EPRI Flex MOSAIC Turns Data Center Power From Grid Constraint Into Grid Asset

EPRI Flex MOSAIC has moved the data center power debate beyond one fixed assumption: every new campus must receive firm electricity around the clock. The framework instead asks operators to specify when, how much, and how long their facilities can reduce grid demand. That shift turns flexibility into a measurable service, not a vague promise.

The Electric Power Research Institute introduced Flex MOSAIC in March 2026 through its broader DCFlex initiative. More than 65 utilities, grid operators, regulators, hyperscalers, and technology providers contributed to the framework, according to its launch announcement. Their goal is faster access to power without weakening reliability or transferring infrastructure costs to existing customers.

The proposal arrives as data center developers face long interconnection studies, limited transmission capacity, and uncertain generation availability. Traditional firm service requires the grid to support a facility during stressed hours, even when those conditions occur rarely. Flexible service reverses that relationship by asking the customer to help manage those hours.

This is not a claim that data centers eliminate the need for new power plants or transmission lines. It is a narrower and more consequential proposition. A data center that can reliably reduce demand might connect before every long-term grid upgrade is finished.

The central contest is therefore clear. The established model treats uninterrupted power as a prerequisite for energizing a campus. EPRI Flex MOSAIC treats verifiable operational flexibility as another path to connection, provided utilities can depend on it when conditions tighten.

EPRI Flex MOSAIC Gives Flexibility a Common Language

The immediate change is standardization: EPRI Flex MOSAIC describes large-load flexibility through measurable performance instead of project-specific promises.

Utilities already offer interruptible rates, demand-response programs, and special contracts for large industrial customers. Yet data center proposals differ widely in design, workload, backup generation, batteries, and tolerance for interruption. Utilities often evaluate each campus as a unique engineering case.

That case-by-case process becomes difficult when proposed facilities reach hundreds of megawatts. EPRI says some emerging large loads range from 500 megawatts to 1.5 gigawatts at one site. Each request can trigger extensive studies of transmission limits, generation adequacy, ramping behavior, and local equipment.

Flex MOSAIC creates a shared classification based on four practical dimensions: magnitude, timing, duration, and frequency. Magnitude describes how much demand a facility can reduce. Timing measures how quickly that response becomes available.

Duration identifies how long the reduction can continue. Frequency describes how often the grid may call on it. Together, those dimensions give utilities a clearer picture than a general statement that a campus is “flexible.”

That distinction matters because different resources solve different grid problems. A battery can respond rapidly, but its stored energy eventually runs out. A backup generator can operate longer, although fuel supply, emissions permits, and operating restrictions constrain its use.

Compute workloads also behave differently. Some training jobs can pause, slow, or move to another region. Real-time inference services, financial systems, and customer-facing applications often have tighter latency and availability requirements.

The framework does not require one technology. A facility might combine workload scheduling, batteries, thermal storage, backup generation, or a temporary reduction in computing performance. The relevant question is whether the complete system can deliver its contracted response.

EPRI presents the framework through its flexibility model, which aims to replace inconsistent terminology with performance categories. That common vocabulary can help developers describe a project before entering detailed negotiations.

It can also help utilities compare proposals across sites. A campus offering a fast, brief response serves a different need from one capable of sustained curtailment during extreme weather. Treating both offers as identical would create planning risk.

The framework emerged from DCFlex, an initiative designed to test how data centers can support grid operations while improving interconnection efficiency. DCFlex organizes work around facility design, utility programs, power supply, and system planning.

Its planned field demonstrations include different markets, climates, facility types, and technologies. EPRI says the demonstrations examine dynamic load management, backup power, power quality, fault response, and carbon-aware operation.

The important event is not another pledge to use energy efficiently. Flex MOSAIC creates a structure that utilities can place inside tariffs, contracts, planning models, and operating procedures. That makes flexibility legible to institutions that decide whether a campus receives power.

Still, classification alone does not guarantee a connection. A utility must determine whether the proposed response helps at the relevant location and time. It must also establish consequences when a facility fails to perform.

