Summer Peak Demand 2026: Can the US Grid Handle AI Plus Air Conditioning

The Pulse

The North American Electric Reliability Corporation released its 2026 Summer Reliability Assessment on May 19, confirming that every region it monitors has adequate resources for a normal summer.

That headline sounds reassuring. The detail underneath it is not. Aggregated peak demand across NERC’s assessment areas has grown by more than 11 gigawatts since the Summer 2025 projection, a faster increase than the 10 gigawatts seen heading into last year.

The grid added 58 gigawatts of new capacity in twelve months, the largest single-year buildout on record. Data center demand grew faster than that buildout could comfortably absorb. The 2026 story is not whether the grid survives summer. It is how thin the margin has become.

Core significance

Why it matters:

• NERC flagged four regions at elevated risk this summer, down from six last year, but the risk has not disappeared, it has concentrated:  NPCC New England, MRO SaskPower, WECC Northwest, and a section of west Texas face potential shortfalls if extreme heat coincides with low wind or hydropower output.[NERC 2026 Summer Reliability Assessment official PDF]

NERC director John Moura was explicit on the assessment’s release webinar that improved conditions should not be read as declining risk, since demand growth from data centers and electrification is outpacing even record resource additions.

• US data center power demand is on a path to more than double in two years:  Goldman Sachs Research projects demand climbing from 31 gigawatts in 2025 to 41 gigawatts in 2026 and 66 gigawatts in 2027.[Goldman Sachs US data center power demand 2027]

Data centers’ share of total US peak summer power demand is projected to rise from 4.1% in 2025 to 5.3% this year and 8.5% by 2027, a shift large enough on its own to move regional capacity markets.

• A 5% interconnection slowdown helped, not hurt, near-term grid stability:  Several NERC assessment areas revised their 2026 load forecasts downward from mid-2025 projections specifically because data centers are coming online slower than developers originally announced.[Utility Dive NERC interconnection delays demand forecasting]

That slower pace bought the grid breathing room this year. It does not change the trajectory. The same megawatts are still coming, just a few quarters later than the most aggressive announcements suggested.

Deep Context: How two decades of flat demand became this

US electricity demand grew at well under 1% annually for roughly twenty years before 2023. Utilities planned, built, and priced power on the assumption that this slow growth would continue indefinitely.

That assumption broke specifically because of data centers. ERCOT, the Texas grid operator, now projects peak summer demand could approach 145 gigawatts by 2031, up from 85 gigawatts in 2024, with over half of the new demand coming from data centers and cryptocurrency miners.[Belfer Center AI data centers US electric grid watershed]

Virginia tells a similar story at the state level. Dominion Energy’s resource plan calls for nearly 27 gigawatts of new generation by 2039, split between 21 gigawatts of renewables and small modular reactors and 5.9 gigawatts of natural gas, driven almost entirely by data center growth in Northern Virginia’s Data Center Alley.

NERC’s own summer snapshot frames this same tension as a shoulder season problem as much as a peak season one: early summer heat is increasingly colliding with the spring maintenance windows utilities rely on to take generation offline for repairs, narrowing the margin for error even before the hottest weeks of July and August arrive.[NERC 2026 Summer Reliability Assessment snapshot record growth]

Why some data centers are actually helping the grid this summer

One counterintuitive finding in NERC’s report deserves more attention than it has received. In parts of Texas, data center operators are voluntarily curtailing their own load during peak demand windows, in exchange for lower electricity rates the rest of the time.[NRDC solar AC summer NERC takeaways data center flexibility]

NERC explicitly credited this demand flexibility with lowering Texas’s near-term peak forecast. Not every data center can do this, training workloads are far less flexible than this kind of voluntary curtailment requires, but it shows that the AI-versus-AC framing is not purely zero-sum if operators are willing to shift load.

As covered in our AI data center power consumption report, the gap between training load, which can sometimes pause, and inference load, which increasingly cannot, is becoming the central technical question in how much flexibility AI operators can realistically offer grid operators during a heat emergency.

