From data overload to decision advantage: Operationalizing AI for CJADC2 at the tactical edge

• Share on X
• Share on LinkedIn

Story

August 11, 2026

The shift toward Combined Joint All-Domain Command and Control (CJADC2) for the front lines represents a major architectural evolution in modern military history. It’s a transition from a platform-centric force, where a fighter jet, a tank, or a destroyer operates as an island of intelligence; to a network-centric force, where every sensor and effector is an interconnected node.

CJADC2 is essentially the U.S. Department of Defense’s (DoD’s) mission to transform a fragmented military into what is sometimes described as a unified, lightning-fast Internet of Military Things (IoMT). Historically, the Army, Navy, and Air Force operated on separate networks that didn’t always talk to one another, leading to dangerous delays in communication. CJADC2 replaces these isolated silos with a series of interconnected capabilities that link every sensor (like a drone or satellite) to every shooter (like a missile battery or jet) across all land, sea, air, space and, cyber domains, with connectivity extended to international allies.

The primary vehicle for developing and validating these capabilities is the Global Information Dominance Experi­ments (GIDE), an iterative series co-sponsored by the Chief Dig­ital and Artificial Intelli­gence Office (CDAO) and the Joint Staff. Held quarterly, the GIDE exercises bring together leaders from all service branches, all 11 combatant commands, and international allies and partners to test how data and software capabilities can be matured and linked to acquisition pathways. Rather than using a traditional waterfall development cycle – that is, linear and sequential, often slow – GIDE operates on a continuous “experiment, assess, and field” loop to ensure CJADC2 evolves with the threat environment rather than reacting to it.

As of 2026, the Pentagon has moved beyond mere experimentation, successfully achieving minimum viable capability and shifting toward enterprise-wide deployment. The goal is to achieve decision advantage, by which the DoD means the ability to see the battlefield, process the chaos, and strike with precision faster than any adversary can react.

However, connectivity is only half of the battle. The data deluge has reached a breaking point. Modern sensor suites generate terabytes of data every hour, and backhauling this to a centralized cloud is virtually impossible in contested, bandwidth-constrained environments.

According to recent Pentagon modernization updates, the military is now doubling edge storage capacities annually (for instance, scaling from 75 TB to 300 TB in single-drive modules) just to keep pace with the tactical data being generated. To achieve decision advantage, the industry can no longer rely on collect-and-store strategies. Intelligence must move to the edge, where it’s needed.

Artificial intelligence (AI) is the cognitive connective tissue that makes CJADC2 viable – but the algorithms are only as good as the foundation beneath them. For AI to be operationalized in mission-critical environments, it needs a trusted software substrate that does three things the enterprise cloud never had to do: Connect nodes autonomously when the network is contested, isolate workloads by security domain, and let very different classes of software safely share the same hardware.

In practice, that rests on three pillars – trusted communication infrastructure that enables autonomous mesh networking between nodes; multilevel security domain isolation; and mixed criticality, so that enterprise, agile, and autonomous workloads can coexist with safety-critical functions on a single platform.

The market imperative: by the numbers

The urgency behind CJADC2 is reflected in recent industrial and budgetary trends. The JADC2 technology market, valued by one study at $12.8 billion in 2024, is projected to surge to $28.4 billion by 2034. Furthermore, the U.S. AI in defense market specifically is anticipated to rise to approximately $10.9 billion by 2031, progressing at a 22.1% combined annual growth rate (CAGR).

For the embedded systems engineer, this trend means heterogeneous compute combining CPUs [central processing units], GPUs [graphics processing units] and FPGAs [field-programmable gate arrays] on tactical platforms is a must. The DoD signaled this priority clearly by allocating $3.2 billion for JADC2 initiatives in the 2024 fiscal year alone.

The architectural challenge: Why the cloud isn’t enough

The fundamental challenge of deploying AI for CJADC2 is the conflict between enterprise AI models and mission constraints deployed at the edge. Enterprise AI is designed for elasticity, scale, and throughput. It doesn’t matter if a facial-recognition algorithm takes 100 ms or 200 ms in a commercial app.