Those details lead directly to the bigger question. If flexibility becomes dependable enough for system planning, it changes which grid investments must be completed before a data center can operate.

Data Center Power Delays Are Now a Competitive Constraint

Power availability has become a scheduling problem for AI infrastructure, placing developers, utilities, and electricity customers under competing pressures.

A data center operator can order servers faster than a utility can build a transmission line. Large transformers, substations, generation projects, permitting reviews, and interconnection studies follow different schedules. That mismatch can leave completed computing infrastructure waiting for electricity.

The pressure falls first on developers. Their customers want AI capacity quickly, while financing and construction plans depend on a credible energization date. A project with land and permits can still lose value when firm grid service remains years away.

Utilities face a different problem. They must assess large requests without knowing whether every proposed campus will be built. Developers sometimes explore several possible sites, which can inflate forecasts and consume study resources.

Pacific Northwest National Laboratory examined these problems in its interconnection framework. The report identifies data centers as the largest current driver of emerging large-load growth and calls for more consistent review processes.

Grid planners cannot simply accept every forecast. Building too little risks reliability problems and delayed economic development. Building for projects that never arrive can leave customers paying for underused infrastructure.

Existing electricity customers face the affordability question. New transmission and generation can increase system costs, especially when regulators allow utilities to recover investments through broader rates. The allocation of those costs has become central to state proceedings.

Data centers can also improve utilization when their demand remains high during otherwise underused hours. More electricity sales can spread some fixed network costs over a larger volume. That benefit depends on location, tariff design, and the investments required for each campus.

The distinction between annual energy and peak capacity is critical. A grid might have enough total generation across a year while lacking sufficient capacity during a few extreme hours. Firm service forces planners to prepare for those hours.

Flexible service offers a bargain. The developer receives an earlier or larger connection, while the utility gains the right to reduce the facility’s grid draw under defined conditions. The contract converts speed to power into compensation for operational flexibility.

That bargain pressures hyperscalers to reconsider how they schedule computing. Cloud platforms traditionally optimize around server availability, latency, hardware utilization, and customer demand. Grid conditions now become another input.

Some work can move across time. Batch analytics, selected training stages, software builds, and nonurgent processing can run when capacity is available. Operators can also shift work between regions when network capacity and data rules permit.

However, moving computation is not free. Data transfers consume energy and network capacity. Specialized accelerators may not be available elsewhere, while synchronized training clusters can be difficult to pause without losing progress.

The utility also needs visibility. A promise to reduce demand has little planning value without telemetry, which is continuous operating data shared with the grid operator. Utilities need to know the current load and the available response.

They may also require testing before treating flexibility as a dependable resource. A facility must prove that controls, backup systems, and workload schedulers respond together. Performance must remain predictable during the exact conditions that stress the grid.

Federal regulators have begun pushing regional operators toward clearer large-load rules. In June 2026, the Federal Energy Regulatory Commission ordered six regional grid organizations to justify or reform their tariffs.

The FERC orders specifically included new transmission services for flexible large loads. They also addressed co-located generation, cost transparency, application studies, and generation adequacy.

That action increases pressure on PJM, MISO, SPP, CAISO, ISO New England, and the New York ISO. Each organization has a different market structure and resource mix. FERC did not prescribe one national tariff.

The outcome will affect where data centers build. Regions that offer predictable flexible service can shorten development uncertainty. Regions without workable rules risk losing projects or receiving proposals centered on isolated on-site generation.

This makes data center power a competitive factor across jurisdictions. The winning region will not necessarily have the most unused electricity. It may have the clearest process for connecting new demand without exposing other customers to unmanaged risk.

Flexible Data Centers Reverse the Build-First Model

The core reversal is simple: a data center can connect before every upgrade is complete if it manages the grid’s scarce hours itself.

Traditional interconnection follows a build-first logic. The utility studies the full requested load, identifies required upgrades, completes those investments, and then provides firm service. This sequence protects reliability but can produce long delays.

