Data center grid flexibility and demand response

Grid flexibility, and why a load that can flex now connects at all

Grid flexibility for a data center is the ability to vary how much power the site pulls from the grid on demand, instead of drawing a flat block every hour of the year. A flexible site can shave its peak with on-site generation or batteries, shift compute that can wait to cheaper or cleaner hours, or curtail its grid draw on the few days the grid is most stressed. The critical IT load keeps running through all of it. The flex comes from the power and storage side and from workloads that tolerate a delay, never from dropping the racks.

Data centers were treated as the opposite of flexible for years. A 24/7 load that wanted firm power and never moved. That was fine when the grid had room. It does not have room now. With utilities unable to add capacity as fast as the load is showing up, a load that can flex has become a way to connect sooner, at lower cost, and as a better neighbor on the system. The grid-and-substation guide covers the interconnection wait that drives this. The on-site-generation guide covers the plant that does the shaving.

This guide is about the flex itself: what it is, where it comes from, the demand-response and capacity programs that pay for it, and the hard rule that none of it touches the critical load.

Why is grid flexibility the issue in 2026?

The grid cannot add capacity as fast as the AI load is arriving, and that mismatch is the whole reason flexibility moved to the front of the conversation. New transmission and generation take 5 to 10 years to permit and build. A data center shell goes up in 18 to 24 months. The interconnection queue for a large new load runs years in the busy markets, and in ERCOT alone grid planners have been studying well over 200 GW of large-load requests, more than 70 percent of it data centers. The grid-and-substation guide walks through that queue in depth.

A flexible load changes the math the utility runs. A site that will agree to curtail on the worst days asks far less of the firm system than one that demands firm power every hour. Duke University's Nicholas Institute put numbers on it in its Rethinking Load Growth work: the existing US grid could absorb roughly 76 GW of new load at an average curtailment of 0.25 percent of annual hours, about 98 GW at 0.5 percent, and 126 GW at 1.0 percent, across the 22 largest balancing authorities. The average curtailment event in that analysis ran about 2 hours. Read those as modeled potential, not a guarantee for any one site, but the direction is clear. A small amount of flexibility opens up a large amount of headroom.

That is why the offer to flex now buys a connection date. Firm-only interconnection has been estimated to add on the order of hundreds of millions of dollars in system supply cost per gigawatt of new demand, while conditional or flexible service can cut that materially and pull a large site's full energization years forward. The exact figures and the program names are utility- and ISO-specific, so confirm them for the market you are in.

What is data center grid flexibility?

Data center grid flexibility is the ability to change the site's grid draw when the grid needs it, and it shows up in four moves: shave, shift, shed, and curtail. Shaving means trimming the peak by supplying part of the load from on-site generation or batteries so less comes through the meter at the worst hour. Shifting means moving deferrable compute in time, running batch and training work when power is cheaper, cleaner, or less constrained. Shedding means briefly pausing non-urgent workloads during an event. Curtailing means cutting the site's grid draw on the handful of days the grid is most stressed, usually by some combination of the other three.

The contrast is with the old inflexible load: a flat 24/7 draw that wanted firm power and treated any variation as a failure. That load is the hardest thing for a constrained grid to serve, because the utility has to build for its peak even though the peak only bites a few hours a year. A flexible load gives the system somewhere to breathe.

Two of the four moves come from the electrical plant, two from the compute. Which ones a given site can actually use depends on its generation, its storage, and how much of its workload can tolerate a delay. The sections that follow take each one apart.

The uptime line you do not cross

Uptime is the one thing a data center will not trade, and every credible flexibility scheme is built around that. The critical IT load, the racks running customer workloads, the cooling that keeps them alive, the controls, does not get dropped to help the grid. A site that risked the workload to shave a peak would be solving the utility's problem by creating its own, and no operator running a Tier-rated facility will do it.

So the flexibility has to come from somewhere other than the critical load, and there are three honest sources. The power side: shifting part of the load onto on-site generation or batteries so the grid sees less, while the racks see no change. The storage side: discharging a battery or a grid-interactive UPS to ride a peak or an event. And the workload side: the compute that genuinely can wait, training and batch jobs that checkpoint and resume, which is a real and growing share of AI load but is not the same as the latency-sensitive inference that has to answer now.

