AI Datacenters and the Power Grid: Are Electrical Networks Ready for the Next Energy Surge?
- May 28
- 7 min read
The grid is not ready everywhere.
AI data centers have moved from an IT planning issue to an energy-system issue. According to the IEA’s Energy and AI executive summary, data centres consumed around 415 TWh in 2024, and the United States accounted for the largest share of that demand. The U.S. Department of Energy says data centers consumed 4.4% of U.S. electricity in 2023 and could reach 6.7% to 12% by 2028.
That growth is not evenly spread. NERC says new AI data centers account for most projected North American demand growth over the next decade, with steep increases in Texas, PJM and WECC. In other words, the real question is not whether electricity demand is rising; it is where the infrastructure is still too slow, too thin, or too inflexible to absorb it. (nerc.com)
Why AI Data Centers Are Different
On the IEA’s AI topic page, traditional data centers are described as 10 to 25 MW loads, while hyperscale AI centers can exceed 100 MW. The same page notes that the largest announced site could consume as much electricity as 5 million households, which helps explain why these projects are no longer “just” digital infrastructure.
The 2024 Berkeley Lab report shows why this matters now: U.S. server electricity use rose from about 30 TWh in 2014 to nearly 100 TWh in 2023, while GPU-accelerated AI servers grew from less than 2 TWh in 2017 to more than 40 TWh in 2023. The report also projects a wide 2028 range because equipment shipments, operating patterns and cooling assumptions can push total demand in very different directions. (eta-publications.lbl.gov)
Cooling is part of the equation too. Berkeley Lab notes that future outcomes depend on choices such as liquid cooling and moving away from evaporative cooling, while PUE and WUE can shift with climate and design. That is why the discussion is broader than racks and GPUs; it is an infrastructure question, which is also explored in our article on infrastructure, resilience and energy in 2026.
Are Electric Grids Ready for the Surge?
Short answer: partially, but unevenly. NERC’s 2025 Long-Term Reliability Assessment says emerging large loads such as AI data centers present unique challenges in BPS planning and operations, and that planned data centers in WECC can reach as high as 40% of the demand forecast in some balancing areas.
The IEA also warns that concentration matters: nearly half of U.S. data center capacity sits in five regional clusters. That means a project can overwhelm local feeders, substations or transmission paths long before the national grid runs out of generation.
Regulators are already reacting. FERC launched a review of co-location issues for large AI-enabled data centers in PJM to clarify rules and protect grid reliability and consumer costs, while DOE’s Speed to Power initiative is designed to accelerate large-scale transmission and generation buildout. (ferc.gov)
The practical takeaway: AI data center readiness is a regional engineering problem, a market-design problem, and a continuity problem at the same time. It is not enough to have national capacity on paper; the load must be deliverable, controllable and recoverable at the site level.
The Main Pressure Points
The pattern is consistent across the sources: local bottlenecks, not national averages, decide whether a project moves on time. The table below summarizes the pressure points that show up most often in AI datacenter planning.
Grid-readiness checklist for AI data centers
Pressure point | Why it matters | Typical response |
|---|---|---|
Concentrated growth | Nearly half of U.S. data centre capacity sits in five regional clusters, so one metro area can hit a bottleneck before the broader grid does. | Stage the build, study feeder and transmission impacts early, and avoid assuming that national spare capacity is available at the project site. (energy.gov) |
Load swings and model uncertainty | NERC says rapidly changing and often unpredictable power use raises balancing, frequency and voltage concerns. (nerc.com) | Require measurable flexibility, clearer ramp limits and better operating models before energization. (dcflex.epri.com) |
Interconnection and cost allocation | FERC says co-location can have huge ramifications for grid reliability and consumer costs, which makes tariff design and responsibility sharing critical. | Use transparent interconnection rules and cost-allocation frameworks so other customers do not subsidize the project by accident. |
Cooling and water | Berkeley Lab shows that liquid cooling, evaporative cooling and climate conditions can materially change energy and water outcomes. | Match the cooling architecture to local climate, water risk and operating density instead of copying a generic design. |
Time to power | DOE says large-scale demand growth is already placing significant burdens on the grid, while EPRI says many projects are slowed by long and uncertain timelines. | Use onsite power, storage and phased energization to bridge the gap between project launch and firm utility delivery. (energy.gov) |
Across these risks, the common theme is simple: the closer a project gets to 100 MW and beyond, the more it behaves like a system resource rather than a standalone building. That is why readiness must be assessed region by region, substation by substation, not by a national headline number.
