Is a Data Center Really an Inflexible Load? — What Can Move and What Cannot

Choosing Sites in Advance Cuts Down the Years

The power grid is a system connecting power plants and electricity-consuming locations with power lines to run as a single whole, and the industry calls it the grid. There is a study that calculated what happens if you choose potential data center sites in advance and grade them on a virtual grid modeled after the Texas grid, namely a synthetic grid. This synthetic grid consists of 2,000 buses. A bus is a junction in a grid model where generators, lines, and electricity consumers all meet. The order of calculation is as follows. The grid operator scans candidate sites to filter out those that fall short of reliability standards. For each remaining site, it calculates in advance what happens to the entire grid if a facility comes in under specific conditions, and ranks them.

Instead of assuming data centers use a fixed amount of electricity all the time, when they were allowed to adjust their operation, the range of available sites widened by 9 to 17% for a 1-gigawatt scale, and by 19 to 21% for a 2-gigawatt scale. Electricity prices barely moved. In the 2-gigawatt case, the median of the average prices across all hours of the day was $24.32 per megawatt-hour—a unit for measuring the amount of electricity—which was effectively unchanged, and under the condition of shifting operating hours, the spread of prices during the busiest hours decreased by about 3.4%. The point raised by the authors is time. First energization is when a new facility receives electricity from the grid for the first time. The authors estimate that using pre-certified condition tables makes first energization possible in 12 to 18 months by skipping major transmission grid reinforcements. The conventional procedure the authors used as a comparison benchmark, which is the process where a developer applies first and proceeds, takes 5 to 8 years, and the calculation is that even with conservative estimates, this is 3.5 to 4 years shorter (arXiv:2605.14714v1).

Two caveats come with these numbers. First, these values came from simulations on a synthetic grid modeled after the Texas grid, and are not values measured on the actual Texas grid. Most of the values in this article are like that. Many come from models mimicking the US grid or standard models used by researchers, and the Korean grid appears in one study as a comparison counterpart to the US grid. Therefore, for each value, we record together which grid it came from and under what assumptions.

Second, this entire proposal rests on a single assumption. It is the assumption that data centers can choose one of three condition tables. They can fix their power use to always be the same, pause temporarily when requested, or shift the hours during which they use power. What expanded the pool of site candidates was not grid reinforcement, but this assumption that places data centers as a selectable party.

This quality of being able to choose has a name. Demand flexibility is the degree to which electricity consumers can change when and how much they use. If it does not matter whether you run the same laundry tonight or at dawn tomorrow, that laundry is flexible demand. An operating room light that is only meaningful if it is turned on at this exact moment is not flexible. The object moving here is the load, namely the power that the power grid must supply at that very moment. Demand flexibility refers to how far that load can be pushed toward time or toward location.

Where Was the Notion That You "Can't Reduce It" Written?

Electricity consumed by data centers in the U.S. accounted for 4.4% of total power consumption in 2023, and is projected to reach 6.7 to 12% by 2028 (arXiv:2605.18517v1). Not only is the amount large, but the pattern of use is demanding. Data centers run heavily around the clock and cluster tightly in specific regions. A concern has been raised that these two characteristics overlap to strain the remaining capacity of the distribution networks that distribute electricity down to neighborhoods and buildings, forcing data centers to compete for space with other demands like housing developments or EV charging (DOI: 10.1093/ooenergy/oiad014).

For a long time, therefore, this load was treated as something "incapable of being reduced." The reason sounds plausible: servers run 24 hours a day, operators have promised response speeds, and delaying computation means losing money.

However, the place where this notion was written down was in planning models. One study addressing the problem of plugging large loads into the grid diagnoses the starting point like this: treating data centers as rigid, inflexible loads has been the traditional framework of grid planning research, and that framework is now becoming increasingly difficult to sustain economically, mathematically, and operationally (arXiv:2604.05376v1). The statement that "data centers cannot be reduced" was closer to a default setting that planning models chose when including data centers than it was to a law of physics.

Settings can be changed. The remaining question is what actually moves along when you change the setting.

