Compare horizons, uncertainty ranges and defined model scenarios. These are research estimates, not official outage predictions.
Odds come from base rates adjusted by today's signals, and they move when you run a stress test. The bar and range show the model's uncertainty for the horizon you choose.
The most likely path is more cost and more close calls. These shares follow the model's 2030 odds and shift with stress tests.
Data-center forecasts come in low, fast-track capacity arrives on time and large loads learn to flex. Grid Strategies argues the grid could be more reliable in five years than it is today.
Bills keep rising. Emergency orders and alerts become routine. Data centers run backup generators or cut load at peaks. A few rotating outages last hours during extreme cold or heat. Storms still cause the long outages.
A long Arctic event overwhelms gas supply and generators in PJM, MISO, the Southeast or Texas. Millions lose power for up to several days. The 2021 Texas failure killed at least 246 people and is the closest precedent.
A war-linked attack or an extreme solar storm damages high-voltage equipment across a region. Power returns in stages over weeks to months, because the largest transformers take years to build.
Sort all 51 state/DC entries. Select a row to return to its map and state details. Values reflect the active horizon and scenario.
Inspect automatic observations alongside the dated research assumptions behind each signal.
Each signal is scored from dated evidence: Calm, Watch, Elevated or Severe. Open one to see the evidence behind it and what it measures.
Current NOAA alerts, newest first. Other evidence remains in the dated research assumptions below.
Each one comes with the strongest evidence against treating it as a crisis.
The average U.S. home paid 18.31¢ per kWh in July 2026, about 22% above the 2022 average of 15.04¢. PJM's capacity auctions have cleared at the price cap three times running. The 2028/29 auction alone costs $16.4B, with about $6.3B tied to data centers.
The other side. The Congressional Research Service reports that states with the largest data-center growth generally saw prices fall from 2019 to 2025, so the link is uneven.
NERC attributes 80% of customer interruptions to local distribution lines, which its long-term forecast does not cover. The average customer lost 10.2 hours in 2024, a heavy hurricane year, and 4.4 hours in 2025. Without major events the figure stayed near 2 hours in both years.
The other side. Everyday reliability is holding steady. The swing comes from weather.
Power transformers averaged about 128 weeks of lead time in 2025, and the largest can take four years. Generator step-up units passed 160 weeks by early 2026. Wood Mackenzie estimated a 30% supply shortfall, with imports covering about 80%. Executive Order 14420 now lets DOE block or unwind some foreign equipment tied to adversary countries, with rules due around December 24. Tariffs on grid equipment run 15% until 2028, then 25%. Scarcity matters most after a failure, because it slows recovery.
The other side. Manufacturers have committed about $1.8B to new North American capacity, and the order reaches covered foreign entities, not all imports.
Wood Mackenzie · pv magazine · Vinson & Elkins on EO 14420 · Reuters via AJOT · White & Case on tariffs
The industry's security center logged more than 3,500 physical incidents in 2025, and about 3% disrupted electricity. The worst recent U.S. case is the 2022 gunfire attack in Moore County, North Carolina, which left 45,000 customers out for days. Federal reports obtained by Wired now track "anti-tech" extremism and monitor data-center protests, with no dated plot described.
The other side. About 97% caused no outage. Many are theft, vandalism or suspicious activity.
NERC rates 13 of 23 North American areas elevated or high within five years. It projects peak demand up 224 GW in summer and 245 GW in winter within ten. PJM, the largest U.S. grid, has missed its reserve target twice, reaching 14.7% against about 20% and falling 6,831 MW short for 2028/29. Its backstop to close that gap slipped: FERC suspended the plan, and the refiled version moves bids to March 2027.
The other side. Only about 5 GW of the 12 to 16 GW of data-center capacity slated for 2026 is under construction. PJM and ERCOT have already seen near-term projects slip. MISO's fast-track program could erase its reserve gap. Grid Strategies, in a report for environmental groups, calls NERC too pessimistic.
POWER · Power Engineering on PJM · APPA on the backstop · BIC · Grid Strategies
Risk is moving to winter mornings and evenings, when solar output is near zero and homes burn gas for heat. NERC's John Moura says "there's no reliability regime over the gas industry." In January 2026, generator outages peaked at 68,169 MW and gas production fell by as much as 14%. The lights stayed on, helped by DOE emergency orders, yet NERC's CEO said the system came "close to the edge."
