This is the condensed version of the deep dive of the same name. Every key number has been checked against primary sources; for the full argument and citations, read the deep dive. Figures as of July 2026.
"Is AI causing shortages" is really two questions: hardware (chips, memory) is one shortage, electricity is another — and they work completely differently.
The hardware shortage is a price-clearing one: the goods exist, prices have doubled, and the people priced out are the lower bidders (like you, buying a PC). The power shortage is a queue-rationing one: in parts of the US grid, prices have hit the regulatory ceiling and there still isn't enough — money can't buy it — and the cost lands on everyone's electricity bill, including people who have never used AI.
Both shortages are real. But how long they last, and who pays, differ.
Most of the memory industry's 2026 capacity sold out before the year was half over: SK Hynix confirmed DRAM, NAND and HBM all gone, Micron confirmed its entire output of HBM — the high-bandwidth memory every AI chip requires — is under contract, and Samsung confirmed its HBM is sold out. Cloud giants are signing multi-year contracts at elevated prices just to lock in supply, textbook shortage-cycle behavior.
The primary price record: in Q1 2026, DRAM contract prices for large buyers rose 90–95% in a single quarter (yes — the "+90%" in headlines is the real contract price, not spot froth); Q2 added another 58–63%, with NAND up 70–75%. The passthrough to you: HP says memory and storage now make up a third of a PC's cost; the big PC makers have all raised prices 15–20%. The mechanism is simple: memory makers moved production lines to AI-grade memory, and consumer parts got crowded out — the shortage isn't stopping AI from getting chips; it's stopping you from affording a RAM stick.
How long? Tight through the end of 2027 — that's the consensus (new factories physically can't arrive sooner). The split is 2028: if AI demand keeps surging, gradual relief; if AI demand slows, the memory industry's old script is a hard flip into glut — every memory supercycle in history has ended that way. Note who's talking: the loudest "shortage will last years" voices are the memory makers profiting from the price spike.
The hardest evidence on power isn't a forecast — it's what already happened. PJM, America's largest grid (13 states, including data-center-capital Virginia), auctions "future power capacity" every year. That price went up 11× in three years: from $28.92 to $329.17, then $333.44, then $325 — three consecutive auctions pinned at the regulatory price cap, and the latest (July 2026) still came up 6,831 MW short. A price stuck at the ceiling means: not expensive — unavailable.
Who caused it? PJM's independent market monitor did the math: 45% of the costs of the last three auctions ($21.3 billion) trace to forecast data-center load — much of it from data centers that don't exist yet.
Who pays? Everyone. Per state consumer advocates' calculations: Baltimore residents pay about $21/month more, Washington DC about $20, Maryland jurisdictions $14–18, and Ohio's utility itself reports about $27 — increases cluster at $14–27/month, roughly half attributed directly to the capacity-price spike.
Texas's grid (ERCOT) has 438 GW of connection requests queued — while the entire Texas grid's all-time peak is 85 GW. The queue is five times the whole system's peak. It cannot physically be built. The reason: developers file the same project in multiple states, build in one, and leave the rest as "phantom demand" in the queues. Experts estimate requests run 5–10× actual construction.
Regulators have started squeezing the water out: Chicago charges $1 million just to apply for a big connection, Ohio makes data centers pay for 85% of forecast usage whether they use it or not, Texas mandates disclosure of duplicate filings. Once raising your hand becomes placing a bet, watch how fast the queue shrinks — that's the best gauge of true demand over the next two years.
Data-center power forecasts have two famous priors: 1999's "the internet will eat half of US electricity within a decade" (reality by 2020: 1–2%), and the EPA's 2007 forecast that actual usage came in far below (efficiency gains plus the financial crisis). Both times, equipment sellers spread the demand myth, and the person with independent data was right. So today's forecasts — "global data-center power doubles to 945 TWh by 2030," "12% of US electricity by 2028" — deserve a discount: the official reports themselves publish ~2× ranges. Panic pieces quote the ceiling; reassurance pieces quote the floor.
But this time has one material difference: in the prior scares, the predicted peak never arrived; this time the constraint has already cashed out — the auctions were gaveled, the bills were mailed, the turbine orders were signed. You can doubt 945 TWh. You can't doubt $325/MW-day. "The forecasts are unreliable" and "the shortage isn't real" are two different claims.
So the conclusion is structural: demand signs up in months, supply delivers in five to eight years. America's power crunch most likely runs to around 2030.
Efficiency won't ride to the rescue — but it bends the slope. Google says one Gemini request now uses 33× less energy than a year ago — true; and the same Google's total data-center power use doubled over the same era (the classic Jevons paradox: the cheaper each use gets, the more everyone uses). Efficiency decides whether the forecast lands high or low; it does nothing for the queue of transformers and power lines.
This is America's shortage, not the world's. China grew its power supply 8% a year for two decades and parts of its grid run at a surplus — data centers just plug in. The US spent twenty years with flat demand, calibrated its whole system to flatness, and got caught flat-footed. At bottom, the power shortage is a mismatch between America's infrastructure cycle and AI's demand cycle.
Hardware: really short, tight through 2027, and 2028 depends on AI demand (keeps surging = gradual relief; slows = flips to glut). Power: really short, US-specific, structural to around 2030 — and this one doesn't hinge on any forecast: it's already written into auction results and your electricity bill.