What is Being Moved — Postponing and Moving Are Different Questions

There are batch reconciliation jobs left to run overnight and checked in the morning, and there are search buttons just pressed. The former can wait, but the latter cannot. Delay-tolerant jobs are computations that do not need to finish right now—in other words, work where the person waiting for the result is not sitting in front of a screen. When the schedule of only such jobs—termed "best-effort" in the paper—was adjusted on a 2,000-bus power system model of the United States (ACTIVSg2000), those jobs migrated to times when electricity was cheap. These were mostly periods with high renewable generation. Jobs that people wait for, which are sensitive to response latency, were left untouched, maintaining service quality. Greenhouse gas and pollutant emissions fell, and transmission congestion—a state where electricity cannot be fully delivered because lines hit their capacity limits—also decreased (arXiv:2604.23128v1).

The "times when electricity is cheap" where jobs gathered also have a name. Node prices are wholesale electricity prices priced differently depending on where in the grid the electricity is consumed, even at the same time. They are cheap near inexpensive power plants, and expensive in neighborhoods where the lines carrying that electricity are full, forcing the grid to run other, more expensive generators. Node prices vary both by time and by location. The reason data centers can make money by moving workloads lies in the differences along these two directions.

Applied to artificial intelligence work, this is commonly summarized as follows. Training is the process of building a model, and inference is using a finished model once. Training can finish a few days late, but inference must deliver an answer right now; therefore, training can be postponed, and inference is what people wait for.

A model that jointly plans the computation and power of geographically dispersed AI data centers turns this summary upside down. In this model, training jobs fall on the "stiff" side, and inference jobs on the "elastic" side. Training has constraints that its start time can be delayed but it must run continuously without interruption, making it hard to stop midway or chop up. Inference cannot be delayed, but the answer is the same to the user no matter which regional facility processes it. Therefore, as long as latency limits are met, it can be sent elsewhere. In this model, a major source of benefit pointed to was the freedom to decide where to send inference tasks (arXiv:2605.03751v2, synthetic case-based simulation).

"Can it be postponed?" and "Can it be moved?" are different questions, and the same work can be flexible along one axis while being rigid along the other. Here, the time axis is when to do it, and the space axis is where to do it. Viewed through this model's classification, training is soft on the time axis and stiff on the space axis, while inference is the opposite. However, this classification is an assumption set as input by the synthetic case model. Aiming at the structural constraints involved in moving AI workloads, approaches such as migration using checkpoints that periodically save the state of a job are also emerging (DOI: 10.63278/jicrcr.vi.3723, literature review). If you lump the two together under a single word, "flexible load," what the power system can actually request becomes invisible.

Four Drawers

Laundry can be run at dawn, lighting can be dimmed temporarily, and laptops can survive on batteries. All three are called "flexible," but they do different things. Data center flexibility must also be opened across four drawers—time, location, output, and its own batteries—to see what is inside.

The first drawer is shifting time, meaning doing the same work later. A study that examined the room for postponement using 2020 data from Germany, the UK, France, and California tested what shakes the results. It is how much it can be delayed, in what order they are chosen, and how accurate the forecast is for how clean the electricity will be several hours later (arXiv:2110.13234v1). There is one premise attached to this drawer. The standard model grid that researchers use together to compare results is called a test grid. In a 24-bus test grid, the model covering time shifting is designed to preserve the total daily energy balance (arXiv:2605.18517v1). Within that design, postponed work is definitely done later. In this drawer, this means the total amount of electricity used remains the same.

The second drawer is shifting location, meaning doing the same work in a facility in a different region. In a 73-bus test grid, placing data centers without flexibility caused line overloads to rise up to 30.1%. When workloads were allowed to be moved between regions, those violations were mitigated in the scenarios examined, and the grid was established again. Solar curtailment, which is the amount of solar power that the grid could not accept and threw away, also fell by up to 61.0% (arXiv:2511.08759v2, test grid simulation).

Underneath the word "shifting" lie several physical parts. They are the procedure of pre-copying memory contents, preserving network connection states, reliance on storage structures, and checking hardware compatibility. A review paper on workload shifting for maintenance and load balancing inside corporate data centers notes that once these parts are in place, shifting actually runs. It adds that while structural constraints remain for response-time-sensitive workloads and artificial intelligence workloads, approaches like checkpoint-based shifting or topology-aware placement are emerging to address those constraints (DOI: 10.63278/jicrcr.vi.3723, literature review). Another constraint enters the power grid models that shift between regions. One model that operates computing resources and power together explicitly included the point that workload shifting is tied to the routes and bandwidth of communication lines. Even with that constraint included, numerical cases showed that inter-regional shifting reduced the mismatch between where renewable energy is generated and where work requests crowd (DOI: 10.2139/ssrn.7218438, preprint, numerical case).

