Nvidia's 100 GW Promise: Can Flexible AI Data Centers Fix the Grid?
Leading AI chipmaker Nvidia and software company Emerald AI will work with a number of energy supply companies to “power and advance a new class of AI factories” that can connect to the grid faster and “operate as flexible energy assets that can support the grid.”
The approach will use a new reference design with Nvidia’s latest chip and its DSX software to help manage power use in real time, modulating demand and coordinating flexible load.
Nvidia envisions factories using on-site, co-located generation and storage as a bridge before they connect to the grid, and then on-site assets will “flexibly support the grid.”
Emerald AI’s Conductor platform will orchestrate computational flexibility, combined with onsite resources to deliver power flexibility.
“Power-flexible AI factories” Nvidia claims, “can help unlock up to 100 gigawatts of capacity across the U.S. power system.”
For perspective, the U.S. hit an all-time peak of 759 GW last July, and has 1,300 MW of installed generating capacity.
Data center flexibility is important because there’s limited transmission and supply available on the grid and new infrastructure can’t be built fast enough.
Plus, the grid has an estimated load factor (the percentage of the energy we use versus the amount we could use if we ran at 100%) of about 60%. It’s very inefficient.
Supply also gets expensive. PJM’s capacity market prices have soared 7 or 8x over average historical numbers, with data loads costing ratepayers an estimated $23 billion in the past three capacity auctions.
By creating more flexibility during grid scarcity, one can meet more demand without building new infrastructure, AND flow more energy across the same grid, lowering the per unit delivery price.
Two recent studies on flexibility suggest that flexible operations can greatly increase ability to add load and result in economic efficiencies: 76 GW of new load could be integrated with just an average annual load curtailment rate of 0.25% and 98 GW of new load could be integrated at a curtailment rate of .5%.
And avoiding just 1% or 2% of the peak hours would reduce utilities’ new natural gas combined cycle construction costs by 10% to 15%.
But the available information doesn’t really tell us all that much. We don’t how flexible the operation of large language training models will be, nor do we know the potential flexibility in the inference function, where the models perform on demand to undertake the work new need on a daily basis.
We have limited empirical data: an EmeraldAI data center in Arizona cut
power consumption by 25% during three hours of peak grid demand. As of late March 2026, Emerald AI confirmed it has demonstrated power flexibility capabilities at five different commercial data centers around the world. But actual performance numbers are limited, for durations and percentages.
Likewise, Google announced it has surpassed 1 GW of demand response but didn’t share the details that matter.
If AI data centers are anxious to connect to the grid, flexibility and ability to curtail should be a pre-requisite. And treated cautiously.
Grid operators such as PJM don’t have the availability to enforce precise real-time load curtailments for individual data centers in real-time, so the system risk is large.
If only 10% of forecasted data loads don’t curtail power during a grid emergency, the shortfall could cost billions.
The PJM Independent Market Monitor states that data load bring their own new generation, which could speed up interconnection. Without it, they should be curtailable before other current demand side customers, and not be paid as demand response – this should be a pre-condition for interconnection.
In summary, the Nvidia Emerald AI and Google announcements on flexible load are interesting, and the economic incentive and technical potential may be there. But we don’t know how it will work, at what scale, and for how long. Until we do know, significant skepticism and caution is warranted.