A data centre can buy racks, servers and cooling equipment, yet remain unable to sell a dependable megawatt. In Russia's 2026 market, the decisive constraint moved upstream to grid connection, engineering capacity and the ability to operate power, cooling, physical protection and cyber recovery as one system. The location map consequently became an operating model rather than a real-estate choice.

Demand did not guarantee a connection

On 11 August 2026, RBC Trends reported that energy and engineering infrastructure had become binding constraints for Russian data-centre development. It said Moscow grid companies had begun rejecting some new connection requests in early 2026 because capacity was unavailable.

The qualification matters. The report did not say that every Moscow project was rejected or that the city had stopped building data centres. It identified a gating risk: a developer could have demand, land and financing while the required electrical connection was not available on an investable schedule.

That changes development logic. A rack is a purchasable object; a firm connection requires network studies, reserved capacity, substations, routes, permits, protection settings and coordinated commissioning. The commercial product is therefore not floor area or installed cabinets. It is reliable, cooled and secure computing capacity at an agreed date.

The annual rack count revealed a slower conversion

The source cited about 4,700 new rack places in 2025 and growth of 5.8%. It also noted a CNews estimate close to 5,300. Both estimates were nearly three times below the roughly 11,000 racks added in 2024. Different methodologies explain why one should preserve both figures rather than merge them into false precision.

A lower addition does not automatically prove weak digital demand. It can reflect delayed connections, expensive capital, higher construction and engineering costs, sanctions-related procurement friction, commissioning delays or a shortage of suitable powered sites. The reported high interest rate compounded those barriers and stretched estimated payback to eight to ten years.

For an investor, the useful question is not how many racks were announced. It is how many reached energised, tested and contracted service. A project can pass through land control, design, equipment ordering, mechanical completion and electrical readiness without producing a billable kilowatt. Each stage needs its own probability and cash-flow date.

A megawatt became the unit of strategy

Conventional enterprise racks may consume roughly 5 to 10 kilowatts, while artificial-intelligence racks can require 30 to 60 kilowatts and sometimes more than 100. These are workload ranges, not universal specifications, but they demonstrate why cabinet counts alone increasingly mislead.

At higher density, the same white-space plan can demand several times more utility power, distribution copper, conversion equipment and heat rejection. Floor area may remain available while the electrical and thermal envelope is exhausted. Sales must therefore qualify the intended load profile before promising a position.

Capacity planning should distinguish connected power, usable information-technology load, reserved redundancy, measured peak, contracted customer load and actual coincident demand. Treating these as interchangeable invites overselling. A facility with impressive nameplate input can still have a much smaller dependable customer envelope after losses and resilience margins.

A power-first investment gate

  • Confirm the connection point, capacity, voltage, route, cost and binding delivery milestones.
  • Model staged load growth instead of assuming full demand on the opening date.
  • Reserve space and fault duty for transformers, switchgear and future density.
  • Define redundancy without placing both paths inside one physical failure domain.
  • Test the financial case against delayed energisation, lower utilisation and higher energy cost.
  • Link customer commitments to demonstrable commissioning gates rather than construction headlines.
A bright physical relief map of Russia contrasts dense constrained power and data-centre infrastructure near Moscow with several smaller regional campuses across the Volga region, Urals and Siberia
Regional distribution offered energy options, but every location still had to reconcile power, connectivity, customers and operating capability.

Regionalisation exchanged one bottleneck for a portfolio

The report described interest shifting towards Yekaterinburg, Kazan, Novosibirsk, Samara, Nizhny Novgorod and Rostov-on-Don. Energy access may be easier in some regional markets, especially near major generation, but digital demand, carrier diversity, specialist labour and customer proximity may be less concentrated than around Moscow.

The answer is not a simplistic migration from one city to another. A regional portfolio can separate failure domains, serve latency-sensitive users locally and match computing loads with available generation. It can also fragment operations and leave capacity underused if customer demand was inferred from national totals rather than contracted locally.

Site scoring should combine connection certainty, electricity price structure, fibre routes, carrier independence, flood and climate exposure, logistics, staffing, security response and expansion rights. Weightings must follow the workload. A disaster-recovery site, public cloud region and latency-sensitive transaction platform do not value distance in the same way.

Policy proposals to place facilities nearer nuclear or hydroelectric generation recognise the power constraint. Proximity alone is insufficient: transmission topology, substation capacity, maintenance outages and delivery routes still determine dependable service. Generation on a regional map is not identical to an available firm connection at a plot boundary.

Reservation without utilisation could strand capital

A developer wants enough headroom for customer growth; the power system cannot reserve scarce capacity indefinitely for speculative demand. This creates a portfolio discipline problem. Projects need staged connection rights and credible load ramps so that scarce infrastructure is neither unavailable to others nor installed far ahead of revenue.