The third drawer is using slightly less, meaning lowering performance on the spot without shifting. A review sweeping this field points to protection equipment, pre-cooling, adjusting processing unit operating speeds, and task sequence adjustment as pathways for flexibility (DOI: 10.1007/s40518-025-00258-9). Methods tailored to artificial intelligence inference have also emerged. This is a method of processing the same requests with less electricity by lowering the numerical precision of the model. A study modeling this as a power control means reports that operating costs for data centers fell by 34.3% without reducing the amount of tokens, which are the text fragments processed by the model (arXiv:2606.18851v1, numerical case of the proposed model). Whether a lower-precision answer is the same answer as the original remains a question to be examined separately.

The fourth drawer is using its own batteries. Data centers already have batteries. An uninterruptible power supply is a battery placed inside a data center to bridge the gap from the moment electricity cuts out until the backup generator takes over. It has the same role as the battery attached to a laptop, but it is large enough to fill a room and does nothing most of the time. Proposals to repurpose data center batteries toward the grid date back at least to 2020, in research covering microgrids, which are small-scale power grids that can run separately on a local community scale (DOI: 10.1049/iet-rpg.2020.0217). This study did not specifically single out uninterruptible power supplies.

At least carbon-aware workload placement literature has mostly opened only one drawer at a time. Among 28 such papers, 11 looked at two axes together, and 2 created optimizations combining the two (DOI: 10.3390/su17146433, review).

Price Is Determined by Speed — From $30 to $3,000

Even inside a single building, flexibility comes with a tiered price tag. In a study that broke down and priced out the electrical headroom created by scheduling workloads in a data center packed with graphics processing units, the order went like this.

The cheapest option was delaying cooling. For a short time, power could be reliably reduced for a reward of around $30 per megawatt-hour of electricity cut. Next came shifting tasks that had been slotted in to fill empty spaces, which was often possible between $30 and $300. This is where the price jumps. Rescheduling or outright postponing tasks started at $600 because of the revenue missed, and truly took off past $3,000. The same study adds one more thing. Flexibility can be created even without knowing what workloads will arrive next, but it becomes much larger when the exact queue of incoming workloads is known (arXiv:2603.27831v2, optimization model compared against a first-in, first-out baseline).

Flexibility is a commodity measured by how much of it you get for a given price. What divides the price is what sacrifice the adjustment demands. Warming up a room is cheap; delaying a customer's task is expensive.

The price has one more axis. Speed. Some of the things the grid buys are not electricity itself. Ancillary services is the collective name for jobs that prop up the grid from the side to keep it from collapsing, distinct from the business of producing and selling electricity. Frequency is a count of how many times a second the direction of alternating current swings, and a moving frequency means the amount being generated and the amount being used are out of balance. Like the job of changing output within a few seconds when the frequency wobbles, ancillary services are closer to selling the quality of electricity than its quantity. In some products of this market, how fast you react within a few seconds determines the price.

Lining them up by speed reverses the order. There is a study that built dynamic models of a large data center's cooling equipment and uninterruptible power supplies to plug into a grid stability analysis. Inter-area oscillation is a phenomenon where groups of generators in two distant regions swing out of step with each other. Demand response using uninterruptible power supplies helped calm this oscillation. Cooling equipment was fundamentally useless for that job because the speed at which heat reacts is slow (arXiv:2607.27575v1, 39-bus test system). Cooling, the cheapest resource on the price ladder, does not even make the candidate pool here. Cheap flexibility and fast flexibility are not the same thing.

On the faster end, data centers are beginning to be calculated as a resource that props up grid frequency. In a model that factored in the lowest point the frequency drops to after a contingency as a constraint, the cost-efficiency of grid operation steadily improved and the accommodation of renewable energy also increased as the share of flexible data center load grew. However, this gain showed clear diminishing returns, meaning the additional yield shrank the more you added. Therefore, the authors introduced a metric that separately measures the marginal value of adding one more unit of flexibility. At the same time, the authors note that as the data center's share of the grid load grows, its participation becomes increasingly indispensable (arXiv:2510.01050v1, baseline grid and 2030 scenario simulation). There is also a study simulating a setup that bundles data center batteries with electric vehicle batteries to prop up a microgrid's frequency, and the authors present this as a conceptual framework for future discussion (DOI: 10.1049/iet-rpg.2020.0217).