That framing resolves the apparent contradiction. A data center can be both a firm home for the critical load and a flexible participant on the grid, because the flexibility is drawn from generation, storage, and deferrable work, not from the mission. Anyone selling flexibility that dips into the critical load is selling risk, not flexibility.

What is data center demand response?

Data center demand response is an arrangement where the site reduces its grid draw when the utility or grid operator calls for it, in exchange for a payment or a lower rate. The utility or ISO sends a signal, the site cuts its metered demand by a committed amount for the duration of the event, and the reduction is measured against a baseline and paid. Events are usually rare and short, called on the hours when the system is tightest, a hot summer afternoon or a cold winter morning, and the site is told in advance how much notice it will get.

The reduction can be made any of the ways flexibility comes: roll part of the load onto on-site generation, discharge batteries, or shed deferrable compute, or some mix. What the program cares about is the megawatts off the grid at the meter during the event, not how the site produced them. The payment structures vary by program and market, but utilities have commonly compensated commercial and industrial participants somewhere in the range of tens to a couple hundred dollars per megawatt-hour curtailed, with separate value for committing capacity in advance.

The relationship is the point. Demand response is a contract with the utility or the ISO, with a defined event window, a notice period, a committed reduction, a measurement method, and a penalty for not delivering. The specific terms, the payment, and the penalties are program-specific, so they live in the agreement, not in a rule of thumb.

What is peak shaving?

Peak shaving is reducing the grid draw during the site's highest-demand hours by supplying part of the load from on-site generation or batteries, so the meter never sees the full peak. The reason it pays is the way large electricity bills are built. A big share of a commercial power bill is the demand charge, billed on the highest demand recorded in the period, not on total energy. Shave a few megawatts off the monthly or seasonal peak and the demand charge falls even if total consumption barely moves.

In some markets the peak that matters is set by the grid, not the calendar. ERCOT's transmission charges for large loads are allocated on the four coincident peak hours of the summer, the 4CP, so a site that can predict those windows and trim its draw during them cuts its transmission cost for the following year. That mechanism is exactly why flexible loads in Texas watch the system peak so closely, and it is part of what Senate Bill 6 and the PUC's 4CP review are reworking. The on-site-generation guide covers the generation and battery plant that does the shaving.

Peak shaving and demand response overlap but are not the same. Peak shaving is something a site does for its own bill, on its own schedule, against the peaks it can see coming. Demand response is something it does on the utility's call, for a payment, when the grid is stressed. The same on-site generation and batteries serve both.

On-site generation for a demand-response event

Running the on-site generation during a peak or a demand-response event is the most direct way to drop the grid draw without touching the critical load. The plant picks up part of the campus load, the meter sees less, and the racks never know. On a site already built with prime or continuous generation for the connection itself, the same machines do double duty, carrying the site normally and shaving the grid on the worst days. The on-site-generation guide covers prime versus standby duty, gas turbines, engines, and fuel cells in detail.

The catch is emissions and permits. A generator run for grid flexibility is running for an economic reason, not an emergency, and that distinction matters to the air authority. The lighter emissions tier that covers a true emergency standby engine does not cover a machine run to shave a peak or earn a demand-response payment, and running gensets for non-emergency hours can quickly hit the permitted-hour limit or pull the plant into stricter review. This is the trap with the simplest flexibility move. The diesel standby plant already on site looks like free curtailment capacity until the air permit says those hours are not allowed.

Confirm what the permit actually lets the machines do for non-emergency duty before counting on-site generation as flexibility. Where the hours are tight, the flexibility shifts toward batteries and deferrable compute, which carry no emissions and no permitted-hour ceiling.

Batteries and the grid-interactive UPS

Batteries flex a data center without burning anything, which is why they have moved to the center of the flexibility story. A battery energy storage system charges when power is cheap or clean and discharges to carry part of the load during a peak or a demand-response event, holding the grid draw down while the critical load runs unchanged. The discharge is fast, clean, and free of any permitted-hour limit, so it covers the cases where running gensets would hit an emissions ceiling.

