Reading time: 5 min
Table of Contents
Key takeaways
- Production failure pattern: 3.1 GW of compute load dropped in 30 seconds because data centers reacted individually, not as a coordinated system. Grid stability depends on orchestrated behavior, not siloed instincts.
- Real cost: Voltage spikes, flickering lights, and risk of blackout are symptoms. The real damage is eroding trust in grid infrastructure for mission-critical workloads — and the contractual penalties that follow.
- Incremental fix: Ride-through systems like ON.Energy’s batteries and control logic can smooth the disconnect/reconnect sequence. ERCOT is already mandating it. PJM should too.
The numbers that matter
Let’s drop the hand-waving and talk about what actually happened. A power line fault near Washington, DC triggered data centers across Northern Virginia—the highest density of compute load on the planet—to switch to backup power. Within 30 seconds, 3.1 GW of demand vanished from PJM’s grid. That’s roughly the output of three nuclear reactors or, put more bluntly, the load of half a million homes. This isn’t theory. The demo worked—until the power line failed.
The grid didn’t black out, but lights flickered from Chicago to Northern Virginia because voltage surged as supply outstripped demand by 3.49 GW at the peak. It took 11 minutes of instability before PJM stabilized. For context, in 2024, 60 data centers dropped 1.5 GW. This event was twice as large, and the trend line is clear: by 2040, data centers will make up 24% of PJM’s load. Most people get this wrong—they assume the grid has enough headroom. The real cost isn’t the flicker—it’s the uncoordinated behavior that scales.
Why this is a structural problem, not a one-off
Here’s what actually happens in production when voltage dips: every data center’s power system senses an anomaly and decides to isolate in milliseconds. They don’t coordinate. They don’t sequence. They act like a thousand drivers slamming brakes simultaneously on a highway—except here, the highway is the grid. That’s not automation—that’s a liability. Each facility optimizes for its own survival, but the collective effect is dangerous. Ali Zain Banatwala from the Independent Electricity System Operator put it correctly: we need these loads to sequentially disconnect or reconnect. Most people miss this: the problem isn’t the line fault—it’s that data centers amplify it by disappearing simultaneously.
Two scenarios: ride-through vs. run-away
The fix isn’t complicated in principle, but execution matters. I’ve seen two approaches in systems I’ve built at Rebirth Distribution:
- Run-away behavior: Backup power kicks in, load drops, grid sees a spike. This is what happened. It’s fragile, reactive, and scales poorly.
- Ride-through behavior: Behind-the-meter batteries and power electronics absorb the fluctuation. The grid sees a consistent load profile. ON.Energy’s approach—essentially hiding the entire data center campus behind a bank of batteries and inverters—does exactly this. It allows computing workloads to ramp up and down without bothering the grid, and can follow grid signals in milliseconds to prevent sag or surge.
That’s not speculative—they’re installing 3 GW of these systems at four campuses right now. ERCOT is already requiring ride-through for large loads. PJM should follow, but the industry can’t wait for mandates alone.
What this means for your infrastructure stack
If you’re running automation, agent systems, or any production workload in Northern Virginia or any densely-packed data center market, you need to think beyond uptime. The grid is a shared resource that your neighbors can compromise. This isn’t theory—I’ve seen n8n pipelines crash during these events because a node’s AWS instance went dark. The failure domain expands from your power panel to the whole transmission corridor. The practical takeaway: design for intermittent grid availability. Use multiple regions, buffer energy costs into your architecture, and demand ride-through capability from your colocation provider or hyperscaler. That’s the incremental path—you don’t need to rebuild from scratch, but you do need to ask the right questions during procurement.
Hard numbers for comparison
Let me be specific. The 2024 event dropped 1.5 GW over 60 data centers. That was 6% of PJM’s load at the time. This week’s event dropped 3.49 GW at peak, from ~3% of current load in a cleaner loss—but the impact was less proportionally because grid operators have better tools. Still, the trend is exponential. If data centers make up 24% by 2040, a single simultaneous fault could trigger a cascade that takes down multiple states. The clock is ticking. The demos worked—now we need production-grade design.
Bottom line
This event is a warning. Data centers are not passive loads—they are active participants in grid dynamics, and they’re behaving like uncoordinated agents. The industry needs ride-through, sequential logic, and smarter orchestration at the power level. At Rebirth Distribution, we build automation that actually holds—not demo-grade—and this is exactly the kind of fragility we root out. You should too.