The Grid Decides the Axis

There is a study that applied the same flexibility model to two grids with different characteristics. One is the US PJM grid, which is operated as a market and has a high share of fossil fuels, and the other is the Korean grid, which is subject to a carbon cap and is centrally coordinated. The authors divided data center load into three categories—fixed, flexible, and interruptible—and put them into a capacity expansion model. A capacity expansion model is a model that calculates which power generation facilities to build in the future and in what amounts, aiming for the lowest overall cost.

The axes where gains emerged differed between the two places.

In PJM, the value was in space. Shifting work between regions reduced total grid costs by 6% in 2028 and 19% in 2038, and avoided the construction of 4.4 gigawatts and 8.9 gigawatts of gas and nuclear power, respectively. In Korea, the value was in time. Shifting load to midday solar hours made it worthwhile to build 0.5 gigawatts more solar power in 2028, and by 2038, the construction of 1.2 gigawatts of gas and 0.3 gigawatts of batteries could be avoided. These are values produced by the model, not observed events. Both sides are given the same caveat by the authors: if the shapes of events that trigger curtailment requests are realistically constrained, the value of reducing or stopping load drops (arXiv:2608.19622v1, capacity expansion model).

The same prescription works along the spatial axis in one grid and the temporal axis in another. The authors' conclusion is that the method of purchasing flexibility and its value are determined by the grid's characteristics and policy goals. The value is not determined solely by what a data center can do; rather, the type of grid and its goals determine the value together.

Other capacity expansion models point in the same direction. The flexibility of AI data centers reduced grid investment and operating costs by 3% to 21%, but those gains did not appear consistently. They varied depending on site location, the scope of flexibility, and the grid's load state, and increasing flexibility did not necessarily reduce the required power generation facilities. There were also diminishing returns to extending the allowable delay time (arXiv:2604.05376v1).

The shape of the gains also repeats. In a 24-bus reliability test grid, a model that turned on both time shifting and space shifting eliminated all normal and post-contingency transmission violations—meaning instances where lines exceeded their designated limits—at a flexibility ratio of 40%, which is the proportion of load that can be shifted. At a flexibility ratio of 30%, renewable curtailment was reduced by up to 84.4%. A sensitivity analysis in the same paper concludes that a moderate flexibility of around 20~30% already captures most of the obtainable gains (arXiv:2605.18517v1, test grid simulation).

Three studies in different grids turn different knobs to draw a similar shape. The knobs turned are the proportion of flexible load (24-bus test grid), the length of the delayable time (capacity expansion model), and the share that flexible data centers occupy in grid load (frequency constraint model). For any knob, the gain is large at first, and the incremental share shrinks the more it is added. This shape fits neither the claim that "data centers must be made fully flexible" nor the claim that "flexibility is futile."

When viewed through the lens of carbon, the axis twists once more. Marginal emissions are the emissions produced by the generator that is turned up further to make up for using a little more electricity. When regional and hourly marginal emissions in the US were estimated using data from 2007 to 2009, the upper Midwest's emission rate was more than 3 times that of the West, and even within the same region, some hours were more than 2 times higher than others. There is a fundamental tension pointed out by this study: the times when electricity is cheapest to produce are generally the dirtiest times. However, that observation belongs to 2007–2009 data, and the US generation mix has changed significantly since then, so this data cannot tell whether the same direction holds today (DOI: 10.2139/ssrn.2153929).

In the US grid model earlier, times when electricity prices were low overlapped with periods of high renewable generation, which reduced emissions together, but the old US data showed tension in the opposite direction. Shifting based on money and shifting based on carbon may not move in the same direction depending on the grid and timing. What is used as a signal changes the answer.

But in Reality, They Do Not Respond Well to Prices

A review summarizing the current state of data center flexibility research lists the pathways seen earlier, noting that adjusting task sequences is the most frequently covered one in academia. The sentences of the summary are heavy. While many researchers propose ways to increase flexibility, we know relatively little about what the scale and cost of that flexibility actually are. The review notes two more things. Some studies suggest operators are unwilling to provide flexibility. And the authors show, using actual energy use data, that data centers are unlikely to be sensitive to electricity prices except during extreme events (DOI: 10.1007/s40518-025-00258-9, review).

