NVIDIA and Six US Power Producers Launch Dispatchable AI
- AI Energy Demand Data Centers Grid Capacity News
Marin Holt
- 2
- 5 minutes read
- NVIDIA estimates the approach could unlock 100 GW of additional US grid capacity without new transmission.
- Six major US power producers will develop AI data centers that ramp GPU power up or down in seconds to act as dispatchable grid assets.
- The first commercial deployment is NVIDIA Aurora, a 96 MW data center in Manassas, Virginia, expected to launch later in 2026.
- The US grid ran at roughly 60% average load factor in 2025, meaning latent capacity exists if large loads shift in time.
NVIDIA and Emerald AI announced at CERAWeek on March 23, 2026 that six major US power producers will develop AI data centers that ramp GPU power up or down in seconds to act as dispatchable grid assets. NVIDIA estimates the approach could unlock 100 GW of additional US grid capacity without new transmission.
What Is Being Deployed
The program uses NVIDIA Vera Rubin DSX AI Factory reference design with DSX Flex software, which modulates GPU compute load in real time — briefly slowing batchable AI workloads when the grid is stressed and resuming within SLA guardrails, per the March 23 NVIDIA announcement. Emerald AI Conductor platform orchestrates computational flexibility and coordinates with on-site generation and storage. In the first phase, AI factories use co-located generation and storage to reach commercial operation without waiting for full grid interconnection.
The first commercial deployment is NVIDIA Aurora, a 96 MW data center in Manassas, Virginia, connected via PJM Interconnection and Digital Realty, expected to launch later in 2026, per Peter Kelly-Detwiler analysis of April 2, 2026.
Why It Matters for Grid Load Balancing
PJM capacity market prices rose 7-8x over historical averages in recent years. Grid interconnection queues stretch to 2030 in most US regions. If AI factories operate as demand-response assets, curtailing GPU workloads during grid stress, the effective capacity available to connect new facilities expands without new transmission or generation. The US grid ran at roughly 60% average load factor in 2025, per Kelly-Detwiler, meaning latent capacity exists if large loads shift in time.
Critical Perspective
The 100 GW figure is a theoretical ceiling, not a committed capacity target. DSX Flex modulates batchable workloads — training runs and batch inference — not latency-sensitive applications like real-time inferencing or fraud detection. How much of any given data center load is actually flexible depends on workload mix, which varies widely. The six partner energy companies have not disclosed contract terms, capacity commitments, or compensation rates. How FERC and state PUCs will classify AI factory demand response as a grid service is unresolved. Aurora is one facility; scaling to 100 GW requires regulatory frameworks that do not yet exist.
