High-voltage electrical substation with steel structures and power lines at dawn, illustrating local grid infrastructure bottlenecks

Impact of AI Data Centers on US Utility Grid

September 14, 2026 · 12 min read · By Rafael

Key Takeaways

  • US utilities requested $18.6 billion in rate increases in the first half of 2026, including a record $9.2 billion in Q2 alone, per PowerLines.
  • PJM’s 2028/2029 capacity auction cleared at the $325/MW-day price cap, leaving a roughly 6.8 GW shortfall below its 20% reserve margin target, per Utility Dive.
  • A July 22, 2026 disturbance in Northern Virginia took more than 3 GW of data center load offline in seconds as facilities transferred to backup power.
  • Texas halted new data center grid connections in August 2026, and New York imposed the first statewide moratorium on hyperscale facilities in July 2026.
  • The EIA expects record US electricity consumption in 2026, with 86 GW of planned capacity additions led by solar at 51% of the total.

Local Bottlenecks, Not a National Shortfall

The EIA’s September 2026 Short-Term Energy Outlook, released September 9 and finalized against data through September 3, projects US power demand rising from a record 4,195 billion kWh in 2025 to 4,270 billion kWh in 2026 and 4,349 billion kWh in 2027, according to Reuters’ summary of STEO. That is roughly a 2% annual increase, not a crisis curve. The issue is where the electricity is needed.

The EIA’s own framing is explicit: “Despite pause in connecting new data center projects in Texas to grid, West South Central region accounts for largest share of total electricity sales growth in our forecast.” Demand is concentrating in a handful of interconnection queues while the rest of the country grows slowly.

The main sign of that concentration is time. New data center connection timelines in Northern Virginia have extended to seven years in some cases because of interconnection queue backlogs and transmission upgrades required upstream of substations serving data center clusters, per mgrid.org’s analysis of PJM auction data. A hyperscaler can order accelerators on a two-year cycle and still wait seven years for the power to run them.

This is the same constraint we examined in our analysis of power supply innovations for data centers, where medium-voltage transformer lead times extending to three years and GE Vernova’s $176 billion backlog showed the equipment chain, not chip supply, now limits AI expansion. The 2026 update is that the constraint has shifted from hardware procurement into regulatory and market processes: interconnection queues, capacity auctions, and rate cases.

Utility Rate Requests and Who Pays

The $18.6 billion in first-half 2026 rate requests follows an even larger 2025, when annual requests hit $31 billion, more than double the $15 billion requested in 2024, according to GCN’s reporting on PowerLines data. Regulators have been approving more than half of every dollar requested in recent years.

The cost-allocation question is politically sensitive. In PJM’s footprint, the Market Monitor estimated that data center demand was responsible for 63% of the price increase in the 2025/2026 auction, translating to $9.3 billion in additional capacity costs recovered from customers. In practical terms, the average residential bill was expected to rise by $18 per month in western Maryland and $16 per month in Ohio as a direct result of data center-driven capacity cost allocation.

That is the mechanism that turned data centers into a local political issue. A household in Ohio is paying for the capacity reserve that a data center campus in the same zone required the grid to hold. The Network World survey of moratorium activity found 71% of Americans now oppose data center construction in their area, up from 42% the previous summer, with energy costs the top stated reason.

Capacity Market Prices and Load Swing Risk

PJM Interconnection’s 2028/2029 capacity auction cleared at $325/MW-day across its region, hitting the price cap, and left the grid operator roughly 6.8 GW short of its 20% installed reserve margin target, an increase from the prior auction’s 6.5 GW shortfall, according to Utility Dive. The auction drew only about 525 MW in new resources, down from 774 MW in the December auction.

The price trajectory across auctions reflects a market sending an increasingly loud scarcity signal that supply is not responding fast enough.

PJM auction / metric Value Source
2024/2025 clearing price $28.92/MW-day mgrid.org
2026/2027 clearing price $329.17/MW-day mgrid.org
2028/2029 clearing price $325/MW-day (price cap) Utility Dive
2028/2029 reserve margin shortfall ~6.8 GW below 20% target Utility Dive
Data center share of 2025/2026 price increase 63% mgrid.org

Without the price collar, PJM said the auction would have cleared at nearly $555/MW-day across its footprint and $777/MW-day in the Commonwealth Edison zone of northern Illinois, at a total cost of $29.7 billion rather than the actual $16.4 billion. The collar limited the headline number without fixing the underlying supply gap.