A flexible grid connection separates firm and conditional service. Firm capacity remains available under ordinary contract terms. Conditional capacity is available when the network can support it and becomes interruptible during defined constraints.

Bring-your-own capacity, often shortened to BYOC, addresses a related generation problem. Under this arrangement, the data center procures qualifying capacity rather than relying entirely on the utility or regional market. That capacity might come from contracted plants, storage, or approved on-site resources.

A major 2025 study by Camus Energy, encoord, and Princeton University’s ZERO Lab modeled both approaches. It used utility transmission data, site-level optimization, and system-level capacity planning.

The flexible data center study examined a 500-megawatt campus. Its modeled flexible connection and BYOC combination reached full operation in roughly two years.

That was three to five years faster than the traditional interconnection paths studied. Grid power remained available for more than 99 percent of hours across the constrained sites in the analysis.

On-site resources supplied or reduced demand for 40 to 70 hours annually. Transmission constraints caused between 7 and 35 curtailed hours, while generation shortfalls added about 32 hours.

These are modeled results, not a guarantee for every campus. The study selected particular sites, assumptions, load profiles, and planning scenarios. Local outcomes will depend on network topology, weather, generation, and actual operating behavior.

Still, the mechanism explains why flexibility has attracted regulatory attention. A transmission system often carries less than its maximum rating during most hours. Conditional access can use that available headroom while preserving the operator’s ability to respond during congestion.

The study also modeled system costs in PJM. It estimated that each gigawatt of traditional firm-only data center demand added $764 million in supply costs under its assumptions. Meeting that load required 2.17 gigawatts of nameplate generation across several technologies.

A scenario with 20 percent conditional firm service avoided 273 megawatts of new capacity per gigawatt of data center demand. The study associated that reduction with $78 million in avoided system costs.

BYOC internalized an estimated $326 million in capacity costs per gigawatt. Energy payments contributed another $329 million, according to the model. Combined contributions reached about $733 million for each gigawatt of new demand.

Those figures should not become universal forecasts. They reflect a specific model of PJM, a defined capacity price, and a particular data center load profile. Market prices and resource requirements can change significantly.

The more durable conclusion concerns sequencing. Flexible interconnection lets a utility serve additional energy through existing assets while long-term reinforcements proceed. BYOC assigns more capacity responsibility to the incoming customer.

That arrangement can transform a campus from a passive load into a dispatchable partner. During normal conditions, it buys electricity and improves infrastructure utilization. During stressed periods, it reduces imports or supplies itself.

Some facilities may also export power, although export rights add interconnection and market complexity. A generator capable of serving an adjacent campus does not automatically qualify to inject electricity into the wider grid.

Data center batteries provide another possible resource. Operators already install uninterruptible power supplies to protect equipment from outages and power-quality events. New control systems can expose part of that capacity to grid programs without compromising essential backup reserves.

On-site generation can extend the response duration. Fuel cells, gas engines, turbines, and renewable systems each present different emissions, permitting, fuel, and reliability profiles. A technology-neutral framework lets contracts focus on delivered performance.

Workload control brings flexibility inside the computing stack. A scheduler can slow selected jobs, delay low-priority tasks, or redistribute work across available hardware. That approach avoids starting generators but requires close coordination with customer commitments.

A United Kingdom DCFlex demonstration tested this idea with National Grid, Emerald AI, Nebius, Nvidia, and EPRI. EPRI reported that the pilot reduced AI data center demand by 30 to 40 percent within seconds.

The European pilot reportedly maintained priority service levels while modulating selected workloads. Its result supports rapid response as a technical possibility, although broader commercial performance still requires validation.

A 130-kilowatt demonstration is not equivalent to a gigawatt campus. Scaling introduces more hardware, customers, contracts, failure points, and network dependencies. The control method must also work during unusual system conditions.

EPRI Flex MOSAIC therefore matters less as a new piece of hardware than as a coordination layer. It connects computing controls, facility equipment, utility signals, and commercial obligations. That mechanism is the path from grid constraint to grid asset.

