Converting miners' 3,450MW of contracted but not-yet-billing AI capacity could cost $28bn to $52bn

Divide the roughly $1.1bn of annualised AI and HPC revenue the listed miners collect by the roughly 550MW actually billing and a converted megawatt earns about $2m a year ^[20][21][1]. Contracted megawatts are being priced as though they will all get there. Apply the 12.9 times enterprise value to next-12-month sales that CoinShares found for companies with contracted capacity to that $2m and you get $25.8m per MW ^[13][2]. CoinShares marked stabilised Northern Virginia space at about $27m per MW in a $3.5bn purchase of three leased facilities ^[9].

Getting from one to the other is a construction project. CoinShares puts the cost of upgrading mining infrastructure for AI at $8m to $15m per MW against $0.7m to $1m for mining infrastructure, so the AI version costs 8 to 21 times as much ^[19][4]. Start at the $3m per MW where energised but unleased miner capacity trades ^[10], add the conversion, and the all-in number is $11m to $18m against a $27m mark. That is a spread of $9m to $16m, if the buyer pays the stabilised price and the build lands on budget ^[3]. Across the roughly 3,450MW contracted but not billing, the conversion bill is $28bn to $52bn, or 25 to 47 times the revenue the sector bills now ^[5][6].

What makes an energised site scarce is paperwork. CoinShares puts the US interconnection queue at about 2,600 GW ^[11], which makes OpenAI's Stargate goal of 10GW of US capacity by 2029 ^[2] equal to 0.4% of the line ^[10]. New York in July became the first state to prohibit construction of new hyperscale data centres ^[12]. The IEA expects data centre electricity use to double to almost 945 TWh by 2030 and says grid limitations might delay up to 20% of intended projects ^[5][6]. VanEck says electricity, not chips, is becoming the main limiting factor ^[7]; CoinShares lists "land, power and shell" ^[8].

So the miners are paying to hold megawatts and to stop hashing. Core Scientific paid $41.9m to cancel about 15 EH/s of next-generation equipment, roughly $2.8m per EH/s it will now never run ^[15][7]. Keel, formerly Bitfarms, stopped mining on June 29 and is expected to report no mining revenue in the third quarter ^[16]. IREN's transition date is December 31, 2026, and Cipher Digital is likely out of mining by the end of 2027 ^[18][17]. S&P Global has tracked the same move across listed miners ^[14].

OpenAI's own figures are the demand side: about $278bn of cash use from 2026 to 2030 and about $856bn on computing infrastructure, the cash number being 32% of the compute number ^[1][8]. The link from that budget to a miner's substation is an inference about scarcity, and the reporting does not name the counterparties behind the miners' backlog. OpenAI is also buying power directly, having put $500m into SB Energy with SoftBank in January and chosen it for a 1.2GW site in Milam County, Texas ^[4]. It said in April that it had added 3GW in the previous 90 days, past its original Stargate goal ^[3].

In my view the 12.9x is a financing multiple: a signed contract is what lets a miner raise the $8m to $15m per MW ^[19] that the 3.7x cohort cannot ^[13]. The counter-thesis is arithmetically simple. If the full 4GW bills at $2m per MW, sector AI revenue is $8bn a year against $1.1bn now, and 12.9 times forward sales will look cheap in hindsight ^[9]. Two things would settle it: the 550MW figure climbing on schedule at unchanged revenue per MW, and conversion capex printing near $8m instead of $15m. If greenfield partnerships like SB Energy deliver megawatts cheaply, the $3m mark falls toward construction cost ^[10].