AI training clusters and hyperscale campuses require power that is continuous, high quality, and delivered at speed. Those demands are colliding with a grid designed for gradual load growth and planning cycles measured in years. When utilities effectively say no, developers still face an unmovable business imperative to build. As a result, a growing share of projects are advancing as vertically integrated sites, equipped with on-site generation and islandable microgrids.
“Data center developers, builders, utilities and ISOs have maybe talked past each other a little bit,” says Tad W. Piper, founder and president of TWP Strategic.
When “No” Becomes a Design Choice
Piper frames the current standoff as a communications failure with real infrastructure consequences. Utilities and independent system operators have tended to treat interconnection requests as binary, approve or deny, while developers treat power as a solvable engineering problem with a deadline. “A lot of utilities believe that if they said no, that that was kind of the end of the road and the data center providers said well I’m just going to go buy my own turbines and put my own power on site,” he says.
This “through or around” dynamic is now shaping the market. Piper notes that “about a third of data center development projects are proceeding as vertically integrated data centers plus power solution with their own microgrid.” In those cases, the facility is not simply a customer. It becomes a self-supplied energy system with the option to interact with the grid only on its own terms.
The irony is that the move to self-supply is neither an ideal outcome for utilities nor for developers. The grid loses a major load it could have served and planned around. The developer inherits a new operational competency it did not set out to own. “Most large data centers didn’t really want to get into the energy business,” he says. “They all wanted to be grid connected even as recently as a year ago, and now they’re going it alone.”
Flexibility Is Not a Daily Problem, It Is a Peak Problem
Part of the misunderstanding is the way flexibility is discussed, often treated as a day-to-day modulating capability, like a thermostat for a building. For data centers, the operational reality is different. “It’s more about peak grid demand flexibility and oftentimes it’s understanding what the utility is trying to solve for.”
A 24/7/365 load is, in effect, a request for the grid to match generation one-to-one with incremental demand. Utilities may hesitate because they are not always convinced the load will materialize at the scale and timing promised, especially when the interconnection queue is crowded with speculative projects. Developers respond by demonstrating commitment the only way they can, by writing checks. “You need no further demonstration of your commitment you’re going to show up than when data centers build their own generation,” Piper says. “They’re taking the capital risk.”
In a recent analysis for Energy Central co-authored with Peter Asmus, Piper contrasts hyperscaler spending with the utility sector’s ability to keep pace, arguing that technology companies operate on months-to-years timelines while utilities and regulators operate on years-to-decades cycles. The mismatch matters because data center investment decisions do not pause while grid planning catches up.
The Technical Playbook: Generation, Storage, and Dispatchable Load
Can data centers actually support grid reliability without compromising uptime? “In many instances, yes. In certain instances, no,” he says. The yes case is built on two levers that are becoming standard design assumptions: on-site generation and batteries. Data centers are “running quickly towards both,” Piper notes, because they provide controllability during tight grid conditions and reduce dependence on interconnection timelines.
The no case emerges during extended peak events. Batteries can cover a typical late-afternoon peak on a hot day, when demand rises and then abates as businesses close and solar output declines. The stress test is multi-day scarcity. Piper points to winter events in Texas and elsewhere as the template: “When your battery runs out, for example, after four hours, what are you prepared to do at that point in time?”
In those scenarios, flexibility becomes an operational plan, not an asset purchase. If an event is forecast, developers can prepare by lining up extended use of backup generators, fuel logistics, and load-shedding protocols. They can also reduce computing load by shifting workloads geographically or conducting an orderly reduction. “Can you reduce your computing load by either shifting it to another geographical location or doing an orderly reduction in that load, stopping models from running, reducing inference workloads, etc.”
He adds that during extreme events, customer tolerance for certain service degradations may increase as other systems shut down. The strategic question becomes which services must remain fully available and which can be throttled without unacceptable consequences. As Piper notes, “If energy costs increase 100x versus normal operating costs, are there steps operators would be willing to take to temporarily reduce energy usage?”
Design Flexibility Before the First Server Arrives
For developers, flexibility has to be built in from day one, not bolted on later. “Every data center developer needs to develop a proactive metering plan. And microgrid and dispatch controls,” he says. Translated into execution, that means installing the instrumentation and software controls required to measure, verify, and dispatch load and behind-the-meter resources. Participation in flexibility programs depends on measurement and verification, often referred to as M&V, because compensation requires proof.
Piper emphasizes the economics: for a 500-megawatt facility, additional meters and control devices are “a rounding error relative to the capital expenditures that you’re making.” The alternative is costly and disruptive retrofits. “Trying to retroactively do that is quite costly and disruptive,” he says, noting that adding meters after commissioning can require power shutdowns.
Utilities and ISOs have a role here as well, and Piper argues they should publish best practices for metering and controls so large-load sites can design to interoperability standards, lowering friction and accelerating program participation.
The Business Model Shift That Makes Collaboration Work
Policy is moving, but the pace is uneven. For example, FERC Order 2222 is a federal rule issued by the Federal Energy Regulatory Commission that allows distributed energy resources such as storage and demand response to participate in wholesale power markets. It’s a foundational step for distributed energy resource participation, even as implementation across regions stretches out for years. Meanwhile, grid operators like ERCOT are exploring flexibility requirements as a gating factor in interconnection, with large loads crowding the queue.
For developers, the incentives have to be explicit. First, does flexibility move a project “to the head of the line” so it gets built? Second, when the site is called upon, “am I getting paid appropriately to offset my cost of putting that infrastructure in?” Third, what are the technical requirements, from communication protocols to controls and metering, that make settlement possible?
The broader implication is that the utility sector cannot rely solely on the traditional rate-base model if it wants to keep these loads on the system. “Utilities, regulators and ISOs” need to recognize that technology timeframes are faster, Piper says, and evolve toward being rewarded not only for capital deployed but for efficient service delivery. The alternative is continued load defection: developers will keep building around the grid.
Piper sees a pragmatic endpoint. Data centers want a pathway back to the grid, but “they’re just not willing to wait.” Collaboration becomes less a slogan and more a necessity, because both sides have a strong incentive to avoid duplicated infrastructure and stranded investment. If utilities can offer clear pathways, standardized controls, and fair compensation, data centers can become a reliability resource rather than a stressor. If not, the industry will keep solving the problem alone, one microgrid at a time.
Follow Tad W. Piper on LinkedIn or visit his website for more insights.
