Across the eastern United States, the physical limits of a power grid built for another era are now pressing against the boundless appetite of artificial intelligence. PJM, the interconnection serving 65 million people, has approved a plan to temporarily cut electricity to data centers during peak demand — a quiet but consequential admission that technological ambition and infrastructure reality have arrived at a reckoning. The decision is less a solution than a signal: the age of assuming unlimited, uninterrupted power for the machines of the digital economy may be drawing to a close.
PJM Grid May Cut Power to Data Centers to Avert Blackouts Amid Demand Surge
The grid was designed with different assumptions about how electricity would be used
Mark
Why would a grid operator choose to cut power to data centers rather than, say, ask them to reduce consumption voluntarily?
Mimi
Because voluntary measures don't work when demand is this urgent and the stakes are this high. If a blackout is coming, you need a mechanism that actually works. Curtailment is enforceable.
Mark
But doesn't that seem unfair to data center operators who've already invested billions?
Mimi
It does, and that's the real tension. They built on the assumption of reliable power. Now the grid is saying that assumption has changed. It's a breach of an implicit contract.
Mark
Could data centers just build their own power plants?
Mimi
Some are exploring it, but it's expensive and slow. A battery system or generator costs millions. Most operators would rather push the grid to solve the problem.
Mark
What happens if curtailments actually start happening? Won't that break AI training runs?
Mimi
Yes. A data center losing power mid-training loses weeks of computation. That's why the threat of curtailment might be more powerful than the curtailment itself—it forces everyone to the table.
Mark
Is this a sign that AI growth is unsustainable?
Mimi
Not unsustainable, but it's unsustainable at the current pace without massive new infrastructure investment. The grid needs time to catch up. AI companies need to accept that time exists.
Mark
What's the endgame here?
Mimi
Either new power plants and transmission lines get built fast enough to meet demand, or data centers become more efficient, or both. The curtailment plan is what happens if neither occurs quickly enough.
Le Pouls
- AI and cloud computing have driven data center electricity demand to levels the grid was never designed to absorb, creating a genuine risk of widespread blackouts across the eastern US.
- PJM's board approved curtailment protocols in late July — meaning data centers could have their power deliberately reduced during periods of peak grid stress, a measure once considered unthinkable for critical digital infrastructure.
- Technology companies that invested billions in regional facilities under the assumption of reliable power now face an unsettling new variable: conditional electricity supply tied to grid conditions.
- PJM is developing specific protocols for how curtailments would work — including notice periods, prioritization of facilities, and potential compensation — while also exploring whether data centers should build their own backup capacity.
- The deeper tension is structural: new generation and transmission infrastructure cannot be built at the speed AI is scaling, leaving curtailment as a stopgap rather than a lasting answer.
Across the eastern United States, the physical limits of a power grid built for another era are now pressing against the boundless appetite of artificial intelligence. PJM, the interconnection serving 65 million people, has approved a plan to temporarily cut electricity to data centers during peak demand — a quiet but consequential admission that technological ambition and infrastructure reality have arrived at a reckoning. The decision is less a solution than a signal: the age of assuming unlimited, uninterrupted power for the machines of the digital economy may be drawing to a close.
The Pennsylvania-New Jersey-Maryland Interconnection, which powers roughly 65 million people across the eastern US, is preparing to deliberately cut electricity to data centers during peak demand periods — a measure designed to prevent the broader grid from collapsing under the weight of surging AI infrastructure.
The PJM board approved the plan in late July after grid planners found themselves caught off guard not by the existence of data center demand, but by its velocity. Major technology companies have been clustering massive AI training and inference facilities within PJM's territory, each drawing power at the scale of a small city. The grid was simply not designed for this pattern of consumption.
The response is two-pronged: a backstop capacity auction to keep power plants available during tight supply, and targeted curtailment protocols that would temporarily reduce power to data centers when the grid is under stress. These would not be permanent shutdowns, but their approval marks a meaningful shift in how grid operators are thinking about the relationship between digital infrastructure and physical limits.
