There’s a buzz going around the AI infrastructure world that the term “flexibility” in interconnecting to the grid is bad for business. The misconception is that this means the utility will pull the plug on the compute workload, and therefore nobody should seriously consider this as an option.
But this is not what a flexible grid connection does. Anyone evaluating this method as a path to power deserves a clear picture on what operations actually look like in a flexible grid.
In fact, across the six data center sites we modeled in our flexible interconnection study with Google last December, on-site resources ran 40 to 70 hours per year on average. The compute stack kept running the whole time.
Let’s dispel the myth by starting with the basic fact that the data center, not the utility, decides how to flex.
A flexible interconnection contract requires the data center to stay within a defined grid import limit for a small number of hours per year. How that’s achieved is the data center's choice.
The options are well-established: on-site battery storage, on-site gas generation, off-site accredited capacity (solar, wind, virtual power plants), and optionally, compute shifting for operators who want it. In our study, the optimal site-level portfolios mixed up to 155 MW of battery storage with up to 48 MW of gas generation, selected by the optimization model to minimize cost against each site's specific constraint profile. Compute flex was available as a potential lever at every site, and at almost every site the economics favored putting the flexibility load on batteries and generators.
These options are being deployed in the field today. Calibrant provides battery storage at aligned data center sites. Verrus successfully demonstrates how it’s flexible data centers can be fully grid responsive. These are proven assets the industry already trusts to deliver power, and they are the same assets that meet the brief, less than 1% annual flexibility needed when a utility is willing to offer conditional firm service.
The lowest-cost choice in almost every case is a portfolio built around physical infrastructure the site was going to install anyway.
The operational mechanism matters. Most descriptions get the process wrong, suggesting the utility can simply dial down a data center's power or force an immediate load drop. This is simply not the case.
In reality, the utility sends a capacity signal to the data center's energy management system. That signal says something like "in X hours, the grid can deliver Y megawatts to your site." The data center's own systems decide how to respond with a combination of batteries, generators, compute load shifting, or other constructs that have been designed to fit the flexible stack.
The utility never reaches across the point of interconnection to dictate what happens in order to reduce the load.
There are some operators who do choose to include compute flexibility in their portfolios, and the public evidence suggests the customer impact is zero.
Google has published detailed accounts of how its hyperscale platform shifts non-urgent workloads across regions and time windows to match available clean power and grid conditions. The program started as a carbon-matching effort and has since become a general tool. Emerald AI's recent field demonstration showed that an AI compute cluster can reduce peak power demand substantially while continuing to serve priority customer workloads without material impact.
Both examples show that when compute is in the flexibility stack, it is because the operator designed it that way, built the orchestration software to handle it, and chose that pathway because it lowers the cost of the on-site portfolio.
For operators who do not want compute in the mix at all, the Camus-Google study result is the clearest data point we have: up to only 70 hours of on-site dispatch per year, across six modeled sites, with the compute stack untouched and grid power available more than 99 percent of the year.
If you’re looking into flexible interconnection and the conversation has drifted toward whether the utility will interrupt your customers, the conversation is off track.
Flexible grid connection is simply a faster path to firm power than waiting for a traditional interconnection in a congested queue. That timeline difference is the competitive variable. It is the difference between a facility that comes online in time for the next wave of AI demand and one that does not.
The questions worth asking are: How many annual hours of flexibility are required? How much advance notice does the utility provide? What fail-safes govern the contract? And what on-site resource mix produces the lowest total cost of flexible capacity for your specific load profile?
These are solvable questions that Camus can answer. If you are working through interconnection options and want to understand what a flexible approach would look like for your site specifically, reach out to me, Laura Cummings, directly at lcummings@camus.energy. That is the conversation worth having.
