The grid in one picture
The U.S. power grid is the largest machine people have ever built, and it has a peculiar rule: supply must equal demand every instant, because there is almost nowhere to keep electricity once it is made.
The model behind this page is one grid about an eighth the size of the lower 48 states, roughly Texas-sized, stocked with the national mix of plants. On a typical day its generation shares track the real U.S. shares, and a single plant tripping moves its frequency the way it does on a real island grid.
4,528 TWhU.S. generation in 2025, a record (EIA)
1,281 GWutility-scale capacity, end of 2025, plus ~60 GW of rooftop solar
759 GWrecord lower-48 demand, July 29, 2025
~5%of power lost as heat in transmission and distribution
How to use it
- Look around: drag to orbit, scroll or pinch to zoom, right-drag or two-finger drag to pan.
- Click anything, a plant, a substation, the town, to read about it here.
- Turn the knobs at the bottom: demand, time of day, and under More, gas price, wind and clouds. The readouts at the top left show frequency, price, reserves, losses and the generation mix.
- Pick a scenario: heat waves, winter storms, an eclipse, a hurricane, blackouts and the restart that follows.
- Explore (top right) lists every topic. The Next link at the bottom of each one follows electricity from the plant to your outlet.
The coloured dots moving along the wires are power flow. Their speed shows how hard each line is working, and their colour shows its voltage. Switch on Voltages to colour the wires too.
Natural gas
The workhorse, and the plant that usually sets the price.
40.2%of U.S. generation in 2025 (1,807 TWh)
~40%of utility-scale capacity, about 500 GW
$3.52per MMBtu, the 2025 Henry Hub average (2024: $2.19)
$51–129per MWh, new combined cycle, unsubsidized (Lazard 2026)
How a combined cycle works
A combined-cycle plant runs two engines on one fuel. A gas turbine, essentially a jet engine bolted to the floor, burns gas and spins a generator. Its exhaust, still around 600 °C, boils water in a heat-recovery steam generator, and that steam spins a second turbine. Getting two generators' worth of work from one fire pushes efficiency to roughly 55–60%, versus about 33% for an old steam plant.
Efficiency is quoted as a heat rate, the fuel burned per unit of electricity. A good combined cycle needs about 6.5 MMBtu per MWh. At $3.52 gas that is about $23 of fuel per MWh, plus a few dollars of operations. That figure is roughly what the plant offers into the market.
Why it sets the price
In most U.S. markets gas plants are the last ones needed in most hours, so their fuel cost becomes everyone's price (see How the market clears). That is why wholesale power prices rise and fall with gas. Open More and drag the gas price to watch the clearing price follow it.
Trade-offs
- For: flexible (combined cycles ramp in minutes, peakers start in ten), relatively cheap to build, and about half the CO₂ of coal per MWh (roughly 0.37 versus 1 tonne).
- Against: fuel arrives just in time by pipeline, with little on-site storage, so the same cold snap that drives demand for gas heating can starve power plants (see Winter Storm Uri). Methane leaks erode the climate advantage, and prices swing.
- Now: demand growth has set off a rush of new gas projects, but turbines are scarce. Lazard's 2026 report notes that constrained turbine supply is stretching development timelines well beyond historical norms, and new combined-cycle costs are at a 15-year high.
Peakers: the expensive last megawatt
Plants built to sit idle most of the year, then save the day for a few hours.
A simple-cycle combustion turbine, or peaker, is a gas turbine without the steam half. It burns 9.5–12 MMBtu per MWh, nearly twice a combined cycle, but it is cheap to build and starts in about ten minutes. Many can also burn oil from a tank on site, which is why oil still shows up in New England's mix during cold snaps when gas pipelines are full.
$144–276per MWh, new gas peaker over its life (Lazard 2026)
$52–122per MWh just to run an existing one
0.4%of 2025 generation came from petroleum
Peakers run few hours, so each start and each hour has to recover a lot of fixed cost. In the model their offers include that start-up cost, which is why the price jumps from the $30s to $50 or more when demand climbs past what the combined cycles can cover. Push the demand knob up on the summer day and watch the marginal unit change from "Combined cycle" to "Gas peaker".
The third unit at the model's peaker site is the one you can synchronize by hand in Synchronizing a generator.
Coal
Declining for fifteen years, then a 2025 comeback.
16.3%of 2025 generation (737 TWh)
+13%more coal power than in 2024, as gas got pricier
~165 GWof capacity, about half its 2011 peak
$34–69per MWh to run an existing plant (Lazard 2026)
How it works
Coal arrives by train; a unit train of about a hundred cars carries 10,000-plus tons, roughly one day's burn for a 1,000 MW plant. The coal is ground to powder finer than flour and blown into a boiler the height of a 20-story building. The fire turns water into steam at more than 500 °C and about 2,400–3,500 psi, and the steam drives a turbine. About a third of the coal's energy becomes electricity. The rest leaves as warm water through the cooling tower and warm gas up the stack.
Trade-offs
- For: weeks of fuel sit in a pile on site, a real advantage in a winter storm when gas pipelines are constrained. Big spinning machines also provide inertia (see Frequency).
- Against: the most CO₂ per MWh of any major source (about a tonne), plus sulfur dioxide, nitrogen oxides, particulates, mercury and ash. Plants are old (many are 40–50 years) and slow: they ramp slowly and cannot go much below about 40% output, which is why coal often runs through negative-price hours rather than shut down.
- Now: new coal is not being built (Lazard puts it at $72–177/MWh), but retirements slowed in 2025, and the Department of Energy began using emergency orders to keep some retiring plants running.
Nuclear
About a fifth of U.S. electricity, from under 8% of its capacity, running nearly all the time.
17.3%of 2025 generation (785 TWh), the largest carbon-free source
~97 GWfrom about 94 reactors at 54 plants
~92%capacity factor, the highest of any source
$31.76per MWh, average total cost to run the fleet in 2023 (NEI)
How it works
Uranium enriched to 3–5% U-235 sits in fuel rods in a reactor core. Splitting atoms heats water. In a pressurized water reactor (about two-thirds of the U.S. fleet), that water is kept under such high pressure that it can't boil, and it heats a separate loop that makes steam. In a boiling water reactor, the core boils the water directly. From there it is a steam plant like coal, at about 33% efficiency, which is why the cooling towers are so big. Those towers release water vapor, not smoke.
