The defence of data in the age of AI

The defence community is late to recognise that the data centre is now both weapon and target. For most of the last two decades it was treated as logistics — a building full of servers, a target of opportunity rather than a centre of gravity. That framing is now obsolete. The data centre is where the cognitive capacity of the state is stored. Damage it, and you damage the state's capacity to think.

A hyperscale region under contest is not an inconvenience. It is a partial lobotomy of the global cognitive system. When a single availability zone goes dark, the effect is not localised. Models in production around the world degrade. Agents in mid-task stall. The compounding inference loops that businesses have come to rely on unwind in unpredictable ways. The blast radius is no longer geographic.

Three categories of attack have already been demonstrated in open literature. The first is physical: cooling sabotage, power interdiction, fibre cuts. The second is logical: poisoning of training data, exfiltration of weights, model inversion. The third is hybrid: a physical event whose forensic trail is obscured by a logical one. The third category is the one that keeps serious people awake.

Defence in this domain has to be vertical. The perimeter is no longer a fence. It is a stack — from the substation to the cooling loop to the silicon to the firmware to the weights to the inference API to the agent to the user. Every layer is a potential attack surface. Every layer is governed by a different vendor, a different regulator, a different jurisdiction. The integration is the vulnerability.

There is a strategic question underneath the tactical one. A state whose cognitive infrastructure is hosted in another state's hyperscale region has, in a meaningful sense, outsourced its sovereignty. The hosting state can, at any moment, throttle, audit, or terminate the relationship. The hosted state has, at any moment, the option of being suddenly less intelligent. Few governments have done the math on what that option is worth.

The defence of data must become a doctrine, not a department. A doctrine names what is to be defended, against whom, at what cost, and with what acceptance of risk. It survives changes of administration. It informs procurement. It shapes the curriculum of the war colleges. The current posture in most countries is the opposite — a department, a budget line, a quarterly report. That posture will not survive contact with a serious adversary.

For small states the doctrine is necessarily different. We cannot defend hyperscale because we do not own it. What we can defend is the layer above — the application, the agent, the audit trail. We can also defend the layer below — the cable landing station, the submarine repeater, the cross-border interconnect. Between those two layers is the work of the next decade.