The other side. Eastern generator outages reached 49,701 MW, far below the 90,500 MW of Winter Storm Elliott in 2022. Preparation is working.
In May 2026, NERC issued a Level 3 alert, its highest tier, over computing loads. Large sites can drop off the grid together during a voltage disturbance and cause gigawatt-scale swings. About 3,800 MW dropped off in Northern Virginia in July, the largest such event in PJM, and a 1,500 MW event hit the Dominion-FirstEnergy area in April. Data centers already use about a quarter of Virginia's power, and EPRI's scenarios reach 39% to 57% by 2030.
The other side. The same loads can help. Texas can now curtail them in emergencies, and DOE counts more than 35 GW of unused backup generation nationwide.
Morgan Lewis · alert coverage · APPA · ERCOT · EPRI · DOE
Dragos reported in February that China's Volt Typhoon is still embedded in U.S. utilities and that some breaches may never be found. U.S. officials describe the goal as positioning for disruption during a conflict. Iran-linked groups have disrupted industrial controllers in energy, water and local government since March 2026. Since late July, attackers have hit water systems in at least seven states, with no formal attribution yet. Cuts have affected about a third of CISA's staff. In Poland, Russia-linked attackers hit about 30 wind, solar and heat sites in December 2025.
The other side. Poland's customers kept their power. This review found no confirmed cyberattack that caused a U.S. customer outage. Access is a long way from a sustained blackout.
The Record · CISA advisory update · RISI on Poland · The Record on water · Nextgov on CISA
A Carrington-class solar storm could damage large transformers across a region. Published odds run from about 1% to 12% per decade, and Lloyd's put the U.S. cost at $0.6T to $2.6T. An electromagnetic pulse attack requires a nuclear detonation at high altitude, which means a war is already under way.
The other side. May 2024's G5 storm caused no major grid failure, and the 1989 Quebec collapse lasted 9 hours. Experts disagree widely on how much equipment would actually be lost.
Stony Brook summary of Riley and Chapman · Make Tech Easier on the odds range · POWER
Review NERC area assumptions, reserve margins and the distinctions between planned resources and customer reliability.
NERC, the grid's reliability regulator, rates each area by year. Elevated means shortfalls in extreme conditions. High means planned resources miss adequacy targets even in normal conditions.
NERC's data dates from July 2025 and leaves out most fast-track capacity approved since then. NERC's John Moura called the outlook "not a prediction of failure but an early warning." In its May 2026 summer assessment, NERC rated ERCOT normal for the season and flagged only New England, the Northwest and the far-west Texas zone in the U.S.
PJM's auction figures and these 2030 numbers measure different years and resources, so they differ. Two margins mislead on their own. ERCOT's counts large loads it can curtail in an emergency. Basin and Northwest clear their margins, yet NERC's hourly modeling shows unserved energy by 2029. NERC's report notes that WECC's probabilistic results are not weighted averages like other regions', so they run high.
Sources: NERC 2025 Long-Term Reliability Assessment, Power Engineering on the NERC briefing, GridBeyond on the 2026 summer assessment.
Explore practical household preparation and the source model’s extended-outage scenario.
That means hours to a few days, usually from a storm. Differences between states are real but too small to justify moving for this alone.
Water, food, cash, medication and a battery radio. Britain's new Be Ready campaign uses the same three-day baseline.
Add more if you live on a long rural line or in a hurricane or ice zone, where repairs take longest.
Anyone on oxygen, dialysis or refrigerated medicine needs a written backup plan and should ask their utility about medical-priority programs.
A Georgia Tech-led model estimated that a multi-day blackout during a Phoenix heat wave could send about 800,000 people to emergency care. Local officials call that combination remote. Most gas furnaces also stop without power, because the blower and ignition need it. Know where you would go.
Rooftop solar without a battery usually shuts off in an outage. A home battery covers the most likely outages, from hours to a few days. Run generators outdoors only.
A 90-day survival kit answers a scenario this page rates under 1%. A modest stock you actually rotate is more useful.
