A 9,000-mile data gap is reshaping how the military builds AI

Bala Selvam, a former Marine who became chief technical officer of Special Operations Command Pacific in early 2025, didn't realize how wrong-shaped the Pentagon's AI plumbing was until he counted the distance. Indo-Pacific Command's headquarters in Hawaii sits roughly 9,000 miles from the command's second-largest data center. The pipelines he inherited were built for the continental United States, where bandwidth is fat and compute is abundant, not for the disconnected islands and littorals where Pacific troops actually operate.

Publicly, the Pentagon sells the AI program as a model story. The contracting and the memos say it is a data-plumbing story.

The vehicle for the rebuild is the War Data Platform, a branch of the Pentagon's Advana data system. WDP ingests thousands of streams, from satellite imagery to logistics feeds, and reshapes them into formats downstream AI applications can consume. According to Defense One, it is the connective tissue the rest of the AI stack will sit on.

The Pacific broke that connective tissue. The 9,000-mile distance from Hawaii to the command's second-largest data center is not a curiosity. It is a constraint that runs through every pipeline choice, from where a model is trained to where it is queried. Without a fiber backhaul, the model cannot sit in a hyperscale data center and assume the answer will reach the user. The whole architecture has to be redesigned so lightweight AI models and curated data are pushed out to the disconnected troop, and so the small models are designed for the form factors they will actually run on. Selvam told Defense One that on the continental US side, the same architecture "didn't matter because you had all the compute you needed." In the Pacific, that assumption is wrong.

The reset is now in writing and on contract. In January 2026, the DoD published a memo titled "Transforming Advana to Accelerate Artificial Intelligence and Enhance Auditability", a program-level acknowledgment that the data layer needed to be rebuilt before the AI layer could scale. The auditability language matters: battlefield AI decisions need traceable data lineage, and that has to live in the platform. In early July 2026, Accenture received an $821 million War Data Platform integration task order, the contract that will do the wiring. These are procurement vehicles for the architectural shift, not research pilots.

The shift has a second-order consequence that is easy to miss on the wire. If the operational reality is disconnected troops with constrained form factors, the procurement priority stops being frontier-scale models and starts being small, ruggedized models that can run on whatever compute is at the edge. That is what Selvam and his colleagues are working toward. Project Maven, the DoD's standing algorithmic warfare effort, has been pushing the same direction in computer vision and adjacent domains for years. WDP turns that bet from exception to default.

The push is also being made publicly. At the AWS Summit in Washington in the week of the Defense One story, Selvam took the stage. At the same event, Stanley discussed new model additions for GenAI.mil, the DoD's internal generative AI portal, which now serves almost 1.7 million users. On the same stage, Stanley named "Operation Epic Fury," the U.S. war on Iran, as a live use case where dozens of new data feeds were folded into the workflow in real time. The codename appears only inside that on-stage remark and should be read as reported speech, not adjudicated nomenclature. Defense One hedges the operational claims with the line that "some of these new workflows have already been used."

Earlier DefenseScoop reporting describes agents being built on top of GenAI.mil, including workflows that pull from internal DoD data stores. That is the user-visible surface of the same data-plumbing argument. The portal only scales if the platform underneath it does.

What to watch next. The Accenture integration contract has years of runway, and the Advana memo names specific auditability goals that are concrete enough to measure. The first signal that the reset is real will not be a new model release. It will be whether the data layer underneath starts delivering curated feeds to tactical formations on a predictable schedule. If it does, the procurement priority for small, edge-ready models will harden. If it slips, the next phase of the AI program will inherit the same disconnected-troop problem Selvam walked into.