Data Center Demand Response: Grid Services, Dry Cooling, and the Utility Partnership Model

Data center demand response is the practice of a facility temporarily reducing, shifting, or self-supplying its electricity draw when a grid operator needs relief, in exchange for payment, priority interconnection, or both. Google has signed 1 GW of data center demand response commitments with utility partners since 2025, and a Duke University Nicholas Institute study found that data centers accepting curtailment for as little as 0.25% of annual hours could free up roughly 76 GW of new US grid capacity without a single new power plant (Duke Nicholas Institute, 2025). The same flexibility logic applies to water: Vantage Data Centers’ closed-loop cooling design at its Wisconsin campus is projected to draw about 22,000 gallons a day, versus roughly 5,000,000 gallons a day for an equivalent campus running conventional evaporative towers (Vantage Data Centers, 2026).

This post covers why the “energy burden” framing on data centers is incomplete, what demand response and virtual power plant participation require of an operator, how the main grid-flexibility mechanisms compare economically, which cooling architectures cut water draw and what they cost in return, where heat reuse and desalination fit into this picture, and why factory-built modular data centers are positioned to deliver both.

Why the “Energy Burden” Framing on Data Centers Is Incomplete

Most public commentary on data centers still treats them as a subtraction problem. Megawatts taken from a fixed pool. Water pulled from a shared basin. Nothing given back. That framing was defensible in 2018, and somewhere around 2024 it stopped being accurate.

The global average PUE across the industry sits at 1.54 (Uptime Institute, 2025). Cooling still eats 30–40% of a facility’s total energy draw (IEA/McKinsey, 2024). Both figures get cited constantly, usually as evidence of waste. The fuller efficiency picture is covered in how PUE is calculated and improved. What the conversation keeps leaving out is that a facility’s electrical and thermal footprint stopped being a fixed liability. A utility can dispatch it during a stress event, or lean on it instead of building new transmission — and on the thermal side, the same facility can sit on a cooling tower drawing water all summer or reject heat with almost none.

Two of those levers, heat reuse and desalination, already have a home elsewhere on this blog. This piece is about the two that get less attention: the electrons, and the water that never gets evaporated in the first place.

What Data Center Demand Response Actually Means for an Operator

Demand response is not a backup generator sitting idle in case the grid fails. It’s a contractual commitment: reduce, shift, or self-supply load on a signal, and get paid or prioritized for it.

Google and Voltus signed what both companies describe as the first large-scale “Bring Your Own Capacity” agreement in the industry: a three-year deal in which Voltus assembles up to 100 MW of flexible capacity from batteries, smart thermostats, and distributed energy resources across the PJM Interconnection footprint (Data Center Knowledge, 2025). Separately, Google has now signed 1 GW of data center demand response agreements with utility partners, a figure the company frames as core infrastructure strategy, not a pilot program (Google, 2026).

The mechanics vary by market. PJM’s Economic Demand Response program pays large loads to cut consumption below their normal baseline during high-price hours, settled through a Curtailment Service Provider (PJM, 2026). Texas took a harder line: Senate Bill 6, effective June 21, 2025, requires any large load over 75 MW to curtail on ERCOT’s instruction during declared grid emergencies. No negotiation, no opt-out (Texas Legislature, 2025). Portland General Electric took a third path entirely, partnering with startup GridCARE to model hourly demand and unused capacity. That work freed up 80 MW of incremental data center capacity in Hillsboro, Oregon for 2026, with more than 400 MW slated by 2029, years earlier than the original interconnection timeline (Utility Dive, 2025).

Carbon Direct calls this a “ghost battery” — a facility that behaves like a grid-scale battery at its interconnection point without owning a single cell (Carbon Direct, 2026). At 40 GW of data center buildout, curtailing an average of 5% of demand for under 1% of operating hours could prevent an estimated $5.5 billion in annual consumer welfare losses tied to forced load shedding (Carbon Direct, 2026). A utility regulator deciding which load gets interconnected first runs exactly that math.