Fortune on Be Ready · Heat & Cool on furnaces · Stone et al. 2023 · Cronkite News on the response
This timeline assumes a wide-area blackout with no restoration and no outside help, which is a worst case. Almost every real outage ends in the first columns. Outlined cells rest on a documented event. The rest are Claude's estimate.
| System | Hour 1 | Hour 8 | Day 1 | Day 3 | Day 7 | Day 14 | Day 30 |
|---|---|---|---|---|---|---|---|
| Home heat and cooling | Home holds temperature | Furnaces and AC stop | Dangerous in extreme cold or heat, as in Texas in 2021 (documented) | ||||
| Home medical devices | On battery | Batteries running out | Users head to hospitals for electricity (documented) | ||||
| Food at home | Safe if closed (documented) | Fridge past USDA's 4-hour limit (documented) | Half-full freezer done (documented) | Full freezer done at 48 hours. Shelf-stable food only (documented) | |||
| Stores | Open with generators (documented) | Small shops close, large ones keep going (documented) | Stock thinning | Empty without deliveries | |||
| Cash and cards | ATMs fail in minutes (documented) | Card payments halve, many shops cash only (documented) | Cash only, and cash runs short | ||||
| Cell service | On backup | Some sites going dark (documented) | Many sites down | Half or more of sites down. Slow to recover (documented) | |||
| Home internet | Out unless the home has backup | Mostly out (documented) | |||||
| 911 centers | On generators | Some reroute, fewer phones can reach them (documented) | Depend on fuel deliveries | ||||
| Gas stations | Most pumps off without power (documented) | Long lines, a third or more out (documented) | Rationing, supply by truck only | ||||
| Water and sewer | Gravity and tanks hold (documented) | High-rises lose water (documented) | Pressure falling, plants on generators | Boil notices spread, sewage overflows (documented) | |||
| Hospitals, nursing homes | On generators (documented) | Crowded by people needing power (documented) | Fuel deliveries critical | Evacuations if fuel fails | |||
| Traffic signals | Dark unless battery-backed. Intersections become four-way stops | ||||||
Home heat. Texas counted 246 deaths from the 2021 storm, 161 of them from cold exposure, and at least 19 from carbon monoxide.
Medical devices. In New York in 2003, ventilator and nebulizer users went to emergency rooms for electricity. HHS counts over 3 million Medicare beneficiaries who rely on electricity-dependent equipment or services.
Food. USDA's limits are 4 hours for a closed fridge, 24 for a half-full freezer and 48 for a full one. In Spain, larger retailers with generators stayed open and small ones generally closed. New York saw diarrheal illness rise by day 3.
Cash and cards. In Spain's 2025 blackout, ATM batteries lasted 10 to 20 minutes, under 4% had other power, and card payments fell 55% that day.
Cell and internet. Five days after Helene, 48.7% of North Carolina's cell sites were out, 338 of them from lost power, and 796,999 cable and wireline subscribers were down across the disaster area. After Maria, 95.6% of Puerto Rico's sites went out and 74.1% were still out 23 days later. Storm damage is mixed into those figures. Battery figures are old: 80% of surveyed California sites had 4 hours or more in 2008.
911. After Helene, eight North Carolina call centers rerouted their calls. After Maria, Puerto Rico's centers kept working while St. Croix's went down.
Fuel. After Sandy, stations had fuel but no power to pump it. 38% of New York-area stations had no gasoline on November 3, and 11 of 57 terminals were shut on November 5.
Water. In New York in 2003, gravity held pressure, high-rises lost water for days and 500 million gallons of sewage were released when plant generators failed. In Texas in 2021, 9 million people had water problems and 853 systems were on boil notices.
Hospitals. CMS requires facilities with onsite fuel to plan to keep emergency power running. This review found no reliable figure for how much fuel hospitals typically hold. In 2003, 4 of New York's 75 hospitals temporarily lost power.
Not found. Traffic-signal battery coverage, home device battery runtimes and store inventory depth. Those cells are estimates. So are the later days of most rows, because few modern events have reached them.
The collapse came at 12:33 on April 28. Some regions took up to 16 hours, and almost 24 in places. Some restart units failed to start.
About 45 million people in eight states lost power. Most of New York City was back within about two days.
ERCOT's chief executive said the grid came within "seconds and minutes" of failure, and that blackouts "could have occurred for months." A total collapse would have needed a full restart.
Power transformers run about 128 weeks to deliver. Shared spares help: more than 50 companies have binding access to one program, and another holds its own inventory for subscribers in 23 states. They cover a local loss far better than a regional one.