The interesting twist is the UPS that every data center already has. A grid-interactive or energy-storage UPS can do more than ride the seconds between a grid drop and the generators starting. With the right design and controls, its batteries can discharge to the load to shave a peak or support an event, then recharge afterward, turning a piece of equipment that used to sit idle into a flexibility asset. The constraint is that the UPS exists first to protect the critical load, so any grid-facing use has to leave the full ride-through reserve intact. You flex the headroom, never the reserve.

Sizing is the real decision. A battery sized for ride-through is not automatically sized for hours of peak shaving, and a battery sized to shift energy is a different machine again. Decide what role the storage is actually playing, ride-through, peak shaving, demand response, or grid services, and size and control it for that role, because a battery asked to do all of them on one set of cells does none of them well.

Can you shift data center compute to cheaper hours?

Yes, for the part of the workload that can wait, and shifting it is the cleanest flexibility a data center has. Load shifting means running deferrable compute, large model training, batch processing, fine-tuning, data pipeline jobs, when power is cheaper, cleaner, or the grid is less stressed, and holding it back when power is tight. The work still gets done. It just gets done at a better hour for the grid and often for the electricity bill.

This is temporal flexibility, and Google has been doing a version of it for years through what it calls carbon-intelligent computing, scheduling non-urgent jobs like processing YouTube uploads into the hours when the local grid is cleanest. The same lever that chases clean power chases cheap power and grid headroom, because those hours often line up. A site that can move even a fraction of its compute off the system peak is shaving without any on-site generation at all.

The limit is which workloads tolerate the delay. Inference that answers a user in real time cannot be shifted. It has to run now. Training and batch work can be scheduled, paused, and resumed, and that deferrable share is exactly the flexibility source the next sections build on. Shifting works only to the extent the site knows which of its jobs can wait and has the scheduling to move them.

Load shedding: pause and checkpoint

Load shedding is briefly pausing or throttling non-urgent workloads during a grid event, as opposed to shifting them in advance. When the utility calls an event, the site drops the power draw of its deferrable compute for the duration, then resumes when the event clears. For workloads that checkpoint, training above all, that pause is close to free. The job saves its state at a checkpoint, holds, and picks up where it left off when power frees up, losing time but no work.

The mechanism matters because it is what makes the pause safe. A training run that checkpoints every so often can be parked at the next checkpoint and resumed from it without corruption or lost progress. The platforms that do this in production do not slam the hardware off. They deprioritize and reschedule the lower-priority jobs around natural flex points like checkpoint intervals, while latency-sensitive work keeps running. The Emerald AI and EPRI DCFlex demonstrations showed exactly this: a GPU cluster cutting its power draw on a dispatch signal while every workload stayed inside its service-level agreement.

Shedding is curtailment at the compute layer, and it pairs with the power side. A site can meet a demand-response reduction partly by shedding deferrable jobs and partly by rolling load onto batteries, so the critical load and the latency-sensitive work never feel the event. The reduction is real at the meter. The mission is untouched.

Can AI training be paused for the grid?

Yes, and this is the insight that changed the flexibility conversation in 2026. AI training is far more flexible than inference. A training run is a burst workload, intense and periodic, and it already checkpoints its progress, so it can be paused at a checkpoint and resumed without losing work. Inference is the opposite: continuous, latency-sensitive, and tied to a user waiting for an answer, and by recent counts inference is 80 to 90 percent of total AI compute. The flexible share is the training and batch work, not the live serving.

The demonstrations put hard numbers on it. In a five-day trial near London, a cluster of NVIDIA Blackwell Ultra GPUs run through the Emerald AI platform responded to a series of live grid dispatch signals, cutting power by up to 40 percent and dropping roughly 30 percent in under a minute, while staying compliant with every requested target and keeping workloads inside their SLAs. An earlier US demonstration under EPRI's DCFlex initiative, with NVIDIA, Oracle, and two Arizona utilities, held a 25 percent reduction for three hours without breaking SLA. Research across representative AI workloads has put the power flexibility opportunity somewhere in the range of 18 to 55 percent depending on the job mix.

