Power Infrastructure Requirements for AI Data Centers

Adding AI computing capacity to a data center changes more than the server configuration. It can change the amount of power the facility must deliver, how quickly that demand rises and falls, and how much margin remains in equipment that may have been designed around a very different IT load.

For an existing facility, that makes the electrical infrastructure part of the expansion decision from the beginning. UPS systems, generators, automatic transfer switches, distribution equipment, batteries and the condition of the existing power system all need to be considered before substantial new GPU capacity is placed into service.

CI Group’s Data Center Power Infrastructure services bring those requirements together, from sourcing high-capacity equipment to maintaining and replacing the systems already supporting critical operations.

AI Workloads Can Change the Way a Data Center Uses Power

Rack density gets much of the attention when AI infrastructure is discussed, but density alone does not describe the electrical challenge.

Large AI training environments can produce rapid changes in demand because many compute nodes may ramp up and down at nearly the same time. That is different from a conventional data center filled with mixed workloads, where changes in individual servers tend to occur at different times and produce a comparatively steady aggregate load.

Uptime Institute has examined this behavior in large AI compute environments and notes that synchronized training workloads can create substantial step changes in power demand. In larger clusters, those changes can reach hundreds of kilowatts or even megawatts over very short periods, placing additional demands on the electrical distribution system.

Read Uptime Institute’s research on electrical considerations with large AI compute.

That does not mean every AI deployment requires an entirely new power architecture. It does mean that a capacity study based only on average load may leave out important operating conditions.

Before adding a significant AI or GPU load, operators should understand the current critical load, expected new demand, available UPS capacity, distribution limits and the amount of headroom that needs to remain for redundancy and future growth.

Start With the Existing UPS Capacity

The UPS system is an obvious place to begin, but the useful question is not simply, “How large is the UPS?”

The more important questions are how much of that capacity is already committed, what happens when part of the system is unavailable, and what condition the equipment is in today.

An existing UPS may have adequate nameplate capacity while operating close to a practical limit once redundancy, future growth or maintenance requirements are considered. An older system may also contain batteries, cooling fans, capacitors or other components that should be addressed before the protected load is increased.

For facilities that need additional or replacement capacity, CI Group supplies UPS systems across a broad range of capacities and manufacturers, including large three-phase systems used in mission-critical environments.

Equipment already in service should be evaluated on condition as well as capacity. CI Group’s UPS Maintenance Services include system evaluation and ongoing service intended to identify developing issues before they become failures.

Capacity Is Not the Same as Redundancy

A facility can have enough total electrical capacity and still be vulnerable to the loss of a single component or power path.

That distinction becomes more important when multiple compute systems are working together on the same job. In that environment, losing power to one portion of the cluster can have consequences beyond the individual server that went offline.

NVIDIA’s DGX H100 SuperPOD design illustrates the point. A DGX H100 system contains six internal power supplies, with at least four required to remain energized for the system to operate. NVIDIA therefore specifies a minimum of three power sources for DGX H100 system racks, with those rack feeds supported by the facility’s UPS and generator infrastructure.

NVIDIA also states that the power feeds should originate from separate upstream distribution paths and identifies N+1 as the minimum power configuration for DGX H100 system racks. Its enhanced configuration uses three discrete upstream UPS and distribution paths.

Review NVIDIA’s DGX H100 SuperPOD electrical design guidance.

Those are requirements for that particular NVIDIA architecture, not a universal specification for every GPU installation. They are useful, however, because they demonstrate why redundancy needs to be traced through the complete power path rather than judged only by the number of power cords entering a rack.

An infrastructure review should determine whether supposedly redundant feeds remain independent upstream, whether equipment can be removed from service without exposing the load, and whether the remaining path can actually support the required demand.

For a closer look at redundancy specifically in GPU environments, see Why Redundant Backup Power Matters for NVIDIA GPU Data Centers.

CI Group’s existing resource on distributed redundancy provides additional background on the relationship between power-system architecture and availability.

Battery Runtime Needs to Be Reconsidered When the Load Changes

Available UPS capacity does not automatically mean the facility has the battery runtime it expects.

UPS batteries deteriorate with age, temperature, cycling and operating conditions. At the same time, increasing the protected load changes how long the existing battery system can support that load.

A battery plant that provided an acceptable bridge to generator power under the previous configuration may provide less runtime after a substantial expansion. This is one reason battery condition and runtime should be reviewed before new computing equipment goes online rather than after the first utility interruption.

