The grid added a record 58 gigawatts. The risk NERC just raised to its top alert tier can't be measured in gigawatts at all.

Two numbers from the same regulator, two weeks apart, look like they should cancel out. On 19 May 2026 the North American Electric Reliability Corporation, NERCNERCThe North American Electric Reliability Corporation, the federally overseen nonprofit that writes and enforces the mandatory rules for keeping the bulk power grid reliable., the federally overseen body that writes the mandatory rules for keeping the power grid reliable, published its 2026 Summer Reliability Assessment and led with a record: more than 58 gigawatts of new generation had been added across North America since the summer before. That is a lot of power. A gigawatt is a billion watts, roughly the output of a big power plant; 58 of them is enough to serve tens of millions of homes.

Two weeks earlier, on 4 May 2026, the same NERC had done something it rarely does: it issued a Level 3 'Essential Actions' alertLevel 3 'Essential Actions' alertThe most serious of NERC's three alert tiers, requiring utilities to formally acknowledge it and file substantive responses on required actions by a set date., the most serious of its three warning tiers. The subject was AI data centers that, after an ordinary faultfaultA brief electrical problem on a power line, such as a lightning strike causing a momentary short circuit; the grid is designed to detect and clear faults in a fraction of a second. on the grid, had been disconnecting themselves in unison, shedding more than a gigawatt of demand in seconds to protect their servers. NERC called it a serious risk to the bulk power system and gave utilities a deadline to respond.

Put the two side by side and the mind reaches for subtraction: a record surplus of supply against a reliability worry, surely the first buys down the second. Our analysis finds that it does not, not a little, but essentially not at all. The 58 gigawatts and the Level 3 alert are measured in different dimensions. One answers whether there is enough electricity. The other answers what the grid does in the split second after a fault. They are orthogonalorthogonalA math word for 'at right angles.' Two quantities are orthogonal when moving one does not move the other; here, adding gigawatts of supply does not change the millisecond stability problem., at right angles, and treating the build as a cushion against the alert is the kind of mistake that turns a reassuring headline into a blind spot.

The confusion is understandable, because both numbers are about 'reliability.' But grid engineers split that word into two questions that almost never touch.

The first is resource adequacy: over the coming hours, days and seasons, is there enough total generating capacity to meet demand, including on the hottest afternoon when air conditioners are all running at once? This is a bookkeeping question, add up the supply, add up the peak demand, check that supply wins with margin to spare. The 58 gigawatts is an adequacy number. It makes the summer's supply-versus-demand ledger look better, which is exactly what NERC said it does.

The second question is transient stability: in the milliseconds to seconds right after a sudden disturbance, a lightning strike, a short circuit, a line tripping, does the grid stay balanced, or does one shock cascade into another? This is not a bookkeeping question but a physics-of-the-instant question. It depends on how machines and electronics behave in the moment, not on how much capacity sits on the books. The Level 3 alert is a transient-stability problem.

Adequacy asks 'is there enough electricity?' Stability asks 'what happens in the first half-second after a fault?' Adding gigawatts answers only the first question.

This is why the two do not net out. You can have a comfortable surplus of supply and still fail catastrophically in the first half-second after a fault, and you can be short on capacity and yet ride cleanly through every disturbance. The 58 gigawatts changes the size of the reservoir. The alert is about the plumbing's reaction to a sudden knock.

If the data-center problem were really a supply shortage in disguise, the grid's own instruments would say so. They say the opposite. The clearest evidence sits in a single measured number from the event that anchors NERC's whole concern.

Grid frequency, the 60-cycles-per-second rhythm of alternating current, is the grid's speedometer for supply and demand. When generation and load are perfectly matched, it sits at 60.000 hertz. When there is too little power for the load, frequency sags below 60. When there is too much power for the load, frequency climbs above 60. Direction tells you which side is short.

