When a data center becomes a utility: Solving the emerging risks behind captive power

According to the IEA, “electricity demand from AI-optimized data centers will more than quadruple by 2030, while total global data-center electricity demand will more than double to around 945 terawatt hours. With booming demand, data center developers can no longer rely solely on public utilities to provide large-scale power. Grid queues are becoming bottlenecks in key markets, and the power demands of data centers are increasingly incompatible with grids.

"For many data center operators, securing power is now more difficult than securing capital", Mark Hiles, Global Head of Power Broking, Willis Natural Resources.

Captive generation is developing as a strategic solution.

To overcome grid constraints and secure power availability, major operators are building behind-the-meter power generation assets, including large H-class gas turbines, fuel cells, battery energy storage systems (BESS) and hybrid power facilities.

By constructing dedicated on-site generation assets, developers can avoid lengthy grid interconnection queues, accelerate project delivery, and secure reliable power supplies capable of supporting rapidly growing electrical loads. Depending on the operating model, facilities may operate as isolated power islands, maintain grid backup capability, or export surplus generation to the grid.

The three types of operating model

Across all operating models, data centers depend on near-zero downtime. Modern AI data centers are increasingly dynamic rather than static loads. Traditional power plants were designed to serve relatively predictable demand, while data center power plants need to cope with rapid, large-scale load changes that stress generation, storage, cooling, and control systems.

Two key issues are on the horizon for data centers that produce their own power

“Even with favorable permitting, the critical challenges become commissioning, full-load testing, turbine reliability, cooling-load interactions, reserve capacity, and potential sub-synchronous torsional interaction (SSTI) risks associated with large turbine-generator installations and power-electronics-heavy systems”, Ian Douglas, Director, Risk Engineering, Willis Natural Resources, U.K.

1. The draw down on power

One of the biggest emerging challenges for utilities and independent power providers is that AI-driven data centers draw power very differently from traditional loads, and this has major implications for testing, commissioning, system design, and grid stability.

Initially, once servers were installed and operating, the electrical load curve was generally smooth and highly predictable with a data haul over several minutes. But now, an AI workload increases the power demand by tens of megawatts within seconds.

• A traditional rack might draw 5-15 kW per rack
• But an AI rack today can draw 50-150 kW, with 300 kW+ in new designs

Large blocks of load can appear very quickly. A generator that performs well under steady-state testing may struggle to accommodate rapid load ramps.

The hidden threat: Sub-synchronous torsional interaction

SSTI occurs when electrical system frequencies interact with the natural mechanical frequencies of turbine-generator shafts, creating damaging stress and vibration. In extreme cases, SSTI can contribute to shaft damage and major equipment failure.

“As more captive generation systems are integrated with sophisticated power electronics, batteries and grid connections, the potential for dynamic stability issues increases. SSTI remains an important consideration, alongside sub-synchronous resonance (SSR), converter-driven oscillations, weak-grid interactions and control system instability associated with inverter-based resources”, Gonzalo Figueroa, Director, Risk Engineering, Willis Natural Resources, Latin America.

Protection and mitigation systems, such as torsional protection modules (TPMs), torsional stress relays (TSRs), remedial action schemes (RASs) together with comprehensive power system studies and dynamic modelling, become critical to ensure reliable operation.

Power is hotting up

But there’s a simultaneous issue. Power demand and cooling demand are directly linked.

When computing hardware for AI centers suddenly moves to full utilization, heat output rises immediately, which means cooling systems respond in parallel. This creates a second power demand increase on top of the IT load itself. Peak AI workloads can also increase cooling water demand, auxiliary transformer loading and emergency power requirements, all of which should be considered during system design and commissioning.

Commissioning therefore increasingly requires testing:

• Maximum thermal load scenarios
• Sudden cooling plant startup
• Integrated power and cooling responses

Solutions:

The strain of sustained operation above approximately 120% of rated load can accelerate insulation aging. The additional wear and tear should be considered an emergency operating condition rather than a normal operating regime. Specialist power facility risk engineers can assess the structural integrity of the asset, model loss scenarios, but also provide guidance on actionable next steps, such as focusing on transformer redundancy (N-1) and spare transformer strategies to build risk resilience.

2. Testing and commissioning

As projects move to the testing and commissioning phase, captive power plants supporting AI data centers must demonstrate they can:

• Move from 50 MW to 150 MW rapidly
• Accommodate fluctuating AI workloads
• Manage different electrical signatures
• Maintain voltage and frequency stability

Although fast-track behind-the-meter (BTM) are often favored by permitting authorities and do not require grid codes, a 7-day continuous run remains a testing and commissioning requirement.

Gas turbines are typically validated through a 72-hour reliability run where the plant must operate continuously without a major trip or forced outage. But for data centers where even brief outages can have extensive financial consequences across entire value chains, the power system is often held to higher reliability requirements to run for 168 hours (7 days) to verify sustained operation under realistic load conditions.

