Shift or curtail? The same data-center flexibility is worth 19% of one grid and 1.2% of another

The same act of flexibility is worth vastly different things depending on where in the world a data center sits. In the US mid-Atlantic, letting a data center run its compute elsewhere can trim the whole region's electricity bill by nearly a fifth by 2038. In South Korea, the identical flexibility barely moves total cost at all. Yet it quietly makes extra solar panels worth building that otherwise would not be. That is the core, counterintuitive finding of a new study by engineers at Johns Hopkins and the Korea Institute of Energy Technology (Khanal et al., 2026): the value of data-center flexibility is not a property of the data center. It is a property of the power grid it plugs into.

The stakes could hardly be larger. Data centers have stopped being a marginal footnote in electricity planning and become a leading driver of new grid investment. Hyperscale campuses now request gigawatt-scale interconnections, and utilities attribute most near-term peak-load growth to them. Governments are scrambling to keep up. But here is the tension the paper digs into: if planners treat all data-center demand as fixed and inelastic — as an immovable block that must be served the instant it appears — they will build generation and transmission for a hypothetical worst-case peak. Much of that capital may never be needed, and its cost gets recovered from ordinary electricity customers who had nothing to do with the AI boom. Flexible operation is the obvious antidote. The catch is that "flexibility" is not one thing. It is at least three, and their worth shifts depending on where you deploy them.

The researchers classify data-center load the way a grid operator sees it — by how its power withdrawal from the grid is shaped, not by whatever internal software trick created that shape. Firm load must be served immediately. Flexible load can be shifted in time (run a training job at 2am instead of 6pm) or shifted in space (run it in a data center hundreds of miles away). Interruptible load can be cut near-instantly — by pausing compute or firing up on-site generation — within strict limits on how deep, how long, how fast, and how often. The genius of this framing is that many different physical mechanisms collapse onto the same three tiers. Throttling a chip and discharging a backup battery both show up to the grid as the same thing: less power drawn, right now.

The Science

To find out what each tier is actually worth, the team built a capacity expansion model — the planning tool utilities and regulators use to decide how much new generation, storage, and transmission to build, and where. They embedded the three flexibility tiers into that model and, crucially, held the entire formulation fixed while varying only the host grid. This is the methodological trick that makes the comparison clean: the model, cost structure, and reliability accounting are identical; only the input data differ.

The two test systems were chosen to be nearly perfect scientific controls. PJM — the huge grid operator covering parts of a dozen US states, including Northern Virginia's Dominion zone with the world's densest hyperscale data-center concentration — is market-organized, fossil-heavy, and governed by a patchwork of state clean-energy requirements. Korea is centrally planned through its national Basic Plan for Electricity Supply and Demand, has a single national carbon cap, is pivoting hard toward solar, and is phasing out coal. Two systems of similar scale, peak demand, and population, facing similar rapid data-center growth — but structurally different in the ways that determine how flexibility pays off. Data-center demand in PJM is forecast to hit 16.3 GW by 2028 and 74.8 GW by 2038; in Korea, 4.9 and 8.9 GW, respectively.

For each grid, the team solved four cases: an all-firm baseline, an all-flexible case, an all-interruptible case, and a "realistic mix" of 30% firm, 50% flexible, and 20% interruptible — proportions that reflect industry estimates that a minority of workloads are must-run, about half can wait or run elsewhere, and a smaller share can be curtailed. They simulated both a near-term 2028 horizon, which captures real interconnection and supply-chain constraints, and a long-term 2038 horizon that allows full co-planning. Reliability is enforced using the industry-standard effective load-carrying capability, or ELCC, accounting, which credits each resource — including flexible and interruptible data-center tiers — only with the firm capacity it genuinely delivers at moments of system stress.

What They Found

The headline result in the US is striking. In PJM 2028, the realistic mix cuts total annualized system cost by 6.34%, even though operating cost rises slightly. By 2038 the saving grows to 19.43%. The money comes almost entirely from deferring capital construction. Gas combined-cycle build falls from 14.09 GW in the all-firm case to 6.37 GW when all load is flexible — a reduction of 4.4 GW by 2028 — and PJM avoids 8.9 GW of gas and nuclear generation by 2038.

PJM 2028: flexibility cuts total system cost

Label	Value
All-firm baseline	34,090
All-flexible	31,190
All-interruptible	27,269
Realistic mix	31,929

PJM 2028 total annualized cost ($M), from Khanal et al. (2026) Table 2. Base 34.09 $B adjusted by reported percentage changes.

For a region convulsing over who pays for the AI build-out, that is real money and real avoided construction.

But the mechanism is the surprise. In PJM, flexibility's value is spatial. The model shifts 0.91 TWh of load between zones in 2028 — the realistic case relocates 1.3% of flexible-tier energy across regions — and this relocation, not any clever rescheduling of hours, is what avoids the new gas. The shifted workload concentrates into ComEd (Chicago's utility) and is drawn fairly evenly from Dominion, PPL, and PSEG. In 2038 this spatial shift balloons to 11.83 TWh under the realistic mix. The reason is capacity accounting: firm hyperscale load in the congested Dominion zone forces capacity and compliance obligations there; moving that load to a cheaper, less constrained zone sweeps those obligations away.

The energy value — shifting between hours of the day — is nearly irrelevant in PJM; the locational value is everything.

Korea tells a completely different story. There, the same realistic mix saves just 0.26% of total cost in 2028. The mechanism is temporal, not spatial: load is shifted into midday hours to line up with the daily solar cycle. That seemingly trivial scheduling adjustment makes 0.5 GW of additional solar PV worth building in 2028 — PV grows from 7.70 GW all-firm to 8.20 GW in the realistic mix — and by 2038, when a carbon cap binds and solar is at its expansion limit, flexibility avoids 1.2 GW of gas and 0.3 GW of batteries.

