Major AI companies are calling for coordinated energy grid policy and infrastructure upgrades to support AI workload growth.

Nearly 60 scientists at Microsoft, Nvidia, and OpenAI have co-authored a paper titled "Power Stabilization for AI Training Datacenters," which addresses critical power management challenges in AI training workloads. The researchers warn that fluctuating power demand during AI training—particularly the oscillating energy consumption between the power-intensive GPU compute phase and the less-taxing communication phase where parallelized GPU calculations synchronize—threatens the electrical grid's ability to handle variable loads and represents a significant barrier to AI model development.

The power variation is extreme, with compute phases approaching the thermal limits of GPUs while communication phases operate near idle energy levels. These oscillations occur at multiple scales: at individual server nodes, across data centers, and ultimately at the power grid level. To illustrate the magnitude, the authors compare this to "50,000 hairdryers (~2000 watts) being turned on at once." As they observe: "At scale, these swings can amount to tens or hundreds of megawatts, occurring at frequencies that, if poorly aligned with the resonant characteristics of power grid components (e.g., turbine generators or long transmission lines), can risk grid instability and mechanical failure." The authors emphasize that "these issues are not theoretical – multiple utility providers have now documented the impact of harmonics induced by synchronized computing loads."

The urgency of this challenge is underscored by broader grid concerns. Schneider Electric expects the US grid will become less stable by the end of the decade due to data center energy demand. A US Department of Energy report published last December stated: "data centers consumed about 4.4 percent of total US electricity in 2023 and are expected to consume approximately 6.7 to 12 percent of total US electricity by 2028."

The Microsoft, Nvidia, and OpenAI team has evaluated three strategies to address power stabilization. Software-based approaches even out power usage by injecting secondary workloads when GPU activity falls below certain thresholds, but they carry performance overhead and require customer-cloud provider collaboration. GPU-level firmware features like power smoothing, supported in the Nvidia GB200, allow developers and cloud providers to set power utilization floors and ramp-up and ramp-down rates, though power smoothing imposes extra energy costs. Data center-level Battery Energy Storage Systems can handle power demand spikes locally without burdening utility grids, but energy storage hardware remains expensive. The researchers argue that an optimal solution requires combining all three techniques and calls for greater coordination among vendors so that rack-level energy storage and GPUs can communicate about workload state changes.

To realize these improvements, the researchers are asking AI framework and system designers to focus on asynchronous, power-aware training algorithms; utility and grid operators to share resonance and ramp specifications and standardize communication channels with data center operators; and the technology industry to establish interoperable standards for telemetry, load signaling, and sub-synchronous oscillation mitigation. As the authors conclude: "Together, we can design for a future where AI training is not only powerful, but also power-aware."