MITRE Federal AI Sandbox: a DGX H100 SuperPOD shared with U.S. agencies

When MITRE announced the sandbox in May 2024, it argued that few federal agencies had access to supercomputers or the staff needed to run them and test AI applications on secure infrastructure, even as an executive order on AI pushed agencies to lower barriers to adoption.

MITRE, a not-for-profit that runs six federally funded research and development centers and may not compete with industry, wanted a shared space where government sponsors could try AI models, estimate the cost of training foundation models and judge mission value before buying anything themselves.¹²

The sandbox, which MITRE named Judy after computing pioneer Judy Clapp, sits in Ashburn, Virginia. MITRE describes it as an NVIDIA DGX H100 SuperPOD with 248 H100 GPUs, 9 PB of VAST storage and high-speed networking. Agencies reach it through existing contracts with any of MITRE's six research centers rather than through a separate purchase. NVIDIA adds that the sandbox pulls together several generations of MITRE's earlier NVIDIA hardware into one central resource and can host open-source models while keeping controlled and classified information secure.

Projects named by MITRE and NVIDIA include high-resolution weather work with NOAA and the National Weather Service using NVIDIA Earth-2, a custom forecast visualization application built with Omniverse technologies for ocean areas, a cybersecurity foundation model for analysts, and models that interpret benefit regulations across departments. MITRE says the system was established in 2025 and, in its first year, trained foundation models and hosted experiments in weather and hazard planning, benefits and services, critical infrastructure defense and imagery understanding.¹²³⁴

What others can take from it: a public-interest operator can pool GPU capacity once and let many agencies reach it through contracts they already hold, which avoids each agency buying and staffing its own cluster for early experiments. Limits: MITRE paid for and runs the system itself and sits in an unusual position between government and industry, so the access model does not copy directly to a single ministry. The published material lists project areas but no measured mission results, named agency deployments or costs per project. For many agencies, an accredited government cloud region is a lower-cost first step before any dedicated hardware.