The DoD AI Bottleneck Is Upstream of the Model

The DoD AI bottleneck is one stage upstream of the commercial AI bottleneck.

Frontier labs are wrestling with GPU economics, distributed training, inference cost, and the challenge of evaluating models that update quickly. Most defense programs cannot get a model into production fast enough to feel those limits. They run into authorization, classified-data movement, acquisition contracts, and evidence gaps first.

The department’s AI-first strategy raises the priority. The official DoW AI-first enterprise page points to seven Pace-Setting Projects and a push to accelerate AI across the force. That pressure is useful. It also makes the infrastructure gap more visible.

The honest fix is not deregulation. The controls exist because mission systems carry real risk. The problem is that many controls were designed around a delivery cadence that machine-learning systems do not have. A model can change because training data changed, a prompt changed, a retrieval source changed, a threshold changed, or an operator feedback loop changed.

That is why MLOps under RMF matters. The program needs data lineage that survives a cross-domain transfer. It needs model provenance that travels with the artifact. It needs continuous evaluation, drift detection, monitoring, and rollback evidence. It needs a way for an authorizing official to see what changed and why the risk remains acceptable.

Classified-data movement is often the real pacing item. Commercial AI tooling assumes easy access to cloud services, package registries, logs, telemetry, and external APIs. Secure mission environments do not work that way. Data may be partitioned by classification, program, coalition release, or operational need. Moving it safely is not a simple upload.

Acquisition adds another constraint. Contracts written for hardware deliverables do not map cleanly to models, data pipelines, evaluation harnesses, and continuous monitoring. A program can buy an AI prototype and still lack the contractual rights, data access, and sustainment plan needed to field it.

The model is only one artifact in the system. The useful AI capability includes data preparation, evaluation cases, monitoring, operator workflow, logging, security controls, and incident response. If any of those are missing, the model can be impressive and still not fieldable.

This is why pilot projects often look better than production programs. A pilot can curate data, accept manual transfer steps, and rely on a small group of experts. Production has to repeat the path, defend the evidence, support users, monitor drift, and recover from failure. The gap between those two worlds is where many defense AI efforts slow down.

The fix is to build the production path during the pilot, not after it. Every early experiment should ask what evidence, monitoring, data controls, and user workflow would be needed if the model were kept alive for years. That turns a demo into a rehearsal for fielding.

Polyrhythm works in that upstream layer: software architecture, secure delivery, test evidence, data movement, and authorization support. The DoD AI bottleneck will not be solved by model performance alone. It will be solved by engineering the path that lets a model become trusted mission software in the environment that must use it.