Article — "AI's Infrastructure Fragility and Layoff Cover: Grid Instability and the 'AI Transformation' Cut": At 7:55 on the morning of July 22, roughly 3.1 gigawatts of load reportedly vanished from the largest power grid in the United States in about 30 seconds, according to PJM and Yes Energy data.
No blackout. No storm. A single transmission line in Ashburn, Virginia — the densest concentration of data centers on Earth — went out of service, and the machines took over from there.
Dominion Energy is emphatic that it did not pull the plug: "the data centers' own control systems transferred them to backup power for a very short period of time." Each facility, sensing the fault, did exactly what its protection logic told it to do. Save the hardware. Drop off.
The problem is that they all did it at once.
Voltage disturbances rippled across 1.4 million sensors from Washington, DC, to Chicago — a thousand miles of flickering lights and refrigerators making sounds they shouldn't. A disturbance that should have resolved in milliseconds reportedly took another eleven minutes to stabilize.
Neil Osnato of Persistence Analytics Group put the diagnosis precisely: "A 3 GW customer response is no longer merely customer behavior. It is grid behavior — and it must be planned, modeled, and verified accordingly."
The mechanism he names is the one worth holding onto: individually reasonable settings, shared across hundreds of facilities, produce a "common-mode response" when everything reacts to the same disturbance at the same instant.
That is the shape of the whole story. Not one bad decision, but thousands of locally rational ones synchronizing into a system-level shock.
Not one bad decision, but thousands of locally rational ones synchronizing into a system-level shock — the shape of the whole story.
Last issue, in the Nscale grid-connection thread, we treated data-center siting as a question — where can this load physically plug in without breaking something. This week the question got answered by a documented incident. The something broke, briefly, for 67 million people.
Now watch the same physics play out in a labor market.
On Wednesday, the same day the grid was flickering, Monday.com filed to cut about 20% of its workforce — just over 600 people — as part of a "restructuring plan" tied to its "AI-driven growth strategy."
Co-founder Eran Zinman told staff the move "was not made to reduce costs or replace people with AI." It was, in his telling, adaptation to an AI-first vision.
Hold that quote next to Dominion's. Both are institutional actors insisting the disruptive thing they just did was not the disruptive thing it looks like. The data centers didn't disconnect — they "elected to transfer load." Monday.com didn't cut costs — it got "leaner" and "more focused."
Each statement is locally defensible. Monday.com is, after all, still projecting up to 20% revenue growth for 2026. This is not a company in distress by the old definition.
But Monday.com is, per TechCrunch's tally, reportedly the 21st tech firm to hang layoffs on AI, and the synchronization is the tell. The Financial Times found nearly 140,000 tech jobs cut since January, with Amazon, Oracle, Meta, and Microsoft alone accounting for almost 50,000 as they pour hundreds of billions into data centers.
And the market has already learned to discount the story. The same FT analysis found that firms citing AI in their layoffs underperformed the Nasdaq by almost 10% in the 30 trading days after announcing.
Investors, in other words, are doing exactly what PJM's sensors did: registering a disturbance the actors involved insist is nothing.
The connective tissue is the same on both sides. In each case, an AI-era operator protects its own fragility by offloading the cost onto a system that cannot refuse it.
The data centers protect their hardware — precision cooling and GPUs worth hundreds of millions — by dumping instability onto the shared grid. Staying online during a fault risks the equipment; disconnecting risks everyone else's lights. Both positions are rational. Neither scales.
The employers protect their equity story. As Revelio Labs managing director Evan Sohn put it bluntly: "Payroll is being converted into capital expenditure."
That is the trick in one sentence. The money that used to pay engineers now pays for chips and buildings, and the relabeling from "cost cut" to "AI transformation" is what makes the swap read as innovation rather than retreat.
Georgetown's Jason Schloetzer described the internal translation: CEOs "may say these cuts are because of AI, when they really mean to say that they don't have enough cash flow generation to fund the investments they would like." For investors, he added, "the reason doesn't matter."
Except it does, eventually — which is why the 10% underperformance matters. The fiction has a discount rate.
The tell in both systems is that the labels don't survive contact with the instruments. Ting Labs' sensors don't care that Dominion calls it a "transfer." Revelio's payroll data doesn't care that Zinman calls it a "vision."
Common-mode response is what makes both dangerous. One data center dropping 50 megawatts is customer behavior. Hundreds sharing the same voltage thresholds and transfer logic is grid behavior. One firm reallocating capital is a business decision. Twenty-one firms citing "AI transformation" in a quarter is a narrative — one that shifts how markets price an entire sector.
And both are scaling toward worse. This July's event was reportedly roughly double the 1.5 GW, 60-data-center disconnection that hit the same grid in 2024, per TechCrunch. Data centers were about 6% of PJM's load then; they're reportedly projected to reach 24% by 2040, according to Synapse Energy Economics.
NERC's John Moura already warned that "the grid is not designed to withstand the loss of 1,500 MW data centers." July was more than twice that.
So the real question is who forces the cost back onto the actors creating it — and on the grid side, that repricing has started.
NERC is developing mandatory ride-through modeling standards and high-resolution monitoring, so operators can see behind the meter before the next event rather than after. Osnato argues hyperscale campuses are now "system-significant" enough to warrant the same performance requirements large generators already meet.
More pointedly, PJM is pursuing regulatory changes to shift transmission and capacity costs directly onto data-center developers instead of spreading them across all ratepayers. That is the externality being repriced onto its source.
Dominion and Virginia data-center owners have also begun retuning control systems so sites ride through brief faults instead of triggering the synchronized drop. The self-protective reflex is being disarmed by rule.
The labor market has no NERC. No task force convenes when 600 people lose their jobs to a "vision," and there is no mandatory modeling standard for whether AI-washing corresponds to real productivity gains — which, per Cognizant's own chief AI officer, may be six months to a year from even appearing.
What labor has instead is the price signal: the 10% haircut, the market quietly refusing to fully underwrite the story. It is slower and blunter than a NERC standard, but it is the same motion — the cost drifting back toward the actor that generated it.
Until it fully arrives, both systems keep absorbing what they're handed. The grid takes the instability. The labor market takes the cuts. And in both cases, the operator gets to call it something else.
The lights stopped flickering after eleven minutes. The narrative discount takes 30 trading days. Neither is the same as the bill being paid.