The same review adds a caveat for cryptocurrency mining. Because mining is work that can be switched on and off, it is inherently flexible and has been covered separately by many studies. Yet mining mostly runs on dedicated hardware rather than running in traditional data centers, and even its price sensitivity fluctuates depending on exchange rates. This means that mining, frequently treated as a representative example of flexible compute loads, may not actually be the story of the facilities this article talks about.

There is no study directly measuring why data centers do not respond well among the literature gathered for this article. However, three things are pointed out as barriers across the literature.

The first barrier is that operators do not reveal their plans. Rescheduling tasks incurs a loss in service quality, and operators are unwilling to disclose that loss. Grid operators must therefore request curtailments without being able to look into those plans. Research addressing this condition creates a method to distribute requests without eliminating the condition and without seeing the plans (arXiv:2605.19116v2).

The second barrier is the lack of public records. An interconnection queue is a line where applications to connect new facilities to the power grid line up in order and wait for review. In the U.S., interconnection queues for generation facilities are public records. Yet equivalent records for large loads like data centers requesting interconnection at a scale of hundreds of megawatts are not public. One policy brief argues that regulators should apply the transparency standards they have established for generation to the load side as well, and presents a path that is possible through tariff revisions alone without new legislation (DOI: 10.2139/ssrn.6614823, policy brief). State-level precedents are emerging in Texas and Georgia, but the brief raises the issue that a federal-level transparency standard does not yet exist for the load side.

The third is institutional barriers. Though not a story limited to data centers, a 2016 policy brief covering the spread of demand-side flexibility in Western Europe noted that its spread is blocked to a considerable extent by market and regulatory barriers, as well as technological uncertainty. It cites the point that storage devices can either complement or substitute for demand-side flexibility as an example of uncertainty. The authors' prescription is to manage adoption by systematically weighing risks and benefits. They recommend using that weighing to gauge how much investment in generation capacity and grid reinforcement—built to prepare for times when demand-side flexibility is used most heavily—can be substituted, and to dismantle the barriers preventing its spread (DOI: 10.2139/ssrn.2831868, Western Europe policy brief).

There are signals that institutions dealing with large loads are being built. U.S. regulatory authorities have begun creating dedicated tariffs that apply only to large loads like data centers. This is a tariff aimed at sharing the financial risks that arise when electric utilities build long-lived facilities for such loads, and its purpose differs from a system that pays a price for flexibility. Among such decisions made across 13 states from 2022 to 2026, a study that examined stock market reactions by gathering the first 15 final decisions confirmed by regulators (excluding bespoke customer-specific contracts) found no evidence that stock prices were repriced. Volume-side signals disappeared when the aggregation method was changed, and the authors themselves state that their conclusions are sensitive to the analysis setup (DOI: 10.2139/ssrn.7367558, event study).

A format for paying compensation has been proposed. However, it remains in the form of a proposed formulation. Research that incorporates data centers' spatial and temporal relocation flexibility into electricity market clearing—the calculation by which the market determines trading volumes and prices—treats data centers as entities that request electricity while simultaneously providing relocation services. That compensation is calculated by the difference in node prices between locations and times. The authors show through case calculations that this proposed clearing method satisfies the properties a market ought to have (arXiv:2105.11416v3, 2021 preprint). This means that a single proposal on how compensation formats can be designed has been put forward. It does not mean that the method has actually entered any market.

Between Planning and Realization

There is a study that asked what remains when a schedule optimized based on forecasts is later resettled using finalized values (workload, weather, price, and carbon). Looking at three scattered artificial intelligence data centers, costs fell by 3.33% and carbon by 4.00% in the main case with constraints applied. The peak imported power—the power drawn most heavily from the grid at a single moment—fell by 6.63%. When the same facilities were simply allowed more workload shifting, the additional gains were 1.45% for cost and 1.80% for carbon. The authors note that the reductions remain conditional and specify that these values should be read as a check on operational domain rather than an evaluation of a specific site (DOI: 10.2139/ssrn.7232709, preprint, simulated settlement over public signal waveforms and scenario workloads).

It is a single-digit percentage. Yet other papers on the same topic feature much larger numbers. In a study optimizing the spatio-temporal flexibility of data centers in a 39-bus grid divided into two layers, the cost efficiency of the data centers improved by 70% and grid operating costs fell by 46%. Renewable energy consumption increased by 26.58% and carbon emissions fell by 63% (DOI: 10.1049/stg2.70020, 39-bus test grid).

