Data Center Flexibility: Smart, Quick 35-Hour AI Grid Fix

Data center flexibility could let new AI data centers plug into grids that are already close to full, according to a paper published in the journal Joule by an international team led by TU Wien. The researchers argue that data centers must adapt to the power grid, “not just the other way around”, and point to a US modelling study in which the sites examined would have needed to cut grid demand for no more than 35 hours a year to stay within existing grid capacity.

That 35-hour figure made the headlines on 28 September 2026, and it deserves a closer look. It comes from a December 2025 study by Camus Energy, encoord and Princeton University’s ZERO Lab, funded by Google. It is the worst case among four constrained sites, and it counts only transmission limits. The same study found another 32 or so hours a year of generation shortfalls, and on-site backup running for 40 to 70 hours a year in total.

This article explains what the Joule paper argues, where the 35 hours comes from, and what the full set of numbers says about batteries, gas generators and pausing AI workloads. It also covers the cost case, how regulators in Europe, Norway and the US are writing data center flexibility into connection rules, and what it all means for utilities, developers and the businesses that buy AI capacity.

Table of contents

What the Joule Paper Says About Data Center Flexibility
Why Grids Cannot Keep Up With AI Demand
Where the 35-Hour Figure Comes From
Reading 35 Hours Correctly: Data Center Flexibility Has Two Parts
What Data Center Flexibility Looks Like on Site
Can AI Workloads Really Pause?
The Cost Case for Data Center Flexibility
How Regulators Are Writing Data Center Flexibility Into Connection Rules
Data, Transparency and What Communities Get
What Data Center Flexibility Means for Utilities, Developers and Businesses
Data Center Flexibility FAQs
References

What the Joule Paper Says About Data Center Flexibility

The paper, “Governing AI infrastructure on decarbonizing grids”, is led by Reda El Makroum of TU Wien’s Institute of Energy Systems and Electrical Drives. It makes a governance case for data center flexibility rather than presenting a new simulation: it draws on existing studies, rules and cases to set out what grid access for AI data centers should require.

A choice between data center flexibility and fossil fuels

The paper opens with a case from the US. xAI powered its Colossus 2 data center in Memphis, Tennessee, with gas turbines parked across the state line in Southaven, Mississippi, classed as mobile units and run without air permits. The researchers warn that AI could drive “a resurgence of fossil fuels” unless data center flexibility becomes part of how new demand is connected.

Four recommendations

The team’s conclusions fall into four groups. Data centers should become flexible consumers under special contracts, with a guaranteed base capacity plus extra capacity that can be reduced when the grid is under strain. Operators should share better data on how their consumption changes over time. Host communities should get binding guarantees that they will not pay for grid expansion. And all of this should be planned early, before sites are chosen.

Who wrote it

The paper has 14 authors. Ten of them spoke at an NTNU workshop on AI and data centres in the energy transition, including researchers from NTNU, the University of Oslo, Oxford, UCL, Cardiff and INESC TEC, plus Salesforce’s head of AI sustainability. That explains the Norwegian examples: the country is a data center hub with almost fully renewable power, yet its grid “does not have enough capacity everywhere”.

Why Grids Cannot Keep Up With AI Demand

Data center flexibility matters because demand is growing faster than grids can be built. The numbers are large and well documented, and they explain why data center flexibility has moved from theory to policy.

Demand is set to more than double

The International Energy Agency estimates that data centres used about 415 terawatt-hours of electricity in 2024, around 1.5% of global consumption, with the United States accounting for 45%. It projects that figure will more than double to about 945 TWh by 2030, “slightly more than Japan’s total electricity consumption today”, with AI the main driver as companies train and serve ever larger AI models.

Waiting three to seven years for power

The Camus study reports that US data centers face grid connection timelines of three to seven years, against 18 to 24 months to build the facility itself. New transmission lines can take seven to ten years. That mismatch is why, as El Makroum puts it, companies ask “Where is the grid well developed? Where can they obtain a suitable grid connection as quickly as possible?” before they ask where power is cheapest.

Gas is filling the gap

The Joule paper reports that 39% of the gas-fired generation capacity under development in the US at the end of 2025 was intended to supply data centers. The xAI case shows the extreme end. In April 2026 the NAACP sued xAI, alleging that it runs 27 gas turbines without an air permit to power Colossus 2. We have covered what happens when power arrives late in our piece on Oracle’s force majeure notice over data center power.