The reason this matters for the grid is scale. AI training is precisely the kind of large, deferrable, checkpointable load that can curtail on the worst days, which is what the Duke headroom numbers assume. The new flexibility is not a trick at the margins. It is a property of how the biggest new loads on the grid actually compute.

Shifting the workload to another region

Spatial flexibility is moving a workload to a different data center, in a different region, where power is cheaper, cleaner, or less constrained at that moment. A company running a fleet of sites can route a deferrable job to wherever the grid has room, the way it already routes for cost and latency. Where temporal flexibility moves work in time, spatial flexibility moves it in place, and the two combine. Run the job later, or run it elsewhere, or both.

This only works across a fleet, and only for workloads that are not pinned to a location by latency or data gravity. A training run that does not care which campus it executes on is a candidate. An inference request from a user in a specific metro is not. So spatial flexibility is mostly a hyperscaler and large-operator lever, available to operators who control where the work lands and have the network to move it.

For the grid, geographic shifting spreads a flexible load across balancing authorities instead of concentrating it on one constrained system. It is one more way the same compute can be a lighter ask on any single piece of the grid, and it is part of why large operators talk about their flexibility as a portfolio across sites rather than a property of one building.

Getting paid: capacity and ancillary-services markets

Beyond a bilateral demand-response deal with the local utility, a flexible data center can earn money in the organized markets the grid operators run. The value splits into three streams. Capacity pays for committing to be available to reduce load when called, whether or not an event happens. Energy pays for actually curtailing when prices spike or the operator dispatches. And ancillary services pay for fast grid support like frequency regulation, where a battery or a fast-responding load helps hold the system stable second to second.

Which markets a site can reach, and how a flexible load qualifies, depends entirely on the ISO or RTO it sits in. PJM, MISO, ERCOT, SPP, CAISO, ISO-NE, and NYISO each run their own programs with their own rules for how demand response and storage participate, and those rules are in active flux as the operators rework how large loads and co-located resources are treated. PJM, for instance, ran a Critical Issue Fast Path process on large-load additions through 2025 and into 2026, and already counts on the order of several gigawatts of demand response in its capacity market.

The honest read is that the payments are real but program-specific, and the rules are moving. A site planning to monetize flexibility in a wholesale market needs to confirm the current participation rules, the baseline and measurement requirements, and the penalty structure with the specific ISO and, often, an aggregator, because what qualifies and what it pays differs by market and by year.

Flexibility as good grid citizenship

A flexible data center is an easier thing for a community and a regulator to say yes to, and that has become part of why operators offer it. The opposition that large data centers now draw is largely about the grid: the fear that a gigawatt of new inflexible load will raise everyone's rates, strain reliability, and force expensive system upgrades that other ratepayers help pay for. A load that agrees to curtail on the worst days defuses much of that, because it asks the firm system for less and leaves headroom for everyone else.

The numbers support the case. The Duke analysis found that a 1 to 2 percent reduction in data center peak demand could lower electricity rates for other customers by roughly half a percent to a few percent, which turns a flexible site from a cost driver into something closer to a grid resource. That is a different political story than an inflexible load that needs the whole system built up around its peak.

The regulatory pressure is moving the same direction. Texas now requires large new loads to provide flexibility during grid emergencies, and other jurisdictions are writing flexibility into how large loads connect. The site-selection calculus is shifting with it. A project that arrives offering flexibility connects with less friction than one that demands firm power and fights the rate case. Flexibility is becoming less a favor to the grid and more the price of admission.

The emissions and permit tradeoff in flexing with generators

Flexing by running gensets has an emissions cost, and treating on-site generation as free curtailment capacity is one of the easy ways to get a site in trouble. Every hour a combustion machine runs to shave a peak or earn a demand-response payment is an hour of NOx, carbon monoxide, particulate, and carbon dioxide that a battery or a shed workload would not produce. The air authority sees those hours, and it sees them differently than emergency hours.

