This Week in Voltage
FERC Chairman Laura Swett didn't mince words at the industry's biggest annual conference this spring. The hyperscalers, she said, "don't speak FERC" — and their complaints about utilities "show a lack of understanding of how the grid functions." That's a remarkable statement from the top U.S. energy regulator, and it points to something deeper than a communication gap. The AI industry isn't just overwhelming the grid with demand. It's introducing a class of electrical behavior the grid was never designed to handle.
This is the story the interconnection queue debates keep missing. The fight over who pays for new transmission lines matters — but underneath it is a more fundamental physics problem: the grid's frequency and voltage stability architecture was built for a world of large synchronous generators and predictable loads. We are rapidly building a world that looks nothing like that.
Deep Charge: When the Grid Meets the AI Load
Here's what's new and genuinely alarming. The Department of Energy's Office of Electricity recently flagged that AI training centers — unlike conventional data centers — run thousands of specialized chips in tightly coordinated cycles. That synchronized activity creates repetitive electrical load oscillations across a wide range of frequencies. Some of those frequencies can adversely interfere with equipment at nearby power plants. The DOE's own measurement adequacy report found that phasor measurement units (PMUs), the workhorses of grid monitoring, excel at detecting low-frequency oscillations but miss or misrepresent higher-frequency events due to their filtering architecture.
Translation: we are adding a massive, pulsing, synchronized load class to the grid, and our primary sensing infrastructure has a known blind spot for the exact frequencies it generates.
This lands on top of a renewable integration problem that was already hard. Hitachi Energy's grid engineering team lays out the core tension clearly: synchronous generators — gas turbines, nuclear plants, coal — naturally provide reactive power that stabilizes voltage. Solar PV at full output has little remaining capacity for reactive power support. Wind generators can provide some, but headroom is limited when they're running hard. As renewables take a larger share of the generation mix, the grid loses the inertia and reactive power reserves that historically absorbed shocks.
The April 2025 Iberian Peninsula blackout is the case study everyone in grid engineering is now studying. A peer-reviewed analysis published in Sustainable Energy, Grids and Networks examined the prologue conditions of that event — the voltage stability dynamics in the hours before the cascade. The full paper is behind a paywall, but the research framing confirms what grid operators suspected: voltage control failures in high-renewable systems can develop faster than conventional protection schemes respond.
Meanwhile, research from the Okinawa Institute of Science and Technology adds a predictive dimension. Using five-plus years of data from 80 wind turbines spread across 20km, the team found that wind farms behave less like collections of independent turbines and more like single turbulent systems — with power fluctuations that correlate strongly with atmospheric turbulence and scale predictably across downwind farms and entire grids. The implication: fluctuation risk isn't random. It's structurally predictable, which means it's also structurally manageable — if grid planners use the right tools.
The problem is that most don't yet. Wood Mackenzie's Ben Hertz-Shargel, presenting at DTECH Data Centers & AI, put the disconnect bluntly: utilities have made commitments to 187 gigawatts in data center load, and the question of whether those loads will ever be meaningfully flexible — curtailable in moments of grid stress — remains genuinely open. "Are we going to see more flexible and essentially curtailable data centers, or is firm service through transmission-enabled generation the ultimate long-term plan?" That's not a rhetorical question. It's the central engineering decision of the next decade.
By the Numbers
- 187 GW — data center load commitments utilities have made, per Wood Mackenzie (announced capacity, not yet operational)
- 5–10 years — timeline frontier AI companies say is too slow for grid interconnection; FERC's current proposal targets faster connection
- 80 turbines over 20km, 5+ years — dataset underlying OIST's new fluctuation-prediction framework, now published in PRX Energy
- PMU bandwidth gap — DOE's NASPI report flags that standard phasor measurement units miss higher-frequency oscillations from AI training loads
What We're Fighting For
Energy abundance doesn't mean energy recklessness. The civilizational case for 100x electricity usage only holds if the grid can actually deliver that electricity without cascading into blackouts. The Iberian event was a warning shot. The DOE's oscillation report is a diagnostic. The OIST fluctuation framework is a tool. What's missing is the urgency to deploy all three simultaneously — better sensors, reactive power compensation through modern power electronics, and load flexibility baked into data center interconnection agreements from day one.
FERC's June proposal on data center grid access is the near-term decision to watch. If it mandates flexibility requirements as a condition of accelerated interconnection, it threads the needle between speed and stability. If it's purely a queue-reform measure that ignores the physics, we'll be writing about the next Iberian event closer to home.
The future is electric. But electricity that oscillates itself into a blackout doesn't power civilization — it interrupts it. Build the sensors. Deploy the reactive compensation. Make the loads flexible. Then scale without limit.