Where the 35-Hour Figure Comes From

The Joule paper cites the Camus, encoord and Princeton study as its evidence that data center flexibility works in practice. That study is published in full, with an appendix, and its site tables are where the 35 hours comes from.

Six sites at one PJM utility

The team modelled six candidate sites for a new 500-megawatt data center, all inside one utility’s territory in the PJM market and within about 125 miles of each other. The sites were given animal names to hide their real locations. The model combined real transmission system data and hourly operation across all 8,760 hours of a year with Princeton’s GenX capacity model and NREL’s REopt site model.

The site-by-site numbers

Two sites, Shark and Snake, near the 500 kV backbone, could take the full 500 MW with no transmission limits at all. The other four could not.

Site	Firm grid limit	Curtailed hours a year	Share of year	Longest event	Events a year
Koala	326 MW	7	0.08%	4 hours	3
Whale	322 MW	11	0.13%	4 hours	4
Pony	314 MW	13	0.15%	5 hours	4
Hare	154 MW	35	0.40%	16 hours	4
Shark, Snake	Full 500 MW	0	0%	None	0

Source: Camus, encoord and Princeton ZERO Lab, Table 2. Transmission limits only.

Thirty-five is the worst case

So “no more than 35 hours” is accurate, but it is the ceiling, not the typical result. Three of the four constrained sites needed between 7 and 13 hours. Hare, which could draw only 154 MW firm, less than a third of its 500 MW, needed 35. The authors warn that these results “do not necessarily indicate” that other sites in the territory, or in PJM more broadly, would stay under 1%.

Reading 35 Hours Correctly: Data Center Flexibility Has Two Parts

A data center can hit two different limits. Transmission limits are local: the wires and substations near the site cannot carry more power. Generation limits are system-wide: the market does not have enough accredited power plants to guarantee supply at peak. The 35 hours covers only the first.

Transmission hours versus generation hours

The study estimates about 32 hours a year when a flexible data center could also be asked to cut demand because of generation shortfalls, usually in two or three extreme-weather events lasting up to about 16 hours. The report did not measure how often the two kinds of event overlap. If they never did, Hare’s total would be about 67 hours, or 0.76% of the year.

40 to 70 hours of on-site backup

The study’s own summary figure is broader than the headline. At each modelled site, on-site or co-located resources were dispatched for 40 to 70 hours a year to stay within transmission or generation limits. That is still less than 1% of the year, which supports the paper’s point. But any data center flexibility contract has to cover the combined hours, not the 35.

Hours a year a flexible data center must reduce grid demand (Camus, encoord, Princeton ZERO Lab)

Koala, transmission only 7 h

Pony, transmission only 13 h

Generation shortfalls, all sites 32 h

Hare, transmission only (the headline figure) 35 h

Upper end of on-site dispatch, all causes 70 h

Short events and long ones

The number of hours matters less than how they are grouped. At Koala, Whale and Pony, events lasted no more than four or five hours. Hare faced events of up to 16 hours, and its longest one needed 2,299 megawatt-hours of energy from somewhere other than the grid. A four-hour battery cannot cover that alone.

Available 99% of the time, not 99% of the energy

Grid power was available at every site for more than 99% of hours. But in the optimised runs, 95% to 96% of each site’s energy came from the grid, because the on-site kit was also used to shave expensive peak-price hours. Data center flexibility, in other words, changes how a site buys power all year, not only in the 35 hours.

What Data Center Flexibility Looks Like on Site

To show what data center flexibility needs in practice, the Camus study modelled cost-optimal kit for the least constrained site, Koala, and the most constrained, Hare. Both combine batteries, small gas generators and “compute flexibility”, meaning the ability to slow or move AI work.

Resource	Koala (7 h a year)	Hare (35 h a year)
Firm grid service	326 MW	154 MW
Battery storage (4-hour)	49 MW / 196 MWh	143 MW / 572 MWh
On-site gas for flexibility	11 MW	60 MW
Co-located solar	26 MW	None
Compute flexibility	25% of load, up to 20 h a year	25% of load, up to 20 h a year
Headroom below 500 MW nameplate	65 MW	65 MW

Source: Camus, encoord and Princeton ZERO Lab, Table 4. The report’s text gives Hare slightly different figures, 155 MW / 620 MWh of batteries and 48 MW of gas, so treat these as approximate sizes.

