Source: S&P Global Media Portal/S&P Global Inc.
As speed to power continues to be the main talking point in the data center industry, a clear road map as to how one might boost their speed on the journey to power is needed. A new study by Camus Energy, encoord and Princeton University ZERO Lab — funded by Google — aims to do just that. In many ways, this study is a spiritual successor to the “Rethinking Load Growth” study by the Nicholas Institute for Energy, Environment & Sustainability at Duke University, which showed that a little bit of flexibility by data center operators could open significant headroom in terms of capacity. This study further asks, “How can various flexibility options be used, and to what benefit at real sites in the US?” Google also looks to have applied many of these principles in a pending contract at one of its latest data center campus expansions.
The Take
The idea that grid flexibility by data centers could open “new” capacity swept the industry in 2025. The natural next step from this idea is to wonder what kinds of flexibility can help, and to what extent. While the pool of similar site selection methodologies is finite, the study does provide a blueprint for faster interconnection times. The path itself, however, will be highly location-dependent and rely on information from utilities for identification. In many ways, the methods identified in the study are a means to sort through the pockets of capacity left around the US. Google’s newest attempt to file a contract with utilities may provide a blueprint (albeit redacted) for how these deals will look going forward. Analysis of this newest study, alongside a working contract, should provide an idea of what form custom deals between utilities and large-load customers may take.
Background and definitions
As outlined by the study, there are two main bottlenecks to grid connection: generation constraints and transmission constraints. Generation constraints are defined as a lack of accredited capacity available to reliably meet existing peak demand, capacity reserve margins and the incremental demand of a new data center. Transmission constraints are defined as situations when lines and substations cannot safely carry additional power without exceeding thermal, voltage or contingency limits.
The authors suggest that the way to avoid these constraints and subsequent delays is to combine a flexible grid connection with bring-your-own-capacity (BYOC). In the article, a flexible data center is one that receives both firm and conditional-firm service. During times of constraint, when some or all conditional-firm capacity must be redirected, the facilities will use on-site or colocated resources, such as batteries, solar, gas generators or compute flexibility, to stay within grid limits.
BYOC is a larger umbrella term that refers to accredited capacity through contracts with new or in-queue resources, such as solar, wind, storage, natural gas, nuclear, virtual power plants or on-site flexibility, to avoid portfolio expansion by the utility. Importantly, the model assumes that new capacity from building generation resources requires roughly five years to navigate the interconnection queue, so the models in the study will use BYOC contracts for generation that is already in the interconnection queue.
Overview and general findings
The study modeled six real sites selected as candidates. The authors also claim the analysis as the first to combine real utility transmission system data, system-level capacity expansion modeling, and site-level capacity optimization to evaluate the relationship between flexibility and speed to power that has been made public.
The study modeled the feasibility of a 500-megawatt nameplate capacity data center at six locations across the PJM Interconnection territory, all within 125 miles of each other. There was a high degree of site specificity with significant variations in available grid capacity. Two of the six sites were able to achieve grid connections without delays, showing that there are pockets remaining with sufficient transmission capacity. These sites were the only two sites proximal to a 500-kilovolt transmission backbone, which consisted of a substation with 500 kV/230 kV and a junction of both 500 kV and 230 kV transmission lines. The other four sites proximal to a 230-kV backbone — consisting of a substation with 230 kV/69 kV voltage levels and a junction of two to four 230-kV transmission lines — all saw some degree of constraint. All sites fell below 1% of the year in terms of curtailment, ranging from 7 to 35 hours total.
The overall general finding is that the utilization of a cost-optimal combination of flexible connection and BYOC, as identified by the model, can shorten interconnection time from five to seven years to about two years. Financially, each megawatt of capacity represents $4 million to $12 million in annual revenue. A site that reaches grid connection three years earlier would generate between $2.3 billion and $3.2 billion across that time. Even with the addition of $1.2 billion to $1.4 billion in life-cycle costs from flexibility and BYOC, positive return on investment becomes evident with an increase of two years in terms of time to power.
Site-specific cost analysis
More detailed analyses were conducted on the sites with the largest (Hare) and smallest (Koala) degrees of constraint. Site Hare was only able to obtain 154 MW of the 500 MW in firm capacity and had to meet a curtailment peak of 281 MW, with 65 MW of headroom and total curtailment of 35 hours per year. In other words, 346 MW of the 500 MW total must be dispatchable for 35 hours of the year. Site Koala was able to obtain 326 MW of the 500 MW in firm capacity and had to meet a curtailment peak of 109 MW, with 65 MW of headroom and total curtailment of 20 hours per year — or 174 MW of the 500 MW total must be dispatchable for 20 hours of the year.
Each site was able to reach interconnection in about two years through portfolios of off-site power purchase agreements (PPAs) and a mix of battery storage, generation and compute flexibility. Overall, the model estimated that each gigawatt of new data center demand equated to about $764 million in supply costs. Using 20% conditional-firm energy can avoid 273 MW of a new build, which corresponds to savings of about $78 million in incremental costs per gigawatt. Essentially, this allows the site to receive power without extensive buildouts, which reduces time to power and the effect on other customers.
Beyond the cost to upgrade transmission infrastructure, the modeling approach also internalizes capacity costs. Recall that the model assumes no extra capacity beyond projects in the queue, so all firm power must be procured from PPAs, virtual PPAs or on-site resources. The model estimates that BYOC can potentially internalize $326 million in capacity costs per gigawatt. The study further claims that flexibility, BYOC and the overall energy costs can cover 91% to 101% of incremental supply costs of the utility.
Key takeaways and limitations
Overall, the study’s findings show that the combination of a flexible approach and BYOC can reduce time to power by three to five years, with this being true by virtue of the data center operator procuring all power via either on-site or accredited in-queue generation. Across all those components, a data center operator can nearly eliminate all net transmission and generation infrastructure upgrades associated with its load.
The authors put much of the onus for improvement on the utilities due to proprietary data. As it stands, that information is opaque, forcing developers to site projects with no clear view of where grid headroom exists. Of course, the utilities might tell a different story. One of the actionable findings of this study is the importance of a proximal 500-kW backbone, both transmission and substation, to attain interconnection in a timely manner. The authors suggest that utilities could help unlock faster, better-informed siting decisions by sharing data that reflects how available transmission capacity fluctuates both spatially and over time.
One limitation to the modeling approach is that PPAs and VPPAs within the queue, which are set to be operational in two years or less, represent a finite resource that will vary significantly depending on location. Once these resources are depleted, on-site power or load shifting will be the only timely option. It is also the case that this approach will be best with AI training facilities, as they are most amenable to large percentage-load flexibility. Sites that must be ready to work at full capacity at any given moment will presumably see more limited benefits from the specified approach.
The model is also limited in its focus on cost optimality. Of course, cost optimality is important, but it would also be helpful to see where the cost optimality ends and what, if any, specific variables emerge as negative drivers. When the PPA pool runs dry for short-term interconnection, what do the numbers look like for fully behind-the-meter options? It is important to know if these findings would hold up in areas with little PPA availability.
What is an AI Datacenter?
Want insights on datacenter trends delivered to your inbox? Join the 451 Alliance.
