Nvidia and AI Infrastructure Executives Push Grid Rules After 3,100-MW Demand Drop

A routine transmission fault can become a much larger grid event when data centers abruptly leave the system. After roughly 3,100 MW of data-center demand disappeared during a July 22 Northern Virginia line failure, executives from Nvidia, Heron Power, Invenergy and Emerald AI are calling for grid-access rules that tie AI facilities’ interconnection terms to how they behave during stress.

The proposal lands as federal regulators examine the rules for connecting very large electricity users. It is an industry prescription from executives whose companies span AI computing, power electronics, generation and grid-response software. Their central argument is that a facility should be assessed not only by how much electricity it needs, but by its demonstrated behavior when the power system is strained.

The failure mode is synchronized disconnection

On July 22, the Northern Virginia line failure led facilities to switch to backup power, removing about 3,100 MW of demand within seconds. PJM said the event did not affect bulk-system reliability. The executives’ article also points to an incident two years earlier in which roughly 60 Northern Virginia data centers disconnected and transferred to backup power.

Roughly 3,100 MWDemand removed after the July 22 fault

Data centers switched to backup power within seconds after a Northern Virginia transmission-line failure, removing roughly 3,100 MW of demand.

A three-part test for new connections

• Avoid worsening disturbances by riding through them instead of adding an abrupt loss of demand.
• Flex consumption during system stress rather than treating data-center demand as fixed.
• Help bring generation, storage and transmission into service as part of the wider grid.

The first point targets both disconnections and volatile computing loads. The authors say workloads can move a facility’s draw by tens of megawatts in seconds. They propose batteries that keep GPUs supplied through a voltage disturbance and can export power to support grid stability, rather than taking an entire site off the system.

Flexibility is the economic case

The proposal’s demand-flexibility case rests on cited estimates, not operating results from these facilities. A Duke University estimate says large new loads cutting use for less than 0.5% of annual hours could let existing infrastructure absorb roughly 100 GW of added demand. A Brattle Group estimate puts a 10% gain in grid utilization at nearly a 4% reduction in average retail rates.

The policy decision is about proof

Co-located generation and storage can serve as bridge power while a fuller grid connection is completed, the authors say. Their distinction is whether those assets ultimately join the shared system or remain permanently isolated for one campus.

The executives want performance-based interconnection rules rather than equipment mandates: facilities demonstrating ride-through, smoother demand, grid support and verified flexibility would receive faster connections and better terms. FERC’s June 18 orders directed all six regional grid operators to justify or reform large-load tariffs, creating the immediate policy venue. The consequential test is whether tariff rules can turn those promised behaviors into enforceable requirements.