⚡ electrical engineering

Grid topology and the spatial cost of AI data-centre siting

By comparing New Zealand's radial grid with Ireland's meshed ring, this study demonstrates that network topology, rather than short-run electricity prices, is the primary driver of AI data-center siting costs, with physical grid bottlenecks creating significant price disparities that vanish only when transmission losses are eliminated.

Original authors: Xiuyue Deng, Stephen Poletti, Timothy Hazledine

Published 2026-07-10

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Xiuyue Deng, Stephen Poletti, Timothy Hazledine

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a giant, hungry robot brain (an AI data centre) looking for a place to plug in. You need a massive amount of electricity, and you are very picky: you want to sit where the power is cheapest. But here's the twist: you don't care where you are geographically, as long as the bill is low. So, you ask the power grid, "Where should I build my home to save the most money?"

A team of researchers decided to play a game of "Grid Tag" to find the answer. They built two digital twins of real-world power systems: one for New Zealand and one for Ireland. They made the two systems look as similar as possible on paper—both have huge amounts of wind and water power, and both rely on natural gas to set the price when the wind stops blowing. Then, they dropped a giant, unmovable 100-to-500 MW "load" (your robot brain) onto every possible spot in both countries and watched what happened to the price.

The Big Surprise: The Shape of the Wires Matters More Than the Price of Gas

For a long time, people thought the cost of building a data centre depended on "parametric" things—like how much gas costs today, how much carbon tax is on the bill, or how windy it is that week. The researchers tested this by changing those numbers wildly in their simulations. They made gas prices go up and down by 50%, changed the weather to dry or wet years, and tweaked demand.

The result? Nothing changed the map.

The paper argues that the cost of where you sit isn't driven by these daily price fluctuations. Instead, it is locked in by the topology—the physical shape and layout of the power lines themselves. Think of it like this: if you are trying to move a heavy box, it doesn't matter if the price of the truck driver's fuel changes; if there is a bridge that is too narrow, you simply can't cross it, and that's the only rule that matters.

New Zealand: The Two-Island Split

New Zealand is like a house with two rooms separated by a very narrow, leaky hallway. The North Island and South Island are connected by a single, high-voltage bridge called the Cook Strait link.

When the researchers dropped their robot brain onto the South Island, the cost was 71.96 EUR/MWh. When they dropped it on the North Island, the cost jumped to 74.96 EUR/MWh. That is a 4.18% difference.

Why? Because the hallway (the bridge) has a "leak." Every time electricity crosses it, a tiny bit of energy is lost in the converter, just like water leaking out of a hose. The researchers found that this 4.18% price gap matched the one-way converter loss of the bridge exactly. Even if they changed the gas price or the wind speed, the gap stayed the same. The bridge was the only thing that mattered. The South Island was a "cheap zone," and the North Island was an "expensive zone," separated by a hard wall of physics.

Ireland: The One-Price Ring

Now, look at Ireland. Imagine Ireland as a giant, perfectly round racetrack made of thick, smooth wires. There are no narrow bridges or leaky hallways. It's a "meshed ring," meaning electricity can flow in many different directions to get to any spot.

When the researchers dropped their robot brain onto any of the five test spots in Ireland, the cost was almost exactly the same. The difference between the most expensive spot and the cheapest spot was a tiny 0.016%.

In this simulation, the price signal was "numb." No matter where you stood on the racetrack, the cost to plug in was identical. The researchers found that even if they changed gas prices, carbon taxes, or demand, the cost remained flat. The shape of the Irish grid was so well-connected that it smoothed out all the bumps. There was no "cheap zone" and no "expensive zone"—just one big, flat price plain.

What This Means for the Future

The paper rules out the idea that we can just fix data centre siting by tweaking market prices or carbon taxes. If the grid is like New Zealand's (bottlenecked), the price signal works perfectly: the market naturally tells you to build in the South because the North is too expensive to feed.

But if the grid is like Ireland's (a meshed ring), the market signal breaks down. The price is the same everywhere, so the market can't tell you where to build. In this case, the researchers suggest that the only way to solve the problem is to physically reinforce the grid (make the wires thicker) or use administrative rules to tell people where to build.

How Sure Are They?

The authors are very confident in these results, but with a specific caveat: these are simulations. They used a computer model called PyPSA-Earth to run thousands of scenarios. They didn't build a new data centre and wait ten years to see what happened. However, they tested their model against real-world data and ran it through extreme "what-if" scenarios (like changing gas prices by 50% or simulating droughts) to make sure the results held up.

They found that the "topology wins" rule held true in almost every single test. The only time the Irish grid showed a price difference was when they artificially froze the grid so it couldn't expand at all and then forced a massive load onto a remote spot. In that specific, stressed-out scenario, the "flat" price broke, and a congestion cost appeared. But under normal, routine operation, the grid stays flat.

The Takeaway

So, if you are a robot brain looking for a home:

In New Zealand, check the map. If you are on the South Island, you save 4.18% because you don't have to cross the leaky bridge. The shape of the wires decides your fate.
In Ireland, it doesn't matter where you stand. The wires are so well-connected that the price is the same everywhere. The market can't help you choose; you'll need a human planner to tell you where to go.

The paper concludes that we can't just rely on price tags to solve where AI should live. We have to look at the physical shape of the power grid first. If the grid is a bottleneck, the market works. If the grid is a smooth ring, the market goes silent, and we need a different plan.

