Global AI data center buildout strains power grids

On July 22, 2026, a single transmission line failure in Northern Virginia caused more than 3 gigawatts of AI data center load - about 3% of PJM's total demand at the time - to disconnect automatically, sending a voltage disturbance that was felt from Washington, DC all the way to Chicago ^[1]. Dominion Energy's system operations team stabilized the situation within about 10 minutes and reported no lasting impact to grid reliability ^[2]. The scale is what stands out: roughly 3.1 gigawatts dropped within 30 seconds, nearly twice the size of a similar July 2024 Virginia incident in which about 60 data centers pulled offline simultaneously and nearly triggered a blackout ^[3][4]. Ting Labs CEO Bob Marshall, whose company serves customers near the affected campuses, said the power-quality hit was immediately visible at the household level - flickering lights and noisy appliances. ON.Energy's Ricardo de Azevedo called the event 'the canary in the coal mine,' and Independent Electricity System Operator specialist Ali Zain Banatwala argued the deeper problem is procedural: grid operators still have no coordinated rule for how clusters of co-located data centers should disconnect or reconnect in sequence, rather than all at once, when something goes wrong upstream.

Behind the July incident is a supply-demand gap that PJM's own independent market monitor calls historic. Monitoring Analytics president Joe Bowring said reserve margins are at 'a crisis stage right now,' with data center load growth outpacing new generation for years running ^[5]. PJM's demand from data centers is projected to grow 5-7 gigawatts annually through 2032, while new power supply is being added at only 2-3 gigawatts a year ^[3]- a gap wide enough that PJM has flagged a 6-gigawatt reliability shortfall as soon as 2027 ^[5]. That mismatch is already showing up in bills: NRDC's Tom Rutigliano projects PJM consumers will pay an extra $163 billion through 2033, roughly $70 a month more per family by 2028, on top of a $9.4 billion capacity charge already passed through this summer ^[3]. The political fallout is starting to show: Bloomberg has reported that the AI-driven strain is fueling talk of breaking up PJM itself ^[6], even as federal regulators are moving in the opposite direction, ordering grid operators in June 2026 to fast-track connection approvals for large AI loads to 90 days ^[7]. That is two branches of policy pulling against each other in the same season - accelerate connections while reserve margins hit a historic low.

None of this is slowing the physical construction boom. Crusoe just announced a new 900-megawatt phase of its Abilene, Texas campus built for Microsoft's AI infrastructure, bringing the site's total planned capacity to 2.1 gigawatts, with the first new building energized by mid-2027 ^[8]. Meta's Hyperion data center, unveiled by Mark Zuckerberg in July 2025 explicitly to get ahead of OpenAI and Google in the AI race, is designed to eventually reach 5 gigawatts of capacity across a footprint large enough to cover most of Manhattan, with 2 gigawatts expected online by 2030 ^[9]. The buildout is now global: HCLTech is committing 142.57 billion rupees ($1.48 billion) with Indian AI startup Sarvam to build its first AI data center in Bhubaneswar, Odisha, alongside a 5,000-seat technology hub set to open by 2028 ^[10]. Each of these projects is being planned and financed on its own timeline, largely independent of whether the local grid can absorb it - a mismatch the Virginia event just made visible in real time.

A recurring theme across this cluster's coverage is that raw electricity generation may not be the tightest constraint - the physical hardware needed to move and process that power is. On the compute side, a surge in AI chip and CPU demand from Chinese AI data center operators has pushed certain server CPU prices up more than 40% since the start of 2026 and stretched Intel's wait times to roughly half a year, forcing Intel and AMD to sign long-term supply agreements that commit volume without locking in price ^[11][12]. AMD has revised its own server CPU market forecast to exceed $120 billion by 2030 on the strength of this demand ^[11]. Data centers represented just 6% of PJM's load in 2024; that share is projected to reach 24% by 2040 ^[1]- a trajectory that risks outpacing not just power plants, but the entire supply chain built to support them.