The Grid Asset Claim Still Faces a Reliability Test

A flexible data center becomes a grid asset only when its response is enforceable, observable, and available during the hardest operating conditions.

The strongest objection concerns coincidence. Extreme weather can stress the grid while also challenging a data center’s cooling equipment, backup systems, and workforce. Flexibility that disappears during those events cannot replace dependable capacity.

Utilities must avoid counting the same resource twice. A battery reserved for emergency backup cannot simultaneously promise its full capacity to a grid program. Contracts need clear minimum reserves and operating priorities.

Fuel-based generation carries similar questions. A generator might have sufficient nameplate capacity but face emissions limits, fuel interruptions, maintenance, or local operating-hour restrictions. Its dependable contribution can be smaller than its advertised rating.

Workload flexibility also has boundaries. Training jobs offer more scheduling freedom than many customer-facing services. An operator cannot curtail every workload without violating service agreements or damaging customer trust.

Repeated interruptions can create cumulative costs. Jobs may restart, checkpoints may fail, and hardware utilization can decline. Customers might also move workloads away from regions with frequent constraints.

Duration presents another risk. A battery can respond in milliseconds but cannot sustain an unlimited event. Workload reductions might last longer, though deferred jobs eventually return and create rebound demand.

A useful tariff must distinguish a short emergency response from a multi-hour capacity obligation. EPRI’s magnitude, timing, duration, and frequency categories help, but utilities must translate them into enforceable operating rules.

Verification will require secure communication between facilities and grid operators. Both sides need agreed baselines for measuring the reduction. Without a credible baseline, a facility could receive compensation for demand that it never planned to use.

Forecasting adds uncertainty. Data center developers may request capacity years before knowing their final hardware mix. New accelerator generations can change power density, cooling design, and workload behavior before a campus opens.

Grid models also depend on project credibility. Duplicate or speculative requests can exaggerate future demand. Reforms must require financial commitments and development milestones without blocking legitimate projects too early.

The Energy Systems Integration Group argues that short-term measures alone are insufficient. Its 2026 work recommends incorporating large-load flexibility directly into resource-adequacy planning, rather than treating curtailment as an emergency fallback.

Resource adequacy asks whether enough dependable supply and controllable demand will exist during stressed periods. A flexible campus must receive an accredited value based on expected performance. That value should decline when uncertainty rises.

Data center operators may resist contracts that give utilities broad curtailment authority. Their businesses often depend on predictable computing availability. A vague interruption clause can become an unacceptable commercial risk.

Utilities may resist the opposite arrangement. A narrow contract with numerous exemptions could provide little help during emergencies. Negotiations must define triggers, notice periods, maximum events, response speed, and penalties.

Ratepayer protection remains another unresolved issue. Flexibility can reduce some required investments, but it does not erase local substation, transmission, or generation costs. Regulators must determine which upgrades serve the campus and who pays.

Communities will also evaluate emissions and water impacts. A connection that avoids grid congestion by running local fossil generation can shift pollution closer to residents. Faster interconnection is not automatically cleaner infrastructure.

The U.S. Department of Energy’s transmission assessment still identifies pressing infrastructure needs. Data centers join manufacturing and broader economic growth as drivers of new transmission requirements.

That evidence prevents an overly optimistic reading of flexibility. The grid still needs investment. Flexible connections primarily buy time, improve utilization, and reduce the amount of capacity needed for rare peaks.

The concept also requires geographic precision. Reducing load in one region does not necessarily relieve a constraint elsewhere. Grid value depends on the facility’s location relative to overloaded lines, substations, and generation.

A campus can therefore be flexible but not useful for a particular bottleneck. Utilities must model the electrical network, not simply count the facility’s total megawatts. Locational value should shape contracts and incentives.

Cybersecurity deserves equal attention. Automated controls connect data center operations with critical grid signals. A compromised control channel could disrupt computing loads or provide inaccurate availability data.

Operators will need authentication, redundancy, and fail-safe behavior. A facility should default to a safe state when communication fails. Utilities should not assume that every digital response remains available during a wider cyber incident.

