Why rapid data-center power swings challenge the grid


Generating enough electricity is only part of the problem. Engineers describe how storage, voltage controls and generation design can address demand that changes faster than the grid can respond.
Data centers running superintelligence (SI)—artificial intelligence—can change their electricity demand in seconds or milliseconds, faster than traditional grid equipment is designed to follow. At the September 21–23, 2026 Data Center World Power conference in Dallas, panelists warned that an unmitigated drop of hundreds of megawatts could cause disturbances across a wide area, Data Center Knowledge reported.
The grid was built around more predictable changes, such as morning and evening demand peaks. Those patterns give operators time to bring generation online and reduce it later. Some turbines supplying additional power take minutes to synchronize with the grid; others need 30 minutes or more. Having enough generating capacity does not by itself solve the mismatch between that timetable and millisecond changes in demand.
Rapid changes can disturb grid frequency—the rate at which alternating current cycles. The engineering challenge therefore concerns both how much electricity a facility uses and how quickly that demand rises or falls.
Nvidia’s principal electrical design engineer, Sai Somayajula, described a sharp increase in the power used by each rack, the frame that houses computing equipment. Data Center Knowledge reported his figures:
| Rack or system generation | Power per rack | Qualification |
|---|---|---|
| Conventional rack | 5–10 kW | Typical a few years ago |

Somayajula said Nvidia has added storage and controls within racks to smooth rapid demand changes and keep the alternating-current input within grid requirements. The purpose is to contain the fastest swings near the computing equipment rather than pass them directly to the grid. Whether, and by how much, these measures reduce grid disturbances remains uncertain.
The panel also said grid operators increasingly ask data centers to remain connected during brief voltage disturbances. This is called “ride-through.” Alongside smoothing demand, it is another requirement the panel identified for how data centers interact with the grid.
At Oracle’s Stargate complex in Abilene, Texas, computing demand was arriving ahead of available infrastructure, according to Rajesh Gopanath, Oracle Cloud’s core infrastructure engineering architect. At the September conference, he said the complex was 60%–70% operational, with campuses planned at gigawatt scale. He cited delays involving substations, transmission and grid availability.
Gopanath argued that supplying a 1-gigawatt campus with a handful of 200-megawatt turbines creates single points of failure. Each large turbine represents a substantial part of the supply. Smaller generation blocks, he said, could improve resilience without making operations unmanageable.
He suggested that 30 megawatts might be a useful generator size, but did not present it as a settled design. The choice also depends on the size of each data-center building and the level of availability needed down to individual racks.
Gopanath also warned about “subsynchronous oscillations”: oscillations below the US grid’s operating frequency of 60 hertz, or 60 cycles per second. If data-center behavior excites these oscillations, he said, they could produce mechanical resonance—reinforced vibration—and catastrophic damage, including to turbine shafts. His warning concerned a possible failure, not a documented incident.
No single technology can address all these problems, Gopanath said. He identified a possible combination of capacitors, which store electrical energy; batteries; uninterruptible power supplies; controls that keep equipment connected through low-voltage disturbances; power smoothing; and fast-responding generation.
David Roop, who led the power systems engineering division at Mitsubishi Electric Power Products as of the October 2026 report, said supporting these workloads requires coordinated investment in substations and the grid. His engineering assessment called for storage suited to both short and longer periods, stable power quality and designs that account for local grid conditions.
Roop said new guidelines, standards and compliance criteria were being introduced to address instability faster. For developers and operators, the work he described extends beyond securing enough electricity: rack controls, on-site generation and grid connections need to be designed together around the utility’s requirements.
Planning and soil surveys can continue, but major site changes must pause for environmental review. Available electricity capacity does not remove that requirement.
4 min

With Java support enabled, a malicious spreadsheet can run attacker-controlled code when opened, without the usual macro warning.
2 min

The regulator also questioned terms users cannot access before signing up, but did not identify the contract version it examined.
2 min
Correction ·
Corrected Armenian wording.