Federal AI Turns to Factories as Cloud Limits Emerge
AI workloads and security demands require infrastructure beyond the cloud, tech leaders say.
April 30, 2026 – Federal agencies scaling artificial intelligence are increasingly turning to “AI factories” as cloud-only strategies fall short on cost, control, and resilience.
Speakers from Leidos, NVIDIA, Dell Technologies and Sterling said Thursday dedicated AI infrastructure is becoming necessary to move from pilot programs into production. Their remarks came during a discussion hosted by the Government Executive Media Group.
“It comes down to scale, security, resilience and efficiency,” said Eric Moore, senior vice president and chief technology officer for digital modernization at Leidos. AI factories, he said, allow agencies to operationalize AI while maintaining control over sensitive data.
The shift comes as AI workloads reshape network infrastructure. Nokia said in March that AI systems are generating 100 trillion tokens per day globally, forcing operators to rethink how compute and connectivity are designed and delivered.
“AI factory” remains loosely defined across the industry.
“Depending on who you’re talking to, you’ll find different definitions,” said Forrest Bussler, senior solution architect at NVIDIA, describing systems ranging from small deployments to large-scale GPU clusters.
For Nathan Bennett, director of innovation and technology at Sterling, the concept centers on “enablement of the stack,” including identity, infrastructure and governance.
Chris Thomas, technical director for AI and data systems at Dell Technologies, said AI systems require tighter integration than traditional IT, with compute, storage and networking no longer operating independently. That complexity, he said, is driving demand for architectures that simplify deployment.
Cloud infrastructure alone cannot meet federal AI demands. “Just being in the cloud is not going to give you the functionality that you need,” Bennett said, warning that continuous workloads can become prohibitively expensive.
Security and resilience also remain concerns, Moore said, noting that some workloads must remain on-premise due to sensitivity, while outages make hybrid approaches essential.
Rather than replacing the cloud, AI factories are being deployed alongside it, with on-premise systems handling steady workloads and cloud environments supporting spikes in demand.
“You don’t want to overprovision to meet peak performance,” Bussler said, pointing to the need to balance capacity across environments.
Control is a key advantage.“If you have an AI factory, you have full control over the models and functions,” Bennett said, adding that control can also improve cost predictability.
“There is a steady-state demand for inference,” Moore said, noting on-premise systems are often more efficient for continuous workloads.
Building AI factories requires planning, particularly around power and infrastructure.
“Making sure that your power and your ability to run this high-performance infrastructure is actually in place, that’s the first step,” Bennett said.
AI traffic is also “bursty and dynamic,” Justin Hotard Nokia’s president said in March, reinforcing the need for distributed, flexible systems.
Looking ahead, panelists said AI factories will evolve into distributed networks across data centers, edge environments and local devices.
“Think of them not as a singular entity but factories across your ecosystem,” Moore said.
But panelists said the starting point is not infrastructure, it is people.
“The AI journey starts with the human element,” Bennett said, emphasizing the need to define use cases and measurable outcomes before investing.
Member discussion