CI Group provides UPS Battery Replacement Services that include testing and replacement for aging battery systems supporting critical operations.

Backup Generation Has to Support the Facility, Not Just the Servers

The UPS carries the critical load immediately after normal utility power is interrupted. If the outage continues, the facility ultimately depends on its backup generation system.

Adding a large computing load therefore affects more than the UPS calculation.

The generator requirement may need to account for cooling equipment, pumps, controls, network infrastructure, security systems and other electrical or mechanical loads necessary to keep the data center operating. The exact combination depends on the facility, but the important point is that the generator should be evaluated against the critical operating load rather than the IT equipment in isolation.

Large generator equipment also has a very different procurement and installation timeline from adding another rack of servers. If expansion will require additional generation capacity, identifying that requirement early can prevent the electrical infrastructure from becoming the project bottleneck.

CI Group supplies commercial and industrial generators for large critical-power applications, including generators in the 2001+kW range for data centers and other facilities requiring multi-megawatt backup power.

If additional generation capacity is part of the project, see How to Source Large-Scale Generators for Data Center Build-Outs for the equipment, rating, permitting and procurement factors to evaluate.

The Automatic Transfer Switch Is Part of the Same System

A generator can be ready to run and the UPS can be functioning correctly, but the facility still depends on the equipment responsible for transferring the load between power sources.

Automatic transfer switches should therefore be included in the same review as the UPS and generator systems. Inspection and operational testing can identify mechanical, electrical and control problems before an actual outage requires the transfer switch to perform.

CI Group’s Automatic Transfer Switch Service & Maintenance includes inspection, testing, diagnostics, preventative maintenance and repair for critical-power environments.

Existing Equipment Condition Matters More as the Load Increases

Expansion often puts attention on the equipment being added. The condition of the equipment already in the power path can be just as important.

UPS systems operate continuously, and several of their critical components have finite service lives. Batteries degrade. Fans wear. Capacitors age. Connections can loosen and operating temperatures can change.

None of those conditions necessarily produces an immediate system failure. That is precisely why they are worth addressing before increasing the load.

A pre-expansion review should look at maintenance history, battery age and test results, alarm records, component replacement intervals, electrical connections, cooling performance and any recurring faults that may indicate a developing problem.

A structured UPS Preventative Maintenance & Inspection program gives operators an opportunity to address those issues while the work can still be planned rather than handled as an emergency.

Plan for How the Equipment Will Be Maintained

The final design needs to work during normal operation, but it also needs to work when equipment is being inspected, tested or repaired.

That means asking maintenance questions during the expansion process rather than after the new load is connected.

Can a UPS be taken out of service without exposing the critical load? Can a transfer switch be tested safely? Is there enough remaining capacity during planned maintenance? Does the facility have a practical way to support the load if replacement work takes longer than expected?

Those questions often reveal limitations that are not visible in a simple nameplate-capacity calculation.

Temporary Power Can Be Part of the Upgrade Plan

Critical infrastructure does not stop needing protection simply because the equipment protecting it is being replaced.

During a UPS replacement, major maintenance event or capacity upgrade, temporary equipment can provide a bridge while permanent infrastructure is offline.

CI Group offers UPS Rental Systems for planned projects, equipment replacements, emergencies and transitional power requirements.

Including temporary power in the project plan can be much easier than trying to source it after maintenance has begun or an unexpected equipment problem has changed the schedule.

Leave Room for the Next Computing Upgrade

One of the challenges in planning AI infrastructure is that the computing hardware can change much faster than the electrical infrastructure supporting it.

A UPS system, generator, transfer switch or distribution system installed during today’s project may still be expected to support several later generations of IT equipment.

That does not mean every facility should dramatically oversize its power system. It does mean the project should identify where future capacity can be added, where hard limits exist and which components are likely to require expansion or replacement if the computing environment continues to grow.

A useful infrastructure plan considers today’s deployment while preserving a practical path for tomorrow’s.

Evaluate the Critical Power System Before the New Load Arrives

The most useful time to discover a power-infrastructure limitation is before the new computing equipment is installed.

For an existing data center, that means looking at UPS capacity, redundancy, battery condition, backup generation, transfer equipment, maintainability and future growth as parts of one connected system. For a new build or major expansion, it means making those decisions early enough that critical-power equipment availability does not become the constraint on the project schedule.

Critical Infrastructure Group works with data center operators to source, maintain, replace and support the equipment behind critical operations, including high-capacity UPS systems, large generators, automatic transfer switches, temporary UPS equipment and ongoing power-system maintenance.