- Researchers from WU Vienna, IIASA and KAUST published findings this week suggesting AI data centres can become grid assets by accelerating investment in new energy and storage technologies.
- A 2025 Duke University study found that flexing data centre demand by 25% for two hours could free up 100 gigawatts of U.S. grid capacity, the equivalent of avoiding more than $2 trillion in new infrastructure spending.
- Google, Microsoft and NVIDIA are deploying on-site batteries and AI-powered load management software, including Emerald Conductor, to shift or curtail workloads in real time, with experiments showing power draw can drop to around 66% in under a minute.
A new paper from researchers at WU Vienna, IIASA and KAUST argues that AI data centres, widely treated as a grid liability, can be operated as grid assets. The mechanism is straightforward: AI training workloads can be paused or shifted, on-site batteries can shave peak demand, and the same AI driving electricity growth can be used to optimise the grid itself. Whether policy keeps pace with the engineering is the open question.
AI Data Centres Pivot to Grid Assets
The research, published this week, challenges the assumption that surging data centre demand is purely a problem for grid operators. The International Energy Agency projects global electricity demand from new data centres will double by 2030, a rate that is outpacing renewable deployment. In the United States, data centres could account for up to 17% of national electricity consumption by that year, according to the IEA, up from 2024 projections. The WU Vienna-led team argues that with the right policies, this demand could drive investment in storage and generation technologies that benefit the broader electricity system, not just the facilities consuming power.
Flexible Workloads Ease Peak Demand
The key hardware insight here is temporal flexibility. Model training, large-scale batch inference and similar workloads are not time-critical in the way a customer-facing API call is. They can be paused, throttled or geographically rerouted without degrading user experience, which makes them ideal candidates for demand response programmes.
Google has moved from pilot to operational deployment. As of 2025 and 2026, the company has demand response agreements with Indiana Michigan Power and the Tennessee Valley Authority, targeting machine learning workloads specifically, building on an earlier demonstration with the Omaha Public Power District. These are reported to be the first agreements of their kind to explicitly target ML compute as the curtailable load. The response times are faster than most grid operators expect from industrial loads: experiments have shown a data centre operator can reduce power draw to around 66% of capacity in under a minute, and hold at roughly 10% for extended periods in some cases.
A 2025 Duke University study put a number on the aggregate opportunity. If new AI data centres flexed consumption by 25% for two hours at a time, for fewer than 200 hours a year, the study estimated 100 gigawatts of additional U.S. grid capacity could be freed, equivalent, the researchers calculated, to more than $2 trillion in otherwise-necessary infrastructure spending.
NVIDIA is approaching the same problem at the hardware-software stack level. Working with Emerald AI and energy companies including AES, Constellation, Invenergy, NextEra Energy, Nscale Energy & Power and Vistra, NVIDIA is developing what it calls “power-flexible AI factories.” The setup pairs Emerald Conductor software with NVIDIA’s DSX Flex to orchestrate compute flexibility alongside on-site energy resources: time-sensitive workloads maintain full throughput while lower-priority tasks shed load during grid stress events.
On-site Energy Storage Enhances Grid Capacity
Researchers analysing energy data from 96 UK data centres, sourced from UK Power Networks, found that on-site batteries used to meet peak demand could cut peak electricity imported from the grid by 10-15%. For a constrained distribution network, that reduction defers costly external infrastructure upgrades.
Google is scaling this approach with capital. In July 2026, the company announced backing for the Steel River Energy Center in Arkansas, described as the largest solar project in the United States upon completion in 2029. The facility is designed for 2.5 gigawatts of peak generation and 2.9 gigawatt-hours of battery storage, sized specifically to supply power during evening peak hours when solar panels are offline. That storage capacity feeds back into the regional grid, not just Google’s own facilities.
Microsoft is working at the rack level. The company is exploring rack-level energy storage and is co-developing power smoothing features with NVIDIA for GB200 GPU deployments. The engineering problem being solved is specific: synchronous AI training jobs produce correlated power spikes across thousands of GPUs simultaneously. Left unmanaged, those spikes can couple to grid frequencies in ways that stress infrastructure. Smoothing them at the rack is cleaner than trying to manage the effect at the substation.
AI Optimises the Grid Itself
Google has also partnered with PJM Interconnection, North America’s largest grid operator, serving 67 million people, and Tapestry, an Alphabet-incubated venture, to build AI tools for managing energy interconnections. The target is PJM’s backlog of new energy projects queued for grid connection; the tools aim to accelerate that process and lower the cost of bringing new generation online. For more on how AI infrastructure is reshaping energy demand, see our coverage of AI’s projected 945 TWh annual energy footprint by 2030.
Microsoft is running a parallel effort with the Midcontinent Independent System Operator (MISO), covering U.S. Midwest power. The collaboration centres on a unified data platform for predicting grid conditions and accelerating data-driven dispatch decisions. Meta is taking a portfolio approach, using AI to optimise energy procurement across smart grid and renewable sources, adjusting data centre energy usage in real time based on grid signals.
The pattern across all three companies is the same: the compute that is straining the grid is also the compute being used to manage it. That creates a genuine feedback loop, though how much of it is operational today versus planned is difficult to verify from public disclosures alone.
Policy Reforms and Infrastructure Requirements
The WU Vienna research team argues the technical capability already exists; the regulatory framework has not caught up. Power-system planning still largely treats large computing facilities as fixed, inflexible peak loads. That classification drives costly infrastructure upgrades and extended interconnection queues, the same queues that AI tools are now being deployed to clear.
The researchers’ proposed fixes are specific: grid-connection rules that reward measurable net benefits to the energy system, streamlined pathways for data centres to deploy on-site batteries and clean generation, and removal of the barriers that currently prevent large loads from offering grid services commercially. They also call for financing mechanisms to scale next-generation storage at the pace AI demand is growing.
The commercial incentives are beginning to align, demand response agreements reduce operating costs, and on-site generation hedges against power price volatility, but until interconnection rules and grid-service markets treat flexible compute loads the same way they treat flexible generation, the full capacity benefit identified by the Duke study will remain theoretical. For more coverage of AI chips and infrastructure, visit our AI Hardware section.