What Strategies Actually Keep Growth and the Grid in Balance?
The best answers combine new supply, smarter demand, and better site design. DOE says data centers can become grid assets through onsite generation and storage, energy community reuse of retired coal facilities, innovative rate structures, and next-generation options such as geothermal, advanced nuclear, long-duration storage and efficient semiconductors.
EPRI’s Flex MOSAIC framework adds something equally important: a common language for flexibility. It describes large-load performance in terms of notification time, duration, depth of adjustment, ramp behavior and availability, which gives utilities a way to plan around the actual behavior of a data center instead of worst-case assumptions alone.
AI can also help the grid. The IEA says widespread adoption of existing AI applications in the electricity sector could save up to $110 billion annually and unlock 175 GW of transmission capacity. So the challenge is not whether AI belongs in the energy system; it is whether the sector uses AI to expand capacity, predict constraints and coordinate demand faster than those constraints arrive.
At Score Group, this is where Noor ITS, Noor Energy and Noor Technology intersect: the digital backbone, the energy envelope and the automation layer need to be designed together for high-density AI workloads. Our IT infrastructure solutions for networks, servers and storage matter when server density rises, while energy management services keep consumption visible and controllable. For resilience planning, the logic is the same as in DCIM software for unified energy and capacity management and in automated PRA testing for a more reliable recovery.
A Simple Operating Roadmap
Measure the real load shape, not only the nameplate MW, because NERC warns that rapidly changing large loads can distort balancing, frequency and voltage planning.
Start interconnection and transmission studies early, because DOE says large-scale load growth is already placing significant burdens on the U.S. grid and demands faster infrastructure delivery.
Build flexibility into workloads, onsite power and storage, because EPRI treats flexibility as a measurable capability rather than a vague promise.
Choose cooling based on climate, water availability and operating density, because Berkeley Lab shows that liquid cooling and evaporative cooling have different energy and water trade-offs.
Test continuity and recovery regularly, and use automated PRA testing for a more reliable recovery so recovery assumptions are validated instead of hoped for.
FAQ
Are AI data centers ready for the next energy surge in the electrical grid?
Not uniformly. Some regions can absorb new load faster than others, but NERC says emerging large loads such as AI data centers create unique planning and operational challenges. FERC has also opened a review of co-location issues because reliability and consumer-cost questions are real, not theoretical. The short version is that readiness depends on local transmission, reserve margins, interconnection timing and the amount of flexibility the site can provide.
How much electricity do AI data centers consume and will it spike grid costs?
The scale is already large. The IEA estimates that data centres used about 415 TWh globally in 2024, while DOE says U.S. data centers consumed 4.4% of U.S. electricity in 2023 and may reach 6.7% to 12% by 2028. Whether costs spike depends on where the load lands and how it is priced. Without new supply, transmission and flexibility, local congestion can raise system costs; with better design, onsite storage and fair tariffs, the impact can be contained.
What are the biggest grid reliability challenges posed by hyperscale AI data centers?
The biggest challenges are concentration, variability and speed. Hyperscale AI sites can exceed 100 MW, and NERC says rapidly changing loads can create balancing, frequency and voltage issues if models are too simplistic. The IEA adds that a large share of U.S. capacity is concentrated in just a few regional clusters, which means a single project can strain a local grid even if the national system still looks comfortable. Reliability breaks first at the edge, not the average.
Can the U.S. power grid handle the rapid build-out of AI-focused data centers?
Yes, but not at the current pace of infrastructure delivery. DOE’s Speed to Power initiative exists because large-scale electricity demand growth is already creating burdens on the grid, and NERC says many data center projections are large enough to materially change regional planning. That means the answer is not a simple yes or no. The grid can handle the build-out if generation, transmission, permitting and flexibility move faster than they do today.
What strategies exist to balance AI compute growth with grid stability and affordability?
The most effective strategies combine flexibility and infrastructure. DOE points to onsite generation and storage, smarter rate structures, reuse of retired plant sites and new firm resources such as geothermal, nuclear and long-duration storage. EPRI adds a flexibility framework so operators can define what a data center can actually do on ramp, duration and availability. In practice, that mix helps speed up connections, reduce surprise costs and keep the grid stable while demand grows.
And now?
If you are planning a new facility, a retrofit or a power roadmap, start with Score Group’s home page and explore DataCenters solutions for critical infrastructure. From there, you can connect Noor ITS, Noor Energy and Noor Technology into one operating model that supports performance, resilience and efficiency.
