Artificial Intelligence has become the new buzzword in military technology. Almost every new weapon system, drone, robot, or battlefield system is now described as “AI-enabled” or “AI-powered”. The impression is that AI will fundamentally transform warfare and that autonomous machines may soon be capable of replacing human decision-making across the battlefield.

There is certainly enormous potential in AI. But there is also a tendency to overlook a basic fact:

AI is not Weightless

AI requires computing power. Computing power requires processors, memory, electricity and, increasingly, cooling. When this capability is brought from a data centre to a battlefield platform, it has to be physically carried. This creates very real limitations of size, weight, power and cost—what is increasingly referred to as the SWaP-C constraint: Size, Weight, Power and Cost.

This is particularly relevant to drones and robots.

From Automation to AI

It is important to distinguish automation from AI. Automation can sense, calculate and act without being AI. The sophistication of the calculation, or the speed at which it is performed, does not by itself make a system AI.

In our own work, we used electronic viscosity controllers to maintain the consistency of ink in printing processes. Sensors continually monitored the viscosity of the ink and, when required, automatically added ink or thinner to maintain the desired consistency according to a pre-programmed table. There was obviously no AI involved. It was simply basic automation: sensing, comparing the measured value with a predetermined parameter, and taking a programmed corrective action.

A somewhat more advanced example was the automation of artillery fire data. Range-table data were curve-fitted to obtain the required firing data, which were then combined with pre-fed meteorological information to determine the required trajectory for the projectile. The relevant meteorological parameters were measured independently and fed into the system before firing. The projectile itself did not sense altitude or changing meteorological conditions during flight and altered its trajectory accordingly.

These examples illustrate the progression from basic automation to more sophisticated automated systems. In both cases, however, the system operated within defined parameters and processes.

Processing information, performing complex calculations and taking an automated action do not, by themselves, make a system AI.

This distinction is sometimes lost in the enthusiasm surrounding autonomous weapons.

A drone, for example, can have autonomous navigation, obstacle avoidance, stabilisation, target tracking or pre-programmed responses without requiring anything approaching the computing capability associated with large AI models. Some of these functions can be performed with relatively modest processing power.

Onboard AI

AI introduces another level of capability. It can be used for image recognition, classification, interpretation of sensor data, identification of patterns and other tasks that go beyond straightforward programmed automation. But the sophistication of such AI is constrained by the computing resources available to the platform.

Consider the simple problem of putting AI into a small drone. A current edge-AI processor such as NVIDIA’s Jetson Orin NX can provide up to around 100 TOPS of AI performance while operating within a configurable 10–25 watt power envelope. A larger AGX Orin can provide substantially greater performance, but its power requirement can rise to around 60 watts. These are impressive capabilities for compact computing hardware, but they remain a very different proposition from the enormous computing resources available in a data centre.

And those watts do not exist in isolation.

A drone has to generate the power required for propulsion. It has sensors, communications equipment, navigation systems and a mission payload. Every additional watt devoted to computing competes with the energy available for these functions. Additional computing also produces heat, which has to be dissipated. Cooling adds further weight, size and power requirements.

Thus, the AI capability that can actually be carried by a drone or robot is constrained by the platform itself.

This does not mean that AI cannot operate onboard. It certainly can. Small AI models can be optimised, compressed and quantised, while specialised AI accelerators can provide substantial inference capability within a relatively small power budget. But there remains a fundamental difference between running a limited AI model at the edge and having access to the computing power of a large data centre.

AI Training

This brings us to another important distinction: AI training and AI inference are not the same thing.

Training sophisticated AI models can require enormous computing resources and large data centres. Once trained, however, a model can be deployed for inference—the actual process of using the model to recognise, classify or interpret new data. That inference can take place onboard a drone, robot or weapon if the platform has sufficient computing capability.

Alternatively, the data can be transmitted elsewhere, processed on a much more powerful computer or data centre, and the result sent back to the battlefield system.