Data Insights

By the numbers:

All figures from NERC official assessments, Goldman Sachs Research, and the US Energy Information Administration cited inline.

• 58 gigawatts:  New generation capacity added to the North American bulk power system in the twelve months before the 2026 summer assessment, the largest single-year addition on record, composed of 16.4 gigawatts solar, 14.7 gigawatts battery storage, 6.7 gigawatts natural gas, and 1.6 gigawatts wind.
• 127 gigawatts:  MISO’s forecast peak internal demand for summer 2026, up more than 2.6 gigawatts from the same projection made in 2024, with data centers identified as the largest single contributor to that accelerated growth.[Electric Coop NERC generation growth summer strain Moura]

MISO expects that figure to climb further to 143.7 gigawatts by 2035, a trajectory almost entirely driven by large load interconnection requests rather than population growth or general electrification.

• 4,250 billion kilowatt hours:  The EIA’s projected total 2026 US electricity consumption, up from roughly 4,097 billion kilowatt hours in 2024, with both 2025 and 2026 setting consecutive new annual records.[Enline Energy AI data center grid capacity 2026 EIA IEA]

The International Energy Agency’s parallel global projection shows data center electricity consumption alone roughly doubling from about 415 terawatt hours in 2024 to around 945 terawatt hours by 2030.

• Four at-risk regions, down from six:  NERC’s 2026 assessment narrowed its elevated-risk list to NPCC New England, MRO SaskPower, WECC Northwest, and western Texas, an improvement attributable directly to the 58 gigawatts of new resources rather than reduced demand pressure.

Table 1: Regional grid operators facing the steepest summer 2026 load growth

Grid operator	2026 summer peak forecast	Primary driver	NERC risk status	Trend to 2031 plus
ERCOT (Texas)	94,650 MW	Data centers and crypto mining	Elevated risk, west Texas only	154,077 MW by 2035
MISO (Midwest)	127,000 MW	Large load interconnections, data centers	Normal risk, watching closely	143,700 MW by 2035
PJM (Mid-Atlantic)	Not separately disclosed here	Northern Virginia Data Center Alley growth	High risk beginning 2029	Capacity prices already up over 10x
NPCC (New England)	Not separately disclosed here	Limited generation, weather extremes	Elevated risk this summer	Ongoing concern, less data center driven

Table 2: AI data center load versus residential air conditioning load

Dimension	AI data center load	Residential AC load
Load pattern	Continuous, near-constant draw for training and inference	Sharp afternoon and evening peaks tied to temperature
Flexibility	Training can sometimes pause, inference increasingly cannot	Largely inflexible once heat index crosses comfort threshold
Growth driver	New facility construction and chip density increases	Population growth and rising average temperatures
2026 to 2027 trend	Share of peak demand projected to nearly double	Relatively stable, grows with extreme heat frequency

The Business Case: What this Means for Enterprises Running AI Workloads

For enterprises operating or contracting AI compute in any of the four NERC-flagged regions, summer 2026 carries genuine operational risk, not just a public relations talking point.

West Texas data center operators in particular should expect occasional curtailment requests during extreme heat events, the same flexibility NERC praised in Texas this year is increasingly likely to become a contractual expectation rather than a voluntary courtesy as 2027 demand arrives.

Enterprises negotiating new data center capacity contracts in PJM territory specifically should treat the region’s already-elevated capacity market prices as a preview. PJM’s capacity clearing price for the 2026 to 2027 delivery year rose to 329.17 dollars per megawatt, more than ten times the 28.92 dollar price just two years earlier, with data center growth named as a primary driver.[arXiv electricity demand grid impacts AI data centers PJM capacity]

As covered in our nuclear AI power plants report, several hyperscalers are responding to exactly this kind of capacity price pressure by locking in dedicated nuclear generation rather than competing for grid capacity during peak summer windows.

What to ask before signing a new data center power contract

Enterprises evaluating new AI infrastructure capacity in 2026 should treat grid risk as a procurement question, not just an engineering one. The four NERC-flagged regions are not abstract, they map directly onto where new data center capacity is actually being built.

