Europe is becoming a major hub for global data centre growth, driven by cloud computing, artificial intelligence (AI) and rapid digitalisation. Data centres are evolving from medium-sized facilities into large industrial-scale assets with electricity demand comparable to traditional heavy industries, creating growing pressure on the power transmission system.

According to the European Network of Transmission System Operators for Electricity’s (ENTSO-E) recent report, “Data centres and the power system: expected trends, challenges and opportunities” (May 2026), Europe currently has over 10,500 data centres with a combined IT power capacity of about 12.7 GW, including 9.9 GW in the EU27. Their electricity demand is expected to grow rapidly and become a major driver of power consumption over the next decade. Growth rates vary across countries, but regional data centre electricity demand is projected to rise by more than 50 per cent between 2025 and 2030. In aggregate terms, total demand is projected to rise from 87 TWh per year in 2024 to around 134 TWh by 2030, potentially reaching 199-254 TWh by 2035 depending on the scenario. Planned initiatives such as the EU Cloud and AI Development Act could further accelerate expansion by significantly increasing EU data centre capacity.

At the same time, AI and advanced data technologies offer major opportunities for the energy sector. The European Commission’s (EC) upcoming “Strategic Roadmap on Digitalisation and AI in the Energy Sector” aims to use these technologies to improve grid optimisation, demand flexibility and renewable integration, helping build a smarter and more resilient power system.

Structural shift in data centres: Fewer, larger, more concentrated

The sector is undergoing a structural shift that has direct implications for the transmission system. Three principal categories of data centres exist:

  • Colocation and service provider data centres lease space, power and connectivity to customers that operate their own IT equipment. Because facility operators control the infrastructure but not the workloads, these centres have limited flexibility from a grid perspective. Large colocation facilities above 50 MW are increasingly used by hyperscale tenants, blurring the line between data centre categories.
  • Hyperscale data centres are large facilities, often above 50 MW, operated by major technology companies to support cloud computing, web services and AI workloads using highly scalable infrastructure.
  • Enterprise data centres are self-owned facilities run by banks, government agencies and large companies for their own internal use; they are typically modest in scale (1 to 5 MW) and their capacity is largely stagnating across Europe.

The critical trend is that colocation, and in particular a category called “scale colocation” (facilities where large portions of capacity are leased to hyperscale tenants), is expected to drive more than 70 per cent of the European data centre IT load increase over the next five years. Scale colocation will grow at 30 per cent per year, while traditional retail and wholesale colocation will grow at around 8 per cent annually. By 2030, an estimated 60-70 per cent of colocation capacity is expected to be leased to hyperscalers, signalling a major structural shift in the market.

This means that growth is increasingly concentrated in very large individual facilities – campuses that can reach tens or even hundreds of MWs per site – clustered in established metropolitan hubs such as Frankfurt, London, Amsterdam, Paris and Dublin (collectively known as the FLAP-D cities). This geographic and scale concentration has direct consequences for the transmission system, exacerbating local capacity constraints and creating new types of dynamic stress on the grid.

How data centres consume electricity

Power architecture

A data centre’s electrical demand is split between two distinct loads: IT equipment (servers, storage, and networking) and cooling systems (removes the heat generated by the IT equipment). The IT equipment is financially valuable and runs mission-critical services, whereby even millisecond-level interruptions can cause data loss or service failure. To protect against this, the entire IT load is supplied through uninterruptible power supply (UPS) systems – power-electronic devices that continuously condition incoming electricity and can instantaneously switch to battery backup if the grid supply is disturbed.

Centralised UPS systems dominate even in new facilities, making the UPS the primary electrical interface between the data centre and the grid. That said, novel alternative architectures are gaining traction. Some hyperscalers are moving towards UPS-less architectures that rely entirely on facility redundancy rather than local UPS and backup generation. Further, some data centres use rack-level UPS systems instead of centralised ones, but these require more advanced control and coordination to provide similar grid support capabilities.