Utilisation should be measured at several layers. Contracted kilowatts indicate commercial coverage, metered computing load indicates real use, and site input captures cooling and conversion overhead. A healthy contract book can coexist with low physical load during customer deployment, while a heavily loaded hall may have poor margin under the wrong tariff.

The commonly used power-usage-effectiveness ratio compares total facility energy with computing-equipment energy. It is useful when boundaries and measurement periods are consistent, but it should not become a decorative single number. Climate, redundancy state, partial loading and measurement scope can change the result without changing service quality.

Cooling architecture followed density and climate

Traditional air cooling can remain appropriate for moderate rack densities if containment, airflow and control are sound. Higher-density accelerators may require rear-door exchangers or direct liquid cooling. The decision affects server compatibility, water loops, leak detection, maintenance skills, floor loading and recovery procedures.

A cooling plant should be designed around realistic weather bins and load stages, not only a nominal peak. At partial load, oversized equipment may cycle or operate inefficiently. At extreme conditions, simultaneous equipment failure and maintenance must not push temperature beyond the recoverable envelope.

Water strategy also belongs in site selection. Evaporative systems can reduce electricity in suitable conditions but increase water dependence; dry systems reduce water use while changing electrical demand and hot-weather performance. Operators need a local water balance, quality plan and contingency rather than a universal technology preference.

Heat reuse can improve system value where a nearby customer needs a compatible temperature and schedule. It is not free efficiency. Heat pumps, pipes, contracts, seasonal mismatch and backup arrangements add capital and operating duties. The computing service must remain safe when the heat customer is unavailable.

Redundancy had to survive real failure domains

Two electrical paths drawn on a diagram are not independent if they cross one room, share protection logic, rely on one fuel delivery route or are maintained by one mistaken procedure. Resilience reviews should trace physical and organisational common causes from grid intake to the server power supply.

Uninterruptible power supplies bridge disturbances and generator start, but batteries age and performance depends on temperature, state of charge and maintenance. Generators require fuel quality, starting systems, exhaust, cooling and load acceptance. A monthly no-load start does not demonstrate a sustained building load.

Commissioning must test integrated sequences: loss of one utility feed, transfer under load, failed generator start, cooling restart, control-network loss and restoration to normal. The aim is not spectacle. It is to expose timing conflicts and hidden dependencies before a customer incident does.

A black-building or full-load test carries risk and requires careful boundaries, but representative failure injection is essential. Test evidence should record expected sequence, measured response, deviations, ownership and retest. A passed component certificate cannot substitute for the behaviour of the assembled facility.

Captive capacity and colocation carried different economics

The source cited a Cloud X estimate that 80% to 82% of Russian information-technology capacity remained in captive corporate data centres. It contrasted this with a worldwide captive share said to have fallen from 58% in 2020 to 43% in 2025. Definitions and market coverage should accompany any direct comparison.

A captive site can give an enterprise direct control, familiar governance and proximity to legacy systems. It can also trap capital in small facilities that lack scale in procurement, staffing, carrier choice and modernisation. Colocation aggregates demand, but customers must understand what the operator controls and what remains theirs.

International estimates cited by the report placed Tier III construction around 10 to 12 million dollars per megawatt of computing load. About half could relate to power and reserve infrastructure, with cooling representing 15% to 25%. Those are orientation ranges, not a Russian project quotation.

Currency, equipment origin, import routes, local standards, ground conditions, density and scope can move a project materially. A useful benchmark normalises the denominator and exclusions: land, tax, network connection, fit-out, financing, spare parts and commissioning may or may not be included.

Building and server clocks did not match

A data-centre shell and utility plant may operate for decades, while computing generations turn much faster. Designing only for today's rack can create an early ceiling; designing every hall for an uncertain extreme can waste capital. Modular mechanical and electrical blocks offer a way to stage both risk and density.

Interfaces are the key. Reserved pipe routes, switchboard positions, floor capacity, ceiling clearance, control addressing and isolation points determine whether a future module can be added safely. Expansion that requires shutting down the live facility defeats much of the intended optionality.

Configuration control must cover more than servers. Protection settings, firmware, cooling set points, valve positions, alarm routing and access permissions all describe the operating facility. Unrecorded changes erode the meaning of commissioning evidence and make incident diagnosis slower.

Physical security was an engineering subsystem

Perimeter barriers, vehicle controls, mantraps, cameras, thermal detection and restricted rooms work only when integrated with procedures and response. More sensors do not automatically improve protection. Alarm quality, coverage, retention, staffing and authority determine whether an anomaly becomes action.

