer, utilizing its Joint Warfighting Cloud Capability (JWCC) contract vehicle with AWS, Google, Microsoft, and Oracle to procure cloud services. In July 2025, the DoD's Chief Digital and AI Office (CDAO) also awarded contracts with a $200M ceiling each to Anthropic, Google, OpenAI, and xAI to accelerate the adoption of advanced AI. - This push is part of a wider trend; the number of generative AI use cases reported by federal agencies surged nine-fold from 32 to 282 between 2023 and 2024. The overall federal AI market for prime contract obligations reached $4.6 billion from FY 2021-2023. - Achieving the required Federal Risk and Authorization Management Program (FedRAMP) compliance is a major undertaking, typically costing $2-3 million and taking 18-36 months to secure an Authorization to Operate (ATO). Most AI applications that handle sensitive but unclassified data will likely target the FedRAMP Moderate level, which requires satisfying 325 security controls. - Specific government-focused platforms are emerging, such as Google's "Gemini for Government" and the GSA's "USAi" suite, which leverages models from multiple providers including Meta, Microsoft, OpenAI, and Anthropic.
