We Asked AI to Solve Its Own Power Crisis. Here’s What Happened.

We conducted an unusual experiment: we asked artificial intelligence to diagnose the very problem that could destroy it.

We fed Google’s most advanced AI system over 10,000 pages of electrical engineering knowledge—everything from fundamental textbooks to cutting-edge energy research. Then we posed one question:

What the AI told us was startling.

AI’s Self-Diagnosis

The first thing AI did? Admitted weakness.

Then it made the problem real:

The AI wasn’t just worried about its own performance. It was thinking about human safety.

The Hidden Problem

Analyzing the data, AI discovered something most people miss: data centers already consume 24% of some utility companies’ electricity sales. By 2030, demand could surge 160%.

But quantity isn’t the real issue. It’s quality.

Data centers experience an average of 2 hours of preventable downtime annually from power quality issues. The AI calculated the cascading costs: millions in lost revenue, corrupted data, equipment failures, system crashes.

AI’s Verdict on Traditional Solutions

When analyzing conventional power infrastructure—backup generators, capacitor banks, UPS systems—the AI was diplomatic but firm:

Why? These solutions lack the speed and intelligence modern AI demands. They measure and respond in seconds or minutes. Electrical problems happen at near light-speed.

70% of our electrical infrastructure dates back to the 1960s. We’re running tomorrow’s technology on yesterday’s grid.

The Solution AI Chose for Itself

Given comprehensive data on various power solutions, the AI consistently chose one approach: Software-Defined Electricity.

The AI analyzed the numbers: electrical networks scoring 30-50% on power quality could jump to 80-99% with real-time optimization.

What We Learned

When you give AI enough information and ask it to solve its own problems, something remarkable happens. It doesn’t just optimize for performance. It considers:

• ▸ Human safety implications
• ▸ Environmental sustainability
• ▸ Economic efficiency over brute-force expansion
• ▸ Systemic risks in critical infrastructure

The AI even showed environmental awareness:

The Urgent Truth

The AI’s analysis led to one conclusion: the gap between what AI needs and what our power infrastructure delivers is widening, not closing.

But here’s the good news: the technology to fix this exists right now. It’s not theoretical. It’s being deployed in real facilities, delivering measurable results.

The question isn’t whether we can solve this.

It’s whether we will.

AI loads swing, and each swing is a step in current

Engineers at Microsoft, OpenAI and a leading GPU maker describe AI training jobs that alternate between two states. Compute runs near the device’s thermal design power, and communication runs near idle. On the Data Centers and Compute page, the aggregate swings reach tens of megawatts in one data center, concentrated between 0.2 and 3 hertz. Each swing is a step in power. By I = P / V, it is also a step in current at the panels between the rack and the meter.

Correction at the panel splits each swing in two. The reactive and harmonic share of each swing stays at the panel, and the source carries the step in working current. The step itself remains, and it is the source’s job: the battery, the rack capacitor unit or the grid.

Measuring those swings is a problem of its own. A North American SynchroPhasor Initiative report found that filtering inside phasor measurement units can miss or misrepresent the high-frequency oscillations of AI workloads. The report concludes that combined phasor and point-on-wave monitoring is the most reliable pathway for assessing oscillations from large loads. The About page records the finding.

Where the case for correction rests

The Data Centers page frames the return from correction as capacity. In a data hall, that capacity is a rating: the kilovolt-amperes of each transformer, uninterruptible power supply and generator. Where the limit is a contract in megawatts, correction adds little capacity under that limit. The page rests the case for correction on rating headroom, neutral current and the record, not on cooling.

Where this stands today

The Data Centers and Compute page now covers the data hall. It describes how server power supplies draw more current when the voltage dips, to keep the same power. More current means a deeper dip across the wires and transformers feeding them.

The same page states a limit of our controller: it does not hold the voltage the source delivers and does not ride through an interruption. The SDE Summary page describes what it does instead: it measures the current waveform at the panel and corrects it within the same cycle. Start with a one-week measurement at one power distribution unit.

Related reading

• Resilience: what a short sag costs a plant, and what correction does not change
• Electric Optimization: current that heats conductors and fills transformers without doing work