On 10 July 2024, when about 1,500 megawatts of Northern Virginia data-center load abruptly switched itself off, grid frequency did not fall. It rose, to 60.047 hertz, and grid operators had to scramble to shed surplus power before the over-frequency itself caused problems. Over-frequency is the fingerprint of a momentary surplus, suddenly there was more generation than load, because the load had walked out of the room. That is the precise opposite of the shortage that building 58 gigawatts is designed to cover. Adding still more generation to a grid that just lurched into surplus does not help; if anything it is more of the thing that already had to be dialed back.

This is the crux of the orthogonality. The alert describes a demand-side reflex, loads bailing out, that shows up as too much supply, not too little. No amount of new supply addresses a problem whose signature is surplus.

It is worth walking through the July 2024 event slowly, because it is the one case NERC has taken apart fault-by-fault in a public incident review, and it shows exactly why more gigawatts are beside the point.

Around 7 p.m. Eastern, a lightning arrester, a protective device, failed on a 230-kilovolt transmission line in the Eastern InterconnectionEastern InterconnectionOne of North America's three big synchronized grids, covering roughly the eastern two-thirds of the US and Canada; Northern Virginia sits inside it., the large synchronized grid covering the eastern two-thirds of the continent. The failure created a fault. The line's automatic reclosing controls, designed to test whether a fault has cleared by briefly re-energizing the line, cycled several times. The result was six successive faults in an 82-second window, each lasting between 42 and 66 milliseconds. A millisecond is a thousandth of a second; 42 to 66 of them is faster than a blink, repeated six times. The protection system did its job and cleared each fault correctly. Voltage in the affected area briefly sagged to roughly a quarter to two-fifths of normal during the dips.

That should have been a non-event. Instead, about 60 data centers sensed the voltage dips and did what their internal protection is built to do: they flipped to backup power to shield their servers, disconnecting roughly 1,500 megawatts of load from the grid all at once. The grid had planned, for a century, around losing big generators suddenly. It had not planned around losing a mid-sized city's worth of demand in an instant. Frequency jumped to 60.047 hertz, and PJM and the local utility, Dominion, had to pull back generation and switch out capacitor banks, equipment that props voltage up, to bring the system back to 60.000 hertz over about four minutes.

NERC's aggregate reporting counts several more of these simultaneous load-rejection events across the Eastern and Texas grids since 2022. The chart above shows the five largest for which sizes are documented. Note two things. First, even the biggest, at 1,500 megawatts, is under three percent of the 58-gigawatt build, the two live on different scales entirely. Second, the trigger in every case is a fault, not a shortage. These loads did not leave because the grid ran out of power. They left because they saw a voltage dip and chose to protect themselves.

There is a subtler version of the 'the build helps' argument, and also a subtler version of its opposite. Both deserve a fair hearing, because the honest answer is 'mostly neither.'

The stabilizing thing a power plant can offer in the first instants after a fault is inertiainertiaThe stabilizing effect of heavy spinning generators that momentarily resists changes in grid frequency, buying operators time to respond to a disturbance., the physical resistance to sudden frequency change that comes from heavy spinning machines like gas and steam turbines. A grid with lots of inertia changes frequency slowly, buying operators time. Solar panels and batteries connect through power electronics called inverters, which have no spinning mass and, in their conventional form, add little inertia on their own. So the composition of the build matters more than its size.

NERC breaks out 39.4 of the 58-plus gigawatts by type: 16.4 GW solar, 14.7 GW battery, 6.7 GW gas, 1.6 GW wind. Only the 6.7 gigawatts of gas is the inertia-providing, spinning-machine kind, a sliver of the build. The 32.7 gigawatts of solar, battery and wind is inverter-based. The remaining roughly 18.6 gigawatts sits in an unlabeled bucket, so we assign it to neither camp.

You could read that chart as 'the record build is mostly inverters, so it adds brittleness.' Resist it. Batteries here are grouped by how they connect, not by how they behave. A growing share of grid batteries use grid-forming inverters, a newer design that can actively set voltage and frequency and supply fast, synthetic inertia, making them a stabilizing resource, not a passive one. So the 32.7-gigawatt 'inverter' bar overstates the genuinely destabilizing share. The build is not making the fault problem meaningfully worse, and its small slug of new gas is not making it meaningfully better. It is, again, orthogonal.

