Navy Brings US Military’s First AI Supercomputer of Its Kind Online

The Naval Postgraduate School in Monterey, California, has brought the first Nvidia DGX GB300 supercomputer deployed within the U.S. military online. Students, faculty and approved research partners will use it to study defense applications involving weather and ocean modeling, cybersecurity, disaster response and operational analysis.

High-performance storage and networking will determine how efficiently researchers can use its advanced processors.

Moving Data Fast Enough for AI

Kevin Delane, President and Chief Revenue Officer of the data infrastructure company DDN, has worked in storage, cloud computing and data management for more than three decades. He joined DDN in April after holding senior positions at EverPure (formerly Pure Storage), Cohesity and OwnBackup (now Own), which became part of Salesforce.

“Whoever has the most data, and is able to process it the fastest, generally has an advantage,” Delane said in an interview with Military.com.

An AI model learns by repeatedly analyzing examples. Graphics processing units, or GPUs, can perform many calculations at the same time, allowing researchers to train models faster. Those processors need a steady flow of information from storage. Delays in loading training data, saving a model’s progress or retrieving information can leave expensive processors waiting for work.

“There’s always a bottleneck when you look at all areas of the compute stack,” Delane said. “First is to get the data in. Second is to be able to disperse that in the network. And then the third is you have to be able to store and bring data back as fast as possible.”

DDN supplied high-performance data infrastructure that helps researchers access, manage, protect and expand the data available to the system. Nvidia donated the DGX GB300 and supplied the computing platform and software. VAST Data provided a data-management platform, while Vertiv supplied racks, cooling, power equipment and commissioning support.

Research for the Navy and Joint Force

The DGX GB300 combines 72 Nvidia Blackwell Ultra GPUs with 36 Nvidia Grace central processing units. It divides large computing jobs among many processors and handles different parts simultaneously.

NPS says researchers can train large AI models, adapt language models for naval planning and intelligence, create synthetic training data, study fleet tactics with multiple AI agents and examine how systems respond to adversarial attacks.

Delane described a basic research process that begins with identifying a problem and finding the relevant information. Researchers assess that data, use it to train a model and study the results. The new infrastructure allows them to process larger datasets and revise their models more quickly.

NPS designed the installation as a research and education resource. Many of its students are military officers and defense professionals who will return to positions across the Navy, Marine Corps and Joint Force. Their work at the school can help them understand where AI performs well, where it fails and how its limits may affect military decisions.

Researchers at NPS plan to use the system for environmental and ocean modeling, autonomous maritime systems, mission planning and decision support, disaster response and scientific simulation. They may also build digital twins, which are detailed computer representations used to study how a vessel, facility or operating environment may behave under different conditions.

NPS also plans to train large AI models, adapt language models for naval planning and intelligence, create synthetic training data, study fleet tactics with multiple AI agents and test systems against adversarial attacks.

Protecting Training Data

Poor or deliberately corrupted information can cause an AI model to learn false patterns. Attackers can also poison training data by inserting information designed to alter a model’s behavior.

“The accuracy and integrity definitely revolve around the input of where you get the data,” Delane said. “Are we feeding the AI the right information?”

Delane said DDN’s system includes cybersecurity features that protect stored information, maintain access and alert users when data appears corrupted or compromised. Storage protections secure the files feeding a model. Researchers still must check where the information came from, test the model’s behavior, control access and monitor its performance.

Defense Department guidance calls for continuous monitoring, validation and verification throughout an AI system’s operational life. The department’s testing framework also emphasizes data integrity, system resilience, security and performance under realistic operational conditions.

Credit: DDN

The DGX GB300’s computing components can draw roughly 135 kilowatts at peak use and rely on liquid cooling. Ships and expeditionary units have tighter limits on power, space and connectivity, so technology developed at NPS may require smaller hardware and additional engineering before operational use.

“With each kind of wave of technology, we’re able to process more and more information and data in a smaller footprint,” Delane said.

NPS now has the computing capacity to examine military AI projects at a scale it could not previously support. Delivering useful tools to sailors and Marines will require secure data, sustained testing and designs that work under operational conditions.