AI Data Centres Could Become Flexible Grid Assets Instead

Last Updated 2 hours ago by Kenya Engineer

The rapid expansion of artificial intelligence is creating a new problem for electricity systems. Data centres are consuming enormous amounts of power, and the largest facilities being planned today can require hundreds of megawatts. In some locations, utilities are struggling to build generation and transmission infrastructure quickly enough to connect them.

But the same facilities may also offer something that conventional electricity consumers rarely provide: the ability to change how much electricity they consume, and when they consume it.

That possibility is now being tested in several markets.

Utilities, grid operators and technology companies are exploring ways to make data centres respond to electricity-system conditions by temporarily reducing, shifting or rescheduling computing workloads. The approach is known as demand response, and it could turn some data centres from rigid electricity loads into flexible grid resources.

The idea is particularly significant as AI pushes electricity demand higher.

The US Energy Information Administration estimates that US data-centre electricity consumption could rise from roughly 177–192 TWh in 2024 to between 383 and 793 TWh by 2030. At the same time, research cited by the Electric Power Research Institute indicates that some data centres could reduce peak electricity demand by between 10% and 30%, depending on the facility and workload.

The engineering opportunity lies in connecting those two trends.

A data centre does not have to consume the same amount of power all the time

A conventional industrial load is generally treated by the electricity system as something that consumes electricity when it needs it.

An AI data centre can be different. Not every computation has to happen at exactly the same second.

AI training, model development, data processing and other workloads can sometimes be moved to another time or location without affecting services delivered to users.

That creates a degree of flexibility.

If a grid operator knows that electricity supply will be tight at 6pm, for example, an AI operator could reduce selected computing workloads during that period and restore them later when the grid is less constrained.

The data centre has not stopped operating. It has changed the timing of some of its electricity consumption.

Research published in Nature Energy demonstrated the principle on a 256-GPU cluster running representative AI workloads. Researchers were able to reduce power consumption by 25% for three hours during periods of peak demand while maintaining quality-of-service requirements. The approach relied on software-based workload coordination rather than installing additional batteries or modifying the underlying hardware.

That is an important development. The flexibility is not necessarily coming from a giant battery sitting outside the building. It can come from the computers themselves.

The data centre becomes part of the control system

The conventional relationship between a data centre and the electricity grid is straightforward. The grid supplies electricity. The data centre consumes it.

The proposed model is more interactive. Grid conditions become an input into the data centre’s energy-management system. When the grid is under stress, software can determine which workloads can be delayed, shifted between machines or moved to another facility.

When conditions improve, those workloads can resume.

Lawrence Berkeley National Laboratory identifies four major mechanisms through which AI data centres can provide flexibility: shifting computational workloads in time or between locations, adjusting supporting facility infrastructure, using energy storage and deploying on-site generation.

This creates what could be described as a grid-interactive data centre. The facility still has to meet its computing obligations, but it gains another operating parameter: the amount of electricity it is consuming at any particular moment.

Google is already doing it

This is not entirely theoretical. Google has been developing demand-response capabilities at its data centres for several years. In March 2026, the company said it had reached 1GW of demand-response capacity integrated into long-term energy contracts with utility partners in the United States.

Google’s system can limit or shift a portion of machine-learning workloads in response to grid requirements. The company says this can help utilities balance supply and demand while reducing the need for additional capacity.

The concept builds on an earlier Google programme that shifted non-urgent computing tasks between times and locations depending on electricity-system conditions.

A data centre does not necessarily have to shut down critical services to participate in demand response. The flexibility can come from workloads that have some tolerance for delay.

The AI workload itself becomes an energy-management tool

AI computing is particularly interesting because its electricity demand is not always uniform. Training a model can involve large computing loads for extended periods. Other workloads may be less time-sensitive.

A software layer can potentially determine:

• which workloads are critical;
• which can be delayed;
• which can be moved to another server;
• which can be transferred to another data centre;
• how much computing capacity can be reduced;
• and how long the reduction can be maintained.

That information can then be coordinated with electricity prices or signals from a utility or grid operator. The result is a form of computational demand response.