The plan raises hard questions for the technology industry. Companies that built facilities in the region on the assumption of reliable power now face the prospect of conditional supply. How often curtailments might occur, how much warning operators would receive, and whether compensation would follow are all details still being worked out. PJM is also exploring whether data centers should invest in on-site backup capacity — batteries, generators — to reduce their grid dependence during peak periods.
What this moment ultimately reveals is a collision between two forces moving at very different speeds: the relentless scaling of AI computing and the slow, capital-intensive work of building electrical infrastructure. PJM's curtailment plan is an acknowledgment that the grid cannot simply absorb unlimited growth — and a signal to the technology sector that the economics and assumptions of AI infrastructure may be entering a new and more constrained chapter.
The Pennsylvania-New Jersey-Maryland Interconnection, which supplies power to roughly 65 million people across the eastern United States, is preparing to do something it has rarely done before: deliberately shut off electricity to data centers during peak demand periods to keep the broader grid from collapsing.
The PJM board approved the plan in late July as a direct response to a problem that has crept up on grid operators with surprising speed. Data centers—the massive facilities that house the servers running artificial intelligence systems, cloud storage, and the digital infrastructure of modern life—are consuming electricity at rates that the grid was not built to handle. The surge is real and measurable. As AI companies race to build out their computing capacity, they are plugging in facilities that draw power like small cities. The grid operator now faces a choice between allowing data centers to consume whatever they need and risking blackouts that would affect hospitals, homes, and businesses across the region.
The board's solution is a two-part approach. First, PJM will establish a backstop capacity auction—essentially a market mechanism to encourage power plants to stay online or come back online during periods of tight supply. Second, and more dramatically, the grid operator will implement curtailment protocols that allow it to reduce power to data centers when the grid is under stress. The curtailments would be temporary and targeted, not permanent shutdowns. But the fact that such a plan is now on the table signals how acute the tension has become between the explosive growth of AI infrastructure and the physical limits of the electrical system.
What makes this moment significant is not just the policy itself but what it reveals about the speed of technological change outpacing infrastructure. Data center operators have long known that their facilities are power-hungry. What caught grid planners off guard was the velocity of growth. The demand is not coming from a handful of new facilities scattered across the region. It is coming from a concentrated push by major technology companies to build massive AI training and inference centers, often in clusters, within PJM's territory. The grid was designed with different assumptions about how electricity would be used and where.
The curtailment plan raises immediate questions about fairness and feasibility. Data center operators will want to know how often they might face power cuts, how much warning they will receive, and whether they will be compensated for the disruption. Technology companies have invested billions in facilities across the region based on the assumption of reliable power supply. A sudden shift to conditional reliability could reshape investment decisions and operating costs. There is also the question of whether curtailment is actually a long-term solution or merely a stopgap while the grid catches up to demand through new generation and transmission capacity.
For now, PJM is moving forward with implementation details. The board has directed staff to develop specific protocols for how and when data centers would be curtailed, what notice they would receive, and how the system would prioritize which facilities lose power first. The grid operator is also exploring whether data centers themselves might invest in backup power systems—batteries, generators, or other on-site capacity—to reduce their dependence on the grid during peak periods. Some data center operators are already moving in this direction, but it is expensive and not yet widespread.
The larger story here is about the collision between two powerful forces: the relentless demand for computing power driven by AI and the finite capacity of physical infrastructure. PJM's decision to plan for data center curtailment is an acknowledgment that the grid cannot simply absorb unlimited growth. It is also a signal to technology companies that the era of assuming infinite, uninterrupted power supply may be ending. What happens next—whether new generation comes online fast enough, whether data centers adapt their operations, whether curtailments actually occur or remain a theoretical threat—will shape the economics of AI infrastructure for years to come.
Citations marquantes
The grid was not built to handle the electricity consumption rates that data centers now demand as AI infrastructure expands— PJM grid analysis
Envie de l'histoire complète ?
Lire l'original sur Google News ↗
Technology