Nuclear turbine-generators often have four poles and spin at 1,800 rpm rather than 3,600, because the huge low-pressure turbine blades would fly apart at the higher speed. A fuel pellet the size of a fingertip holds roughly the energy of a ton of coal. Plants refuel every 18–24 months, swapping about a third of the core.
Costs: cheap to run, very expensive to build
Existing plants are cheap to run: about $32/MWh in 2023, of which fuel was only about $5. Building new ones has been the problem. Georgia's Vogtle units 3 and 4, the only new U.S. reactors in a generation, came in years late and at more than $30 billion. Lazard's 2026 estimate for new nuclear, based on Vogtle, is $175–255/MWh. Restarting a closed plant is far cheaper, and a few are being restarted, largely on contracts with data-center operators.
Trade-offs and quirks
- Carbon-free, firm, compact, and plants last 60–80 years with license renewals.
- U.S. reactors run flat out and rarely follow load. In the model nuclear is must-run: it is paid whatever the price is, even when that price is negative.
- Plants cannot black start: after a blackout they need reliable off-site power before they restart, so they are among the last to come back (see Black start). Losing off-site power triggers emergency diesel generators to keep the core cooled.
- Spent fuel has no permanent U.S. home yet; it stays on site in pools and dry casks.
Hydropower
The original renewable, and the grid's favorite black-start plant.
5.5%of 2025 generation (247 TWh)
~80 GWconventional hydro, plus ~22 GW of pumped storage
6.8 GWGrand Coulee, the largest U.S. plant of any kind
The physics in one line
Power = water density × gravity × flow × height × efficiency. One hundred cubic meters of water per second falling 100 meters through a 90%-efficient turbine gives 1,000 × 9.81 × 100 × 100 × 0.9 ≈ 88 MW. Hydro turbines are the most efficient machines on the grid, often above 90%.
Water turbines turn slowly, so their generators have many poles: speed = 7,200 ÷ poles in rpm at 60 Hz, so a 40-pole machine turns at 180 rpm and still makes exactly 60 Hz.
Its role on the grid
- Flexible: a hydro unit can go from idle to full in minutes, so operators save water for the hours when it is worth the most. The model gives hydro a daily water budget and spends it on the highest-demand hours.
- Black start: it needs only batteries to open its gates and excite its generator, so hydro is the classic plant for restarting a dead grid.
- Limits: output depends on rain and snowpack, and droughts cut it sharply. River flows also have floors and ceilings for fish and downstream users. In the spring scenario, snowmelt keeps the river high, so hydro can't hold back.
- Costs: nearly free to run, very hard to build anew. Most good dam sites are taken, and dams reshape rivers, block fish and flood valleys.
Wind
One-tenth of U.S. power, from the windy middle of the country.
10.3%of 2025 generation (464 TWh)
~159 GWcapacity at the end of 2025
~34%typical capacity factor
$37–99per MWh, new onshore wind, unsubsidized (Lazard 2026)
How it works
Power in the wind grows with the cube of wind speed: double the speed and there is eight times the power. A typical new U.S. turbine is about 3.4 MW on a tower around 100 m tall, with a rotor wider than a football field. No turbine can capture more than 59.3% of the wind's energy (the Betz limit); good ones get about 45%.
A turbine starts turning at about 3 m/s, reaches full output around 11–13 m/s, and feathers its blades to shut down above about 25 m/s to protect itself, which is why wind output collapses in a hurricane (see Hurricane). The rotor turns at about 10–15 rpm, so it is not synchronized to the grid. Power electronics convert its variable-frequency output to 60 Hz, which is also why it adds no inertia unless its controls are programmed to.
Why wind bids negative
Wind has no fuel cost, and many farms earn the federal production tax credit, roughly $30 per MWh generated. A farm earning it can accept a negative price and still come out ahead, so on windy nights with low demand, wholesale prices go below zero (see Negative prices). In the model, 70% of the wind fleet bids −$25.
Trade-offs
- For: among the cheapest new energy; no fuel, no emissions while running, land between turbines stays farmable.
- Against: variable and only partly predictable. The best wind is far from cities, which makes long transmission lines essential. Turbines also ice up: in Uri, frozen turbines were one of many failures. Other concerns are birds and bats, and neighbors' objections to noise and views.
- Policy: under the 2025 tax law, wind and solar projects that began construction after July 4, 2026 must be in service by the end of 2027 to get the federal credits.
Solar
The fastest-growing source on the grid, and the reason midday prices fall.
8.6%of 2025 generation, including rooftops (389 TWh)
+28%growth in solar generation in 2025
~151 GWutility-scale, plus ~60 GW on rooftops
$40–98per MWh, new utility solar, unsubsidized (Lazard 2026)
How it works
Each silicon cell makes about half a volt of direct current when light knocks electrons loose. Cells are wired into panels of about 400–600 W, panels into strings, and strings into inverters, which chop DC into 60 Hz AC that follows the grid's voltage. The panels in the model sit on single-axis trackers that turn east to west through the day, as most U.S. utility plants do. Watch them follow the sun.
Plants are usually built with more panel capacity than inverter capacity (a DC/AC ratio around 1.3), so they hit their inverter limit for a few midday hours and produce a fuller curve the rest of the day.
Trade-offs
- For: cheapest new energy in much of the country; quick to build (months, not years); modular from a rooftop to a gigawatt.
- Against: output is zero at night and lower in winter, so solar mostly adds energy rather than firm capacity unless paired with batteries. It concentrates in the middle of the day, flattening midday prices and leaving a steep evening ramp (see The duck curve). Large plants need land, and tariffs and supply-chain rules have raised panel costs.
- Grid physics: inverter-based, so it adds no inertia unless programmed to; newer "grid-forming" inverters can help hold voltage and frequency.
Try the eclipse scenario: a known, minute-by-minute ramp down and back up.
Geothermal
Small today, always on, and suddenly interesting again.
