- April 28, 2025: 15 GW of power vanished from the Iberian grid in 5 seconds. 60M people went dark. The final ENTSO-E report just landed — and the cause wasn't "too much renewable." It was a blind spot. The grid's own monitoring couldn't see the failure forming. 🧵
- The part that should keep operators up at night: at 12:31, the 400 kV transmission readings showed 418 kV. Within limits. Nothing wrong. Meanwhile 220 kV collector substations were hitting 242 kV. Tap-changers couldn't keep up. Nobody was watching that level.
- With sub-synchronous oscillations at 0.21 and 0.63 Hz already building, one major facility injected reactive power into an already-overvolted grid — when the rules (P.O. 7.4) required it to absorb. The feedback loop cascaded: 15 GW gone in 5 seconds. The hole won.
- The US has the same disease in a different form. PJM just fell 6,625 MW short of its reliability target — the first time the entire region missed. Capacity prices slammed the $333.44/MW-day cap. One auction cost $16.4B. Projected 2028-2033: $163B in capacity costs.
- What's driving it? Data centers. Of PJM's 5,250 MW forecast demand increase, 5,100 MW is data center load. ERCOT's large-load queue hit 233 GW — up 269% in a year, 77% data centers. Only 23 GW of new generation synchronized in 2025. The math doesn't close.
- The national interconnection queue is now 2,600 GW. Median time to operation: 5 years. Of projects that asked to connect 2000-2018, only 20% ever made it. You cannot build your way out of this fast enough. The grid has to get smarter, not just bigger.
- The cheapest capacity already exists — stranded on lines run too conservatively. Dynamic Line Rating proves it: LineVision + AES unlocked 61% more capacity on a 345 kV line at 76% lower cost than reconductoring. GETs run at 5-7% the cost of traditional upgrades.
- So why isn't everyone doing it? Nobody owns the analytics layer. LineVision sells the sensors. GE Vernova and Siemens own the SCADA. The missing piece is the AI that decides WHEN and WHERE to deploy GETs — and watches the collector level the Iberian grid couldn't.
- A reality check on the hype: physics-informed neural nets jumped from <10 papers in 2019 to 820 in 2025 — but there's still no production PINN grid controller in early 2026. We use them where they're real: fast advisory + planning models, not autonomous control.
- This is also a compliance cliff. Under the EU AI Act, grid management AI is "high-risk" — compliance due Aug 2, 2026, penalties up to €15M or 3% of global turnover. The agentic-copilot wave (Argonne's GridMind) only works if it can explain itself. Black boxes fail.
- Honest question for grid engineers: after the ENTSO-E report, is your monitoring blind at the collector level too — and would you trust an AI copilot to flag it in real time, or only ever a human in the loop? #GridResilience
- We wrote up the full failure chain, the queue math, and where AI actually fits (and where it doesn't) for ISOs and utilities: https://veriprajna.com/solutions/power-grid-ai