Virtual Power Plants: Why Data Centers Are Becoming Grid Assets, Not Just Loads

A virtual power plant, or VPP, aggregates distributed resources (batteries, backup generators, flexible loads, sometimes rooftop solar) and dispatches them as a single controllable resource, the way a utility would dispatch an actual power plant. The Google-Voltus deal is functionally a VPP built around a data center’s flexible edges rather than its core IT load.

CERRE, the Brussels-based think tank, sorts data center grid contributions into three buckets: Demand Response (adjusting or shifting consumption), Electrons (battery storage the facility owns or hosts), and Molecules (fuel-based generation the facility can run for the grid’s benefit, not just its own backup) (CERRE, 2026). A facility playing in more than one bucket is worth more to a utility than one that just sits on the grid drawing power.

Crusoe and Redwood Materials built a working example of the “Electrons” bucket without waiting on a utility contract at all: a solar microgrid running on secondhand EV batteries, powering a modular data center that held 99.2% uptime over its operating months, beating both companies’ internal targets (Redwood Materials, 2025). Separately, SIP’s Verrus is designing new facilities to segregate critical from non-critical load specifically so a microgrid can flex the non-critical portion without touching uptime commitments, a resilience-and-flexibility approach covered in more depth in modular data center resilience strategy (Microgrid Knowledge, 2025). Grid participation baked in at the drawing stage, in other words, instead of retrofitted onto an existing hall.

None of this requires AI training workloads or extreme densities to matter. If a facility has batteries sized for backup and a controls layer that can take a utility’s dispatch signal, it is most of the way there. Segment the load by criticality and it becomes a VPP candidate at almost any scale.

Why Utilities Need Data Center Demand Response More Than They Need New Transmission

Start with the queue. Over 2,060 GW of generation and storage capacity is sitting in US interconnection queues as of the end of 2025, more than the country’s entire installed power capacity (Lawrence Berkeley National Laboratory, 2026). The median wait from interconnection request to commercial operation for projects built in 2025 exceeded five years, up from under two years in 2008 (LBNL, 2026).

The load side is just as backed up. ERCOT’s large-load interconnection requests rocketed past 230 GW in 2025, nearly four times the 63 GW on the books at the end of 2024, with more than 70% of that coming from data center developers (LBNL/ERCOT data, 2025). In Pennsylvania, PPL Electric Utilities is fielding over 60 GW of prospective data center demand, with 13 GW already in advanced planning (PPL, 2025). PJM’s December 2025 capacity auction came up 6,625 MW short, triggering record clearing prices (PJM, 2025).

A utility facing that math has two options: spend a decade and tens of billions on new transmission, or find loads willing to flex. Duke University’s Nicholas Institute modeled the second path and found that if new large loads accept curtailment for just 0.25% of annual hours, roughly 85 hours a year, the existing US grid could absorb at least 76 GW of new demand with no additional generation or transmission build-out, rising to 126 GW if loads accept curtailment for up to 1% of annual hours, roughly 366 hours (Duke Nicholas Institute, 2025). Average curtailment events run about two hours. For a facility with UPS and on-site generation already built in, that’s a rounding error — and exactly why “we’ll curtail” has become the fastest sentence in a data center’s interconnection application.

Europe is legislating toward the same math. The EU’s Cloud and AI Development Act, proposed June 3, 2026, aims to triple the bloc’s data center capacity within five to seven years and introduces a tripartite framework binding operators, grid stakeholders, and public authorities to coordinated flexibility commitments (European Commission, 2026). CERRE estimates the EU could free up 50–60 GW of demand-side flexibility by 2035 through exactly this kind of grid-user integration (CERRE, 2026), and ACER’s newly approved methodology requires every member state to complete a national flexibility assessment by July 2026 and set indicative targets by January 2027 (ACER, 2025). More on the surrounding compliance calendar is in the EU data center regulations overview. None of that is hypothetical for a facility being spec’d today. The filing deadlines are already on the calendar.

Data Center Grid Services: Comparing the Flexibility Mechanisms

Not every flexibility mechanism pays the same way, or asks the same of the operator. The table below lines up the main options.

Grid-flexibility mechanisms compared

Across all five, the ask is the same. Utilities don’t need data centers to use less power overall; they need predictability about when the power gets used, and they will pay or expedite to get it.