These limitations do not invalidate EPRI Flex MOSAIC. They define the evidence required before planners can rely on it. The framework succeeds only when performance data replaces optimistic assumptions from both developers and utilities.

Three Signals Will Show Whether Data Center Grid Flexibility Scales

The next phase depends on tariffs, full-scale operating evidence, and transparent cost allocation, not additional statements of support.

The first signal is how regional grid operators answer FERC’s 2026 orders. Their tariff filings should reveal whether flexible large loads receive a distinct service path. The most important details are study procedures, curtailment rights, and performance obligations.

A useful tariff will connect flexibility with a concrete interconnection benefit. Developers need to know how a qualified response changes available capacity or timing. A program without a faster path offers little reason to accept operational risk.

The filings should also address co-located generation. Many proposed campuses plan to combine grid imports with on-site power. Operators need clear rules for studying those configurations and preventing unintended exports.

If several regions adopt comparable services, the grid-asset thesis becomes stronger. Developers could incorporate flexible design before selecting a site. Utilities could evaluate proposals through repeatable procedures.

If the filings remain highly bespoke, adoption will stay slow. Every project would still require extended legal and engineering negotiation. Flex MOSAIC would provide vocabulary without removing the administrative bottleneck.

The second signal is performance from larger DCFlex demonstrations. EPRI says its initiative includes field projects across several locations and use cases. Results should report actual response magnitude, duration, notice time, and recovery behavior.

Scale matters here. A small cluster can show that software adjusts computing demand. A commercial campus must coordinate thousands of servers, cooling systems, storage assets, generators, and customer obligations.

Demonstrations should include difficult conditions rather than ideal test windows. They should examine long events, repeated calls, communication failures, and simultaneous equipment constraints. Results should separate planned flexibility from emergency backup capability.

Independent or utility-reviewed measurement would increase confidence. Vendors have incentives to emphasize successful demonstrations. Grid planners need complete performance distributions, including failures and unavailable periods.

The third signal is whether regulators prevent cost shifting. Flexible service gains political durability only when existing customers can see how risks and expenses are allocated. New tariffs should identify upgrade costs, capacity obligations, and nonperformance penalties.

The Princeton-led modeling offers a possible framework for evaluating those effects. Its results suggest flexibility and BYOC can internalize most incremental supply costs under specific assumptions. Regulators now need evidence from real projects.

Customer bills provide one test, but they move for many reasons. More direct measures include required network upgrades, accredited capacity purchases, campus payments, and utility revenue from added electricity sales.

Transparency will also discourage exaggerated claims. Developers should not call every battery or generator a grid asset. Utilities should not treat every interconnection request as a fully committed future load.

Beyond those three signals, readers should watch how cloud providers expose energy-aware controls to enterprise customers. Flexible computing works best when workload priorities are explicit. Businesses must distinguish urgent services from tasks that can wait.

That requires better operational records. Engineering teams need to connect application requirements with infrastructure decisions, energy events, and service outcomes. A searchable technical knowledge base can help teams preserve those decisions across facilities and vendors.

The shift also changes procurement. Enterprise buyers may eventually compare cloud regions using electricity availability, carbon intensity, and exposure to curtailment. Providers will need to explain what flexibility means for service guarantees.

EPRI Flex MOSAIC will not make every data center dispatchable. It creates a common way to identify which facilities can respond, how they can respond, and what that response is worth.

Its larger contribution is institutional. Data center developers, utilities, regulators, and computing vendors have historically planned around different time horizons. A performance-based framework gives them a shared object for contracts and engineering.

The old path required the grid to finish expanding before a major load arrived. The emerging path allows selected loads to connect earlier, provided they manage scarcity and fund the capacity they require.

That bargain remains unproven at nationwide scale. Yet the underlying question is no longer whether data centers consume substantial power. It is whether operators can make part of that demand dependable enough to support the system.

Over the coming months, follow the tariff language, not the slogans. Look for measured field performance, explicit penalties, and cost protections for other customers. Those details will decide whether data center grid flexibility becomes infrastructure or remains a promising pilot.