0.3%of 2025 generation (16 TWh)
~4 GWmostly in California and Nevada
$67–111per MWh, new geothermal (Lazard 2026)
Wells reach hot water or steam a mile or more down. Dry-steam plants, like those at The Geysers in northern California (the largest geothermal complex in the world), send that steam straight to a turbine. Flash plants let very hot water boil as pressure drops. Binary plants pass moderately hot water through a heat exchanger to boil a second fluid with a lower boiling point. Spent fluid is usually injected back into the reservoir.
- For: firm, low-carbon power around the clock (capacity factors of 70–90%) on a small footprint.
- Against: it works only where heat is near the surface, and drilling is expensive and risky before you know what is down there.
- What's new: "enhanced" geothermal borrows horizontal drilling and fracturing from the oil and gas industry to create reservoirs in hot dry rock, which could open up much more of the West. Unlike wind and solar, geothermal keeps the federal tax credits for projects starting construction through 2033.
Storage: batteries and pumped hydro
The newest way to break the "made the instant it's used" rule.
43.6 GWof grid batteries at the end of 2025 (EIA)
+58%battery capacity added in 2025
~22 GWof pumped storage, most built decades ago
~85%round-trip efficiency for lithium-ion
Grid batteries are shipping-container rows of lithium-ion cells, mostly lithium iron phosphate, typically sized to run four hours at full output. Pumped storage moves water uphill between two reservoirs when power is cheap and lets it fall back through turbines when power is dear; Bath County, Virginia, at about 3 GW, is sometimes called the world's largest battery.
What storage does
- Arbitrage: charge when power is cheap (midday solar, windy nights), discharge when it is expensive (the evening peak). In the model storage trades only when the price spread beats its ~18% round-trip loss. Watch the battery yard's lights: green while charging, amber while discharging.
- Fast frequency response: a battery can go from zero to full output in a fraction of a second, faster than any spinning machine. Toggle Battery fast response under More and trip a unit in Frequency to see the difference.
- Peak capacity: four hours of storage can cover the evening net peak, which is exactly what failed in California in 2020 and held in 2022 (see Heat wave). In 2026 California's batteries discharged more than 12 GW at the evening peak.
Limits: four hours is not four days, so a long winter storm still needs other plants. Lithium-ion batteries also carry a fire risk; a large fire at California's Moss Landing battery plant in January 2025 sharpened scrutiny of siting and safety.
Comparing costs
Two numbers matter, and they answer different questions.
The levelized cost of energy (LCOE) is the all-in price, over a plant's life, that would repay building and running a new one. It decides what gets built.
The marginal cost is what it costs to make one more MWh from a plant that already exists: mostly fuel. It decides what runs each hour, because that is what plants bid into the market.
The two tell different stories. New wind and solar are among the cheapest energy you can build, but they produce only when the weather allows, so LCOE alone undersells the value of firm plants. Lazard's report adds a "cost of firming" analysis for that reason. Meanwhile an old, paid-off gas or nuclear plant is often cheaper to keep running than almost anything new, which is why rising demand has made existing plants more valuable.
Inside the generator
Almost every power plant ends the same way: something spins a magnet inside coils of copper.
A generator's rotor is an electromagnet fed with direct current by the exciter. As the turbine spins it, its field sweeps past three sets of stator windings spaced 120° apart, inducing three alternating voltages, one per phase, a third of a cycle apart. That is three-phase power, and it is why transmission lines come in threes.
3,600 rpmtwo-pole gas and steam generators
1,800 rpmfour-pole machines, common at nuclear plants
~13.8–25 kVtypical generator output voltage
~24,000 Aper phase from a 1,000 MVA unit at 24 kV
The frequency is fixed by the speed: frequency = rpm × poles ÷ 120. A two-pole machine must turn exactly 3,600 rpm to make 60 Hz. Every synchronous generator in an interconnection turns in lockstep with every other one, electrically speaking, which is the subject of Frequency, inertia and torque.
Each generator has two control knobs. The governor sets how much steam, gas or water reaches the turbine, which sets real power (megawatts). The exciter sets the rotor's field strength, which sets voltage and reactive power (see Voltage and reactive power). Big units are cooled with hydrogen gas, which carries heat away better than air and causes less drag.
Those tens of thousands of amps cannot go far: losses grow with the square of the current. So the first thing outside every plant is a transformer.
Transformers: trading current for voltage
The reason the grid is AC at all.
A transformer is two coils wound on a shared iron core. Alternating current in the first coil makes an alternating magnetic field in the core, which induces a voltage in the second. The voltage ratio equals the ratio of turns: a winding with 20 times as many turns gives 20 times the voltage and, since power in equals power out, one-twentieth the current. It has no moving parts, and large units are more than 99% efficient. It works only on AC, which is a big part of why Westinghouse's AC beat Edison's DC.
The voltage ladder
- ~24 kV leaves the generator
- 345–765 kV after the generator step-up transformer (500 kV in the model)
- 69–138 kV subtransmission, after a transmission substation
- 4–35 kV distribution, most often 12.47 kV, after a distribution substation
- 120/240 V at your house, after a pole-top or pad-mounted transformer
The model's wires and particles are coloured by this ladder when Voltages is on. Big power transformers weigh hundreds of tons, are mostly custom-built and often imported, and in the 2020s have had lead times of years, which is one quiet bottleneck on building the grid faster (see Building more).
Transmission lines and losses
Bare aluminum wires on steel towers, moving bulk power hundreds of miles.
High-voltage lines are uninsulated: the air is the insulation, which is why the wires hang on long strings of insulators and keep their distance. Each phase is often a bundle of two to four conductors, which carries more current and reduces the corona discharge that crackles on wet days. The thin wires on top are shield wires that catch lightning.
69–765 kVU.S. AC transmission voltages
~240,000circuit-miles at 230 kV and above (DOE)
~5%of electricity is lost in transmission and distribution (EIA, 2018–2022)
Why voltage cuts losses
A wire turns power into heat at a rate of current squared times resistance (I²R). For a fixed amount of power, current falls in proportion to voltage, so losses fall with the square of voltage. Send the same power at twice the voltage and you lose a quarter as much.
Because losses grow with the square of the flow, they rise faster than demand. In the model, losses are about 5% at average load and climb toward 6–7% at peak. Each extra megawatt at peak costs roughly twice the average loss rate (see Locational prices).