Review base rates, mappings, source claims, original research calls and the limitations of the model.
FORTRESS America, announced September 30, aims to let major bases run at least 90 days without outside power, fuel or water in wartime. That is a military resilience target. It forecasts nothing about civilian power, and the Pentagon has released no cost or timeline. Fortune · Snopes · Prepared States
A 2019 peer-reviewed article found no published model behind the figure and could not validate a primary source. A long regional outage would still be deadly, especially for people on medical devices and in extreme heat or cold. Energy, Sustainability and Society
NERC's high-risk areas do cover much of the country by 2029. In this rating, high means modeled shortfalls such as about 10 hours a year in PJM, before counting most fast-track capacity. That points to rotating outages at peaks, and nothing like the country going dark. The claim · NERC
Executive Order 14420 covers foreign-made equipment. The separate DOE order for PJM ran October 13 to 16 and covered a maintenance-season crunch, with backup generation as a last resort. Neither one reports a failing grid. Morgan Lewis · DOE
In September a federal appeals court vacated DOE's first order for a Michigan coal plant and called the power a "limited backstop." Later orders have not been reviewed and remain in force, and DOE has asked for a rehearing. Utility Dive · APPA
The access is real and serious. Taking down the whole country would mean disabling the Eastern, Western and Texas grids together and keeping them down. Nothing public shows that capability. A regional disruption during a war is the credible version. Dark Reading
Predictions of national collapse have failed since at least 2013. That record says nothing about the auction shortfalls, price increases and NERC ratings, which are measured facts. The argument · Utility Dive
Original research calls from the supplied file. These remain dated records and require manual outcome review. Automatic reports archive model observations separately.
Everything here can be checked. The model is simple on purpose, because the events it covers are too rare to fit anything fancier.
Each outcome starts from how often it has happened. Ten ordered-outage events since 2011 touched 14 grid areas. Five reached 100,000 customers. Attack outcomes have no large U.S. precedent, so they start from small assumed rates.
Each area's rate is multiplied by its rating for that year: 0.8 for Normal, 1.1 for Elevated and 1.6 for High. Each area also has its own season, so California's risk sits in summer and the Carolinas' in winter.
Eighteen signals are scored from dated evidence. Each one pushes the odds up or down from its usual level, and its effect fades over time. A cold forecast fades in months. A capacity shortfall lasts years.
Automatic inputs use published alert and geomagnetic-scale rules and may change by multiple levels. Feed status and timestamps are recorded hourly. Other inputs remain dated assumptions. Original research calls require manual outcome review; they are not automatically scored.
Monthly hazard = base rate × season × rating × eΣ weight × level change × fade Chance = 1 − e−Σ hazard
What it can't do. No model can predict an attack. Attack odds are base rates nudged by tension, and they say nothing about anyone's actual plans. The weights are Claude's judgment and have not been statistically fit, because there are too few events to fit them. State odds use approximate load splits. Treat a number here as a reasoned estimate with a wide range.
What stress tests mean. A stress test shows the odds if that scenario happens. It doesn't say how likely the scenario is.
| When | Where | Cause | Size | Source |
|---|---|---|---|---|
| Feb 2011 | Texas and the Southwest | Cold, unit failures | 4.4M customers | POWER |
| Jan 2014 | SERC-East | Polar vortex | Under 300 MW | POWER |
| Aug 2020 | California | Heat, supply shortfall | 492,000, then 321,000 | CAISO |
| Feb 2021 | ERCOT, SPP, MISO | Freeze, gas failure | 23,418 MW, 4.5M people | NERC |
| Jun 2022 | Columbus, Ohio (PJM) | Heat after storms, line overloads | Nearly 250,000 | Ideastream |
| Sep 2022 | Northern California | Emergency alert, misread order | Thousands, 46 MW | CBS |
| Dec 2022 | SERC-Central and East | Cold, unit outages | 5,400 MW, 500,000 at one utility | FERC |
| Apr 2025 | SPP, Louisiana area | Storm damage, protection | 24,209 | KSLA |
| Apr 2025 | SPP, northwest Louisiana | Voltage, maintenance outages | 140 MW, 30,000 | SPP |
| May 2025 | New Orleans (MISO) | Transmission emergency | 600 MW | MISO |
That is 10 events in 15.8 years, and 8 since 2020. This list may be incomplete, and this review found none in 2026. Left out: the automatic September 2011 Southwest cascade, which no one ordered, and smaller utility-ordered cuts in New York in 2019 and Spokane in 2021. This review found no cyber or physical attack that cut 100,000 customers. The largest was about 45,000 in North Carolina in 2022. GAO said in 2019 that cyber incidents "reportedly have not caused a domestic power outage."
