loading
loading
AI Power Spikes Demand Chip-to-Grid Design Changes AI data centers’ rapid power swings threaten grid stability, requiring coordinated chip-to-grid strategies across chip makers, operators, utilities, and infrastructure providers. AI data center developers and operators are being forced to confront not only how much power their facilities draw, but how quickly that demand can surge and collapse, creating frequency excursions that can ripple across the grid. Panelists at the recent Data Center World Power conference in Dallas (Sept. 21-23, 2026) warned that extreme events, such as a sudden drop of hundreds of megawatts in AI power draw, could trigger wide-area disturbances if not mitigated. The modern grid was designed for predictability and stability. Morning and evening peaks are forecastable, giving grid operators time to bring additional generation online and then ramp it down. Some turbines that provide the extra power can synchronize within minutes, while others require 30 minutes or more. That cadence is not compatible with AI workloads, which can spike and subside in seconds or even milliseconds. The panelists emphasized that no single company can solve this challenge; it will take a coordinated effort among chip providers, data center operators, utilities, and infrastructure providers. Nvidia: Managing Rack-Level Load Swings Nvidia’s rising rack power density has upstream consequences, said Sai Somayajula, principal electrical design engineer at Nvidia. A few years ago, a conventional data center rack typically used 5-10 kW. With Nvidia’s H100 era, that climbed to about 40 kW per rack. Blackwell-class systems pushed it to roughly 150 kW, and the Vera Rubin generation targets around 240 kW per rack. “We are projecting up to 1 MW per rack when 800-volt DC distribution appears in the future,” Somayajula noted. “Power density within the rack is increasing with every generation as AI evolves and solves more complex tasks.” To blunt the impact of rapid load swings, Somayajula said Nvidia has added rack-level storage and controls to smooth fast excursions within the rack and keep the AC input within grid requirements. According to the panel, grid operators are increasingly asking data centers to provide ride-through performance – remaining connected through short-duration voltage disturbances – and to buffer the grid from volatile AI workloads. Earlier this year, Nvidia introduced its DSX architecture, which combines chip, thermal, system, and software technologies to maximize AI factory throughput. It includes high-speed telemetry and advanced analytics to characterize and mitigate oscillations, allocates unused power headroom to GPUs, and reduces cooling power via warmer-water liquid cooling. However, Somayajula stressed that Nvidia cannot solve all these problems alone. “AI workloads, rack density, and power dynamics require a rethink of how we have been designing the data center,” he said. “That is why we partner with the industry, openly share information, and co-design all the way from rack to the grid instead of designing in silos.” “AI workloads, rack density, and power dynamics require a rethink of how we have been designing the data center.” – Sai Somayajula, principal electrical design engineer, Nvidia Oracle’s Abilene Stargate: Right-Sizing On-Site Generation for AI Some of these ideas are being implemented at Oracle’s AI facilities in Abilene, Texas, said Rajesh Gopanath, core infrastructure engineering architect at Oracle Cloud. Oracle’s Stargate data center complex is 60% to 70% operational, with campuses planned at a gigawatt scale, he said. As compute demand outpaces grid build-out, Oracle has been compelled to bring more energy and power-generation expertise in-house. “Compute was needed yesterday, but infrastructure is delayed due to substation, transmission, or grid unavailability,” Gopanath said. “We can’t get enough turbines, reciprocating engines, or natural gas.” Gopanath argued that supporting a 1 GW campus with a handful of 200 MW turbines creates single points of failure. Smaller building blocks improve resilience without making operations unmanageable. “Maybe 30 MW would be a good size. But you have to work out the best size of each data center building: ‘Should we build 100 MW buildings, 300 MW buildings, or 500 MW buildings?’” he said. “We have to find the sweet spot to make sure that we have the right level of availability down to the rack level.” Subsynchronous Oscillations and Ride-Through Requirements Gopanath warned about subsynchronous oscillations, which are frequencies below the US grid’s 60 Hz operating frequency. These oscillations, if excited by AI data centers’ behavior, can induce mechanical resonance that leads to catastrophic failure, including damage to turbine shafts. “That kind of catastrophic failure is possible, so we must make sure that it does not happen,” Gopanath said. “All these problems cannot be solved by one technology; you might need capacitors, batteries, uninterruptible power supplies, low-voltage ride-through, power smoothing, and fast-responding generation.” Mitsubishi Electric: A Layered, Grid-to-Chip Approach to Reliability David Roop, who leads the power systems engineering division at Mitsubishi Electric Power Products, said supporting AI workloads requires coordinated investments at the substation and grid levels. Success depends on a chip-to-grid approach that accounts for short- and long-duration energy storage, harmonic performance, power quality, and local grid characteristics. “Data centers are going to play a much larger part in terms of grid reliability by understanding utility requirements, control needs, and the architecture criteria required for a layered approach that will deliver faster projects with reduced project risk,” Roop said. “New guidelines, new standards, and new compliance criteria are being introduced so we can deal with instability much faster.”
document fetched 2026-10-24 09:57 · canonical https://datacenterknowledge.com/energy-power-supply/ai-power-spikes-demand-chip-to-grid-design-changes