Instead of asking a factory to switch off a machine, the grid operator is effectively asking a computing facility to change its workload profile.

The electricity reduction can happen within the digital infrastructure.

The economics could be significant

There is a strong financial incentive behind the idea.

The rapid growth of data centres is forcing utilities to consider new generation, substations and transmission infrastructure. Building enough infrastructure to meet every possible peak in demand can be expensive.

A more flexible data-centre load could reduce some of that requirement.

A Duke University study cited by Reuters estimates that greater flexibility from data centres could save between US$40 billion and US$150 billion in capital expenditure over the next decade.

The benefits could extend to the data-centre operator itself.

Utilities and regulators are increasingly dealing with enormous new electricity connections. A data centre willing to reduce consumption during system constraints could potentially receive more favourable connection arrangements or gain access to capacity that would otherwise be unavailable.

In other words, flexibility could become part of the commercial proposition for getting a new data centre connected.

It could change the meaning of a 1GW data centre

A data centre advertised as having a 1GW electrical load does not necessarily need to behave like a 1GW constant load.

There could be a difference between: maximum electrical demand and firm electrical demand.

A facility might have a maximum requirement of 1GW but agree to reduce that demand by 100MW or 200MW during specified grid events. For the grid operator, that difference can be significant.

If a new data centre can guarantee that part of its load will be curtailed when the system is under pressure, the operator has more options for managing peaks.

This does not eliminate the need for generation and transmission. It can reduce the amount of infrastructure that has to be built specifically to accommodate the facility’s highest possible demand.

Storage makes the proposition stronger

Batteries can take the concept further. A data centre with on-site battery storage can potentially use the batteries to reduce its demand from the grid without immediately reducing computing activity.

The battery can discharge during a grid event while the computing workload continues. If the event lasts longer, workload flexibility can take over.

This creates a layered response: battery → workload adjustment → additional generation or grid support.

The combination could make large data centres considerably more flexible than conventional industrial loads.

Lawrence Berkeley National Laboratory identifies energy storage and on-site generation alongside computational workload flexibility as potential mechanisms for integrating AI data centres more effectively with the electricity system.

The data centre can also move the workload

One of the more unusual possibilities is geographical flexibility.

If an AI company operates data centres in different regions, some computing workloads can potentially be moved between facilities depending on electricity availability, grid conditions and other constraints.

A workload that does not need to be completed immediately could run where electricity is more readily available.

Australia is already testing this concept.

In September, Australia’s CSIRO announced a pilot using software to shift GPU-based AI workloads across time and locations in response to grid conditions. The project’s early modelling suggested that the power consumed by certain GPU workloads could potentially be adjusted by 20–50% within seconds of receiving a grid signal.

If such systems mature, a future AI company could operate its computing infrastructure almost like a portfolio of flexible electricity loads. The location of computation would become an energy decision as well as a computing decision.

But flexibility has limits

The idea should not be oversold. Not every data-centre workload can be interrupted. Cloud services, real-time applications and latency-sensitive computing cannot necessarily be moved or delayed without affecting customers.

Even AI workloads that appear flexible have deadlines.

Training a model for three days rather than two may have commercial consequences. A cloud customer running a time-sensitive process may not accept an interruption simply because electricity prices have risen.

Data-centre operators also have strict requirements around reliability.

The industry has traditionally been built around redundant power supplies, backup generators, batteries and multiple network connections precisely because interruption is unacceptable.

Demand response introduces another layer of operational complexity. The challenge is to identify what can be made flexible without compromising the service-level agreements that customers are paying for.

The grid also needs to change

Making a data centre flexible is only half the problem. Utilities and grid operators need mechanisms to recognise and reward that flexibility.

There must be clear rules governing how much load can be reduced, how quickly it must respond, how long the reduction must last and how frequently the operator can be called upon.

There also needs to be a financial incentive. If a data centre is being asked to invest in software, batteries, backup systems and operational capability to support the grid, it needs to receive some value in return.

The International Energy Agency says demand response remains significantly underused globally. It estimates that only around 100GW of demand response was being utilised worldwide as of 2024, despite much larger potential. The agency identifies market design, regulation and incentives as important barriers to wider deployment.

