EIA sees US power demand climbing to a record 4,356 billion kWh by 2027

Each record beats the last by a small margin: about 2.2% in 2026 and 1.6% in 2027 ^[11][12]. Over both years the US grows 3.8% ^[13], the rate the IEA expects for global demand in 2027 alone ^[8]. Data-center load moves faster. Gartner projects global data-center use rising from 447 TWh last year to 565 TWh this year ^[6]. That one-year increase of 118 TWh is larger than the 93 billion kWh, or 93 TWh, the EIA adds to all US demand in 2026 ^[14]. Berkeley Lab's reference case has US data centers at 649 TWh by 2030, or 11.8% of national use, inside a 9.5% to 15.3% range of scenarios ^[7].

The 11% rise in wholesale power to $52 per MWh ^[1] is thin support for an AI-driven squeeze, because the EIA attributes it largely to extreme weather ^[2]. The regional forecasts say more. PJM prices are expected to rise 41% while Mid-Columbia falls 23% ^[4], a 64-point spread inside one national average ^[15].

Deal prices per megawatt are where a power constraint would show up as money. According to CoinShares, one recent transaction valued stabilized AI facilities at roughly $27 million per megawatt, against below $3 million for bitcoin miners' energized but unleased capacity ^[10]. The finished site fetches more than nine times the connected one, a gap of over $24 million a megawatt ^[16]. There are three ways to read that gap. If power is the binding input, energized land is underpriced and should rerate toward the facility value. If the gap is mostly buildings, cooling and a signed lease, the scarce inputs are capital and tenants, and a connection is worth roughly what it trades for. If the constraint is regional, only connections in markets with PJM-style price rises rerate.

I think the second reading fits the evidence best. A buyer paying under $3 million for a megawatt of live connection ^[10] is pricing an option on a future tenant. The counter-case is the queue. Around 2,600 GW is waiting to connect in the US ^[9]. Cryptopolitan's summary of Gartner, Berkeley Lab, EPRI and IEA work names power access and connection times as the real limits on AI growth ^[18]. If those waits lengthen, an existing connection is worth more than one deal suggests. The view is wrong if energized, unleased miner sites start selling well above $3 million a megawatt with no tenant signed.

Miners are allocating as if the first reading holds. Core Scientific, Cipher and IREN are moving toward AI and high-performance computing ^[20]. Each megawatt they lease to an AI tenant is a megawatt that stops mining bitcoin. Cryptopolitan's headline calls electricity the new constraint ^[19], but its text is narrower: chips are no longer the sole hurdle, with grid connections, energy contracts, generation capacity and permitting now shaping where compute gets built ^[17].