The line is emergency versus economic. A standby engine run to ride a genuine grid outage falls under a lighter emissions tier in most jurisdictions. The same engine run to make money in a demand-response program, or to dodge a demand charge, is running for an economic reason and generally does not get that treatment, which can push it into permitted-hour limits or stricter review. Regulators have been tightening exactly this distinction as data center generation has grown. The on-site-generation guide covers the air permitting in depth.

The practical consequence is that the cleanest flexibility, batteries and deferrable compute, has a permitting advantage on top of its carbon advantage, because it adds no combustion hours at all. Where a site does plan to flex with generators, confirm the permitted non-emergency hours and the emissions tier first, and size the battery and the shed-able workload to carry the flexibility the permit will not let the engines provide.

What is an interruptible rate?

An interruptible or curtailable rate is a utility tariff that gives a customer a lower price for power in exchange for agreeing to cut load when the utility calls for it. The customer commits to drop to a firm contracted level, or off entirely, within a notice window on the utility's signal, and in return pays less per kilowatt or per kilowatt-hour than a firm-service customer. The discount is the utility paying in advance for flexibility it can count on when the system is tight.

The terms are the whole deal. An interruptible contract specifies the notice period, how often and how long curtailments can be called, the level the site must drop to, the measurement method, and the penalty for failing to curtail when called. That penalty can be severe, because the utility planned around the reduction, so a site that signs an interruptible rate has to be genuinely able to deliver the cut every time, not just on a convenient day. This is where the flexibility plan meets reality. The rate is only worth taking if the on-site generation, the batteries, and the deferrable workload can actually hold the contracted reduction on demand.

These tariffs vary by utility and are often negotiated for a large load rather than taken off a published schedule. The newer conditional-firm and flexible interconnection service arrangements are a related idea applied to getting connected at all: accept curtailability and connect sooner and cheaper than firm-only service would allow. The specific rate, the curtailment terms, and the penalties live in the contract with the utility, and that contract is the document that controls.

Reliability first: the critical load is never the flex

The redundancy that protects the critical load does not get spent on flexibility. Whatever a site does to flex the grid, the N+1 or 2N power chain, the UPS ride-through reserve, and the backup generation that exist to keep the racks alive stay intact and uncommitted. You flex the headroom above the mission, never the margin that protects it.

That means a battery used for peak shaving still has to hold its full ride-through reserve for an actual outage, and an on-site plant used for demand response still has to be available to carry the critical load if the grid drops. Flexibility uses the non-critical capacity and the deferrable work. It does not borrow against the reliability budget. A scheme that only works because it dips into the redundancy is not flexibility. It is a reliability cut waiting to be discovered on the day both the grid and the flex are needed at once. Test it before you trust it.

The controls that make flexibility automatic

Flexibility at the speed the grid needs is a controls problem. The grid operator's signal can call for a reduction in minutes or seconds, and a human reading an email cannot dispatch a multi-megawatt cut on that timescale. The site needs automated demand response: a control system that receives the grid signal, decides how to meet the reduction across generation, storage, and deferrable compute, and executes it without dropping the critical load. OpenADR is the common open standard for the signaling between the utility or aggregator and the site.

On the electrical side, the electrical power monitoring system, the EPMS, and the building management and battery management systems have to coordinate so that a flex action, discharging a battery, picking up load on generation, trimming cooling within limits, happens as one orchestrated move rather than a set of disconnected manual steps. On the compute side, a workload orchestration layer is what actually pauses, throttles, or reschedules the deferrable jobs on a dispatch signal, the way the Emerald AI platform did in the DCFlex demonstrations. The two layers, power and compute, have to act together.

The controls are also where the safety interlocks live. The automation has to know what it is not allowed to touch, the critical load and the reserve, and enforce that limit even under a dispatch signal. A flexibility control scheme is engineered and commissioned, not configured once and trusted, because the failures show up in the dispatch sequences, not in the steady state.