The two results differ in what they measure, the grid they use, and their assumptions. If you strip away just the percentages in this field, you learn nothing. You can only read those percentages if you know which grid was used, what was set as the baseline, and whether forecasts were assumed to be correct.

There are cases that build this discipline directly into the paper. A study estimating the carbon effects of workload allocation using only public data for a computational hub in China divides claims into grades, separating the main scenario from exploratory stress tests. The baseline emissions are 20.30 megatons. Simply improving the power usage effectiveness—the value obtained by dividing the total power used by a facility by the power used by its computing equipment—reduced emissions by 3.63%. In the case confirming 10% flexibility, emissions fell by 4.47%, and in the main scenario of workload allocation, they fell by 11.78%. In an upper-bound sensitivity analysis that included renewable energy and storage together, the figure reached 29.23%, but the authors explicitly note that they do not treat this value as a representative result (DOI: 10.2139/ssrn.7052138, preprint, public data-based estimation).

There are also studies that examine the upper bound of what can be gained by shifting. Carbon intensity is the amount of carbon produced when generating a single unit of electricity. Using carbon intensity data from 123 regions—covering most major cloud regions—the study calculated the ideal and realistic limits of temporal and spatial shifting, separating batch workloads from interactive workloads. The conclusions were threefold. Temporal and spatial shifting can reduce emissions, but the realistic upper bound is currently far from the ideal value. Simple policies generally capture most of the reductions, and adding sophisticated techniques yields almost no extra benefit. Furthermore, as electricity supply becomes cleaner in the future, the relative benefit of shifting based on carbon will actually shrink (arXiv:2306.06502v2).

The middle conclusion weighs heavily. The upper-bound study says that, at least when it comes to shifting work to reduce carbon, there is little to be gained from making algorithms more sophisticated. The situation may be different when it comes to grid-side flexibility. There are calculations showing that if we know the queue of incoming work perfectly in advance, flexibility becomes much greater (arXiv:2603.27831v2). Still, the remaining homework on the carbon side lies elsewhere: how to classify what counts as shiftable work, and how to verify that promise.

Verification is harder than it sounds. When a data center claims it 「reduced output,」 confirming that requires knowing consumption in the counterfactual case where it did not reduce output, but that value is never observed. Therefore, a conservative conversion procedure is separately designed to translate the decision to delay work into a verifiable amount of power reduction on the grid side (DOI: 10.2139/ssrn.6546460, preprint, case study). Whether the promise can be kept is the same problem. A work queue is the line of tasks waiting to be processed on servers. If a data center sells frequency regulation capacity in advance, separately from its work schedule, the state of the work queue can make the promised regulation amount unrealizable or cause it to last only briefly. Co-optimizing the schedule by placing constraints on the queue state increases the regulation capacity that can actually be delivered (arXiv:2602.01508v1, 68-bus test grid).

The value of flexibility varies by facility. There is a study that calculated the cost of providing flexibility using actual work records from seven high-performance computing facilities for artificial intelligence and seven general-purpose facilities, along with computing prices from three cloud platforms. The AI facilities were able to generate greater flexibility for various types of grid services at a 50% lower cost than general-purpose facilities. When comparing those costs to market prices for flexibility, providing flexibility made financial sense for the AI facilities (arXiv:2410.17435v2). Placed side by side with the empirical finding from the previous section that they 「do not respond well to prices,」 it becomes hard to explain why they do not participate solely by saying 「it doesn't pencil out,」 at least within this calculation.

What Flexibility Solves and What It Does Not

The queued power facilities, a day of spreading blackouts on the Iberian Peninsula, and two carbon footprints written side by side for the same training session all appeared in previous primer articles. Those articles were hung in different places. In "How Much Electricity Does AI Use? — Why Every Source Gives a Different Number", the reason estimates diverge was a matter of boundaries and denominators. "The Few Seconds a Generator's Spinning Mass Gave Away for Free — What Gets Harder as Grid Inertia Falls" was about the properties that generators used to provide to the grid. "Plugging It In Is Harder Than Building It — What an Interconnection Queue Actually Decides" was a matter of the queue's structure, and "When Air Is No Longer Enough — Where the Heat from Computing Goes" was a problem inside the facility.