The same flexibility is worth 19% in PJM, 1.2% in Korea

Label	Value
PJM 2028	6.34
PJM 2038	19.43
Korea 2028	0.26
Korea 2038	1.24

Total-cost savings from the realistic flexibility mix (% change vs all-firm), Khanal et al. (2026) Table 2.

Korea's near-term need is not to relocate load away from congested spots, as in PJM; it is to reshape when load appears so cheap solar can be used instead of gas. The value of flexibility tracks the character of the grid it serves.

There are three other findings worth holding onto. First, fully interruptible load is the cheapest tier in both systems — in 2028 PJM it cuts cost by 20.01% and in Korea by 16.35%, dwarfing every other option. But that outsized value evaporates under realism. When the model imposes event-shape limits — caps on how deep, how long, and how often curtailment can happen — and the opportunity cost of idle hardware, the all-interruptible case's value collapses. This is the paper's most cautionary result: the temptation to treat data centers as giant dimmer switches is real, but the realistic physics of curtailment reins it in sharply.

Interruptible load is cheapest — until real limits apply

Label	Value
All-firm	0
All-flexible	0
All-interruptible	20.01
Realistic mix	6.34

PJM 2028 cost savings by case (% change vs all-firm), Khanal et al. (2026) Table 2.

Second, the value of each tier relative to the others is host-grid dependent — there is no universal "best" flexibility mechanism to procure. And third, the realistic mix in 2038 PJM raises operating cost by 5.83% even as it slashes total cost, a reminder that deferring construction and running the grid differently are two different games being played at once.

Why This Changes Things

The practical implication is that grid operators, utilities, and regulators should stop asking "how much flexibility is data-center load worth?" and start asking "what kind, where, under what rules?" The answer depends on whether the host system is congested in space (like PJM) or constrained in time (like Korea's solar-heavy, centrally planned system). This matters enormously for policy in both countries.

In the US, PJM is currently the epicenter of a furious debate over flexible interconnection. FERC issued show-cause orders in June 2026 requiring all six jurisdictional regional grid operators to justify or reform their large-load tariffs and develop new transmission services for flexible loads. Texas's Senate Bill 6 already requires large loads to install remote-disconnect capability and curtail during emergencies. These reforms are contested — mainly over whether loads can hold firm uptime via on-site generation — and the paper's spatial finding suggests they should focus on giving data centers the ability and incentive to run elsewhere during congestion, not just to shed load. A data center that can shift work to another zone is worth more to PJM than one that can merely pause, because it relieves capacity obligations where they bite.

In Korea, the lesson is different. Central planners have focused on siting incentives to push new data centers out of the Seoul metropolitan area, where 60.4% of them now sit amid congestion and grid-stability constraints. The paper suggests an equally valuable lever: let data centers shift their consumption into the solar-rich midday window, which makes clean capacity worth building without new transmission. Korea's 11th national plan projects 6.2 GW of data-center peak demand by 2038 — roughly 8.9 GW of contract capacity — and it has yet to develop any non-firm interconnection product. This study makes the case that doing so could be the cheapest way to accommodate that growth.

There is also a broader conceptual point. Most capacity expansion models reduce data-center flexibility to a single mechanism or study mechanisms in isolation. This paper shows why that is a dangerously incomplete view: the three tiers are not interchangeable, and a model that assumes only one exists will produce systematically wrong answers about what to build. The finding that event-shape limits crush curtailment's value is a direct warning to planners who model interruptible load as an unlimited dial. When you impose the real disciplines of depth, duration, ramp, and recovery, the "free" curtailment option stops being free.

A prior system-level study found PJM alone could host roughly 18 GW of additional flexible load at a 0.5% annual curtailment rate — but that assessment treated new load as a fixed block, ignored transmission limits and plant start-up constraints, and did not co-plan generation and storage. This paper closes exactly that gap, and its numbers are the payoff: 4.4 GW of avoided gas in 2028, 8.9 GW by 2038, 1.2 GW of avoided gas in Korea by 2038, and millions of dollars. The nuance — that curtailment's headline value is real but fragile — is the honest part that makes the whole thing credible.

What's Next

The model holds flexibility parameters fixed across both grids by design, which is what makes the comparison fair — but it also means the numbers are directional rather than a precise road map. Real costs of shifting and curtailing compute vary by data-center operator, hardware, and workload, and the opportunity cost of idle GPUs is a moving target as chip prices and utilization shift. The 2028 and 2038 snapshots are deliberately separate rather than a continuous trajectory, so the study does not say exactly when flexibility pays for itself along the way.

What it opens up is more interesting than what it settles. The results flag that the preferred flexibility mix changes with grid structure, interruption limits, and policy constraints — which invites the obvious follow-up: design flexibility products and interconnection tariffs per grid, tuned to where and when that grid is tight, rather than adopting one national template. For the US, that means asking which zones benefit from which tier and building spatial-flexibility products around them. For Korea, it means testing whether temporal-flexibility products can substitute for storage. The paper also points toward the deeper question of who gets paid for this flexibility and how — the value it computes is system-wide, but the costs (and benefits) land unevenly on data-center operators, utilities, and ratepayers.

Most of all, the study is a reminder that the AI-driven electricity surge is not a monolith to be fought but a shape to be understood. The same data center, run the same way, is worth 19% of one grid's cost and 0.26% of another's. The difference is not the machine. It is the grid — its congestion, its fuel mix, its policies, its solar hour. Flexibility should be designed like a key, shaped to fit the lock it is meant to turn. The engineers behind this paper have handed planners in two very different countries the precise dimensions of their own locks.