A useful first question for any colocation or cloud provider is whether their facility sits inside one of the four elevated-risk zones, and if so, what curtailment obligations are written into their own interconnection agreement. A provider that has already agreed to demand flexibility with its utility is a fundamentally different risk profile than one that has not.

A second question worth asking directly is how much of the facility’s committed capacity is firm versus interruptible. The gap between those two categories is where summer reliability risk actually lands, and it is rarely disclosed unless a customer specifically asks for it during contract negotiation.

Expert Nuance: Training and inference strain the grid differently

The technical distinction between training load and inference load matters more for summer grid planning than most coverage acknowledges. Training runs are burst workloads, intense, periodic, and largely predictable months in advance.

Inference load behaves nothing like that. It is continuous, globally distributed, and driven by real-time user activity, producing short, unpredictable bursts layered on top of a high sustained baseline. Inference now accounts for 80 to 90% of total AI compute load across major vendors, a structural shift that makes AI demand look increasingly like the inflexible residential AC load it sits next to on the grid, rather than the pausable industrial load utilities originally expected.

That shift is the real story behind the AI-versus-AC framing. A training cluster could theoretically power down during a heat emergency. An inference fleet serving live ChatGPT, Claude, or Gemini queries across millions of users cannot, without breaking the product millions of people are actively using at that exact moment.

Strategic outlook

1. Watch whether winter risk overtakes summer risk in planning priority:  NERC’s January 2026 long-term assessment found that for many regions, winter load growth now outpaces summer load growth for the first time in decades, reversing a pattern that has held since air conditioning became widespread.[PureSky Energy NERC January 2026 long term reliability winter]

Thirteen of twenty-three NERC assessment areas may face capacity shortfalls within the next decade on current trends. If winter electrification and AI demand continue compounding together, utilities planning purely around summer AC peaks will be solving yesterday’s problem.

2. Capacity market prices will keep functioning as the real-time signal to watch:  PJM’s tenfold capacity price increase already happened. Similar early pressure is emerging in New York ISO and ERCOT forward markets as data center expansion gets priced in ahead of physical construction.
3. Demand flexibility programs will likely shift from voluntary to contractual:  Texas showed in 2026 that data center curtailment during peak events genuinely helps. Expect grid operators in other constrained regions to start requiring similar flexibility commitments as a condition of new interconnection approval, rather than treating it as a bonus.

Key question answered

Can the US grid handle AI data centers and summer air conditioning demand at the same time in 2026?

Yes, under normal conditions, but with a meaningfully thinner margin than a year ago. NERC’s 2026 Summer Reliability Assessment confirms all monitored regions have adequate resources for normal peak demand, after the grid added 58 gigawatts of new capacity, the largest single-year increase on record.

Four regions, NPCC New England, MRO SaskPower, WECC Northwest, and western Texas, face elevated risk specifically during extreme heat combined with low wind or hydropower output. Data center demand is the primary driver of accelerated load growth, with US data center power demand projected to climb from 31 gigawatts in 2025 to 66 gigawatts by 2027, and its share of summer peak demand rising from 4.1% to 8.5% over the same period. The risk is not a 2026 blackout. It is the trajectory beyond 2027, when the margin NERC describes today disappears under current growth rates.

The Takeaway

Summer 2026 will almost certainly pass without the widespread blackouts that headlines about AI power demand sometimes imply. NERC’s own assessment is clear that normal conditions are manageable this year.

The more useful number in the entire report is not this year’s adequacy finding. It is the 11 gigawatt year-over-year jump in aggregated peak demand forecasts, growing faster than the prior year’s already-elevated pace, even as 58 gigawatts of new capacity came online. The grid is adding resources at a record rate and still falling further behind the curve of its own demand growth.

For enterprises planning AI infrastructure investment in 2027 and beyond, the message from this summer’s data is not that the grid is fine. It is that the grid is fine for now, in most places, because of exceptional one-year resource growth that utilities cannot realistically repeat every single year. The regions where that growth slows, or where data center announcements convert to actual interconnections faster than expected, are the regions to watch for real strain.