Technical Summary: Grid Topology and the Spatial Cost of AI Data-Centre Siting

Problem Statement
The rapid growth of energy-intensive artificial intelligence (AI) data centers creates a conflict between the need for flexible, clean-energy supply and the imposition of large, inelastic, "footloose" loads. A critical unresolved question in power system planning is whether the spatial cost of siting these loads is driven by short-run parametric variables (wholesale prices, fuel costs, carbon charges) or by network topology (line impedances, transfer capacities, and grid geometry). Existing literature fails to isolate these factors, often treating facility locations as fixed, using synthetic test feeders that lack inter-regional constraints, or aggregating data to zonal resolutions that obscure nodal cost signals. Consequently, it remains unclear whether market reform (price signals) or physical reinforcement (grid upgrades) is the primary lever for managing the integration of gigawatt-scale, inflexible loads.

Methodology
The authors conduct a controlled cross-country experiment to decouple network topology from parametric variation, leveraging the contrasting grid geometries of New Zealand and Ireland. Both systems are modeled as national, electrically isolated networks (Ireland is islanded by removing cross-border interconnectors to match New Zealand's isolation) with comparable populations (~5.3–5.4 million) and high renewable shares, yet both rely on Combined-Cycle Gas Turbines (CCGT) to set the marginal price—a condition termed the "flexibility trap."

Modeling Framework: The study uses the open-source PyPSA-Earth model to solve linear Direct Current (DC) Optimal Power Flow (OPF) problems. The AC network is treated as a lossless transport system, with the exception of the New Zealand Cook Strait HVDC link, which includes a specific converter loss.
Experimental Design: A firm, constant, inelastic point load (representing a hyper-scale data center) is injected sequentially at every candidate node across both networks.
- New Zealand: 8 candidate nodes (5 North Island, 3 South Island) across a 55-bus network.
- Ireland: 5 candidate nodes (county-level archetypes) across a 26-bus network, plus an 8-bus analytical limit.
- Load Sizes: 100 MW, 300 MW, and 500 MW.
Perturbation Suite: The study tests robustness against variations in gas and carbon prices (±50%), demand (±20%), weather years (dry/wet for NZ; high/low wind for Ireland), transmission capacity, and converter efficiency.
Metric: The "siting cost" is defined as the change in the optimal annual system cost divided by the injected capacity. Nodes are classified into "equivalence classes" if their marginal siting costs differ by less than 1%.

Key Results
The experiment yields qualitatively opposite spatial structures for the two systems, demonstrating that topology, not parameters, dictates the cost surface.

New Zealand (Bottlenecked Topology): The network splits into exactly two equivalence classes separated by the Cook Strait HVDC link.
- South Island nodes exhibit a siting cost approximately 4.18% lower than North Island nodes.
- This gap corresponds precisely to the one-way converter loss ( $\eta = 0.96$ ). The price difference is analytically derived as $(1-\eta)/\eta$ .
- The gap persists across all load sizes, fuel price scenarios, and weather years. It vanishes only when the converter loss is removed in the model, confirming the cost signal is purely topological and physical.
Ireland (Meshed Ring Topology): The network collapses into a single equivalence class.
- Across all candidate nodes, the generation-cost component of the siting cost varies by less than 0.1% (often as low as 0.016%), effectively rendering the cost surface flat.
- This uniformity holds even under extreme parametric perturbations (gas price swings, demand changes) and weather variations.
- The single class persists because the meshed AC ring operates below congestion thresholds in routine dispatch; line-flow constraints carry no shadow prices, and locational marginal prices (LMPs) are "numb" to location.
- Boundary Condition: The flat surface is contingent on the freedom to expand transmission. When transmission expansion is frozen and load is increased to 300–500 MW, the meshed ring congests, and a multi-class structure with significant cost spreads (up to 76%) re-emerges at remote nodes.

Key Contributions

Parametric Invariance Principle: The paper establishes that within a fixed transmission network, the spatial structure of siting costs is set by network topology and is insensitive to short-run dispatch parameters (fuel prices, carbon costs, weather).
Isolation of Topological Effects: By matching two systems with similar marginal technologies (CCGT) and renewable shares but divergent geometries, the study isolates grid geometry as the sole determinant of spatial cost variation.
Mechanism Identification: The study demonstrates that in bottlenecked systems (like NZ), price signals naturally reflect physical losses and constraints. In highly meshed systems (like Ireland), the topology inherently masks locational scarcity, rendering price signals ineffective for guiding siting decisions without physical reinforcement.

Significance and Claims
The authors claim this work resolves a long-standing indeterminacy in infrastructure planning: whether market signals or physical constraints drive the economics of large load siting.

Policy Implication: The findings suggest that market-based interventions (e.g., locational pricing) are effective only where the network is bottlenecked (as in New Zealand), where price signals can naturally steer demand to lower-cost regions. Conversely, in highly meshed grids (like Ireland), price signals are inert; the "flat" cost surface is a market-design failure imposed by topology, requiring physical grid reinforcement or explicit administrative zoning to manage concentrated load growth.
Relevance to AI: The study highlights that for gigawatt-scale, inflexible AI loads, the decision of where to site is a live economic decision driven by grid physics rather than a fixed geographic given.
Limitations: The authors modestly note that the model uses a lossless DC OPF approximation, excluding N-1 contingencies, voltage limits, and reactive power. However, they argue these omissions do not alter the macro-scale topological findings, as the New Zealand gap is driven by the specific HVDC loss the model does include, and the Irish flatness is a structural property of the ring topology. The study does not claim to solve operational security but rather to define the spatial structure of economic costs.

Technical Summary: Grid Topology and the Spatial Cost of AI Data-Centre Siting

More like this