Measuring the curtailment: baseline and M&V

A demand-response payment is only as good as the measurement behind it, because the program pays for a reduction against a baseline, and the baseline is a counterfactual: what the site would have drawn if no event had been called. The measurement and verification, the M&V, compares the metered demand during the event to that baseline and certifies the megawatts reduced. Get the baseline wrong and the site is either underpaid for real curtailment or accused of delivering less than it did.

Baselines are estimated from the site's recent load history, often an average of comparable days adjusted for weather, and the exact method is set by the program. That is why the metering matters. Interval meters at the grid tie, and submetering that can separate the flexed load from the critical load, are what let the site prove a reduction and defend it. A site that cannot show its baseline and its event-hour demand cannot collect, however real the curtailment was.

The measurement obligation runs both ways. The utility needs proof the reduction happened. The site needs proof it delivered, both to get paid and to avoid the non-performance penalty in an interruptible contract. Keep the interval data, the event logs, the baseline calculation, and the verified reduction for every event, because the payment and any dispute over it ride on that record.

Testing the flexibility before you sell it

A flexibility commitment has to be proven before it is sold, the same way backup power is proven before it is trusted. The test is whether the site can deliver the committed reduction on a signal, hold it for the required duration, and do it all without disturbing the critical load. That means actually transferring load to generation or batteries, actually shedding the deferrable compute, and watching the critical bus stay steady through the transition, not assuming it will from the one-line.

The sequences are where the failures hide. The transfer to on-site generation that bumps the bus, the battery dispatch that does not ramp fast enough to meet the program's required rate, the workload pause that does not release cleanly, the recovery that slams the grid with a rebound spike when the event ends, these only surface under a real test. The demonstrations that have made the news, the GPU clusters hitting their dispatch targets and holding for hours within SLA, are commissioning exercises proving exactly these sequences before the capability is offered into a market.

Commission the flex like you commission the backup. Script the events, run them under real conditions, confirm the reduction and the ramp rate and the recovery, and prove the critical load never moved. A flexibility commitment that has not been tested end to end is a promise the site does not yet know it can keep.

The data center as a grid asset

The direction is clear even where the specifics are not: the large flexible load is becoming a resource the grid plans around rather than a problem it absorbs. The same AI buildout that strained the grid is producing the most flexible large load the system has ever had, because so much of AI compute is deferrable training that can pause and shift. The constraint and the answer arrived together.

If the modeled headroom holds, flexibility lets the existing grid take on far more load than its firm peak suggests, on the order of 100 GW in the Duke work, while new generation and transmission catch up. That reframes the data center from a pure consumer into something closer to a participant: a load that shaves its own peak, shifts its work to help the system, and gets paid for curtailing on the worst days. The pilots that ran in 2025 and 2026 were the proof of concept. The open questions are how fast the ISO programs, the tariffs, and the interconnection rules adapt to count flexibility as the firm-equivalent capacity it can be.

None of it changes the core constraint. The flexibility comes from generation, storage, and deferrable work, and the critical load stays firm. The site that gets this right connects sooner, costs the system less, and runs its mission unbothered while it does the grid a favor. That combination, not the favor alone, is why this moved from pilot to strategy.

What to document

A flexibility program lives or dies on its records, because every payment, every curtailment, and every contract obligation has to be provable after the fact. Write down what the site committed to, how it meets the commitment, and what actually happened on each event, because the demand-response payment and any non-performance penalty both turn on that record.

Capture the flex method for each source, the committed reduction and the program or tariff it serves, the baseline method, the notice period and event limits, the measured reduction per event, and the confirmation that the critical load was untouched. Record the assumptions too, the permitted generator hours, the battery reserve held back, the workloads classified as deferrable, because those are what change when the site or the program changes. A field tool like FieldOS is a reasonable place to keep the event logs, the baselines, and the M&V records tied to each site and each program, so the proof of curtailment is in one place when the payment or the dispute arrives.