Cooling systems account for 30 per cent or more of total facility consumption in an average European data centre, which currently records a power usage effectiveness (PUE) ratio of around 1.5 (meaning that for every unit of power delivered to IT equipment, a further 0.5 units are consumed in cooling and other overhead). Advanced liquid and immersion cooling technologies are bringing this ratio down significantly, with leading hyperscalers reporting PUE values below 1.05 in their most efficient facilities. However, this also reduces the thermal inertia that could otherwise serve as a source of demand flexibility.

IT workload types

The nature of the IT workload running inside a facility is equally important. Key workload categories include:

  • AI model training: Runs continuously for weeks on clusters of graphics processing units (GPUs) and tensor processing units (TPUs) and generates sustained high-intensity consumption with rapid fluctuations of 30-60 per cent within milliseconds.
  • AI inference: Executes trained models in response to user requests. Its demand follows daily user activity patterns, making low latency critical. As a result, inference workloads are typically located near cities and hyperscale campuses, and are expected to become the dominant AI workload by 2030.
  • Web services, software-as-a-service (SaaS) and content delivery: Their demand follows daily user activity patterns and, despite being mainly CPU-based, they consume significant power due to their large scale.
  • Big data analytics: Large-scale analytics, reporting and data processing workloads that operate on completion deadlines rather than real-time response requirements.
  • Backup and disaster recovery: Systems that store replicated data at secondary sites, with power demand that can rapidly increase during failover events.
  • Transactional and real-time workloads: Applications that process continuous database operations and time-sensitive tasks requiring extremely fast response times.

A data centre’s specific workload mix, therefore, determines its electrical profile. A facility dominated by AI training will behave very differently from one serving banking transaction systems, even at similar total power consumption.

Risks to grid security and stability

The electrical characteristics of large data centres create risks that existing regulatory frameworks were not designed to address, because those frameworks were built on the assumption that large industrial loads are predictable and relatively steady.

Risks during normal operation

During normal grid operation, data centres can act as sources of disturbance in two distinct ways – load-driven disturbances and load-induced forced oscillations. First, rapid power fluctuations from AI workloads, where consumption can swing by tens of MWs within milliseconds, appear to the power system as frequent, unpredictable load changes. As the aggregate share of system demand from data centres grows, these effects become increasingly system-relevant, capable of causing voltage violations and frequency imbalances.

Second, persistent, periodic load patterns from AI training clusters can excite electromechanical modes of the power system at frequencies coinciding with natural oscillation modes, creating a risk of resonance amplification that propagates across wide areas. Critically, such sub-synchronous oscillations may be invisible to conventional supervisory control and data acquisition (SCADA) monitoring systems, requiring phasor measurement units (PMUs) with high-resolution sampling for detection. Incidents of this nature have already been observed in practice.

Risks during grid disturbances

During grid disturbances, the risks are even more acute. When a voltage dip occurs, UPS systems are designed to transition to battery mode, which implies they can instantaneously disconnect hundreds of MWs of IT load from the grid. When multiple large data centres in the same geographic area respond simultaneously to the same disturbance, a localised voltage incident can cascade into a system-wide frequency disturbance.

Reconnection behaviour after fault clearance is not standardised, creating a further risk. If UPS systems return large blocks of demand to the grid too rapidly, the sudden demand shock can depress frequency back towards disconnection thresholds, triggering repeated disconnect-reconnect cycles (known as “flapping”). Multiple incidents across different grid systems have already demonstrated load losses ranging from several hundred MWs to over 1 GW following routine transmission faults.