Compute-based data center flexibility does real work

Both sites assume the data center can cut up to 25% of its load, a maximum of 109 MW, for up to 20 hours a year. At Koala, compute flexibility alone could have covered all seven curtailed hours, though the authors note that would leave “no operational margin for error”. That is why even the easy site carries batteries.

Headroom most sites already have

The modelled load peaks at about 435 MW, or 87% of the 500 MW nameplate, leaving 65 MW of headroom. The authors call this conservative, since most data centers run at lower utilisation. Some data center flexibility, in other words, already exists in the gap between what a site is built for and what it draws.

Gas does not disappear

On-site gas plays two roles. It fills long events such as Hare’s 16-hour one, and at Koala 202 MW of on-site gas nameplate provides 168 MW of accredited capacity so the site can be classed as firm. The study treats clean energy contracts as the preferred source and gas as the fallback, but it is honest that gas is part of the answer.

Can AI Workloads Really Pause?

Batteries and generators are the familiar tools of data center flexibility. The newer idea, and the one the Joule paper leans on, is that AI computing itself can bend around the grid.

Training can move, serving users is harder

The paper notes that “the training of large AI models, for example, could take place when particularly large amounts of renewable energy are available.” Training runs last weeks and can be checkpointed. Serving live chatbot or agent requests is less forgiving, because users expect answers in seconds. Most of a site’s data center flexibility will come from batch and training work.

The Phoenix field test

The best public evidence comes from Emerald AI. In a field demonstration published in Nature Energy, its software cut the power of a 256-GPU cluster running AI workloads in a hyperscale cloud facility in Phoenix, Arizona, by 25% for three hours during peak grid events, while keeping quality-of-service guarantees and without batteries or hardware changes. We covered Emerald’s industry coalition in our piece on Emerald AI and the grid.

Moving work between regions

Large cloud computing providers have a further option: moving work to a data center in another region where the grid is not under strain. That turns a local curtailment into a routing decision. It only works for workloads without strict location or latency requirements, which is why data center flexibility contracts need to say exactly how much load can move.

The Cost Case for Data Center Flexibility

The Camus study also asks who pays for data center flexibility, and who pays without it. It modelled adding data center demand to PJM in 2030 and compared a traditional firm-only connection with a flexible one.

$764 million a year per gigawatt

Under a firm-only connection, each gigawatt of new data center demand required 2.17 GW of new generation, including 1.1 GW of gas and 775 MW of batteries, and added $764 million a year in system supply costs. That is the bill other customers could face if the data center does not cover it.

What data center flexibility and self-supply cover

Three pieces offset that bill. A flexible connection with 20% of load on conditional service avoids 273 MW of new capacity, worth $78 million. “Bring your own capacity”, in which the data center contracts its own accredited supply rather than waiting for the utility, internalises $326 million. The data center’s energy payments cover $329 million more.

Who covers the $764M a year of system costs per GW of new demand

Avoided by flexible connection $78M

Paid through bring-your-own capacity $326M

Covered by the data center’s energy payments $329M

Left for other customers $31M

The three pieces add up to $733 million, 96% of the $764 million. The report describes this as reducing the net system cost increase “by nearly 100%”.

Why developers would sign up

For a developer the incentive for data center flexibility is speed. The study estimates a flexible 500 MW site could reach full power in about two years, three to five years sooner than a traditional connection. It values that head start at $4.7 billion to $5.5 billion in present-value earnings per site for five years, against $1.2 billion to $1.4 billion in extra lifecycle costs. Those estimates assume revenue of $8 million per megawatt and a 45% margin.

Who funded it

The study was funded by Google, which reviewed the analysis before publication, and reviewers included Emerald AI and Tesla Energy. The authors say the conclusions are their own. Readers should weigh the findings with that in mind: the numbers come from one utility’s territory, and the sponsors benefit if data center flexibility becomes standard.

How Regulators Are Writing Data Center Flexibility Into Connection Rules

The Joule paper’s main demand is that grid access should be tied to data center flexibility. Several rulebooks already point towards data center flexibility.

Europe’s flexible connection agreements

The EU’s 2024 electricity market reform, Directive (EU) 2024/1711, requires regulators to create a framework for flexible connection agreements in areas with limited or no network capacity. Customers get a connection with agreed limits on how much they draw and when, rather than waiting for reinforcement. The Joule paper cites this directive in its argument.

