Scale or separation? Data centre resilience: lessons from wartime
The question
Data centres have become targets in war. In the week to 14 October 2026, Ukrainian drones reportedly struck three data centres operated by Yandex, followed by Sberbank’s site at Balakovo, described as Russia’s largest. Russia has in turn struck data centres inside Ukraine, and earlier this year Iranian strikes reportedly hit data centres in Bahrain and the United Arab Emirates (UAE).
Some commentators read the strikes as exposing the concentration of Russia’s data centre capacity. That may be so. But the strikes also raise a broader question, and one that applies well beyond a war zone:
‘Do the scale benefits of large campus data centres outweigh the resilience benefits of a more distributed estate?’
This article does not attempt an answer. It sets out what each approach offers, what the strikes do and do not tell us, and the questions that need answering before anyone can reach a view.
What does scale buy?
A large campus spreads its fixed costs: one grid connection, one security perimeter, one operations team. That makes redundancy within the site cheaper per megawatt (MW). Duplicated power and cooling paths, standby generation and a dedicated substation are all easier to justify at scale.
Against random, independent failures, such as a failed transformer or a cooling fault, this approach has a long operating record. A large operator can also afford hardened construction and on-site security that a small site may not.
Some workloads also need scale within a site. Training large artificial intelligence (AI) models typically uses large clusters of processors joined by very low-latency links inside a facility. Such training does not, however, generally need to be close to its users, so the cluster can in principle be located wherever power and space allow.
And scale does not preclude redundancy elsewhere. After the Balakovo strike, Sberbank said its systems and services were operating normally. That is the operator’s own statement and cannot be verified independently. But if it is right, it suggests a large facility can be lost without loss of service, provided the wider architecture allows for it. If the original architecture was 1+1, with each service duplicated in a second facility, losing one facility leaves no redundancy until it is restored, and the service is exposed to a second event.
What does separation buy?
The buildings on a campus typically share their dependencies: the grid connection, the water supply, the fibre routes into the site and the access roads. An event that affects one shared dependency can therefore affect every building at once. Engineers call this ‘common-mode’ failure.
Smaller, separated sites reduce the exposure of a service to any single event. Each site is also a less valuable target, whether for an adversary or simply in terms of what is lost.
The Yandex strikes illustrate the point. The Sasovo site was reportedly one of the company’s five large data centres, hosting tens of thousands of servers. Yandex suspended operations there and said it could not confirm whether the equipment could be restored. The sites struck earlier reportedly hosted all three of Yandex’s operating supercomputers.
But separation has its own costs. More sites mean more interconnect, duplicated operations and a smaller security budget per site. Distance also limits synchronous data replication, which in practice constrains how far apart two sites can be while still acting as one. Spread capacity too thinly and cost and complexity rise; concentrate it too far and a single event can do disproportionate damage.
What do the strikes tell us?
Less than it might seem, at least for now. The reports are days old, and most claims about damage and service impact come from the parties themselves. Wartime physical attack is also a particular threat, and conclusions drawn from it do not transfer directly to countries that are not at war.
Perhaps the more general lesson is that physical attack is one of several common-mode risks. Others include grid failure, extreme weather, flooding and water shortage. Each can affect a campus as a whole.
The power outage across the Iberian Peninsula in April 2025 is a peacetime example. A grid-wide failure affects every site that depends on that grid, however many sites there are. On-site generation can protect against it, but brings a dependency of its own on a single fuel and supply route. The more useful question may be how many independent classes of supply a site can draw upon.
Get Andrew White’s stories in your inbox
Join Medium for free to get updates from this writer.
Interestingly, Ukraine’s own reported response has been to harden rather than disperse: its digital minister has said the country is moving digital infrastructure underground. That suggests the choice is not simply between scale and separation. Many large operators already combine the two, running several large sites far enough apart to fail independently.
Which questions need answering?
Before one can weigh scale against separation, it seems to me that six questions need answers:
- Resilient against what? Random equipment failure, grid loss, weather, water shortage, cyber attack and physical attack each favour a different design;
- How many independent failure domains does a given service actually have, and what do they still share?
- How far apart must two sites be for a single event to affect only one, and does latency allow that separation for the workloads concerned?
- What proportion of a region’s, or a country’s, capacity sits on a single campus or in a single area?
- Who bears the cost of resilience: the operator, its customers, or the wider public through shared grid and water infrastructure? And
- Can operators and their customers find out how resilient the links connecting their sites are, and what those links share? I raised this question in a July 2025 submission to the European Commission.
In summary
Scale and separation each buy something real. Scale buys efficient redundancy within a site and the capacity some workloads demand. Separation buys protection against events that strike everything sharing a dependency.
The strikes in Russia do not settle which matters more, and nor should they be expected to. The answer is likely to depend on the threat, the workload and how the wider estate is designed. But the questions above deserve explicit answers rather than assumptions.
I would welcome views from those who design, operate or depend on this infrastructure.
Disclosure: I am Group CEO of Wray Castle, a telecommunications training organisation, and have given expert evidence on a data centre development proposal. The views here are my own.
Sources
- Reuters, ‘Ukrainian drones hit Russia’s top data center in attack on AI ambitions’, republished by Insurance Journal, 14 October 2026.
- Reuters, same report, republished by Euronext, 14 October 2026.
- Al Jazeera, ‘Ukraine takes aim at Russia’s AI data infrastructure’, 9 October 2026.
- Semafor, ‘Data centers emerge as attractive targets in war’, 11 October 2026.
- Reuters, ‘Ukraine expands drone strikes on data centers owned by Russia’s Yandex’, republished by The Detroit News, 9 October 2026.
- CNBC, ‘Russia’s internet giant Yandex halts operations at major data center after drone attack’, 8 October 2026.
- Wray Castle Limited, Consultation Input: Digital Networks Act, feedback to the European Commission call for evidence, 11 July 2025, ref. F3575305 (registered Ares(2025)5711278, 14 July 2025), Have your say.
All sources accessed 15 October 2026.
Originally published at https://harberry.com on October 15, 2026.