What limits a line
- Heat: current warms the conductor and it sags. Sagging into an untrimmed tree is how the 2003 blackout started (see Northeast blackout). Ratings are lower on hot, still days, exactly when demand peaks.
- Stability and voltage: long AC lines can carry far less than their thermal rating before voltages sag or generators at either end lose step. A useful yardstick is "surge impedance loading": roughly 400 MW for a 345 kV line, 900 MW at 500 kV, 2,200 MW at 765 kV.
- No steering: power on an AC network flows along every parallel path in inverse proportion to impedance, not along the route on a contract, unless there is a DC link (see AC–DC–AC).
Substations: stepping down and switching
The grid's junctions, where voltage changes and where faults get cut off.
A substation is a fenced gravel yard of transformers, breakers, switches and buses. The model's central substation steps 500 kV down to 138 kV, and distribution substations near the towns step 138 kV down to 12.47 kV. There are about 21,500 substations at 100 kV and above in the lower 48, and many more small distribution ones.
- Circuit breakers interrupt fault currents of tens of thousands of amps, typically within three to five cycles (50–80 milliseconds), by stretching and quenching the arc in sulfur hexafluoride gas, vacuum or oil.
- Protective relays watch current and voltage through instrument transformers and decide, in milliseconds, which breakers to open so a fault removes as little of the grid as possible.
- Load tap changers adjust a transformer's ratio in small steps (commonly ±10% in 32 steps) to hold distribution voltage steady as load swings.
- Capacitor banks and reactors add or absorb reactive power to prop voltage up or pull it down (see Voltage and reactive power).
The planning rule behind all this switching is N−1: the system must survive the sudden loss of any single element, any line, transformer or generator, without overloading anything else.
Distribution: the last few miles
Where most outages happen.
Feeders leave the distribution substation at medium voltage. The most common U.S. level is 12.47 kV between phases, which is 7.2 kV from any one wire to ground. Along the street, three-phase mains split into single-phase taps. Every few poles a gray can, a pole-top transformer of typically 25–50 kVA, steps 7,200 V down to the 120/240 V that serves a handful of houses.
12.47 kVmost common distribution voltage (7.2 kV line to ground)
30 : 1pole-top turns ratio: 7,200 V to 240 V
~5.5 millionmiles of U.S. distribution lines (an older DOE estimate)
Distribution is where weather, trees, cars and squirrels meet the grid, and most customer outages start here rather than at power plants. Utilities fight back with reclosers (breakers that try again a few times, since most faults are momentary) and fuses that sacrifice one branch to save the feeder. Rooftop solar has turned many feeders into two-way streets, which older protection schemes weren't designed for.
Try the hurricane: the power plants keep running while the poles come down.
120 volts at the outlet
Your house gets two 120 V wires running opposite each other.
The pole transformer's secondary winding is center-tapped. The middle connects to the neutral (and to ground), and the two ends are the two "hot" legs. Each leg measures 120 V to neutral, and because they swing in opposite directions, leg to leg is 240 V. Lights and outlets use one leg; dryers, ranges, water heaters, heat pumps and EV chargers use both.
- 120 V is an average. It is the RMS value; the actual waveform peaks at about ±170 V, sixty times a second, crossing zero 120 times a second.
- Allowed range: the ANSI C84.1 standard asks utilities to deliver 114–126 V (±5%) at your meter under normal conditions.
- Service size: a typical new house has 200 A service, about 48 kW at 240 V. An average U.S. home uses roughly 850–900 kWh a month, an average draw of about 1.2 kW.
- Why 120 and 60: early U.S. lighting systems settled on roughly 110–120 V and Westinghouse standardized 60 Hz in the 1890s. Most of the world chose 230 V and 50 Hz instead; neither is better enough to be worth switching.
Frequency, inertia and torque
60 Hz is the grid's speedometer: it tells everyone, everywhere, whether supply and demand match.
Electrical load pulls back on every synchronous generator's rotor through its magnetic field, an opposing (counter) torque equal to power divided by rotational speed. The turbine pushes the other way with mechanical torque. When the two balance, the rotor holds speed and the grid holds 60 Hz.
Switch on more load and the counter-torque grows instantly, before any turbine control can react. The rotors slow down, and since every synchronous machine in the interconnection is locked in step, the frequency drops everywhere at once. Lose a generator and the remaining ones each carry a bit more load, with the same effect. Lose load and they all speed up.
Inertia: the shock absorber
Spinning turbines and generators store kinetic energy, typically 2–9 seconds' worth of their rated output (the inertia constant, H). That stored energy feeds the grid in the first seconds after a trip and sets how fast frequency falls:
rate of change of frequency ≈ 60 Hz × power lost ÷ (2 × stored energy)
Wind, solar and batteries connect through inverters and have no inherent inertia. As they take over midday, the same trip makes frequency fall faster. Run the spring scenario at noon, turn off Battery fast response, and trip a unit: the nadir drops well below what the same trip causes on a summer evening. That is the case for "grid-forming" inverters and fast-responding batteries.
What happens after a plant trips
Four layers of defense, each slower and stronger than the last.
| When | What responds | How |
|---|
| 0–5 s | Inertia | Physics: spinning machines give up stored energy. Sets the initial rate of fall. Batteries can now respond inside this window too. |
| 2–30 s | Primary response | Governors open valves in proportion to the frequency error. A typical setting is 5% droop: a 3 Hz drop would call for 100% more output, so 0.1 Hz calls for about 3.3%, outside a small deadband of about ±0.036 Hz. Load also helps a little: motors slow and draw less. This arrests the fall at the nadir. |
| 30 s–10 min | Secondary: AGC | Automatic generation control in the control center computes each area's control error every few seconds and nudges dispatched units until frequency is back to 60.000 and interchange is back on schedule. |
| 10–30 min | Tertiary: reserves | Operators and the market re-dispatch, start quick units and restore reserves so the system can survive the next trip. |
If it isn't enough
If frequency keeps falling, under-frequency load shedding relays disconnect preset blocks of customers automatically. In ERCOT that is 5% of load at 59.3 Hz, 10% more at 58.9 Hz, and another 10% at 58.5 Hz. Shedding some customers on purpose keeps everyone else on. Below about 57.5–58.4 Hz, generators' own protection disconnects them to save their turbines, and a collapse follows. The model uses ERCOT's settings.