Flexibility touches one of those: the condition when joining the interconnection queue. The method by which the synthetic Texas grid study that opened this piece calculated it could broaden site candidates and advance the first energization date was changing what is promised when joining the queue. A load that promises in advance to back off when the grid is congested only needs to use as much as is available when it is not congested. Then, the need to build up facilities to match the most congested moment decreases. That is also why the values from the model we saw earlier emerge as the resolution of transmission violations, reduced curtailments, and power facilities that do not need to be built.

At the same time, what flexibility cannot solve is clear.

On the side of moving time, the total amount stays the same. The fact that the time-shifting model we saw earlier was designed to preserve the total daily energy balance means that postponed computations must definitely be done later. In that drawer, flexibility is closer to a device that makes you build less rather than use less electricity. What decreases is not the amount used, but the grid-wide peak demand—in other words, how much this load contributes to the power at the time of heaviest use. However, not all four drawers are like that. The side that lowers precision touches the total amount itself.

What cannot be moved still cannot be moved. Work that a human waits for in front of a screen is not postponed, it is not sent over a distance exceeding the latency limit, and a training run once started is difficult to scatter midway. The flexible share is not the entire load, but a portion of it.

And all of this is currently happening mostly inside models. Most of the values in this article are calculations on test or synthetic grids, and reviews looking at actual usage data state that data centers are unlikely to be sensitive to electricity prices except during extreme events. Between potential and execution, the records and rate plans for large loads are only just being created, and the way to pay a price for flexibility is still at the proposal stage.

Whether the entire data center load is something that cannot be reduced was not a fact determined by physics. Planning models had chosen that as the default. That does not mean there are no constraints. Among the constraints this article followed, latency limits, training that must continue without interruption, and cooling where heat responds slowly come from physical and service conditions. And in actual usage data, data centers were unlikely to be sensitive to electricity prices except during extreme events. The share that cannot be moved is determined by physical and service conditions, and whether the share that can be moved is actually put to use is determined by models and systems.

To the question of whether data centers can become flexible, there are already many values calculated through models. The remaining questions are three: Who knows what is being moved, who verifies that it was moved, and who pays the price for it? These three are matters of records, verification, and contracts.

Source Papers

Papers published on arXiv, SSRN, and Research Square are marked as "preprint," and whether they were subsequently published in an academic journal has not been separately verified. Most results in this field are simulations on grid models, and the grading is indicated in the text each time.