Supply shortfall. Each NERC rating from 2026 to 2030 scores Normal 10, Elevated 50 or High 100, then averages across the five years. NERC areas don't follow state lines, so each state gets an approximate load split across areas. Claude estimated those splits, and they are the softest numbers here. NERC doesn't assess Alaska or Hawaii.
Outage history. The average of EIA's all-events outage minutes for 2022 to 2025. One big storm can dominate a state's average, which is why the panel shows each year.
Data-center load. Modeled 2023-vintage electricity estimates divided by population. They understate North Dakota, where EPRI put data centers at 15% of state electricity.
Potential targets. The average of four percentile ranks: defense spending as a share of GDP, defense dollars, data-center electricity and population. It measures consequence. It carries no information about anyone's actual plans. State-level incident counts weren't available in a usable public form, so they are not included. This map stops at the state level on purpose and shows no individual sites.
Overall score. 40% supply, 30% outages, 15% data-center load and 15% potential targets by default, each scaled 0 to 100. It ranks states against each other and is not a probability. Change the weights under the map to test it.
Danger if it fails. 70% climate and 30% outage duration. Climate takes the larger of heating demand against 8,000 degree days or cooling demand against 3,500, from EIA's 2024 figures. Duration is the state's percentile on outage hours. One year of degree days is a rough stand-in, and hurricanes, wildfire and ice aren't scored here. HHS publishes state counts of people on home medical devices, and this build could not retrieve them, so they aren't counted.
Safe zones. A state must have at least 75% of its load in NERC areas rated Normal through 2030. It then gets credit for below-median outage hours, for staying outside the top 10 on potential targets and for staying outside the top 10 on outage danger.
The cause is contested, often along policy lines. The evidence supports parts of each case.
Coal utilities and DOE argue dispatchable plants are closing too fast. DOE modeled a 100-fold rise in outage risk in 2030 if 104 GW retires without replacement. Michigan's attorney general and environmental groups are challenging DOE's orders that keep plants open. DOE · POWER
Clean-energy groups say permitting and interconnection delays block the fastest resources to build, and point out that gas plants made up 70% of PJM's unplanned outages in Winter Storm Elliott. NERC also lists siting and permitting among the top causes of transmission delays.
Competitive generators argue that policy interventions distort investment and that forecasts should separate committed load from speculative projects.
Grid Strategies argues NERC undercounts new supply and overcounts data centers. Utility Dive
Dates that will move these odds. Each one tests a claim on this page.
EIA's next energy outlook updates gas storage and winter prices.
The current Michigan coal-plant order runs out, testing how DOE responds to the court ruling.
NERC's winter assessment, usually published in November. Watch PJM, MISO, SERC-East and the Northwest.
The stopgap extending the cyber information-sharing law runs out.
DOE rules under EO 14420 are due. They decide how hard the order hits equipment supply.
NERC's next long-term assessment, which will count MISO's fast-track projects.
PJM takes backstop bids March 1 to 28 and reports results May 10.
PJM's first delivery year below its reserve target begins.
MISO enters NERC's high tier if its plans do not change.
PJM, ERCOT, Basin and the Northwest enter the high tier on the same terms.
APPA on the PJM backstop · APPA on the court case · Federal News Network · EIA · Jenner & Block on EO timing
Adapted from the supplied U.S. Grid Forecast research file, compiled by Claude from public sources. Model assumptions, estimates, and source claims are preserved for inspection and have not been independently certified by Grid Safety.
Hourly source checks, current feed health, recorded model stress, and weekly/monthly reports.
Connecting to public observations… Until connected, the atlas shows the dated research baseline.
U.S. power grid forecast
Choose one or more scenarios. The model shows conditional outcomes—not the chance that a scenario will happen.
Change the weights of the overall score. These controls change relative state rankings, not the probability model.