CENTCOM scales its AI infrastructure, shaping future of AI for combatant commands

After years of using external platforms like the Maven Smart System, Advana and National Geospatial-Intelligence Agency’s computer vision tools, the Central Command is taking artificial intelligence into its own hands. It’s shifting to a self-service AI model and starting to develop its own capabilities.

The command has recently launched CENTGPT, its own AI-powered chatbot and document analysis tool built on open source models like Meta’s Llama. CENTGPT incorporates retrieval-augmented generation, an AI framework that allows users to upload, interpret and analyze documents. The tool has quickly become indispensable for personnel in recent months.

CENTCOM has deployed CENTGPT on both its classified and unclassified network, so personnel can use the tool at different classification levels.

In six months, CENTGPT’s user base has grown to 900 users out of approximately 2,500 personnel at the command’s headquarters.

Meeting AI compute load

Throughout this process, CENTCOM has found that its classified cloud infrastructure is not robust enough to meet the command’s growing demand for high-performance computing at scale.

“We’ve talked to a number of the commercial cloud providers for classified compute, and what we found is the commercial side, on the classified side, does not have the ability to scale for compute at this time. There’s an investment that the commercial cloud providers will be making over the next several years, and the demand signal from the Defense Department is growing,” Dan Leventry, CENTCOM’s chief data officer, told Federal News Network.

To address its computing challenges, CENTCOM is building its own high-performance computing infrastructure. Since commercial cloud providers cannot yet scale classified compute resources, CENTCOM launched a pilot project. It has purchased and installed four Nvidia H100 GPUs, some of the most powerful AI processing units available, on its classified network to support CENTGPT and other AI efforts.

The command has also partnered with the Air Force Research Lab and Lambda through a Cooperative Research and Development Agreement, which will let CENTCOM acquire an additional 24 Nvidia H100 GPUs.

“It will give us a really huge, significant amount of compute capability that no one else — at least that I’m tracking — has in the Defense Department for classified networks,” Leventry said.

CENTCOM effort to help gauge compute load needs

The pilot will serve as a testing ground to determine how much computing power other combatant commands will need in the future. The command is working with Chief Digital and AI Office to determine what combatant commands will need to scale, Leventry said.

“We think we’re going to be good for several years, but it allows us to benchmark what the requirements are for other combatant commands as we move into the cloud and the commercial cloud providers begin to provide that compute power in the cloud,” he said.

The pilot will also help establish baseline costs and help determine whether cloud or on premise makes more sense economically — “to ultimately save the government money,” Leventry said.

As the command works to install the 24 additional Nvidia H100 GPUs in its classified server rooms, it is also shifting some of the Maven Smart System’s AI processing workload from NGA to its own in-house infrastructure.

“The fact that we’ve installed our own GPU farm here at CENTCOM, we’re not seeing the challenges right now. But we will see the growth as we add in the full tool suite of the AI platform, bringing other tools on from commercial vendors. I want to run my own computer vision models, and I’m kind of constrained on doing that right now because I can’t build them and I can’t run them for inference here at CENTCOM, just because we don’t have that computational resource resident. We will within weeks, and that’s really going to be a game changer for us,” Leventry said.

CENTCOM is also accelerating AI model development through initiatives like Desert Sentry, a commercial solutions offering focused on building low-code, no-code computer vision models. Approximately 82 vendors were interested in participating in the Desert Sentry initiative. CENTCOM narrowed it down to five finalists who were invited to conduct a live test in a real-world scenario.

The live test allowed the command to develop high-performing computer vision models within 48 hours, proving that tactical AI tools can be developed and deployed within days rather than months.

“We’re excited to get that off the ground. I’m working on the funding issue right now to get that finalized and then be able to integrate that into our platform,” Leventry said.

“I don’t think we could be conducting the operations today — and the level of coordination at the speed we are doing it at today — without using the most advanced and up-to-date tools available to us as a command. It’s hard to believe, if you go back 18 months ago, some of the manual processes that we had to do, some of the cycles that we had to do emailing PowerPoint slides to different people.”

Addressing early user feedback to CENTCOM AI tools

Since CENTGPT was built on open source software, its interface initially lacked the usability of commercial products. CENTCOM is now working to make the platform more user-friendly.

“Right now, we’re only able to ingest certain types of documents and certain sizes, so we’re getting better at that,” Leventry said.

Beyond CENTGPT, CENTCOM is getting similar feedback on other AI-powered tools, including logistics automation and predictive analytics, with users asking for more advanced capabilities and faster processing times.

“When you get the majors and lieutenant colonels in our formation seeing that value proposition, and it really makes their life easier, they want to adopt it,” Leventry said, adding, “They’re adopting it on their own, and they’re running with it. And that’s been really the neat thing to see — and a very powerful thing.”

Copyright © 2026 Federal News Network. All rights reserved. This website is not intended for users located within the European Economic Area.