Over-frequency is a problem too. When load suddenly disappears, as when data centers jump to backup power, governors must back generators down quickly (see Data centers drop off).
Synchronizing a generator
Matching frequency and phase before closing the breaker: the grid's version of merging onto a highway.
A generator coming online spins up on its own, unconnected. Before its breaker closes it must match the grid in four ways:
- Phase sequence (A-B-C rotation), checked once at commissioning.
- Voltage: within about 0 to +5% of the grid's, slightly high so the unit doesn't absorb reactive power the instant it connects.
- Frequency: a small slip, within about 0.067 Hz, ideally a little faster than the grid so the unit picks up load instead of being motored.
- Phase angle: within about ±10° at the moment of closing (limits from IEEE C50.12/C50.13).
The classic instrument is the synchroscope: a pointer that turns at the slip frequency, clockwise when the generator is fast and counterclockwise when it is slow. The operator or an auto-synchronizer closes the breaker just before 12 o'clock, because the breaker takes a few cycles to make contact.
Close it out of phase and the grid instantly yanks the rotor into step. The torque pulse can exceed what a short circuit at the generator's terminals would produce, enough to crack shaft couplings and twist windings, and large voltage and current surges ride along. That is why synchrocheck relays (device 25 in the standard numbering) block the close when conditions are wrong.
Once connected, the generator can't change its speed: the whole interconnection holds it. Opening the throttle now adds megawatts instead of rpm.
Voltage and reactive power
Frequency is global; voltage is local.
Motors, transformers and long lines need magnetic fields to work, and building those fields every half-cycle means current sloshing back and forth without delivering energy. That sloshing is reactive power, measured in VARs. Too little of it in an area and voltage sags; too much and voltage rises. Unlike real power, reactive power doesn't travel well, so it has to be supplied near where it's needed.
- Generators make or absorb it by adjusting their excitation.
- Capacitor banks supply it; shunt reactors absorb it; static VAR compensators and STATCOMs do both quickly.
- A heavily loaded line consumes reactive power, so voltages sag at peak. A lightly loaded long line produces it (its wires act like a capacitor), so voltages rise. That rise is the Ferranti effect, which matters during black start.
Voltage problems cause blackouts too. Voltage sagging around Cleveland was part of the 2003 cascade. On April 28, 2025, Spain and Portugal blacked out after oscillations and gaps in voltage control led to cascading over-voltage disconnections. Europe's grid operators' expert panel called it the first blackout of its kind in the European system.
How the market clears
Bids, the merit order, and why everyone gets paid the same price.
In the regions run by an independent system operator, every generator submits offers: how many megawatts it will provide at what price, usually based on its marginal cost (fuel × heat rate + variable operations). Every five minutes the operator stacks the offers from cheapest to most expensive (the merit order), finds where the stack meets demand, and dispatches everything to the left of that point.
The last offer needed is the marginal unit, and its offer becomes the clearing price paid to every generator that runs. A nuclear plant that offered −$15 and a combined cycle that offered $28 both get the clearing price. The gap between a plant's cost and that price is what pays back its construction.
Why not pay each plant its own bid?
Because then nobody would bid their real cost. Under pay-as-bid, every plant tries to guess the clearing price and bids just under it. Prices end up about the same, except some guesses are wrong, and the wrong plants run. Uniform pricing makes "bid your true marginal cost" the sensible strategy, and market monitors check offers against cost whenever a unit could move the price.
What the knobs show
- Demand up: the price stays flat through the coal and combined-cycle band, then jumps when peakers are needed, then explodes as reserves run out. The supply curve is a hockey stick.
- Gas price up: most of the stack shifts up, so the price usually follows gas.
- Wind and sun up: zero-cost energy pushes the stack right, so the same demand meets a cheaper unit.
Locational marginal prices
The real market has thousands of prices, not one.
The model treats the grid as a "copper plate", with no limits inside it. Real ISO markets compute a locational marginal price (LMP) at thousands of points: the cost of delivering one more MWh there. It has three parts:
- Energy: the system-wide marginal cost, the same everywhere.
- Congestion: if a line is full, the cheap power on one side can't reach the other, so the far side must use a more expensive local unit, and its price rises.
- Losses: since losses grow with the square of flow, the marginal loss of one more MWh is about twice the average loss. Remote locations pay a little more, or get paid a little less.
Congestion is where much of the money goes. Traders and utilities hedge it with financial transmission rights, which pay out the price difference between two points. Persistent congestion is also the price signal that a new line is needed.
Day-ahead, real-time and reserves
Two markets, a stack of backup products, and one control room.
- Day-ahead market: the day before, hour by hour. It decides which units to commit (turn on), accounting for start-up costs, minimum run times and ramp limits. Most energy is bought and sold here at fixed prices.
- Real-time market: every five minutes. It pays or charges only the differences from the day-ahead schedule, at real-time prices. This "two-settlement" design lets traders place financial "virtual" bids that pull the two prices together.
- Ancillary services: the market also buys capacity held back from producing energy. Regulation follows AGC up and down every few seconds. Spinning reserve is online headroom deliverable within 10 minutes. Non-spinning reserve is offline quick-start units. Newer products, like ERCOT's fast frequency response, pay batteries to react within a fraction of a second.
The model shows one simplified piece of this: the reserves readout counts headroom on running units plus idle quick-start peakers, and when it thins out, the price rises before anyone is shed (see Scarcity).
Scarcity pricing and negative prices
Wholesale power regularly costs less than zero and, occasionally, a hundred times its average.
When it's scarce
As reserves shrink, the risk of having to cut customers off rises, and so does the value of every extra megawatt. ERCOT adds a scarcity adder to prices through an operating reserve demand curve, so prices climb toward the cap before anyone is shed. The model does the same: watch the price below 10% reserves. Caps vary by market. FERC's rule for most ISOs caps offers at $1,000/MWh, or up to $2,000 with cost justification. ERCOT's system-wide cap was $9,000 during Winter Storm Uri and has since been cut to $5,000.
