The vocabulary behind the AI power buildout, in plain English — every definition grounded in the same public EIA / LBNL / ERCOT data as the rest of Gridlas.
The multi-year line every new power plant or battery must wait in before it can connect to the grid. In the U.S. the median wait has roughly doubled to about 5 years, and around 77% of proposed capacity is ultimately withdrawn. It — not a shortage of power plants — is the real bottleneck on powering AI.
A distinct, faster-tracked queue some grid operators (notably ERCOT) use for very large single customers such as data centers. ERCOT's large-load queue jumped from 63 GW to ~226 GW in a single year, about 77% of it data centers.
A very large data center operated by a cloud/AI provider (AWS, Microsoft, Google, Meta) — typically hundreds of MW to over a gigawatt of IT load, purpose-built for cloud and AI compute at scale.
Generation built on the customer's side of the utility meter — e.g. on-site gas turbines or fuel cells — so a data center can power up without waiting in the interconnection queue. A common way operators 'route around' the ~4.5-year grid wait.
The ratio of a data center's total energy to the energy delivered to computing equipment. A PUE of 1.0 is perfect; efficient hyperscale facilities run ~1.1–1.2, meaning only 10–20% overhead for cooling and power distribution.
Liters of water consumed per kWh of IT energy, mostly from evaporative cooling. The industry-average WUE is about 1.8 L/kWh; large campuses can draw several million gallons a day, which is why siting in high-water-stress areas is a growing constraint.
A long-term contract to buy electricity (often from a specific wind, solar, or nuclear source) at a set price. Data-center operators use PPAs to lock in firm, often low-carbon power — including nuclear restarts and small modular reactors.
A compact nuclear reactor (typically under ~300 MW) designed for factory build and faster siting. SMRs are increasingly proposed to give data centers firm, carbon-free baseload power without waiting on the grid — though most are years from commercial operation.
The minimum, always-on electricity supply that runs continuously regardless of demand swings — historically coal and nuclear, increasingly gas. AI data centers want firm baseload they can count on 24/7, which shapes their appetite for nuclear PPAs and on-site gas.
Deliberately reducing a generator's output — or a large load's draw — when the grid can't absorb or deliver the power. 'Curtailable' or flexible data-center load is one proposed way to connect faster by agreeing to pause during grid stress.
The elapsed time from applying to connect a project to actually energizing it — currently about 4.5 years on average in the U.S. Building new transmission on top of that takes 4–8 years, so the power for the next AI generation is set by decisions made half a decade earlier.
Queued projects that will likely never be built — placeholders, duplicates, or bets that inflate the queue. Of ERCOT's ~226 GW large-load queue, only about 1.8% is actually operational, so headline queue figures overstate what will really connect.
Bringing a shut-down nuclear plant back online to serve new demand — e.g. Three Mile Island Unit 1 (the Crane Clean Energy Center) restarting on a 20-year Microsoft PPA, and Palisades, the first U.S. reactor recovered from decommissioning. A fast-tracked way to add firm, carbon-free power this decade.
A power plant — usually gas turbines — that runs only during peak demand. Data-center operators increasingly stand up on-site gas that resembles a dedicated peaker to energize a site before the grid can deliver.
A plain-language way to express a data center's electricity use: how many typical U.S. homes would consume the same power. The Gridlas estimator reports it alongside raw MW so non-specialists can gauge scale.
Generation on the utility's side of the meter — power that flows through the grid to the customer, as opposed to behind-the-meter on-site generation. Meta's Entergy-built 2.26 GW gas plants in Louisiana are front-of-the-meter but dedicated to one campus.
The rise in electricity demand on a grid. After two flat decades, U.S. load growth has returned — the EIA calls it the strongest four-year demand growth since 2000 — driven largely by data centers and electrification.
Hardware and software — dynamic line rating, advanced power flow control, reconductoring — that squeeze more capacity out of existing transmission lines, offering faster relief than building new lines through the interconnection queue.
A data-center model where a provider leases space, power, and cooling to multiple tenants, rather than a single owner-operator. Market inventory figures (e.g. CBRE's) largely track colocation and hyperscale supply.
A measure of how much of an area's available water supply is already being used. About two-thirds of new U.S. data centers built or planned since 2022 sit in high-water-stress areas, making cooling water a siting constraint with few workarounds.
The full report maps where AI is landing, the power it needs, and the grid bottlenecks ahead — every figure cited.
Read the research →