Water-Reduced Cooling: Dry, Adiabatic, Closed-Loop, and Evaporative Compared

Cooling is where a data center’s water story gets written. Most public debate stops at “data centers use a lot of water” and never asks which cooling architecture is doing the drawing.

Traditional evaporative cooling towers are the default because they’re cheap to run. The industry-wide average water usage effectiveness (WUE) sits at 1.9 liters per kWh (Baltimore Aircoil / industry data, 2025). They also carry the lowest PUE penalty, which is exactly the trade-off: a facility relying on non-evaporative cooling runs 6.4% to 10.2% higher annual average PUE than one leaning on evaporative cooling (arXiv water-systems study, 2026). Lower water use costs energy. That’s physics, not a design failure, and the broader tension between AI workload growth and municipal water systems is explored in the water math behind AI data centers.

Adiabatic, or indirect evaporative, cooling narrows the gap without closing it. Ambient air passes through a wetted medium, cools by evaporation, then transfers that cold to a separate closed coolant loop through a dry heat exchanger. No moisture ever touches the electronics (industry cooling guides, 2026). In a cool climate, a 25 MW facility built around this architecture can run on free cooling for 90–95% of the year, drawing water only during the hottest stretches (industry engineering analysis, 2026).

Closed-loop dry heat rejection skips evaporation entirely. Water recirculates in a sealed loop; heat leaves through a dry exchanger to ambient air. Vantage Data Centers’ Wisconsin campus, built this way, is projected to draw about 22,000 gallons a day at peak, against roughly 5,000,000 gallons a day for a similarly sized campus running conventional evaporative towers (Vantage Data Centers, 2026). Microsoft’s 2024 rollout of zero-water, closed-loop cooling for its AI data centers is reporting savings exceeding 33 million gallons per facility annually (Microsoft, 2024).

Full dry cooling sits at the far end of the spectrum: fans and dry radiators only, with WUE approaching zero and a 25–35% energy penalty to compensate (industry cooling analysis, 2026). It makes sense in water-stressed regions where the power premium is cheaper than the political cost of a facility straining the local aquifer. It makes less sense in a country with abundant hydro or nuclear baseload and an already stressed transmission grid, where every extra megawatt of cooling fan load competes with the demand-response headroom described above.

Cooling architecture comparison

This is also the direct line to the water-positive pledges hyperscalers have made public. Microsoft is targeting a 40% improvement in water-use intensity by 2030 and had already reached 25% by 2025 (Microsoft, 2026), while Google has committed to replenishing 120% of the water it consumes by 2030, backed by $500 million in public water infrastructure investment (Google/Computer Weekly, 2026). A water positive data center commitment can’t be met through offsets alone, and both companies’ targets depend on cooling architectures that draw less water at the source rather than accounting tricks downstream.

Where Heat Reuse and Desalination Fit

Two other utility-adjacent contributions round out the picture, and ModulEdge has already covered both in depth elsewhere. Turning rejected heat into district-heating or greenhouse revenue is the subject of Data Center Heat Reuse: Greenhouses, Desalination, and the Grade of Heat You Have, including the 221 TWh annual EU technical potential and why your cooling grade decides which offtaker you can serve. The revenue mechanics and a worked 1 MW calculation live in Data Center Waste Heat Recovery: Turn Cooling Costs Into Revenue. If heat reuse or desalination is the angle you need, start there. This post stays focused on the grid and the water tap.

Why Factory-Built Modular Data Centers Are Positioned for Both

Grid flexibility and water-reduced cooling share a requirement that traditional construction struggles with: the decision has to be made at design time, not bolted on after commissioning.

A modular data center is engineered once, in a factory, and repeated. That means the critical/non-critical load segmentation Verrus is building into new facilities for microgrid flexibility can be a standard module spec rather than a custom retrofit. It means the closed-loop or adiabatic cooling plumbing, the actual pipe runs, heat exchangers, and control logic that separate a 22,000-gallon-a-day facility from a 5-million-gallon one, gets pre-integrated and factory-tested instead of value-engineered out during a site build. And it means battery and controls integration for VPP or BYOC-style participation can ship as a module option rather than an 18-month, multi-vendor site project.