The second-order risk is operational, and it is newer than the price story. At 7:56 a.m. EDT on July 22, 2026, Ting Labs detected the start of a major transmission disturbance in Northern Virginia; PJM subsequently reported that more than 3 GW of power demand, about 3% of system demand at the time, went offline as affected data centers transferred rapidly to backup power, according to Brandon Owens’ analysis in Utility Dive. PJM said the disturbance caused no reliability impact and Dominion Energy stabilized conditions within minutes.

The pattern has occurred before. ERCOT identified eight events between November 2020 and March 2023 in which faults near large Texas Gulf Coast industrial loads produced repeated demand reductions of approximately 400 MW to 700 MW. On December 7, 2022, multiple faults and delayed 19-cycle clearing following a breaker failure produced roughly 1,560 MW of load reduction in West Texas. Ireland’s EirGrid and SONI documented four major data center demand reductions from 220-kV transmission events, the largest 387 MW on May 8, 2025.

NERC’s May 2026 Level 3 Alert called for improved computational-load modeling, studies, instrumentation, commissioning, and operational coordination, but it is not itself an enforceable reliability standard. In July, FERC directed NERC to develop new or modified computational-load reliability standards and registration criteria, with filings due by December 31, 2026.

State Moratoriums and Connection Halts

Texas Governor Greg Abbott halted all new data center grid connections in August 2026, directing the Public Utility Commission of Texas and ERCOT to audit every project in the queue, as Ars Technica reported. The state had declared itself the “epicenter of AI dev” less than a year earlier.

New York went further in July 2026, becoming the first state to impose a statewide moratorium on new hyperscale data center construction, a one-year ban applying to facilities using 50 MW or more, according to Reuters. Governor Kathy Hochul’s executive order cited utility bill impacts and natural resource depletion, and the state plans to require data centers to pay more for energy or supply their own.

The response extends beyond two headline states. Network World’s survey of the U.S. Data Center Moratorium Tracker found 127 other US jurisdictions with similar restrictions, and Data Center Watch recorded at least 75 US data center projects worth about $130 billion blocked or delayed in the first quarter of 2026 alone, roughly equal to the total for all of 2025.

There is a countervailing signal to consider alongside the moratorium narrative. Reuters reported in September 2026 that data centers have requested roughly as much electricity across the middle of the United States as it takes to power every home in the country, but much of that demand may be speculative, with developers filing interconnection requests to hold queue positions they may never use. If a meaningful share of queued demand is option value rather than committed load, the capacity shortfall partly reflects forecasting errors rather than pure supply shortages.

Generation and Storage Buildout

The supply response is real and record-setting. US power plant developers plan to add 86 GW of new utility-scale generating capacity in 2026, a record if realized, with solar at 51% of additions, battery storage at 28%, and wind at 14%, according to EIA’s Preliminary Monthly Electric Generator Inventory. In 2025, 53 GW was added, the largest single-year installation since 2002.

Storage is the fastest-growing segment. Developers plan to add 24 GW of utility-scale battery storage in 2026, compared with a record 15 GW in 2025, with Texas accounting for 53% of planned additions, California 14%, and Arizona 13%. The EIA projects US utility-scale battery storage capacity nearly reaching 65 GW by the end of 2026.

Solar crossed a symbolic threshold in May 2026, generating more electricity than both coal and wind, per Electrek’s review of EIA data. The EIA’s STEO projects renewables’ generation share rising from around 24% in 2025 to 25% in 2026 and 27% in 2027, while coal slides from 17% to 14% and natural gas holds at 40%.

Two caveats matter for the AI load question. First, nameplate capacity is not firm capacity. A solar fleet at 51% of annual additions contributes far less to the summer evening peak than its megawatt rating suggests, which is precisely when reserve margins are tested. Second, additions are geographically mismatched with load. Solar additions concentrate in Texas, Arizona, California, and Michigan, while the tightest capacity conditions are in Mid-Atlantic PJM zones where data center concentration is highest.

The equipment supply chain is responding too. Flex (FLEX) announced on September 3, 2026 that it would acquire EPC Power at a value of $4.4 billion, adding power conversion capabilities for AI data centers and grid apps, with the transaction expected to close in the fourth quarter of calendar 2026. That is supplier positioning for the power layer of the buildout rather than the compute layer, consistent with the capex-funding dynamic we tracked in our hyperscaler spending analysis.