Norway’s connection with conditions

Norway’s energy regulator, RME, allows grid companies and customers to agree a “connection with conditions for disconnection”, in which the customer accepts curtailment without compensation instead of waiting for new grid or paying the full connection charge. Neither side can impose it: both must agree, and the contract must spell out exactly when curtailment can happen.

The US: faster lanes for flexible loads

In the US, the Southwest Power Pool board approved a High Impact Large Load (HILL) process on 16 September 2025 to speed up connections for loads such as AI data centers. A February 2025 Duke University study estimated that the largest US grid areas could absorb 76 GW of new load if it accepted curtailment for 0.25% of its maximum annual use, rising to 98 GW at 0.5% and 126 GW at 1%. We have written about how communities can use regulation to gain leverage over data centers.

Jurisdiction	Mechanism	What the customer accepts
European Union	Flexible connection agreements (Directive 2024/1711)	Limits on withdrawal where network capacity is short
Norway	Connection with conditions for disconnection (RME)	Curtailment without compensation, agreed in advance
US, SPP region	High Impact Large Load (HILL) process	An accelerated, integrated study and operations path
US, PJM study	Firm plus conditional firm service, bring-your-own capacity	About 7 to 35 transmission hours plus about 32 generation hours a year

Data, Transparency and What Communities Get

The Joule paper’s other two recommendations get less attention than the 35 hours, but they decide whether data center flexibility schemes can be trusted.

Better data on consumption

“Another important issue is the availability of better data on the electricity consumption of data centers,” El Makroum says. “To plan grid infrastructure effectively, we need a better understanding of how their electricity consumption changes over time.” The authors suggest anonymised operational data and information about the hardware used. The Camus study makes a similar point: utilities hold data on when grid capacity is available, but developers rarely see it.

Who pays for the grid

Data centers bring investment, the paper notes, but create jobs mainly during construction, while grid expansion can raise costs for everyone. The researchers propose binding agreements so that local residents do not bear the cost of expanding the grid. That echoes rules already appearing in the US, which we covered in our look at California’s new data center energy and water rules.

Waste heat as a local benefit

The paper also suggests using waste heat from data centers, for example in district heating, as a direct benefit to the local community. Nordic cities already do this at scale, and it gives residents something tangible in return for hosting a large new load.

What Data Center Flexibility Means for Utilities, Developers and Businesses

The headline number is less important than the change in thinking behind data center flexibility: a data center connection is becoming a contract with conditions, not a fixed entitlement.

For utilities

Publish time-varying capacity data for candidate sites, as the Camus authors recommend, and define flexible service in concrete terms: the maximum hours a year, the longest single event and the notice period. A figure like 35 hours means little until the contract says whether it is transmission-only or all causes. Treat the curtailment signal as a cybersecurity matter too: a channel that can switch off hundreds of megawatts of load needs the same protection as any other grid control.

For data center developers

Size on-site kit for the longest event, not the annual total. Hare’s 35 hours included a 16-hour event that no four-hour battery could cover alone. Price in generation curtailments on top of transmission ones, and treat data center flexibility as something the operations team must prove, not just promise.

For businesses buying AI and cloud capacity

Ask your providers how their sites are connected and whether any capacity you rely on sits on conditional service. Batch work, model training and reporting jobs can often move in time, which is exactly the data center flexibility grids are starting to reward. Our data center operations team can help plan workloads and infrastructure around these constraints.

Data Center Flexibility FAQs

What does “35 hours of reduced demand” mean?

In a US study of six sites, the most constrained site had to reduce grid demand for 35 hours a year to stay within transmission limits. Three other sites needed 7 to 13 hours.

Does 35 hours include all grid limits?

No. It covers transmission limits only. The same study found about 32 more hours a year of generation shortfalls, and on-site backup ran for 40 to 70 hours a year in total.

What is data center flexibility?

It is a data center’s ability to cut or shift its grid demand when the grid is strained, using batteries, on-site generation or by slowing or moving AI workloads.

Who wrote the Joule paper?

A 14-author team led by Reda El Makroum of TU Wien, with researchers from Norway, the UK and elsewhere. The 35-hour figure comes from a Camus, encoord and Princeton study it cites.

Can AI data centers really cut power use on demand?

A field test in Phoenix cut a 256-GPU cluster’s power by 25% for three hours during peak events without hurting service quality. Training is easier to shift than live user requests.