In a CJADC2 scenario, however, such as hypersonic threat classification or automated electronic warfare (EW) response, latency variability is lethal. If an AI inference engine consumes 100% of a GPU’s cycles and delays a target-tracking loop or an interceptor-cueing decision, the node fails.

Closing this gap means moving toward a faster, deterministic, safe, and secure software substrate. This architecture must provide:

• Trusted communication infrastructure: Enabling nodes to discover one another and form a resilient, autonomous mesh so that the network keeps routing data and sharing tracks even when links are jammed, degraded, or severed.
• Multilevel security domain isolation: Keeping workloads of different classifications and trust levels strictly separated on shared hardware, so data or code in one security domain cannot bleed into another.
• Mixed criticality: The decisive capability for CJADC2, which is in fact enabling enterprise, agile, and autonomous AI workloads to safely coexist with safety-critical functions on the same node, so high-tempo, frequently updated software can never disrupt the certified core.
• Deterministic resource allocation: Guaranteeing time and compute for ­mission-critical loops regardless of the AI load, across heterogeneous silicon – Intel, Arm, NVIDIA, AMD – without a complete system redesign.

Example of operationalizing AI at the UAS edge

In the unmanned domain, the General Atomics Gray Eagle Extended Range (GE-ER) uncrewed aerial system (UAS) provides an example of the modular open systems approach (MOSA) strategy. General Atomics sought to integrate advanced video encoding and AI-driven analytics on modern systems on chip (SoCs). By leveraging a lightweight execution layer – unikernel-based software substrate – it achieved secure partitioning and true mixed criticality.

The compute-heavy video encoding is isolated from the flight-critical networking and control functions. This structure enabled General Atomics to innovate at the speed of software, updating AI models without having to recertify the entire flight-control system. This capability is vital for the JADC2 vision of distributed, autonomous nodes.

The Golden Dome and zero trust at the workload layer

Nowhere are these requirements more stringent than in programs like the Golden Dome, a proposed distributed sensor network and integrated missile defense system whose nodes must operate under constant cyber adversity and across multiple security domains. This is exactly where trusted, autonomous mesh communication and multilevel security domain isolation become decisive: Every node has to defend itself, prove its own integrity, and keep operating even when neighboring nodes are jammed, lost or compromised.

To accomplish this, the software substrate can implement zero trust at the workload layer, which means that it leverages least-privilege enforcement, under which AI workloads should have zero access to the memory of other partitions unless authorized; and attestation, which means that the system must be able to prove its integrity to other nodes in the CJADC2 network in real time.

Certifiable by design

Certification has historically been what kept AI in the lab rather than the field, because standard frameworks like PyTorch – an open-source deep-learning library – are hard to certify when inference is not deterministic. The answer is not to certify the AI itself, but to isolate and bound it. Placing each AI workload in its own partition – surrounded by small, certifiable monitors that check its outputs against defined safety limits – ensures that a runaway model can never reach a safety-critical function. It also produces the evidence trail needed to field AI-enabled systems across both CPU and GPU workloads.

Vendor neutrality needed

CJADC2 will likely fail if it’s built on proprietary black-box systems. To succeed, the industry must embrace vendor neutrality through approaches like the Sensor Open Systems Architecture), or SOSA, Technical Standard; the Future Airborne Capability Environment, or FACE, Technical Standard; and the MOSA strategy.

The goal is to provide an underlying architecture that is a deterministic, safe, and secure substrate that enables any vendor’s AI to run on any vendor’s hardware with the safety and security required for modern conflict. As we move through 2026 and beyond, the winner of the CJADC2 era will not be the side with the most data, but the side that can most reliably transform that data into action at the tactical edge.

Michel Genard is chief strategy officer and board member at Lynx Software Technologies, where he leads strategy for safety-critical and mixed-criticality computing platforms. He has over two decades of experience in embedded systems, avionics, and defense software architecture.

Lynx Software Technologies • https://www.lynx.com/