Even the 6.7 GW of new spinning gas does not stop a data center from tripping itself offline. Inertia slows a frequency swing; it does not change a load's decision to bail out.

Step back and the logic is clean. The failure mode in the alert is a large load choosing to disconnect when it senses a disturbance. The cure has to change that choice, to make the load stay connected and ride through the dip. Supply sits on the other side of the ledger entirely. You can pile generation as high as you like; it does nothing to a data center's internal protection setting that says 'voltage dipped, switch to backup.'

This is what 'orthogonal' means in practice, and why netting the two numbers is a category error. The 58 gigawatts moves the adequacy axis. The alert lives on the stability axis. Moving one leaves the other exactly where it was. Reading the record build as insurance against the alert is not conservative; it is a way of feeling covered against a risk you have not touched.

If this pattern feels familiar, it should, the grid has lived through a near-identical story on the supply side, and how long the fix took is the most sobering part of this analysis.

On 16 August 2016, the Blue Cut wildfire in Southern California caused a normal, correctly cleared fault on a 500-kilovolt line. In response, about 1,200 megawatts of solar generation stopped injecting power, not because anything broke, but because the inverters' protection settings told them to briefly stop, a behavior called momentary cessationmomentary cessationA self-protective behavior in which an inverter briefly stops injecting power during a voltage dip, which can turn a normal fault into a large loss of generation.. On 9 October 2017 the Canyon 2 fire produced the same pattern, dropping roughly 900 megawatts of solar during a fault. In both cases, as in Northern Virginia, ordinary electronics turned a routine fault into a large, sudden loss, just on the generation side instead of the demand side.

The remedy was a rule forcing inverter-based generators to ride through disturbances instead of dropping out: NERC standard PRC-029-1. FERCFERCThe Federal Energy Regulatory Commission, the US agency that approves NERC's reliability standards and makes them enforceable., the federal agency that makes NERC's standards enforceable, approved it through Order 909 on 24 July 2025, and it becomes enforceable for bulk-system inverter resources on 1 October 2026. Count the years: about nine from the first fire to an approved standard, and a tenth to enforcement. Writing a ride-through rule, testing it, and pushing it through the federal process is slow, deliberate work, even when everyone agrees on the problem.

The demand-side version of that rule does not exist yet. It is being written now, and the same slow clock has just started ticking.

On 18 March 2026, NERC's Standards Committee accepted the request that launched Project 2026-02Project 2026-02NERC's standards project, launched March 2026, drafting the first rules for how large computational loads (data centers above 20 MW) must behave during grid faults., the first effort to write ride-throughride-throughThe ability of a generator or large load to stay connected and keep operating through a brief voltage or frequency dip instead of disconnecting. and modeling rules for large computational loads, data centers of 20 megawatts and up. The target is a standards filing by 31 December 2026. That is the demand-side twin of PRC-029-1PRC-029-1The NERC reliability standard requiring inverter-based generators to ride through voltage and frequency disturbances, approved by FERC in 2025 and enforceable from October 2026.: a rule to make big loads stay connected through a fault rather than bail out. It is early, and it will move through the same multi-year gauntlet its supply-side predecessor did.

In the meantime, the Level 3 alert is doing the fast, blunt work that standards cannot do quickly. It ordered utilities to acknowledge the alert by 11 May 2026 and to report on their required actions by 3 August 2026, improving visibility into how these loads behave, installing recorders to capture their performance during disturbances, and coordinating settings. Those are the near-term patches while the real rule is drafted.

The fix for the alert is a rule about how loads behave in a fault, not a pile of new gigawatts. That rule is now being written, on a clock that historically runs close to a decade.

So the answer to our question is a clean no. None of the record 58-plus gigawatts reduces the risk in the Level 3 alert. The two numbers measure different failure modes, adequacy and transient stability, in incompatible dimensions. What to watch is not the next capacity record. It is Project 2026-02's 31 December 2026 filing and the 3 August 2026 utility responses, because those, not gigawatts, are where the actual risk gets bought down.