Microsoft, NVIDIA, and OpenAI presented their ~2.5-year cross-company research on power stabilization for AI training data centers at OCP Global Summit 2025, addressing the problem of GPU clusters creating massive, grid-destabilizing power oscillations at gigawatt scale. The talk covered three complementary mitigations: NVIDIA's Blackwell GPU Min Power Floor (MPF) hardware feature (20% swing), Microsoft's Firefly software approach (<5% swing using NVIDIA MPS), and rack-level energy storage for future deployment. The session concluded with a call to action for OCP community standardization of telemetry, load signaling, and energy storage device specifications.

Key Takeaways

• AI training workloads at 300 MW–1 GW scale create rapid, repeated power swings between near-100% TDP (compute phase) and near-idle (communication phase), producing load oscillations at frequencies of 0.1–30 Hz that can damage grid equipment and de-stabilize interconnections.
• NVIDIA's GPU Min Power Floor (MPF) feature, deployed on Blackwell (GB200) products, constrains GPU power within a programmable band (floor 20%–90% TDP) using hardware-level bounding box and ramp rate controllers, achieving a 20% power swing at data center level.
• Microsoft's Firefly software mitigation runs a secondary power-hungry workload during training communication phases using NVIDIA MPS (<5% impact to primary job), reducing power swing to below 5% — meeting the tightest utility grid specifications.
• Rack-level energy storage is the preferred long-term solution for steady-state swing masking (charges during low-power phases, discharges during high-power phases), but is not yet deployment-ready; the current recommended approach combines GPU MPF + Firefly today.
• The composite mitigation strategy scores GPU MPF as high reliability / medium performance impact / high energy overhead; Firefly as medium reliability / medium performance / high energy overhead; energy storage as high reliability / high performance / low energy overhead but high cost.
• The team calls on the OCP community to develop standards for fine-grained telemetry (FFT-based monitoring), load signaling, sub-synchronous oscillation mitigation, and energy storage device certification to enable industry-wide deployment of these solutions.
• A cross-company research paper covering the full technical details (including energy storage placement trade-offs at rack, corridor, and data center levels) is available at https://aka.ms/PowerStabilization.