The latter architecture has obvious advantages. Heavy computing does not have to be carried by every drone, robot or weapon. A central or remote computing resource can be shared by many battlefield systems. Computing hardware can also be upgraded without necessarily replacing every weapon platform.

Networked for AI

But this introduces another dependency - Communications.

If a battlefield platform depends on a remote computer for AI processing, it depends on the communications network connecting it to that computer. That network requires bandwidth, connectivity and resilience. In a contested electromagnetic environment, communications may be jammed, disrupted or degraded.

A system that is autonomous only while connected to a remote computing resource is therefore not autonomous in quite the same sense as a system capable of performing the required inference onboard.

This is where the military requirement becomes particularly interesting.

A battlefield is not a laboratory or a commercial cloud-computing environment. It is an environment in which communications can be interrupted, sensors can be degraded, satellites can be denied, the electromagnetic spectrum can be contested, and systems can be attacked.

The more a weapon depends on an external computing architecture, the more important the resilience of that architecture becomes.

At the same time, putting sophisticated AI computing into every weapon creates another problem:

Costs of AI Computing

If AI processing is incorporated into an individual weapon, the weapon has to carry its processor, memory, power supply and associated electronics. Depending on the sophistication of the AI, there may also be requirements for thermal management and additional sensors.

This may not matter greatly for an expensive, reusable platform. It matters considerably more when the AI-enabled object is an expendable or attritable weapon.

There is an important economic difference between having one powerful AI computer serving many weapons and putting substantial AI computing into every individual weapon.

In the first case, the expensive computing resource can be shared. In the second, the cost of that computing capability becomes part of the cost of every weapon.

This does not necessarily mean putting a large, expensive GPU into every missile or munition. Modern specialised AI accelerators can perform inference efficiently, and AI models can be reduced in size to suit the available hardware.

But there remains a fundamental engineering trade-off: The greater the AI capability demanded at the edge, the greater the requirements for computing, power, cooling, space and cost.

Data Centre is not Battlefield

This raises a larger question about the current enthusiasm for AI-enabled autonomous weapons.

Are we sometimes confusing the potential capability of AI in a data centre with the capability that can actually be delivered by a battlefield weapon?

A data centre can provide enormous computing power because it does not have to fly, drive, survive enemy fire or carry a warhead. It has access to electricity, cooling, networking and large quantities of computing hardware.

A small drone has none of those luxuries.

The difference is fundamental.

AI can certainly enhance drones and robots. It can improve reconnaissance, target acquisition, image interpretation, navigation, sensor fusion and decision support. It can enable machines to perform tasks that would otherwise require continuous human intervention.

But a battlefield system has only two choices: it must either carry the computing capability required for those tasks or remain connected to a system that provides it.

There is therefore no such thing as unlimited AI capability at the tactical edge.

The physical world imposes constraints that do not disappear simply because the processing is called Artificial Intelligence.

Computing power has weight. Computing power consumes energy. Computing power generates heat. Computing power costs money.

And when that computing has to be distributed across thousands of drones, robots, missiles or other weapons, the economic consequences can become significant.

This does not diminish the importance of AI in warfare. Quite the opposite. It suggests that the military value of AI will depend not merely on how powerful the underlying model is, but on where the processing takes place, how much computing power is required, how resilient the communications architecture is, and what SWaP-C penalty results.

The military system of the future is therefore unlikely to be a simple choice between “AI” and “no AI”. It will involve a combination of automation, onboard AI, remote computing, human decision-making and communications networks.

The important question is not whether AI can perform a particular task in a controlled demonstration. It is whether that capability can be delivered reliably, economically and continuously under the physical conditions of war.

The AI revolution may indeed change warfare. But the laws of physics, engineering and economics have not been repealed. AI may be extraordinarily powerful in the data centre. The real challenge is carrying that capability to the battlefield.

(The author is an Indian Army veteran and a contemporary affairs commentator. The views are personal. He can be reached at  kl.viswanathan@gmail.com )