Those spikes are not a malfunction in an energy-only market like ERCOT's: a few hours a year at very high prices are how peakers and new plants earn back their costs. The fear is the opposite problem, prices that never rise enough to pay for reliability (the "missing money"). Most other markets pay for capacity separately (see Capacity markets).
When there's too much
Prices go negative when generators would rather pay than stop. Wind farms earning a per-MWh tax credit, nuclear and coal units that can't cycle cheaply, and hydro at high river flows all do this. Run the spring scenario: windy nights clear at zero or below (−$25 when tax-credit wind sets the price), sunny middays at −$5, and the surplus gets curtailed.
Capacity markets
Paying plants to exist, years before they're needed.
Most ISOs outside Texas run a capacity market: an auction, typically three years ahead, that pays resources per megawatt per day to be available at the peak, whether or not they end up running. PJM, the largest U.S. market (13 states and D.C.), holds a Base Residual Auction. Its results show how fast the grid's economics are changing:
Prices rose more than tenfold between the 2024/25 and 2025/26 delivery years, then hit a negotiated cap three auctions in a row. Load forecasts jumped with data centers, older plants retired, and new supply was stuck in the interconnection queue. The 2028/29 auction (July 2026) cost $16.4 billion and still fell about 6.8 GW short of PJM's reliability target.
Capacity costs flow through to customers in proportion to their share of the system peak, so a higher capacity price shows up on retail bills a year or two later. Try it in Your bill.
Your bill, and how wholesale prices pass through
Wholesale energy is only part of the price of a kilowatt-hour.
Nationally, about 56% of the average price pays for generation, 31% for distribution and 13% for transmission (EIA). How wholesale costs reach you depends on where you live:
- Regulated utilities (much of the Southeast and West): a state commission sets rates in rate cases. The utility earns an allowed return on the plants and wires it has built, and passes fuel and purchased power through at cost via a fuel adjustment clause. It makes no profit on fuel, but a gas price spike reaches bills automatically, often spread over months or years.
- Retail choice states (Texas, Pennsylvania, Ohio, Illinois, much of the Northeast): retail providers buy wholesale power and sell plans. Most plans are fixed-price, because the provider hedges. Default service is often procured by auction, so capacity and energy price changes reach customers on a lag.
- Wholesale passthrough plans charge the real-time price directly. During Winter Storm Uri, Texas customers on such plans were billed thousands of dollars for a few days of power when prices sat at $9,000/MWh, or $9 per kWh. The best-known provider went bankrupt weeks later.
The U.S. average residential price was about 17–18¢/kWh in late 2025, up roughly 5–6% from a year earlier.
Who runs the grid
Nobody, and several thousand organizations at once.
- FERC (Federal Energy Regulatory Commission) regulates wholesale markets and interstate transmission, approves market rules and enforces reliability standards.
- NERC writes and enforces those reliability standards through regional entities, and investigates major events, including the Uri, Elliott and data-center reports cited here.
- Seven ISOs/RTOs (CAISO, ERCOT, SPP, MISO, PJM, NYISO, ISO-NE) operate the grid and run markets for about two-thirds of U.S. demand. Elsewhere, mostly the Southeast and much of the West, vertically integrated utilities dispatch their own fleets, increasingly through regional markets: the Western Energy Imbalance Market since 2014, and CAISO's Extended Day-Ahead Market since May 2026.
- Balancing authorities, dozens of them, each keep their own area's supply, demand and scheduled interchange in balance.
- About 3,000 utilities: investor-owned, municipal and cooperative. State commissions set their retail rates and approve most power lines.
In the control room, operators watch frequency, area control error, line flows and voltages. Software continuously runs "what if" contingency analysis so that the loss of any single element is survivable. The model's control center label shows its frequency.
Three grids, not one
"The grid" is really three big synchronous machines, plus Alaska and Hawaii on their own.
- The Eastern Interconnection, from the Rockies to the Atlantic, plus much of Canada.
- The Western Interconnection, from the Rockies to the Pacific, plus British Columbia, Alberta and part of Baja California.
- ERCOT, most of Texas. It stays separate largely so that power never flows across state lines in sync, which keeps Texas's grid mostly outside federal (FERC) regulation.
Within each interconnection, every synchronous generator turns in lockstep. Between them, the only connections are DC: seven back-to-back converter stations join East and West, about 1,320 MW in total, a sliver of either side's size. ERCOT has two DC ties to the Eastern Interconnection (about 600 MW and 220 MW) and small ties to Mexico. Québec runs its own grid, linked to its neighbors by DC.
The model sits next to a "neighboring grid" across the river on the right, joined only by a DC link. Its label shows its own frequency, wandering independently of ours.
Connecting grids: AC–DC–AC
To join two grids that aren't in step, turn the power into DC and back.
Two AC grids can't be tied together with an AC line unless they are synchronized: at any moment their voltages sit at different points in the cycle, and an AC tie would carry whatever flow that angle difference drives, regardless of anyone's schedule. A converter station rectifies AC to DC, and a second one inverts it back to AC in step with the receiving grid. DC has no frequency or phase, so the two sides never need to agree, and operators set the flow to whatever they want.
Why else use DC
- Distance: DC lines need two conductors instead of three, have no reactive charging current, and lose less over long distances. Past roughly 400–500 miles overhead (or about 30–50 miles of cable), DC usually wins despite each converter station losing around 1%.
- Control: a DC link's flow is a set-point, so it can't be overloaded by a disturbance elsewhere, and it acts as a firewall that keeps a cascade from crossing.
- Two technologies: line-commutated converters (thyristors) handle the biggest transfers but need a strong AC grid on both ends. Voltage-source converters (IGBTs) can control voltage and even help black-start a dead grid; they're the choice for offshore wind.
U.S. examples: the Pacific DC Intertie, about 846 miles at ±500 kV and roughly 3 GW, from Oregon to Los Angeles since 1970; SunZia, about 3 GW from New Mexico wind to Arizona over roughly 550 miles, fully operating in 2026; and the Champlain Hudson Power Express, 1,250 MW of buried and underwater cable from Québec to New York City, which entered service in 2026.
The duck curve and storage
Plot demand minus wind and solar over a day in a sunny place and you get a duck.
