Run the units and the projections stop being abstract. A 100MW IT campus running flat out consumes on the order of 876GWh a year [6]. McKinsey's 606TWh figure for US data centres in 2030 [2] divides into roughly 690 of those campuses [1]. That is what a terawatt-hour looks like on the ground: several hundred fully loaded halls, each needing its own connection and its own energisation date.
Capacity plans tend to be written in hardware. GPUs land in Q3, cluster in production by Q4. What the schedule actually waits on is the interconnection agreement, and the Data Center Dynamics column that gathers these forecasts describes queues stretching for years and new high-voltage transmission taking most of a decade to permit and build [7]. The Department of Energy and Lawrence Berkeley National Laboratory put data centres at roughly 50GW of the new peak capacity the grid has to add by 2030 [3], about twice what the sector drew in total in 2024 [6]. Everyone in that queue holds a purchase order for the same thing.
Then there is firmness, which is where the cheap answers fail. Wind runs at a 30-40 percent capacity factor and solar at 20-30 percent [10]. Covering 100MW of continuous load from solar alone means sizing 333MW to 500MW of nameplate [3], plus storage for the hours after sunset. Nuclear runs above 90 percent of the time, but a new plant takes five to ten years or more [11], which outlasts the roadmap it would serve. Gas is dispatchable and quick to build, which is why new plants are being announced specifically to serve data centres behind the meter [14], and turbine order books are already backlogged into the late 2020s [13]. The economics run one way only: an idle GPU is a depreciating asset, so curtailment lands on the compute P&L rather than the energy line [15].
Two axes decide the source, per the column: whether it delivers around the clock, and what it costs the climate [17]. A third belongs on the grid for anyone planning capacity, which is time to first electron. Cross firm-versus-intermittent with on-grid-versus-behind-the-meter and the four boxes fail differently. Firm and on-grid means waiting in the queue; firm and behind the meter means gas today, and a carbon number the same team answers for later. Intermittent and on-grid puts curtailment against depreciating hardware; intermittent and behind the meter is a storage bill.
What the column does not supply is the operator's half of the evidence: no queue durations in months, no prices, no named contract with a megawatt figure and a date. Its own 130 percent US demand growth projection is credited only to unnamed "other analysts" [5]. Treat the aggregate as direction rather than schedule, and put three fields beside each compute increment in your own plan: firm megawatts, source and capacity factor, energised-by date. A blank in any of the three makes the GPU date a guess.
