AI Data Centres Are Breaking the Power Grid, and It Is Not About Supply
Two massive blackouts in Virginia exposed a deeper problem: the way AI campuses connect to the grid is the real danger, and a new power architecture aims to fix it.
Key points
- A transmission fault in Ashburn, Virginia, on 22 July 2026 knocked more than 3 gigawatts of power off the grid in seconds.
- A separate 2024 Virginia event knocked roughly 60 data centre facilities and 1,500 megawatts offline at once after a single failed surge arrester.
- The root cause is not a shortage of electricity but a decades-old power architecture that was never designed for AI-scale loads.
- A redesigned system tested at a US Department of Energy facility in early 2026 absorbed full grid faults without the compute side registering a flicker.
- Rearchitecting the power stack can turn AI data centres from grid liabilities into assets that earn money through demand-response programmes.
Picture a single blown component on a power line outside Washington DC bringing down the equivalent of a small city's worth of electricity in the time it takes to blink. That happened twice in Virginia within two years.
The first time was 2024: one failed surge arrester (a device that protects electrical equipment from voltage spikes) tripped roughly 60 data centre facilities and wiped 1,500 megawatts off the grid simultaneously. The second time was 22 July 2026, when a transmission line fault near Ashburn, the world's largest cluster of data centres, yanked more than 3 gigawatts offline in seconds.
So why does this keep happening?
The grid's protection systems are doing exactly what they were designed to do. That is the problem.
Old power grids were built for predictable customers: steel mills, refineries, houses cooking dinner. Those loads draw power steadily and, when something goes wrong upstream, they misbehave one at a time. AI data centres behave very differently. A single AI campus can swing 70% of its power draw in milliseconds when a training job kicks off, then cut itself off from the grid just as fast if it senses trouble, to protect the billions of dollars' worth of chips inside.
One campus doing that is manageable. Dozens doing it at the same instant, at gigawatt scale, is something the grid has never faced before.
What is wrong with the current setup?
The standard data centre power stack has not changed much in decades. Electricity arrives at medium voltage, transformers step it down, and a UPS (uninterruptible power supply, the battery backup system that keeps servers alive during a blip) conditions it before it reaches the computers. At AI scale, that chain cracks in three places.
The batteries are undersized for the job. The UPS spends most of its life switched out of the circuit to save energy, meaning raw power swings from the AI chips go straight out to the grid unfiltered. And the protection logic was written for facilities of 50 megawatts, not 500. When voltage dips hit, the system disconnects itself after the third one, exactly as designed, at the worst possible moment for everyone else on the grid.
| Problem | What it means |
|---|---|
| Undersized UPS batteries | Cannot absorb fast, repeated power swings |
| UPS runs in bypass mode | Chip load swings hit the grid unfiltered |
| Outdated protection logic | Disconnects on third voltage dip, amplifying outages |
| Low-voltage design | Grid sees hundreds of separate devices, not one managed load |
Is there a fix?
A company called ON.energy, whose piece originally appeared via MIT Technology Review, says yes, and it tested the idea at a US Department of Energy facility in early 2026.
The approach has three parts: move the power conversion up to medium voltage (around 13.8 kilovolts, closer to the voltage the grid actually runs at), move the equipment outside the building near the substation, and put it directly in the path of every electron rather than bypassing it. That last part is key. There is nothing to detect and switch because the system is always active.
In testing, real AI workloads hit one side while simulated grid faults, including a complete loss of voltage, hit the other. The compute kept running. The grid saw a flat, stable load. The design also passed the large-load voltage ride-through rules set by ERCOT (the Electric Reliability Council of Texas, which manages that state's grid) with room to spare.
Bonus: equipment like this can qualify for tax credits and earn revenue by helping the grid during peak demand, so the backup power stops being a pure cost and starts paying for itself.
Common questions
Does this affect electricity bills for ordinary people?
Not directly, but grid instability from large data centre outages can cause wider disruptions that affect everyone. Fixing the architecture reduces that risk without requiring new power plants to be built.
How long before this new design becomes standard?
The engineering is proven but adoption takes time. Permitting, procurement and construction cycles mean most of the next wave of AI facilities will be planned over the next two to three years, so the window to build the new architecture in from the start is open right now.