The belly is midday, when solar covers much of demand; the neck is the evening, when the sun sets just as people come home. What's left over, the net demand, is what every other plant has to cover. In California it can climb by roughly 20 GW in the three hours around sunset.
Batteries flatten the duck: they charge in the belly and discharge up the neck. Toggle Storage on the chart, or switch the scenario to Spring or Data-center boom (2030). Click the chart to jump the clock.
Black start: restarting a dead grid
Almost every power plant needs electricity to make electricity. So how do you start from zero?
A big steam plant's pumps, fans, coal mills and controls draw several percent of its output, so after a total blackout it can't restart on its own. Restoration starts from a handful of designated black-start units, usually hydro or small gas turbines with diesel or battery starters, increasingly batteries with grid-forming inverters. These are tested regularly: NERC requires each one to prove it can start without outside power at least once every three years.
What makes it hard
- Small islands are fragile. With little inertia, every block of load causes a big frequency dip, so load comes back a feeder at a time.
- Cold-load pickup: after hours without power, every thermostat calls for heat or cooling at once, so restored feeders draw several times their normal load.
- Voltage: long, empty lines push voltage up, and energizing a big transformer draws a surge of current, so reactors and careful switching come first.
- Nuclear last: reactors need stable off-site power before restarting.
Real examples: after the 2003 blackout, some areas waited up to four days. After the April 2025 Iberian blackout, Spain's operator restored most demand by the next morning, using hydro plants and links to France and Morocco. Pick Black start from the scenarios, or step through it above, to watch the model come back one island at a time.
Typical days
Summer: demand bottoms out around 4 a.m., climbs with the heat, and peaks around 4–6 p.m. as air conditioners work hardest. Solar covers much of the afternoon; gas peakers and batteries cover the evening. Winter: two peaks, one when people wake up and turn on heat and lights, another after dark. The sun sets before the evening peak.
Things to try: drag the demand knob up 20–40% and watch the marginal unit, price and reserves; switch on Voltages; open How the market clears while it runs.
Heat wave: California, August 2020
A heat dome covered the West in mid-August 2020. On August 14 and 15, as the sun set and solar output fell while demand stayed near its peak, the California ISO ran out of reserves, declared a Stage 3 emergency and ordered rotating outages. About 800,000 customers lost power, for up to a couple of hours each, the state's first rotating outages since the 2001 energy crisis. The official root-cause analysis blamed extreme heat beyond what planners had assumed, resource planning that hadn't kept pace with the evening "net peak", and some market practices that made a tight situation worse.
Two years later, on September 6, 2022, California set a record peak of about 52 GW and got through without rotating outages, helped by several gigawatts of new batteries, demand response and an emergency text alert asking people to conserve.
In the model, the heat wave starts with 2020-era storage. When reserves run out after sunset, use the story card's Add today's batteries toggle and run the evening again.
Winter storms: Uri (2021) and Elliott (2022)
Uri, February 2021. Arctic air covered Texas for days. Demand for heat set winter records just as the cold froze wellheads, gas lines, instruments and turbines. FERC and NERC found that 61,800 MW of generation was unavailable at the worst point, with freezing equipment the leading cause and gas supply problems close behind. Every type of plant had failures, gas the most by volume. ERCOT shed up to 20,000 MW of load. More than 4.5 million homes and businesses lost power, many for days in sub-freezing cold, and Texas officials counted 246 deaths.
At about 1:51 a.m. on February 15, frequency fell to 59.302 Hz and stayed below 59.4 Hz for roughly four and a half minutes. Nine minutes there would have tripped generators off by their own protection and blacked out the entire state, which would have taken days or weeks to restart. Operators shed load fast enough to stop it. Prices then sat at the $9,000/MWh cap for days.
Elliott, December 2022. The same failure mode, bigger and further east: 90,500 MW of coincident unplanned outages across the Eastern Interconnection, 13% of expected resources. TVA shed up to 3,000 MW for about eight hours on Christmas Eve, and Duke Energy up to about 2 GW. Most failing units were operating above the minimum temperatures they had been rated for.
The eclipse of April 8, 2024
The path of totality crossed Texas in the early afternoon. ERCOT lost about 8.9 GW of solar output as the moon covered the sun; between 1 and 2 p.m., gas plants added about 6.2 GW while coal and batteries added about 0.8 GW each. Then the sun came back and solar ramped up faster than any thermal plant could ramp down.
It was a rehearsal for a normal day: the same ramp happens every evening, just more slowly, and the grid handled it because the event was forecast to the minute.
Hurricane Beryl, July 2024
Beryl came ashore near Houston as a Category 1 hurricane. Its winds and falling trees took down poles and lines across the region, and about 2.26 million CenterPoint customers lost power, out of roughly 2.8 million. Many waited more than a week, in July heat. Generation and the high-voltage grid were largely fine. The damage was in the distribution system, which is where most storm outages happen.
Notice that prices stay low: plenty of supply, and nowhere to send it.
The Northeast blackout of August 14, 2003
A hot afternoon in Ohio. The regional overseer's state estimator was down. A FirstEnergy alarm system failed silently, so operators didn't know their screens had stopped warning them. Then three 345 kV lines, heavily loaded and sagging in the heat, touched overgrown trees and tripped, one after another. Each trip pushed more flow onto the remaining lines. At 4:05:57 p.m. the Sammis–Star line tripped on overload, and within about seven minutes a cascade tripped more than 508 generating units at 265 power plants.
About 50 million people in eight states and Ontario lost 61,800 MW of load. Some U.S. areas waited four days. The joint U.S.-Canada task force blamed inadequate situational awareness, poor tree trimming and inadequate diagnostic support. Its recommendations led to mandatory, enforceable reliability standards, which is why NERC can levy fines today.
Losing a big power plant
Large units trip several times a year somewhere on every interconnection, and almost nobody notices. Planners size reserves so that the largest single loss is survivable. In the model, a 1.24 GW nuclear unit trips at 4:30 p.m.: frequency dips, governors catch it, AGC brings it back to 60.000 Hz within a minute or two, and the market re-dispatches around the hole.