Modular data centers are category-faster to deploy than traditional construction, and that matters directly here: a utility offering expedited interconnection in exchange for flexibility, the way PGE and GridCARE did in Oregon, rewards whoever can actually build to the timeline. A facility still in design when that offer arrives gets nothing. ModulEdge designs modular data centers at 5–150 kW per rack, with AI inference workloads positioned at 40 kW per rack and above, integrated power, cooling, fire suppression, and monitoring, and builds designed to meet Tier III/Tier IV principles. Siting flexibility means the cooling architecture and grid-interconnection strategy can be matched to the site, whether that’s a Czech industrial park or a water-stressed region abroad, instead of forced onto whatever a legacy building happened to have. The full delivery model behind that flexibility is in the modular data center guide.

Most operators still spec their cooling and power architecture around the building they were given, not the grid and water constraints of the site they’re on. That’s backwards. The interconnection queue is over 2,060 GW deep and the water doesn’t come back once it’s evaporated. Spec for both, or spec twice.

Frequently Asked Questions

What is data center demand response? Data center demand response is a facility’s contractual agreement to reduce, shift, or self-supply its electricity draw when a grid operator signals stress, in exchange for market payments, capacity credits, or expedited interconnection. Google’s 1 GW of demand response agreements with utility partners since 2025 is among the largest publicly disclosed examples (Google, 2026).

What is a virtual power plant in a data center context? A virtual power plant (VPP) aggregates distributed resources (batteries, on-site generation, flexible load) and dispatches them as a single controllable grid resource. The Google-Voltus “Bring Your Own Capacity” deal assembles up to 100 MW this way across the PJM footprint (Data Center Knowledge, 2025).

Why do utilities want data centers to be flexible instead of just adding generation? Because over 2,060 GW of capacity is stuck in US interconnection queues as of the end of 2025 and median wait times now exceed five years (LBNL, 2026). A Duke University study found the grid could absorb at least 76 GW of new flexible load with curtailment for as little as 0.25% of annual hours, rising to 126 GW at 1% of annual hours, avoiding new transmission and generation build-out entirely (Duke Nicholas Institute, 2025).

What’s the difference between dry cooling and adiabatic cooling? Dry cooling uses fans and radiators only, drawing no water but consuming 25–35% more cooling energy. Adiabatic (indirect evaporative) cooling evaporates water to cool incoming air, then transfers that cooling to a sealed loop through a dry heat exchanger, drawing water only during peak heat and running on free cooling for 90–95% of the year in temperate climates (industry engineering analysis, 2026).

How much water can closed-loop cooling actually save? At Vantage Data Centers’ Wisconsin campus, closed-loop cooling is projected to draw about 22,000 gallons a day at peak, versus roughly 5,000,000 gallons a day for a similarly sized campus using conventional evaporative towers (Vantage Data Centers, 2026). Microsoft reports savings exceeding 33 million gallons per facility annually from its zero-water closed-loop rollout (Microsoft, 2024).

What does “water positive” mean for a data center? A water-positive commitment means a company pledges to replenish more water to local watersheds than its facilities consume. Google has pledged to replenish 120% of consumption by 2030, backed by $500 million in water infrastructure investment, and Microsoft is targeting a 40% improvement in water-use intensity by 2030, having reached 25% as of 2025 (Google/Computer Weekly, 2026; Microsoft, 2026).

Do EU regulations require data centers to provide grid flexibility? The EU’s proposed Cloud and AI Development Act (June 2026) introduces a tripartite framework requiring operators, grid stakeholders, and public authorities to coordinate on demand-side flexibility as part of tripling EU data center capacity (European Commission, 2026). Separately, ACER’s approved methodology requires member states to complete national flexibility assessments by July 2026 and set targets by January 2027 (ACER, 2025).

Are modular data centers better suited to grid and water flexibility than traditional builds? Factory-built modular data centers can have load segmentation, closed-loop or adiabatic cooling plumbing, and battery/controls integration for grid participation engineered into the standard design rather than retrofitted on-site. They are also category-faster to deploy than traditional construction, which matters when utilities reward flexibility with expedited interconnection.