Common mistakes

• Flexing in a way that risks the critical load, instead of drawing flexibility only from generation, storage, and deferrable workloads.
• Having no on-site generation or storage to shave the peak, so the only flexibility left is cutting load that cannot be cut.
• Treating all compute as inflexible when training and batch work can pause and resume from a checkpoint.
• Running gensets for demand response or peak shaving without confirming the air permit allows non-emergency hours.
• Signing an interruptible rate the site cannot actually hold, then eating the non-performance penalty when an event is called.
• Skipping the measurement and verification, so a real curtailment cannot be proven and the payment cannot be collected.
• Selling flexibility into a market with no utility or ISO contract behind it, or without confirming the current program rules.
• Treating the flexibility controls as configured-and-forgotten instead of commissioning the dispatch, ramp, and recovery sequences.
• Spending the redundancy reserve on flexibility, leaving no ride-through margin on the day both the grid and the flex are needed.

Field checklist

Standards and references

Flexibility programs are run by the utility and the grid operator, so the controlling documents are theirs. The demand-response and capacity-market rules come from the ISO or RTO the site sits in, PJM, MISO, ERCOT, SPP, CAISO, ISO-NE, or NYISO, each with its own participation rules, baseline methods, and penalties, and those rules have been changing fast as the operators rework how large loads and co-located resources connect. The interruptible or curtailable tariff comes from the local utility, often negotiated for a large load rather than taken off a published schedule. Confirm the program, the payment, and the penalty against the specific utility and ISO, because the terms differ by market and by year.

On the technical side, OpenADR is the common open standard for automated demand-response signaling between the program and the site. Where on-site generation or storage exports to or parallels the grid, IEEE 1547 governs the interconnection of those distributed resources, with the UL 1741 listing for the equipment, as covered in the on-site-generation guide. The air permit for any generator run in non-emergency flexibility duty comes from the EPA Clean Air Act program and the state or local air authority, and the standby-engine emissions exemption does not extend to economic dispatch.

The uptime framework still sits over all of it. The Uptime Institute Tier topology and the project basis of design set the redundancy the critical load requires, and no flexibility scheme overrides them. Treat the modeled headroom figures, the Duke Nicholas Institute curtailment numbers among them, as direction rather than a guarantee for any one site, and let the utility agreement, the ISO rules, the air permit, and the engineer of record control the actual commitments. Cite the controlling document by topic, and hedge every program, payment, and limit to the utility, the ISO, and the operator.

Units, terms, and abbreviations

Grid flexibility borrows vocabulary from the utility, the grid operator, and the data center, and the same word can carry a different meaning in each. Pin the term to the program before acting on it. Power is measured in megawatts and now gigawatts, energy in megawatt-hours, and a flexibility commitment is usually stated as a number of megawatts of reduction for a number of hours, against a baseline, on a defined notice.

The demand charge is billed on peak demand, not total energy, which is why shaving the peak cuts the bill out of proportion to the energy moved. A baseline is the modeled draw the curtailment is measured against. M&V is the measurement and verification that certifies the reduction for payment.

Grid flexibility

The ability to vary a site's grid draw on demand, by shaving, shifting, shedding, or curtailing, without dropping the critical load

Demand response (DR)

Reducing grid draw when the utility or ISO calls an event, in exchange for a payment or a lower rate

Peak shaving

Trimming the grid draw during the highest-demand hours by supplying part of the load from on-site generation or batteries

Load shifting

Moving deferrable compute to cheaper, cleaner, or less constrained hours; temporal flexibility

Load shedding

Briefly pausing or throttling non-urgent workloads during a grid event, then resuming

Curtailment

Cutting the site's grid draw on the grid's worst days, often under a program or interruptible tariff

Capacity market

An ISO market that pays a resource for committing to be available to reduce load when called

Interruptible rate

A utility tariff giving a lower price in exchange for agreeing to curtail load on the utility's signal

Demand charge

A bill component based on the highest demand in the period, not on total energy used

Baseline / M&V

The modeled counterfactual draw a curtailment is measured against, and the verification that certifies it

Related tools

Calculators and readiness checks for this work

Compare your options

FAQ

People also ask

• Why are data centers building their own power?
• How does a data center get power?
• What is power capping in a data center?
• What is a signal reference grid (SRG)?
• Why does a data center need a formal incident process?