When data centers drop off all at once
On July 10, 2024, a lightning arrester failed on a 230 kV line in Virginia. The fault cleared normally, but about 1,500 MW of data centers, spread across 60 connection points, sensed the voltage dip and switched to their own backup power within seconds. No utility equipment disconnected them; their own protection did. NERC's review called it a new kind of risk: planners study sudden losses of generation, not sudden losses of gigawatts of load.
It has happened again. On July 22, 2026, a fault in the same region set off a load transfer of about 3,800 MW in two waves, the largest such event in PJM's history, and PJM's total load fell 3.8% in moments. Grid operators are now writing ride-through requirements for large loads.
Spring: too much of a good thing
Spring brings mild weather (low demand), strong wind, strong sun, high rivers, and the season when nuclear and gas plants schedule maintenance. In windy and sunny regions such as SPP's plains and California, prices regularly go negative and operators curtail wind and solar they can't use. This scenario doubles the model's wind and adds 60% more solar to resemble those regions.
A data-center boom
Ten gigawatts of always-on load arrive in the model (about 16% more energy), along with 60% more solar, more than twice the batteries, and 30% fewer coal plants. Prices settle higher all day, reserves are thinnest just after sunset, and that evening hour is what capacity markets are now paying for.
Data centers and load growth
After fifteen flat years, U.S. electricity demand is growing again.
4.4%of U.S. electricity used by data centers in 2023 (176 TWh)
6.7–12%projected by 2028 (325–580 TWh, LBNL for DOE)
+83 TWhgrowth in U.S. retail sales in 2025
AI training and inference, along with manufacturing and electrification, are driving the first sustained demand growth in decades. Data centers are concentrated (Northern Virginia, Texas, Ohio, Arizona), run around the clock, and want power faster than plants and lines can be built. The strain shows up in capacity prices (PJM), in the interconnection queue, in tech companies' contracts to restart or keep nuclear plants running, and in new rules on how large loads must behave during grid disturbances (see Data centers drop off).
The interconnection queue
Thousands of projects waiting for permission to plug in.
Before a new plant can connect, the grid operator must study what it would do to the grid and what upgrades it would need, and someone must pay for them. At the end of 2024 about 2,290 GW of generation and storage was waiting in U.S. queues, more than all existing U.S. capacity. Most will never be built: of the capacity that applied from 2000 to 2019, only 13% had reached operation by the end of 2024. The median wait from application to operation has roughly doubled, to more than four years. FERC's Order 2023 moved operators to batch "cluster" studies, and the queue shrank to about 2,061 GW by the end of 2025.
Building more: lines, turbines, transformers
- Transmission: only a few hundred miles of new high-voltage lines get finished in a typical recent year, held back by siting, permitting and arguments over who pays. FERC's Order 1920 (2024) requires long-term regional planning, and a few big DC projects are finally finishing (see AC–DC–AC).
- Equipment: large transformers and gas turbines are backordered for years. New gas plants announced today may not run until the end of the decade.
- Getting more from what exists: dynamic line ratings that raise limits on cool, windy days; advanced conductors that carry more current on the same towers; and grid-enhancing software that reroutes flow around congestion.
- Policy churn: the 2025 tax law phases out credits for wind and solar while keeping them for storage, geothermal and nuclear, and the Department of Energy has used emergency orders to keep some fossil plants open.
About the model
A teaching model, not a forecast. Everything runs in your browser (media/powergrid/sim.js).
- Size and fleet: one island grid, about an eighth of the lower 48: 12.4 GW nuclear, 20.5 GW coal, 35 GW combined cycle, 29.5 GW of gas and oil peakers, 9.6 GW hydro, 19 GW wind, 25 GW solar, 7.8 GW (46.8 GWh) of storage, and small amounts of geothermal and biomass. Its average demand is 62 GW. Over a representative year it produces about 39% gas, 17% nuclear, 17% coal, 10% wind, 9% solar and 5.5% hydro, close to 2025's actual U.S. mix.
- Market: a uniform-price auction every frame. Offers are heat rate × fuel price + variable cost (+ start-up costs for peakers). Nuclear and part of coal are must-run; tax-credit wind bids −$25; solar bids −$5. A scarcity adder modelled on ERCOT's lifts prices as reserves fall below 10%. The cap is $5,000/MWh ($9,000 in the Uri replay).
- Hydro and storage: scheduled a day ahead. Hydro spends a daily water budget on the highest net-demand hours; storage buys low and sells high when the spread beats its round-trip loss.
- Frequency: the aggregate swing equation, with inertia from whichever synchronous units are running, 5% governor droop with a ±36 mHz deadband, load damping, battery fast response, AGC and re-dispatch, and ERCOT-style under-frequency load shedding. The Uri and 2003 frequency traces are replays of published values, compressed in time.
- Losses rise with the square of load, about 5% on average.
- Left out: transmission limits inside the grid (prices are the same everywhere), unit commitment and ramp limits, imports, voltage and reactive power, and much else. Real grids are messier.
Sources
- U.S. Energy Information Administration: Electricity generation, capacity and sales; Electric Power Monthly; T&D losses; 2025 record peak demand; 2025 natural gas prices; battery storage growth; the 2024 eclipse; electricity prices.
- Full-year 2025 generation by source, compiled from EIA data: Wolf Street, February 2026.
- Lazard, Levelized Cost of Energy+ (LCOE v19.0, July 2026).
- Nuclear Energy Institute, Nuclear Costs in Context (2023 data).
- FERC and NERC: February 2021 cold weather outages (Uri); Winter Storm Elliott report; NERC, Incident Review: large load loss (2025).
- U.S.-Canada Power System Outage Task Force, Final Report on the August 14, 2003 Blackout.
- PJM: 2027/2028 and 2028/2029 Base Residual Auction reports; the July 2026 large load transfer.
- Lawrence Berkeley National Laboratory: Queued Up, 2025 edition; data center energy use (2024, for DOE).
- NREL, Interconnections Seam Study (East–West DC ties).
- ENTSO-E, Expert Panel report on the 28 April 2025 Iberian blackout.
- CAISO, EDAM launch; Hitachi Energy, SunZia inauguration.
- Synchronizing limits (IEEE C50.12/C50.13) as summarized by Schweitzer Engineering Laboratories.