A successful 7-day demonstration run provides evidence that:

• Controls are stable
• Protection settings are correct
• Turbines and auxiliary equipment do not develop latent defects
• Cooling systems can support sustained operations
• Fuel and electrical supply systems perform as expected

The challenge is cumulative reliability. A power plant may pass a four-hour test but fail after several days due to fuel contamination, cooling issues, software faults, synchronization problems, generator vibration, BESS thermal management problems, among others.

But the critical challenge at this stage is that while plants are typically plugged into grids to offload power during testing and commissioning, H-class turbines necessary to supply the intensive needs of data centers can be too powerful for local grids. This leaves power producers at data centers unable to offload the power, breaching the continuous runtime requirement and unable to move the project into full operation. The critical challenge: how to offload the power produced.

Solutions:

• Some projects are using grid connections from mothballed large power user facilities, such as aluminum smelters, but data center projects that are newly constructed/in construction are unlikely to have a latent power plant at their disposal.
• Investing in BESS is a strategic solution, but requires proper planning and capital allocation at the early stages of the project design. Although interconnection standards require that distributed energy resources (DER) such as BESS disconnect when a grid is experiencing issues with stability, advanced load-forming inverters still offer the grid support capability to enable the ongoing modernization of grids. For projects that are exploring retrofitting or including BESS capabilities during the construction phase, different integration risks and capital barriers can stall projects.
• “Reciprocating engines are capable of synchronizing to the grid within approximately 30 seconds and achieving full load in less than five minutes. The engine's hot ramp capability exceeds 100% load per minute, providing substantially greater operational flexibility and load-following capability than conventional large-frame combined-cycle gas turbines: a developing technology to watch”, Bryan Allcock, Vice President, Risk Engineering, Willis Natural Resources, North America.

Lessons learned

Grid distribution companies have been reluctant to support large gas turbine independent power producers during the commissioning phase and restrict the power connection capacity, to minimize their exposure. But many of these challenges facing data centers have already been encountered by companies across the metals and mining sector, such as aluminum smelters and steel producers.

Mining and metals companies have been operating large, energy-intensive facilities with self-generated or behind-the-meter (BTM) power for decades, often in remote locations where grid power was unavailable, unreliable, or insufficient.

Their experience highlights several recurring themes:

• Commissioning is often riskier than steady-state operation
• Control system integration is frequently a larger source of failure than mechanical equipment
• Grid disturbances can propagate into captive plants and vice versa
• Maintaining spinning reserve significantly improves resilience
• Black-start capability should be validated under realistic conditions
• Long-lead equipment failures can result in business interruption lasting many months

Reliable operations depend on the entire energy ecosystem, building interconnected resilience across:

• Fuel supply contracts
• Fuel transportation logistics
• Water supply
• Energy storage
• Maintenance planning
• Spare parts availability
• Grid synchronization and control systems

GPU clusters can create unusually concentrated and dynamic power demands. As microgrids become more common, operators need sophisticated control systems to manage moment-to-moment changes in power flows.

Mining companies have tackled this by:

• Simulating load swings before startup
• Testing island-mode operation
• Verifying black-start capabilities
• Running generators for extended periods
• Stress-testing control systems

Five questions every data center developer should ask before building captive power

1. How will full-load rejection testing be performed?
2. What happens if the primary turbine fails? Can the facility continue operating after the loss of its largest generating unit?
3. Is grid import capacity sufficient to carry the full facility load?
4. Has SSTI or torsional stress been assessed?
5. Does the insurance program adequately address machinery breakdown, business interruption and commissioning exposures?

The race to secure power for AI infrastructure is transforming the data center sector, bringing unforeseen issues to the fore. Some critical challenges are entirely new, calling on industry specialists and cross-sector collaboration to innovate new solutions.

Looking ahead, some jurisdictions are already updating regulation. Some system operators such as ERCOT in Texas are making changes to their connection process to ensure that large-load users, such as data centers, connect at grid points capable of supporting their demand. How far this regulation will be adopted by other U.S. states and internationally, will be a trend to watch. So far, Ireland are early movers in regulatory reform. Irish policy changes under discussion are indicative of wider European approaches in the future: to allow large-load users such as data centers to connect only where they support system security, provide or fund additional generation and renewables, locate appropriately on the network, and operate with greater flexibility and transparency, rather than relying solely on grid expansion to meet their demand.

Owning generation fundamentally changes your risk profile.

Companies and projects that treat power generation as a core operational discipline, with dedicated governance, maintenance, fuel strategy, resilience planning, and stakeholder engagement, will be best positioned for long-term success. With environmental, regulatory, commissioning, operational reliability, and community relations risks adding new dimensions to traditional property damage and business interruption risks, data centers’ shifting risk profiles need a comprehensive and objective review.

Backed by decades of industry expertise, specialist risk engineers will be a critical asset to building long-term resilience. With onsite risk surveys, assessments, scenario modelling and best practice guidance, risk managers can make decisions that will build resilience and boost operational efficiencies now and into the future.

To find out how to make future-ready decisions, contact our team.