Claim report cards
Are common AI data center cost claims true?
Every claim gets a ruling on the first line: true, false, or unproven. Then the evidence behind it, and the measurement that would change it.
4 of 5 claims are settled outright by the cited evidence. The remaining one is unproven: no published measurement supports them in either direction.
“AI is driving the return of U.S. electricity-demand growth.”
AI data centers are the largest single driver of new U.S. electricity demand.True. After nearly two decades of flat demand, U.S. electricity use is rising again, and data centers are the biggest new driver. EIA names data-center load as the cause of current growth, and LBNL’s server model attributes a large share of that growth to accelerated AI systems. The national numbers settle the direction and the scale; they do not split AI from ordinary cloud computing facility by facility.
Open report card“AI raised my household electricity bill.”
No published analysis puts an AI-attributable dollar figure on a household bill.Unproven — and the popular version of the claim is wrong about the cause. EIA attributes much of the 2025 rise in wholesale power prices to natural gas, not to data centers. A large data center does change a utility’s costs, but whether any of that reaches your bill is decided by its rate agreement, who pays for dedicated upgrades and how regulators split shared costs. No public analysis has yet converted that into a dollar amount on a residential bill, so anyone quoting one is guessing.
Open report card“AI caused the transformer shortage.”
The transformer shortage began years before the AI buildout.False. The transformer shortage was already here before the AI buildout. DOE traces it to post-pandemic demand, aging infrastructure, workforce constraints, raw materials and roughly 40,000 product configurations that slow manufacturing. Data centers add real pressure to that market, but they did not cause the shortage.
Open report card“AI is making ordinary memory and storage more expensive.”
AI memory production is displacing the capacity that makes ordinary DRAM and NAND.True. Micron says building AI memory (HBM) uses about three times the factory capacity of ordinary DDR5, and that there is not enough DRAM and NAND to supply both AI and conventional servers. The same production lines that make your RAM and SSDs are being pulled toward AI, which pushes ordinary prices up. What no public price series pins down is exactly how many dollars of any given retail tag come from AI rather than the broader server cycle.
Open report card“Every AI prompt uses a bottle of water.”
No per-prompt water figure is real, including the bottle.False. Measured water use varies by more than 10,000-fold from one data center to the next, so no single “bottle per prompt” number is real. The same prompt can use almost no water in a cool climate with efficient cooling and far more somewhere hot. Server efficiency, utilization, cooling design, climate and the local power source all change the result.
Open report card