That means the technology is not the only challenge. The electricity market has to be designed to use it.

A new relationship between AI and electricity

The significance of this development goes beyond data centres. AI is rapidly becoming one of the largest new sources of electricity demand in some markets. The traditional response would be to build more generation, strengthen transmission networks and expand substations.

Those investments will still be required. But a second approach is emerging. Instead of treating large new loads as fixed customers that the electricity system must accommodate, utilities can ask whether those customers can participate in managing the system.

That changes the relationship. The data centre stops being merely a consumer of electricity and becomes an active participant in the electricity market.

It can potentially:

• consume electricity when the system has surplus capacity;
• reduce consumption during periods of grid stress;
• shift flexible workloads between time and location;
• use batteries to smooth its grid demand;
• and coordinate on-site generation with the wider network.

The resulting system could be more efficient than one designed around the assumption that every data centre will always consume its maximum contracted load.

An emerging industry alliance

The technology industry is beginning to organise around this idea.

In September, Google, NVIDIA and Emerald AI launched the AI Energy Management Alliance, bringing together companies across the AI and energy industries to promote data centres capable of dynamically adjusting electricity consumption according to grid conditions.

The alliance’s timing is significant. The AI industry is simultaneously facing a shortage of computing infrastructure and a shortage of available electricity connections. Making data centres flexible could address both problems.

A facility that can demonstrate that it will reduce demand during periods of grid stress may be easier to integrate into a constrained electricity network. That could allow some AI infrastructure to connect sooner without waiting for every planned transmission or generation upgrade to be completed.

It is not a replacement for grid investment. It is a way of getting more use out of the grid that already exists.

What this means for Africa

The issue is likely to become increasingly relevant to Africa. Several African countries are positioning themselves as destinations for data-centre investment, cloud infrastructure and AI services.

Kenya is already one of East Africa’s major digital infrastructure markets, while South Africa has a considerably larger data-centre ecosystem. Nigeria, Egypt and other markets are also attracting investment in digital infrastructure.

The electricity question will become more important as facilities become larger and AI workloads increase.

For countries with constrained grids, a data centre capable of reducing its demand during system stress could be easier to integrate than an equivalent rigid load.

That creates an opportunity for policymakers and utilities. Instead of only asking: “How much electricity does the data centre need?” they could also ask: “How flexible can that electricity demand be?”

That could eventually influence tariffs, connection agreements and even the location of new facilities. A data centre that can demonstrate flexible demand might receive different grid-access conditions from one that requires uninterrupted maximum power at all times.

Kenya’s opportunity is not just about building more power

For Kenya, this discussion comes at an interesting time. The country is investing in digital infrastructure while also trying to expand electricity generation and transmission capacity.

Kenya’s relatively high share of geothermal generation gives it an advantage in firm low-carbon electricity, while additional wind and solar capacity increases the importance of managing variable generation.

The lesson from the emerging data-centre model is that the electricity system does not have to be designed around passive loads. Large digital facilities could potentially become participants in managing demand.

That would require appropriate technical standards, tariffs, grid-control systems, data-sharing arrangements and commercial incentives. It would also require data-centre operators to design flexibility into facilities from the beginning rather than attempting to retrofit it later.

The data centre of the future may have two jobs

The first job will remain computing. The second could be helping manage the electricity system that powers the computing. That is a significant change in the way large digital infrastructure is viewed.

For years, the central question surrounding data centres was how much electricity they would consume and where utilities would find enough generation to supply them.

The next question may be more nuanced. How much of that demand can be made flexible? If the answer is significant, AI data centres could become more than enormous new electricity customers.

They could become grid-interactive infrastructure capable of responding to electricity shortages, absorbing surplus power, shifting workloads and helping utilities postpone some costly network upgrades.

The paradox is emerging quickly: the infrastructure creating one of the world’s fastest-growing electricity demands may also become part of the technology used to manage that demand.

For engineers, that changes the design problem. The future data centre may no longer be designed simply to receive electricity from the grid. It may be designed to interact with it.