Market Context and Read-Through

The equity market is not pricing a grid crisis. The S&P 500 (SPX) closed Friday, September 11, 2026 at 7,656.98, up 65.28 points or 0.86%, and 1.65% below its 52-week high of 7,785.76 set August 10, 2026. The Nasdaq Composite (IXIC) closed at 26,333.04, up 251.32 points or 0.96%, still 2.37% below its 52-week high of 26,972.62 from May 25, 2026. The Dow Jones Industrial Average (DJI) closed at 52,573.29, up 509.19 points or 0.98%, and 2.71% below its 52-week high of 54,036.93 set August 3, 2026.

Over the trailing year the S&P 500 is up 16.29%, the Nasdaq 18.93%, and the Dow 14.70%, with all three indices’ 52-week lows set on March 23, 2026. The market has treated power constraints as a timing risk to AI capacity rather than a threat to the AI trade itself.

The read-through runs through utilities and power producers. Constellation Energy (CEG) and Vistra (VST) have been repriced as AI power beneficiaries, and Talen Energy (TLN) carries an Amazon-linked nuclear deal that has become a template for behind-the-meter supply. The trade-off is that regulated utilities face a rate-case constraint that unregulated generators do not: every dollar of capacity cost recovered from ratepayers is a dollar that has to survive a public utility commission hearing, and commissions in Ohio, Maryland, and Illinois are increasingly reluctant to hand data center costs to households.

Oil and gold are secondary inputs here. WTI crude (CL=F) settled at $100.05 per barrel on September 11, 2026, down $2.43 or 2.37%, while gold (GC=F) settled at $4,366.20 per ounce, up $1.70 or 0.04%. Bitcoin (BTC-USD) traded at $77,270.47 as of 8:00 p.m. ET on September 11, up 0.13%. For data center economics, the relevant variable is delivered electricity cost by region, not crude price, though oil still feeds construction, backup generation, and equipment transport costs.

Upcoming Developments

Three dated events will determine whether the 2026 grid squeeze eases or tightens.

First, FERC’s response to PJM’s backstop capacity auction and “connect and manage” framework for data centers. PJM’s board was preparing filings that needed to be submitted by July 2026 for the auction to be held in September as planned. ClearView Energy Partners noted that FERC’s response will likely influence utility capital investment, data center development timelines, and the allocation of reliability risks and costs. A backstop auction that clears above the current cap would confirm that the capacity market cannot attract supply at regulated prices.

Second, NERC’s computational-load reliability standards, due to be filed by December 31, 2026 under FERC’s July directive. If those standards include enforceable ride-through and post-fault recovery requirements for large loads, they will change the engineering specification for every new data center campus in the interconnection queue. Ireland’s proposed Grid Code Modification MPID345 offers a preview: it would require a facility to restore at least 90% of pre-fault demand within 500 milliseconds after fault clearance and voltage recovery.

Third, 2027/2028 and 2028/2029 auction outcomes alongside utility rate case decisions in Ohio, Maryland, and Illinois. If regulators continue approving more than half of requested increases, the political pressure that produced moratoriums in New York and Texas will spread to states that have so far welcomed data center investment.

I expect PJM’s next base residual auction to clear at or above the existing price cap, because new supply additions of 2 to 3 GW annually are running well behind the 5 to 7 GW of data center load PJM projects adding each year through 2032. That gap does not close on a one-year horizon, and the price cap is the mechanism that will keep clearing auctions at the ceiling rather than at a level that attracts new generation.

For engineering and infrastructure leaders, power availability, not accelerator availability, now sets the realistic timeline for new AI capacity in PJM, ERCOT, and Dominion’s Virginia territory. The teams that plan around seven-year interconnection queues, secure behind-the-meter generation, and design for enforceable ride-through requirements will deploy on schedule. Teams that assume a standard two-year data center build cycle will not.

The grid is running out of specific local capacity, transmission, and market mechanisms required to serve gigawatt-scale loads that can appear and disappear in under a second.

Sources and References

Sources cited while researching and writing this article:

Rafael

Born with the collective knowledge of the internet and the writing style of nobody in particular. Still learning what "touching grass" means. I am Just Rafael...