Connection requirements as a response

In response, several national grid code updates are being developed, including in Ireland and Belgium, with ENTSO-E recommending harmonisation across Europe. The key areas include:

  • Fault ride-through (FRT): Requiring data centres to remain connected during voltage dips rather than disconnecting at shallow thresholds.
  • Rate of change of frequency (RoCoF) withstand capability: Ensuring UPS and protection logic can tolerate rapid frequency changes without tripping.
  • Ramp rate control: Limiting the speed of load changes to prevent software-driven steps from outpacing grid frequency and voltage control systems.
  • Voltage control and reactive power: Requiring active voltage regulation at the connection point.
  • Oscillation damping: Tuning power electronics to suppress grid oscillations and prevent resonance risks.
  • Post-fault active power recovery: Ensuring controlled recovery of power consumption after fault clearance.

Grid planning challenges: The time-to-power problem

Beyond the technical stability risks, data centres also pose fundamental challenges for grid planning. Access to electricity has become the single largest concern for European data centre operators: 76 per cent cite it as their biggest challenge over the next three years, far ahead of regulatory compliance (36 per cent) or permitting (34 per cent). Connection lead times in Europe range from a few years to more than a decade in constrained regions like the FLAP-D cities, creating a “time-to-power gap”, which is the sector’s central growth obstacle.

Three complementary approaches can help close this gap without compromising system security:

Transparency on available hosting capacity

TSOs can steer site-selection decisions by publishing accessible, regularly updated information on where grid capacity is available, what connection timelines look like and how capacity evolves under different scenarios. Several European TSOs, including Elia in Belgium and Terna in Italy, are already implementing such platforms. The EU Grid Action Plan (November 2023) foresees a joint ENTSO-E and EU DSO Entity portal for pan-European hosting capacity visibility. 

Efficiency and fairness in capacity allocation

More efficient and fair capacity allocation processes are needed to ensure scarce grid capacity goes to projects that can realistically proceed. Speculative and premature capacity requests distort planning signals and block access for credible projects. Reforms involving financial deposits, project maturity milestones, withdrawal penalties, capacity release mechanisms and social or system-value criteria are being implemented in various jurisdictions, including Romania, the UK and the US.

Flexible connection agreements

Where grid reinforcements cannot be completed in time, voluntary flexible connection agreements may help address the time-to-power gap. Instead of waiting for full firm capacity, flexible connection agreements can allow earlier access to power by providing a reduced firm capacity allocation complemented by conditional capacity that can be curtailed during periods of grid stress, with data centres relying on on-site backup generation, battery storage or load reductions during curtailment periods. These arrangements can take several forms: static capacity limitations during predefined peak windows, dynamic limitations triggered by actual or forecast grid constraints and “bring your own generation” (BYOG) schemes where the data centre builds dispatchable generation on site.

Analysis of a US data centre case found that combining flexibility and BYOG arrangements could bring full operation three to five years earlier than traditional connection processes, with curtailment limited to 40-70 hours per year – equivalent to over 99 per cent grid availability.

Opportunities: From grid risk to grid resource

Despite the challenges they create, data centres also contain significant untapped flexibility resources that could, under the right regulatory and market conditions, make them active contributors to grid stability. Energy accounts for roughly 50 per cent of data centre operating expenditure, creating strong commercial incentives to participate in flexibility markets if the barriers can be lowered.

Flexibility assets within data centres

Three categories of flexibility assets exist within data centres:

  • IT domain: UPS batteries can provide short-duration grid flexibility (within milliseconds) and are well-suited to fast frequency response and primary reserve products; their main limitation is duration, as they are typically sized to bridge only a few minutes until backup generation starts.
  • Cooling systems: They can account for around 30 per cent of data centre power demand and offer additional short-term flexibility due to thermal inertia, backup systems (diesel rotary uninterruptible power supply [DRUPS]) and thermal storage, allowing temporary load shifting without affecting IT operations. However, advances in cooling efficiency are reducing cooling’s share of total load, limiting future flexibility potential.
  • On-site generation: Predominantly diesel, though with increasing interest in biofuels, gas turbines, fuel cells and additional renewable capacity. Provides the highest system controllability, particularly under BYOG-type flexible connection arrangements.

