Google, Nvidia and Emerald AI launch grid alliance citing claim about 100 curtailed hours a year

The person this lands on is whoever takes the utility's call and decides what stops running. Tyler Norris, Google's head of energy market innovation for AI and infrastructure and the alliance's inaugural board chair, wrote that the capacity can come "via compute flexibility, storage, or generation" ^[3][7]. Storage and generation are equipment purchases with a delivery date. Compute flexibility means a job that was running is not running, and somebody owns that job.

A hundred hours is 1.1 percent of an 8,760-hour year ^[8]. Applied to the 1GW Google says it has already integrated, 100 hours of full curtailment comes to 100GWh of consumption dropped or moved ^[9][4]. The alternative Norris names is 5-7+ years of waiting for traditional system upgrades ^[3].

The "dozens of gigawatts" belongs to "multiple studies" that Norris cites in the LinkedIn post without naming them ^[3]. The part with signatures behind it is Google's own: 1GW folded into long-term energy contracts with US utilities as of March ^[4], starting with a demonstration at Omaha Public Power District in 2024 ^[13], then agreements with Indiana Michigan Power and the Tennessee Valley Authority ^[14], and now Entergy Arkansas, Minnesota Power and DTE Energy ^[15], six counterparties named in all ^[16]. Google's 1GW is the only capacity figure the DatacenterDynamics account attaches to any member ^[10].

The two reports Google released alongside the launch point in different directions. The Brattle Group wrote a technical blueprint for turning recent bipartisan direction from the Federal Energy Regulatory Commission into operational reality ^[11]. Aurora Energy Research modelled flexible data centers in ERCOT paired with front-of-meter resources ^[12]. In the Aurora case, some of the flexibility sits out in those front-of-meter resources, beyond the scheduling decisions made inside the building.

For an operator, this is a sizing exercise. Pull last year's load curve, mark the hundred hours the utility would most plausibly have called, and look at what was running in them. If it was batch training you can checkpoint and restart, the flexibility costs delay and nothing else. If it was customer-facing inference under a service commitment, the flexibility has to be bought as batteries or on-site generation, and the comparison becomes that capex against 5-7+ years in the interconnection queue ^[3].