Government and public sector
How ministries, agencies and their research partners use NVIDIA compute, open models, inference software and simulation for document-heavy casework, weather and hazard planning, emergency coordination and secure experimentation, under rules on citizen rights, classified data and public accountability.
The problem
Agencies hold enormous volumes of regulations, case files and correspondence, and staff must answer citizens accurately while the rules keep changing. NVIDIA's account of French public finance work cites more than 100 million documents a year handled by the administration. Emergencies add a different burden: during a wildfire or flood, command centers must merge video, radio, sensor and 911 data from several agencies at once.
Many agencies also lack the compute and specialist staff to try AI on secure infrastructure; MITRE built its shared sandbox on exactly that argument. When projects do go ahead, savings are easy to overstate. France's Cour des comptes reviewed AI at the economy and finance ministry in October 2024 and, as reported by Le Monde Informatique, found that five cost-cutting projects delivered about 20 million euros a year against a 46.6 million target.
Public decisions about benefits, migration, justice and policing carry legal duties to citizens. The EU AI Act treats AI used for access to essential public services, border control and the administration of justice as high-risk.12345
The approach
A first pattern is shared, secure compute. MITRE runs a DGX H100 SuperPOD as a Federal AI Sandbox that U.S. agencies reach through existing research center contracts, and NVIDIA lists a DGX GB300 system at the Naval Postgraduate School for public sector AI education and research.
A second is agents that stay inside government boundaries. Leidos built a command and control agent system for disaster response with NVIDIA AI Enterprise, NIM microservices, a Riva speech NIM and Metropolis, developed with synthetic emergency scenarios. NVIDIA describes France's public finance directorate using ThinkDeep's DeepBrain assistants on premises with NIM, NeMo Retriever, Nemotron models and NeMo Guardrails. Palantir uses Nemotron open models for agencies and critical infrastructure operators, and Nemotron is available on Amazon Bedrock in AWS GovCloud (US), which NVIDIA reports is cleared for FedRAMP High. A third is simulation: NVIDIA promotes Omniverse for training environments and digital twins, and Jetson for robots, drones and video analytics at the tactical edge.
Plenty of public sector work does not need dedicated GPUs. Search over existing document systems, rules engines for eligibility, and assistants bought as a service inside an accredited government cloud cover many needs, and they are faster to procure. Owned hardware fits when data cannot leave agency control, workloads are steady, or several agencies can share one system.12367
Conceptual architecture
Diagram as a list
Applications & solutions
- Casework and command center agents on NVIDIA AI EnterpriseDraft answers, triage incidents and gather evidence for officialsConnects to Official review, decision record and appeal process
- Official review, decision record and appeal processA named official makes and records each decision that affects a person
Models & frameworks
- Open models tuned for agency tasks (Nemotron and others)Retrieve, summarize and reason over agency documents and dataConnects to Self-hosted inference services (NIM)
- Simulation and training environments (Omniverse)Rehearse emergencies and test plans before real operationsConnects to Casework and command center agents on NVIDIA AI Enterprise
Inference & runtime software
- Self-hosted inference services (NIM)Expose models through standard APIs without sending data outsideConnects to Casework and command center agents on NVIDIA AI Enterprise
Operations & orchestration
- Agency records, regulations, sensor and video feeds in a secure enclaveHolds data at its classification level with access loggingConnects to Shared or agency-owned GPU systems (DGX) or accredited government cloud
Accelerated computing
- Shared or agency-owned GPU systems (DGX) or accredited government cloudTrains and runs models where the data is allowed to resideConnects to Open models tuned for agency tasks (Nemotron and others)
Technologies and their roles
NVIDIA DGX6
Shared secure AI compute for agencies
MITRE's Federal AI Sandbox runs on an NVIDIA DGX H100 SuperPOD that agencies reach through MITRE's research centers.
NVIDIA AI Enterprise7
Supported software base for government AI
NVIDIA markets government-ready AI Enterprise software for secure deployment, and Leidos built its disaster response agents with AI Enterprise.
NVIDIA NIM2
Self-hosted model serving for agents
Leidos used NIM microservices for the agents in its command and control system, including a Riva speech NIM for transcription.
NVIDIA Nemotron7
Open models for mission-specific agents
NVIDIA says Palantir uses Nemotron for sovereign AI for agencies and that Nemotron is available on Amazon Bedrock in AWS GovCloud (US).
NVIDIA Omniverse8
Simulation, training and scientific twins
NVIDIA presents Omniverse for agency training environments, and MITRE built a forecast visualization application with Omniverse technologies.
NVIDIA Jetson7
Autonomy and analytics at the tactical edge
NVIDIA's federal page presents Jetson for robots, drones and video analytics at the tactical edge.
What you need first
- A legal basis and documented purpose for each AI use that touches personal data or individual decisions
- An approved hosting option at the right classification level, whether on premises or an accredited government cloud
- A records management plan that keeps prompts, outputs and decisions available for audit and freedom of information requests
- Named officials who remain accountable for decisions, with an appeal route for affected people
- Evaluation sets built from real agency cases, including hard and rare ones, before launch
- Procurement terms covering data ownership, model portability and exit
- A baseline and a method for measuring savings or service gains, agreed before the project starts
Risks and how to reduce them
- Harm to citizens from wrong automated decisions on benefits, migration or justice
- Keep a human decision-maker, give reasons people can contest, test for bias across groups, and meet EU high-risk duties such as logging and human oversight where they apply.
- Leakage of classified or personal data
- Run models inside the agency boundary or an accredited cloud, block external calls by default, and log all access to sensitive collections.
- Overstated savings that do not materialize4
- Set a measured baseline, track results independently of the vendor, and report shortfalls, as audit bodies such as France's Cour des comptes do.
- Lock-in to one vendor or integrator
- Prefer open models and standard APIs, keep data and evaluation sets under agency control, and include migration rights in contracts.
Documented examples
MITRE · Government and public sector (federally funded research and development)
MITRE Federal AI Sandbox: a DGX H100 SuperPOD shared with U.S. agencies
MITRE runs its Federal AI Sandbox on an NVIDIA DGX H100 SuperPOD with 248 GPUs so that U.S. agencies can train and test AI models through MITRE's research centers. In its first year it hosted work on weather models, benefits systems, infrastructure cyber defense and imagery. No mission results are published.
In production
Related
Sources
- ThinkDeep's AI Agents Help Automate Public Services for the Government of France (opens in a new tab)
- Leidos Uses AI Agents to Cut Response Times and Boost Coordination in Disaster Scenarios (opens in a new tab)
- MITRE to Establish New AI Experimentation and Prototyping Capability for U.S. Government Agencies (opens in a new tab)
- Bercy dégage moins d'économies que prévus avec l'IA (opens in a new tab)
- AI Act: regulatory framework for AI (opens in a new tab)
- Top Publication Honors Impact of MITRE's Federal AI Sandbox (opens in a new tab)
- AI for the U.S. Federal Government (opens in a new tab)
- MITRE's Federal AI Sandbox Unleashes AI's Potential for Public Good (opens in a new tab)
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.