AI in Defense: What Ukraine Has Taught the World About Implementing AI on the Battlefield

There is a large difference between an AI model that performs well on a test dataset and an AI system that can be successfully deployed into the field.

In the former situation, the team has a number in a report. In the latter, the question is what this number will be in a week if the imaging conditions change, new target types appear, the connection deteriorates or the model has to be run on a different device.

Ukraine finds itself in the situation where these questions must be answered constantly.

This is why Ukraine's experience is of interest, not only for its borders. Here, AI systems receive what is usually missing in their development in the laboratory: a large flow of real-world data, a constant testing under real-world conditions of reality, and a very close connection between the user and the developer.

In March 2026, the Ukrainian Ministry of Defense announced that it would be opening access to real battlefield data to international partners to train AI models.

Why is this important?

Because for computer vision, the difference between a good dataset and the reality can be huge.

On the test set, the camera sees the object quite clearly. In the field, the same object might be partly covered by smoke, shot from another angle, captured in poor lighting or looks entirely different from the data the model was trained on.

And then a high benchmark score says little about how the system will perform after the deployment.

Data becomes part of the system itself.

This is one of Valkyria, defense AI development company from Ukraine, key development principles.

The company works with machine learning, computer vision, autonomous systems and embedded AI and tests its solutions in Ukraine.

Essentially, developers are able to work not with a hypothetical "combat" dataset, but with data that actually occurs in production.

Instead of training a model once and checking its accuracy, they must constantly monitor the current data, where the model is making mistakes, and what has changed since the latest deployment. A good model may not run where it's needed.

There is another problem that benchmarks usually do not reveal.

Let's say a model runs in 40 milliseconds on a powerful GPU in the lab. This is an excellent result. But if the inference takes 180 milliseconds on the intended edge device, such AI may be useless for a specific task.

Or the model just doesn't fit within the available computing resources.

For defense AI, this is a typical engineering reality. The system must run on a particular platform, with a given power budget, memory, and computing capabilities.

Therefore, optimization does not start after the model is "ready". The hardware needs to be taken into account during development.

Valkyria, for example, works with embedded and edge AI and integrations based on NVIDIA Jetson, Hailo-8, and TI Jacinto/TDA4.

An FP32 checkpoint that works perfectly. In reality, quantization, compiler-level optimization, profiling, and completely different restrictions appear.

The field is changing model requirements.

The Ukrainian experience demonstrates another thing: system requirements can change faster than the typical development cycle.

What worked yesterday may require refinement today. Conditions, data, equipment, and system application methods change.

The US military has explicitly described a similar dynamic when speaking about Ukraine. In 2025, the commander of U.S. European Command noted that systems can quickly lose effectiveness due to environmental changes and that industry must quickly adapt technologies.

Therefore, deployment here cannot be perceived as the final stage of the project.

Rather, it follows a cycle:

field data → training → model → deployment → testing → feedback → new field data.

The faster the team completes this cycle, the faster it can adapt the system to new conditions.

And this is perhaps one of the most interesting lessons from Ukraine for the global defense tech market.

From benchmark to operational readiness

The US Government Accountability Office came to a similar conclusion from a different perspective. In its analysis of AI in weapon systems, the GAO specifically highlights challenges with data quality, AI integration into existing systems, and continuous monitoring of already deployed capabilities. AI, the agency notes, requires large volumes of data and continuous monitoring; a standard approach to software development is not sufficient.

As a result, the key question for Defense AI is gradually shifting.

Not just:

"How well does the model perform?"

But also:

"How well does the entire system perform, on the right hardware, with real data, and under the conditions for which it was designed?"

For Valkyria, this is no longer a theoretical question. The company is building AI/ML software specifically around this gap, from computer vision and autonomous systems to edge deployment and field testing.

The Ukrainian experience shows why this approach is becoming increasingly important.

Winning a benchmark isn't enough.

The model must survive deployment, real data, hardware limitations, and feedback from the field, and then improve.

This is the next level of Defense AI: not just building a model, but learning to quickly transform reality into the next version of a working system.