The source described closer integration of physical security with data-centre infrastructure management, access, climate and fire systems. Integration can accelerate detection, yet it also creates dependencies. A compromised management network must not quietly disable doors, suppress alarms or alter environmental control.

Zones should follow consequence. Visitors, contractors, delivery teams, facility engineers and computing administrators require different routes and permissions. Temporary access needs expiry, escort rules and review. Emergency egress must remain safe without creating an uncontrolled path into protected areas.

Cyber and operational technology shared the same incident

Modern facilities connect building management, power monitoring, cooling controls, access systems, cameras and customer networks. Segmentation should prevent a compromise in one domain from becoming unrestricted movement across the building. Necessary data exchange can pass through controlled, monitored interfaces.

Administrative access deserves strong identity, multifactor authentication, dedicated management paths and recorded sessions where appropriate. Vendor support should be time-limited and explicitly authorised. A permanent shared maintenance account turns convenience into an untraceable resilience risk.

Supply-chain attacks make provenance and update practice operational issues. Operators need inventories of hardware, software, firmware and support status; verified update sources; testing before deployment; and a response for components that can no longer be trusted or maintained.

Patch urgency must be balanced against uptime, but deferral cannot mean silence. Risk acceptance should identify exposure, compensating control, owner and deadline. A maintenance window is a resilience tool when rollback, spare capacity and communication are prepared.

One recovery evidence pack

  1. Map critical services to power, cooling, network, identity and supplier dependencies.
  2. Maintain three copies of important data on two media, with one isolated from the primary environment.
  3. Protect at least one copy against ordinary administrator alteration and destructive automation.
  4. Restore representative systems, not just files, within measured recovery objectives.
  5. Exercise facility and cyber teams together with named decision authority.
  6. Close findings through owned actions, deadlines and repeat tests.
A bright cutaway data centre separates utility and backup power, cooling machinery, secure server aisles and an isolated recovery vault as connected resilience layers
A resilient rack inherited the condition of every supporting layer, including the procedures that restored service after failure.

Backups mattered only when restoration worked

The three-two-one pattern is a useful baseline: three copies, two media types and one copy offsite or otherwise isolated. Modern destructive attacks add a stronger requirement for immutability or separation from ordinary administrative credentials. Replication alone can faithfully copy corruption.

Recovery objectives should be set per service. The permitted data-loss interval and target restoration time determine replication, backup frequency, network capacity, staffing and cost. Declaring one aggressive target for every system usually hides which services have never been tested at that level.

A restoration exercise should rebuild identity, configuration, application, data and connectivity in the correct order. It should also test access to keys, documentation and specialist people. Success means users can perform an agreed transaction, not merely that a backup console displayed a green status.

Responsibility remained shared but not vague

A colocation operator may provide building, power, cooling and physical controls while the customer manages operating systems, applications, identities and data. Cloud and managed-service layers divide the stack differently. Every contract needs an explicit responsibility matrix linked to incident communications and evidence rights.

Suppliers also occupy the resilience boundary. Generator maintenance, fuel, batteries, carriers, security systems and specialist cooling can each become a single operational dependency. Service-level language should connect to spare strategy, escalation, access and tested contingency.

Insurance, certification and audit can support governance, but they do not operate the building. Leaders need current risk registers, capacity views, incident trends, overdue maintenance, failed recovery tests and exceptions that expose the real condition of the service.

People connected the layers

Staffing plans must cover nights, holidays, illness and simultaneous incidents. A named expert who lives far away is not the same as on-shift capability. Training should pair technical knowledge with authority to stop unsafe work, isolate equipment and escalate customer impact.

Runbooks need observable triggers, safe steps, decision points and rollback. They should be used during exercises and routine maintenance so that inaccuracies surface before an emergency. Shift handovers must transfer active alarms, impairments, permits and unusual configurations.

Change control is especially important where facility and information-technology teams meet. A server deployment can alter heat density; a cooling optimisation can change hardware inlet conditions; a firewall rule can block monitoring. Cross-domain review prevents each team from optimising its own metric at the expense of service.

The resilience contour became the product

For data-centre developers and customers in Russia, the 2026 constraint was not a shortage of metal cabinets. It was the ability to secure power, convert it efficiently into dependable computing and recover the complete service when one supporting layer failed.

The stronger location decision begins with a firm connection and a realistic load ramp, then tests networks, cooling, physical protection, cyber segregation, suppliers and people as one contour. Regional sites can diversify that contour, but only when demand and operating evidence travel with the investment.

That is why the market map changed. Energy availability redirected projects; higher rack density magnified engineering requirements; capital cost punished idle capacity; and converged threats made isolated controls obsolete. The winning facility will not be the one with the most announced racks. It will be the one that can prove each contracted kilowatt remains usable through change, failure and restoration.