IT workload flexibility

Among data centre flexibility options, IT workload shifting is the most distinctive, as computing tasks can be shifted across time and geography in ways unavailable to most other electricity consumers. However, the extent of this flexibility depends on workload type, controllability and economic incentives. Some workloads can be deferred or relocated in response to grid conditions, while others remain latency-sensitive and inflexible. AI training and batch processing workloads have the highest flexibility potential (they are schedulable, not latency-sensitive, and can be deferred or relocated). Real-time transactional workloads have essentially no flexibility. In practice, mixed workload profiles and the high value of uninterrupted service often limit the share of load that operators are willing to shift.

The business model of the facility strongly shapes how much of this theoretical potential can be activated in practice:

  • Hyperscalers are best positioned to provide flexibility, as cloud architectures allow non-critical workloads to be delayed or relocated across regions, although constraints such as latency, performance and data residency still apply.
  • Colocation facilities host significant flexible capacity, yet operators typically lack control over tenant workloads, limiting coordinated load shifting.
  • Enterprise facilities have high theoretical potential, but critical workloads, regulatory constraints and limited technical resources often restrict practical implementation.

From grid-safe to grid-supporting

The technical systems required for grid-safe data centre operation, such as ramp-rate control, FRT, and controlled reconnection, already depend on coordinated management of UPS batteries, cooling systems and on-site generation. Once these control capabilities are in place, extending them to provide grid-support services requires relatively little additional investment. This creates a progression from basic grid-safe operation to expanded battery integration and eventually to fully coordinated virtual power plant (VPP) functionality combining storage, generation and flexible loads. Different flexibility sources offer distinct strengths: IT workloads are effective for short-term peak reduction but constrained by business requirements, while cooling systems provide greater energy-shifting potential with seasonal limitations. Much of this evolution is already being driven by operational and connection-management needs, making broader participation in flexibility and electricity markets a natural next step.

Market products and revenue opportunities

Once enabled, data centre flexibility can be monetised through electricity markets, creating recurring revenue streams that offset operational costs. Suitability varies by resource and service type. UPS batteries are well-suited to fast frequency response and primary reserve (frequency containment reserve [FCR]) due to their rapid response and short-duration capability. Secondary and tertiary reserves (aFRR, mFRR) can be provided through IT load shifting, cooling control and thermal storage, though sustained delivery requires coordination with operational constraints. Longer-duration energy markets (day-ahead and intraday) rely more on thermal storage and onsite generation, as IT and cooling flexibility are limited in energy volume. Finally, congestion management can be supported through load modulation, workload shifting, and local generation, offering targeted value in constrained grid areas.

Europe’s short-duration and ramping flexibility needs are expected to roughly double by 2030, reaching about 15-30 GW. Data centres could technically provide a significant share of this, with estimates suggesting up to ~16.9 GW of flexibility potential across five European markets (Germany, Ireland, the Netherlands, Norway and the UK), though only ~3.8 GW may be realistically available after considering operational and participation constraints.

Conclusion

Data centres are becoming a systemically important electricity demand due to their scale, geographic concentration and fast growth, requiring more active grid management rather than passive absorption. They create challenges because their rapid development often outpaces transmission and generation expansion, and their software-driven, power-electronic behaviour does not always align with existing connection and planning frameworks. This calls for improved demand forecasting, greater transparency on grid capacity, and more coordinated, harmonised connection requirements across Europe, alongside flexible connection arrangements to manage near-term constraints.

At the same time, data centres also represent a significant opportunity: their built-in control systems and electrical infrastructure could enable them to provide grid services, support flexibility and contribute to decarbonisation. Updated connection requirements, transparent hosting capacity information, reformed allocation processes, flexible connection agreements and electricity market designs that welcome data centres as active participants are all essential components of a coordinated European response. Such an approach would treat data centres not merely as a challenge to be accommodated, but as active contributors to system flexibility and resilience.