Interoperability is not a feature – it’s a foundation

Story

August 11, 2026

The architecture surrounding edge artificial intelligence (AI) – including sensors, networks, hardware, software stacks and orchestration layers – is still being built in silos. Vendors build closed ecosystems, contracts reward vertical integration, and the warfighter ends up with tools that only function when everything lines up perfectly: the specific connectivity in the specific configuration using specific software, running on specific hardware. That perfect situation is not the reality at the tactical edge, however. Realistically, the edge in this case is often degraded, denied, intermittent, and limited. If one link in the chain requires a cloud connection, the whole chain breaks.

Artificial intelligence (AI) has arrived at the tactical edge: Models are being deployed on vehicles, worn by soldiers, and embedded in autonomous platforms across diverse contested environments. The question is no longer whether AI can operate at the edge; it is whether the systems around AI are built to let it actually work. The honest answer today to the question of whether AI works at the edge is often muddied.

What real interoperability looks like

When defense leaders talk about interoperability, the conversation often stops at compliance: modular open systems approach (MOSA) checklists, application programming interface (API) documentation, and open standards on paper. Standards compliance is not the same as operational interoperability, however. Real interoperability means that a unit with a different sensor suite than the one beside it can still share the data, run inference, and produce a common operating picture without a week of custom integration work.

This situation requires a bidirectional flow of data: from sensor to inference layer, from inference layer to operator output, and back again. Every layer of the system must support that loop while disconnected from cloud infrastructure and under operational constraints. Today, that loop breaks repeatedly, as models fail to translate across hardware configurations, orchestration platforms have no visibility into the actual available compute, or updates are pushed without accounting for what else is already running on the device. These scenarios are not exotic edge cases, but are actually the daily friction of deployed AI.

The orchestration gap in DDIL environments

Compute resources taper as forces advance: A forward operating base has meaningful bandwidth and storage, a vehicle has less than the base, and a soldier-worn device has still fewer resources. What the community has not solved is how to make AI systems adapt gracefully to that narrowing funnel. The problem is orchestration, not hardware. A system needs to know what resources are actually available at any given moment, accounting for what else is running, what the network looks like, and what the mission demands. Without that capability, every new mission readdresses the same problems from scratch for every new deployment: new scripts, new operating system images, new containers, new configurations. Cost rises, time-lines stretch, and the warfighter waits.

The last tactical mile is where this friction is most acute. Updating a model once it is forward-deployed requires knowing whether the platform can run the new version, managing the push while disconnected, and validating performance without the ability to troubleshoot in the field. When AI capabilities, hardware, and software layers are each developed with different assumptions about the rest of the stack, the result is systems that work in isolation and fail during integration. Every closed integration limits future capability. An open, modular architecture is the only approach that preserves the ability to plug in something new without discarding what was already built.

Orchestration is not optional

Edge AI is frequently framed as a deployment challenge. That framing misses most of what actually determines long-term mission success, however. Deployment is not a destination; it is the start of a system’s exposure to conditions it was never fully tested against. The ability to update, retrain, and push a new model version to the right platform at the right time is in fact life cycle management, and it is as operationally critical as the original deployment. Orchestration is the layer that makes that possible at scale, determining what gets pushed where, when, and in what configuration. Without it, every update is a manual operation, every new deployment is a custom integration, and the tempo of change in the field will always outpace the tempo of delivery from the rear.

No single vendor can own the sensor, the network, the compute platform, the orchestration layer, and the AI model simultaneously. Attempts to do so produce the closed ecosystems that have historically stymied true interoperability. The more productive model is an ecosystem of complementary capabilities, each built with open interfaces, where success is measured not by owning the majority of the stack but by how quickly other components can work alongside it.

The business-model argument matters as much as the technical one. Large, vertically integrated contractors are structurally incentivized to maintain closed systems. The government cannot simultaneously demand interoperability and award decade-long contracts to organizations incentivized to prevent it. Commitment to the warfighter and commitment to the stock price are not the same thing, and right now, the market does not reward the former.

The architectural decisions that will determine what scales

The choices being made now in programs, contracts, and vendor selection will determine whether edge AI scales across platforms and services over the next decade. A few principles should govern them:

Design for modularity from the start: Systems that stratify their architecture correctly can swap diverse components without rebuilding from scratch. The goal is an ecosystem in which new capabilities connect because the interfaces were designed to be open, not because every component came from the same vendor.
Treat disconnected operation as the baseline, not a special case: Every layer of the system must function without a cloud connection. AI that requires connectivity at the tactical edge is not edge AI; it is cloud AI with a latency problem.
Build the orchestration layer before you think you need it: Orchestration is the foundation on which life cycle management, resource allocation, and cross-platform deployment are built. Organizations that treat it as an afterthought will rebuild it at significant cost and extended timelines when operational scale demands it.
Demand open interfaces in procurement: Requiring exposed APIs and published software-development kits as contract conditions and evaluating vendors on how well they integrate with others would shift the market.

Building the foundation

The edge AI systems fielded today will define the operational baseline for the next decade. The systems that absorb change without requiring complete rebuilds are the ones built on open, modular, orchestration-capable architectures from the start. Interoperability is not a feature to be added later. It is a foundation that must be designed deliberately by defense leaders, procurement officials, and industry partners willing to build differently and hold one another accountable.

The warfighter at the tactical edge does not care which vendor’s stack is running beneath the interface. They care whether it works when the network is down, whether it updates when the mission changes, and whether the tools available today will still be relevant when the adversary adapts tomorrow. That is the standard; everything else is a means to meeting it.

Dr. Sek Chai is a co-founder and CTO for Latent AI. A 20-year technology veteran, Sek co-developed the AI tech underlying Latent AI products. Michael MacFadden, CTO for Sigma Defense, has more than 20 years of experience in defense; at Sigma Defense he is responsible for seeking out new technologies across multiple technology domains. James Beere, vice president, technology & innovation at Sigma Defense, is a 26-year veteran of U.S. Special Operations. At Sigma Defense he is responsible for R&D in C5ISR, JADC2, SATCOM, and DEVSECOPS.

Latent AI • https://latentai.com/

Sigma Defense • https://sigmadefense.com/