The Department of Defense has moved public money and organizational energy to AI at an unprecedented scale. The FY2024 budget request explicitly called out $1.8 billion for artificial intelligence inside an overall RDT&E portfolio, and that line is being used alongside service experiments and agency programs to build production paths for AI-enabled capabilities.

In practice the department is not funding a single monolithic AI program. Instead it is directing R&D dollars and organizational effort into multiple, complementary vectors: data and model infrastructure, autonomy and attritable systems, sensor-to-shooter fusion experiments, and defenses against AI-enabled deception. Below I spotlight four convergent efforts that together account for how that $1.8 billion plus related research lines are being applied inside the department’s broader technology push. Each spotlight is followed by a brief technical read on risk, integration cost, and what to watch next.

1) Replicator - scaling attritable autonomous systems What it is: Announced publicly in August 2023, the Replicator concept is a Department-level push to field large numbers of relatively inexpensive, attritable autonomous systems across domains to blunt adversary advantages of mass and attrition. The announcement framed Replicator as an 18 to 24 month acceleration effort to change how the department buys and fields autonomy at scale. Why it matters technically: Replicator is an industrial and software integration problem as much as a sensor or airframe one. To get thousands of autonomous platforms to work together requires common messaging standards, resilient onboard autonomy that degrades gracefully under jamming, and a software stack that supports rapid retraining and over-the-air updates while preserving safety constraints. The challenge is not merely producing cheap air and surface frames. It is designing an assured autonomy architecture that supports distributed decision making at the tactical edge. Risks and watch items: supply chain fragility for components, adversary electronic attack that forces fall-back modes, and human-machine teaming rules that govern when and how autonomous agents act. Procurement reform and modular open architectures will determine whether Replicator is affordable at scale or becomes a logistics and maintenance sink.

2) JADC2 and Project Convergence - sensor-to-shooter and AI-enabled fusion What it is: Joint All Domain Command and Control is the department’s architectural effort to move information advantage to the speed of relevance. The FY2024 material plan included major funding for combined JADC2 activity alongside experimentation venues like the Army’s Project Convergence, which explicitly use AI and machine learning to reduce sensor-to-shooter timelines and test multi-service integration. Project Convergence has been the Army’s continuous campaign of learning where AI-enabled data fusion and tactical model updates are exercised in multinational experiments. Why it matters technically: JADC2 and Project Convergence expose the hardest systems engineering problems for defense AI. They demand common data representations, secure low-latency transport across contested links, model lifecycle management at the edge, and validated decision support that commanders can trust. Success here depends on engineering predictable latencies, provenance-aware data pipelines, and model validation regimes that are repeatable in operational conditions. Risks and watch items: brittle interfaces between legacy platforms and modern data fabrics, data labeling and curation at operational scale, and an overreliance on centralized compute that may not survive in contested communications environments.

3) DARPA’s Semantic Forensics (SemaFor) - automated detection and attribution of manipulated media What it is: DARPA has funded semantic forensics efforts since 2019 and selected research teams in early 2021 to develop algorithms that detect, characterize, and attribute falsified multimedia. SemaFor builds on earlier Media Forensics work and focuses on identifying failure modes in automated media generators and constructing robust detectors and attribution tools. Why it matters technically: SemaFor is the defensive complement to offensive uses of generative models. It requires cross-modal reasoning - linking text, audio, image, and video analysis - and building attribution pipelines that combine digital provenance checks with semantic inconsistency detection. These capabilities are indispensable for countering disinformation at scale and for preserving the integrity of tactical intelligence that depends on open source and coalition-sourced media. Risks and watch items: arms-race dynamics with generative model improvements, the risk of false positives in time-critical targeting workflows, and legal limits on using certain public data sets for attribution. Operational adoption will require human-in-the-loop thresholds and explainability mechanisms.

4) Data repositories, the Joint Common Foundation and the CDAO/JAIC lineage - infrastructure for AI R&D What it is: The institutional groundwork for DoD AI is legislative and technical. The NDAA and prior legislative actions directed the department to establish data repositories and to provide shared services and infrastructure to accelerate AI development and transition. Those mandates aim to create curated, accessible data repositories and shared model infrastructure that industry, academia, and the services can use to train and validate algorithms for defense missions. Why it matters technically: High-quality, well-labeled, and provenance-traced data is the single largest gating factor for effective ML deployment. The Joint Common Foundation concept and the data repository directives seek to reduce duplicated effort, accelerate model evaluation, and enable reproducible testing. In systems terms this is about creating secure data fabrics, model registries, and test harnesses that work at multiple classification levels. Risks and watch items: policy and legal constraints on data sharing, software supply chain risk, the cost of long-term curation, and the technical debt of maintaining model and dataset lineage so that field updates remain auditable and testable.

Synthesis and technical verdict Taken together these initiatives show the department applying a portfolio approach to AI. The FY2024 AI line sets a scale and ambition; programs like Replicator apply that ambition to mass-produced autonomy; Project Convergence and JADC2 test the command-level integration; DARPA covers resilience against synthetic media threats; and the data infrastructure work is intended to make the rest repeatable and auditable. Each node in that portfolio has distinct technical bottlenecks. Interoperability, assured autonomy under contested conditions, dataset quality, model lifecycle governance, and human-machine interface design are the five engineering domains that will determine whether the investment produces durable operational advantage.

Practical recommendations for program managers and engineers

  • Treat data as infrastructure not as a side project. Invest time and money in curated labeling, provenance and version control. Without that, model updates will be risky.
  • Prioritize modular open interfaces. Open data models and messaging standards reduce brittle coupling between legacy platforms and new AI services.
  • Build red teams for AI. Programs like SemaFor illustrate why you must test for adversary use of generative AI as well as deploy defensive pipelines.
  • Emphasize degraded-mode autonomy. Systems must be useful when communications and positioning services are degraded or spoofed. Replicator-scale architectures must bake in fallback behaviors that are predictable and auditable.

Bottom line The department’s public budget posture and program announcements through early 2024 make clear that AI-related R&D is being funded at near-$2 billion scale when one aggregates explicit AI line items and closely related initiatives. The technical work required to convert that investment into operational capability is not primarily about new models. It is about dependable data infrastructure, robust validation and testing, assured autonomy in contested environments, and acquisition pathways that reward modularity and speed. Get those pieces right and the DoD’s investment will buy persistent advantage. Get them wrong and the money will create costly point solutions that fail at scale.