One very interesting instance of the "military data moat" is Ukraine's annotated database of drone footage, perhaps the first of its scale from live engagements [0]:

> They can now draw on an enormous pool of real warfare information. Last year alone, Ukrainian drones recorded around 820,000 verified strikes against Russian targets... Meanwhile, the country’s Avengers AI platform detects upwards of 12,000 enemy targets every week. Developers can now access these sources and the data that they gather to train their systems on the movements of a real Russian turtle tank or a camouflaged Lancet launcher.

> “Ukraine currently possesses a unique body of battlefield data unmatched anywhere in the world,” recently appointed Defence Minister Mykhailo Fedorov said in a statement. “This includes millions of annotated frames collected during tens of thousands of combat drone missions.”

With the latency and offline constraints of battlefield technology, smaller models, trained with better data, may prove to have a significant edge. But it's still early days on how data like this might prove advantageous in other environments.

[0] https://resiliencemedia.co/how-ukraine-is-transforming-its-b... (unconfirmed source, this is not an endorsement)