NVIDIA and Emerald AI have announced a multi-partner initiative with major energy companies to develop “flexible AI factories” designed to operate as dynamic grid assets rather than fixed, high-load data centers. The collaboration introduces a reference architecture that combines NVIDIA’s AI infrastructure and orchestration software with Emerald AI’s workload management platform to enable real-time adjustment of power consumption based on grid conditions.
The core shift is architectural. Traditional data centers are provisioned for peak demand and treated as inflexible loads, creating bottlenecks in grid interconnection and increasing infrastructure costs. The proposed model treats AI compute as a controllable variable. Workloads such as model training and batch processing can be slowed, paused, or redistributed across regions, while latency-sensitive inference remains uninterrupted. This allows facilities to modulate energy usage in response to supply constraints or price signals without degrading service-level objectives.
Operationally, the design integrates on-site generation, storage, and grid connectivity with software-driven orchestration. Facilities can begin operating with co-located power resources and scale into full grid interconnection over time, reducing time-to-deployment. The approach also aligns with demand response programs, positioning AI infrastructure as a participant in grid balancing rather than a passive consumer.
The implications are material for enterprise AI deployment. Power availability is emerging as a primary constraint on scaling large models and inference workloads. By enabling partial-load operation and flexible scheduling, the model increases effective grid capacity without requiring equivalent expansion in physical infrastructure. Prior demonstrations suggest that modest, time-bound reductions in data center load could unlock significant additional capacity across existing grids.
For operators, this introduces new trade-offs between compute throughput, energy cost, and reliability. It also creates a dependency on orchestration software capable of aligning workload priorities with external energy signals. Governance considerations extend beyond data and models to include energy policy compliance, grid participation agreements, and coordination with utilities.
The initiative reflects a broader convergence of AI infrastructure and energy systems. As AI workloads scale, infrastructure design is shifting from maximizing utilization at all times to optimizing across compute performance, energy availability, and system resilience. Flexible AI factories represent an attempt to decouple growth in AI demand from linear increases in energy consumption and grid strain, reframing data centers as active components of energy networks.
Commenting on the partnership, Michael Polsky, Founder and CEO of Invenergy, said: “AI is changing how we’re thinking about energy, and our customers need power fast, with the ability to scale over time. Combining near-term generation solutions with a path to full grid connection and flexible operations is an innovative and efficient way to help our customers meet their energy needs faster while keeping the system reliable.”