• "Flexibility-Aware Framework for Efficient Planner-Initiated Siting of Data Center" (2026, preprint) — arXiv:2605.14714v1 — Synthetic Texas 2,000-bus system. Flexible operation broadens candidate sites.
• "Data Center Spatio-Temporal Load Flexibility in Security-Constrained Unit Commitment" (2026, preprint) — arXiv:2605.18517v1 — 24-bus system. 20 to 30% accounts for most of the gains.
• "Stretched grid? Managing data center energy demand and grid capacity" (2023) — DOI: 10.1093/ooenergy/oiad014 — Load concentration strains distribution capacity.
• "To Defer or To Shift? The Role of AI Data Center Flexibility on Grid Interconnection" (2026, preprint) — arXiv:2604.05376v1 — 3% to 21% reduction in investment and operating costs, inconsistent.
• "System-Level Impacts of Flexible Data Center Load Scheduling" (2026, preprint) — arXiv:2604.23128v1 — Only best-effort workloads shift, delay-sensitive tasks remain unchanged.
• "Carbon-Aware Compute-Power Scheduling for AI Data Centers with Microgrid Prosumer Operations" (2026, preprint) — arXiv:2605.03751v2 — Training is rigid, inference is elastic.
• "Let's Wait Awhile: How Temporal Workload Shifting Can Reduce Carbon Emissions in the Cloud" (2021, preprint) — arXiv:2110.13234v1 — Investigation of deferral room and influencing factors.
• "Grid Operational Benefit Analysis of Data Center Spatial Flexibility" (2025, preprint) — arXiv:2511.08759v2 — 73-bus system. Spatial relocation mitigates violations.
• "Enabling Zero-Downtime Maintenance And Dynamic Load Balancing Through Intelligent Workload Migration" (2026, literature synthesis) — DOI: 10.63278/jicrcr.vi.3723 — Components of migration, remaining constraints, and new approaches.
• "Privacy-Preserving Operation Approach for Electricity-Computing Power Systems" (2026, preprint) — DOI: 10.2139/ssrn.7218438 — Migration reduces mismatches even with transmission line constraints.
• "The Potential of Data Center Energy Demand To Provide Grid Flexibility" (2025, review) — DOI: 10.1007/s40518-025-00258-9 — Scarce knowledge of scale and cost, and low real-world price sensitivity.
• "From Tokens to Energy Flexibility: Quantization-Enabled Demand Response" (2026, preprint) — arXiv:2606.18851v1 — Precision adjustment as a means of power regulation.
• "Carbon-Aware Spatio-Temporal Workload Shifting in Edge-Cloud Environments: A Review and Novel Algorithm" (2025, review) — DOI: 10.3390/su17146433 — Most literature opens only a single axis.
• "Quantifying and Attributing Power Flexibility from GPU-Heavy Data Centers" (2026, preprint) — arXiv:2603.27831v2 — Itemized price ladder.
• "Inter-Area Oscillation Damping in Data-Center-Integrated Power Systems" (2026, preprint) — arXiv:2607.27575v1 — 39-bus system. Batteries work, cooling does not.
• "Grid Frequency Stability Support Potential of Data Center" (2025, preprint) — arXiv:2510.01050v1 — Diminishing returns, though participation is increasingly indispensable.
• "Virtual inertia provision through data centre and electric vehicle for ancillary services support in microgrid" (2020) — DOI: 10.1049/iet-rpg.2020.0217 — Conceptual simulation of supporting frequency with batteries.
• "Shift or curtail? How much data-center flexibility is worth depends on the host power grid" (2026, preprint) — arXiv:2608.19622v1 — PJM uses the spatial axis, South Korea uses the temporal axis.
• "Spatial and Temporal Heterogeneity of Marginal Emissions" (2012) — DOI: 10.2139/ssrn.2153929 — 2007–2009. Cheap hours are generally dirtier hours.
• "Robust Restless Multi-Armed Bandit for Data Center Flexibility Services" (2026, preprint) — arXiv:2605.19116v2 — Distributing requests under conditions where the inside cannot be seen.
• "The Missing Half of the Queue: Why FERC Should Require Load-Side Interconnection Queue Disclosure" (2026, policy brief) — DOI: 10.2139/ssrn.6614823 — No records of large load queues.
• "Demand-Side Flexibility for Energy Transitions: Policy Recommendations for Developing Demand Response" (2016, policy brief) — DOI: 10.2139/ssrn.2831868 — Recommendations for barriers and risk governance.
• "Large-Load Tariff Decisions and Utility Market Responses" (2026, case study) — DOI: 10.2139/ssrn.7367558 — 15 decisions, no evidence of re-evaluation.
• "Remunerating Space-Time, Load-Shifting Flexibility from Data Centers in Electricity Markets" (2021, preprint) — arXiv:2105.11416v3 — Proposed market clearing formulation.
• "Forecast-to-Realized Carbon-Aware Dispatch Evaluation for AI Data Centers" (2026, preprint) — DOI: 10.2139/ssrn.7232709 — Simulated settlement over scenario workloads.
• "Bi-Level Optimisation Model for Harvesting Spatial-Temporal Load Shifting Flexibility of Data Centres" (2025) — DOI: 10.1049/stg2.70020 — 39-bus system. Much larger scale of improvement.
• "A public-data decision-support framework for carbon-aware workload allocation in AI data-center hubs" (2026, preprint) — DOI: 10.2139/ssrn.7052138 — Public data estimation with graded claims.
• "On the Limitations of Carbon-Aware Temporal and Spatial Workload Shifting in the Cloud" (2023, preprint) — arXiv:2306.06502v2 — 123 regions. Upper bound is far from ideal.
• "Integrated Low-Carbon Scheduling of Workload Shifting and Demand Response in AI Data Centers" (2026, preprint) — DOI: 10.2139/ssrn.6546460 — Conversion into verifiable reductions.
• "Harnessing Flexible Spatial and Temporal Data Center Workloads for Grid Regulation Services" (2026, preprint) — arXiv:2602.01508v1 — 68-bus system. Increases defensible capacity.
• "AI-focused HPC Data Centers Can Provide More Power Grid Flexibility and at Lower Cost" (2024, preprint) — arXiv:2410.17435v2 — Half the cost, and it breaks even.