AI News · Good news ·
Nvidia, Oracle and Mitsubishi Electric outline fixes for AI power swings
Nvidia, Oracle and Mitsubishi Electric highlighted approaches to managing AI data center power swings at Data Center World Power. Rapid load changes can cause grid frequency excursions, making power variability a concern alongside total consumption. Nvidia projects that rack power density could reach 1 megawatt with 800-volt DC distribution.
Key points
- Rapid AI power swings can disrupt grid frequency, making demand variability a concern alongside total consumption.
- Nvidia projects future rack power of 1 megawatt with 800-volt direct-current distribution.
- Nvidia has incorporated rack-level storage and controls to smooth fast power fluctuations.
- Oracle suggested 30-megawatt generation blocks could improve resilience at gigawatt-scale campuses.
- Mitsubishi Electric called for coordinated investments spanning data center equipment, substations and the grid.
What happened: AI data center planning needs to account for how quickly electricity demand changes, not just how much power a facility consumes. ETDatacenters reported that Nvidia, Oracle and Mitsubishi Electric outlined approaches to that problem at Data Center World Power. Panelists warned that AI workloads can surge and subside in seconds, causing frequency excursions, or deviations in the grid’s operating frequency. Sudden drops in demand could trigger wide-area disturbances if left unmanaged. Their message was that managing these swings requires coordination among chip providers, data center operators, utilities and electrical infrastructure companies.
The details: Nvidia principal electrical design engineer Sai Somayajula described a sharp increase in power concentrated within individual equipment racks. Conventional racks previously used 5 to 10 kilowatts, compared with roughly 150 kilowatts for Nvidia’s Blackwell-class systems and a target of around 240 kilowatts for the Vera Rubin generation. He projected that future racks could reach 1 megawatt with 800-volt direct-current distribution. Nvidia has incorporated rack-level storage and controls to smooth fast fluctuations while keeping alternating-current input within grid requirements. Somayajula stressed that Nvidia cannot address the problem alone: rising rack density and changing power dynamics require a broader rethink of data center design.
Background: Oracle is confronting the same issue at campus scale. Rajesh Gopanath, Oracle Cloud’s core infrastructure engineering architect, discussed the company’s AI facilities in Abilene, Texas. ETDatacenters reported that the Stargate complex was 60 percent to 70 percent operational, with campuses planned at gigawatt scale. Oracle is bringing more energy and power-generation expertise in-house because compute demand is outpacing grid build-out. Gopanath suggested that smaller generation building blocks, such as 30 megawatts, could improve resilience. He also warned about subsynchronous oscillations, which occur below the US grid’s 60-hertz frequency and can cause mechanical resonance and catastrophic failure if excited by AI data centers.
Who it affects: The challenge extends beyond equipment inside a data center. David Roop, who leads power systems engineering at Mitsubishi Electric Power Products, called for coordinated investments at substations and across the grid. His chip-to-grid approach considers energy storage, harmonic performance, power quality and local grid characteristics together. Roop said new guidelines, standards and compliance criteria are being introduced to address instability faster.
What to watch: For business teams choosing AI infrastructure, contracted power capacity is only part of the assessment. Buyers should also ask providers how their electrical systems will handle rapid load changes and increasingly dense racks. Deployment timing for Nvidia’s projected 1-megawatt racks, the costs of these approaches and comparable performance results were not reported.
Our take
Infrastructure planning needs to account for rapid load changes, not just contracted megawatts. Teams should ask providers how electrical systems will handle increasingly dense AI racks.